├── .gitignore ├── INFERENCE.md ├── LICENSE ├── Makefile ├── README.md ├── helpers ├── gradio_demo │ └── app.py ├── model_init_scripts │ ├── init_dummy_model.py │ ├── init_dummy_model_with_encodec.py │ ├── init_large_model.py │ └── init_model_600M.py ├── push_to_hub_scripts │ ├── push_dac_to_hub.py │ └── push_trained_parler_tts_to_hub.py └── training_configs │ ├── librispeech_tts_r_300M_dummy.json │ ├── starting_point_0.01.json │ ├── starting_point_v1.json │ └── starting_point_v1_large.json ├── parler_tts ├── __init__.py ├── configuration_parler_tts.py ├── dac_wrapper │ ├── __init__.py │ ├── configuration_dac.py │ └── modeling_dac.py ├── logits_processors.py ├── modeling_parler_tts.py └── streamer.py ├── pyproject.toml ├── setup.py └── training ├── README.md ├── __init__.py ├── arguments.py ├── data.py ├── eval.py ├── run_parler_tts_training.py └── utils.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Adapted from https://github.com/huggingface/diffusers/blob/main/.gitignore 2 | 3 | # Byte-compiled / optimized / DLL files 4 | __pycache__/ 5 | *.py[cod] 6 | *$py.class 7 | 8 | # C extensions 9 | *.so 10 | 11 | # logs 12 | logs/ 13 | 14 | # Distribution / packaging 15 | .Python 16 | build/ 17 | develop-eggs/ 18 | dist/ 19 | downloads/ 20 | eggs/ 21 | .eggs/ 22 | lib/ 23 | lib64/ 24 | parts/ 25 | sdist/ 26 | var/ 27 | wheels/ 28 | *.egg-info/ 29 | .installed.cfg 30 | *.egg 31 | MANIFEST 32 | 33 | # PyInstaller 34 | # Usually these files are written by a python script from a template 35 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 36 | *.manifest 37 | *.spec 38 | 39 | # Installer logs 40 | pip-log.txt 41 | pip-delete-this-directory.txt 42 | 43 | # Unit test / coverage reports 44 | htmlcov/ 45 | .tox/ 46 | .nox/ 47 | .coverage 48 | .coverage.* 49 | .cache 50 | nosetests.xml 51 | coverage.xml 52 | *.cover 53 | .hypothesis/ 54 | .pytest_cache/ 55 | 56 | # Translations 57 | *.mo 58 | *.pot 59 | 60 | # Django stuff: 61 | *.log 62 | local_settings.py 63 | db.sqlite3 64 | 65 | # Flask stuff: 66 | instance/ 67 | .webassets-cache 68 | 69 | # Scrapy stuff: 70 | .scrapy 71 | 72 | # Sphinx documentation 73 | docs/_build/ 74 | 75 | # PyBuilder 76 | target/ 77 | 78 | # Jupyter Notebook 79 | .ipynb_checkpoints 80 | 81 | # IPython 82 | profile_default/ 83 | ipython_config.py 84 | 85 | # pyenv 86 | .python-version 87 | 88 | # celery beat schedule file 89 | celerybeat-schedule 90 | 91 | # SageMath parsed files 92 | *.sage.py 93 | 94 | # Environments 95 | .env 96 | .venv 97 | env/ 98 | venv/ 99 | ENV/ 100 | env.bak/ 101 | venv.bak/ 102 | 103 | # Spyder project settings 104 | .spyderproject 105 | .spyproject 106 | 107 | # Rope project settings 108 | .ropeproject 109 | 110 | # mkdocs documentation 111 | /site 112 | 113 | # mypy 114 | .mypy_cache/ 115 | .dmypy.json 116 | dmypy.json 117 | 118 | # Pyre type checker 119 | .pyre/ 120 | 121 | # vscode 122 | .vs 123 | .vscode 124 | 125 | # Pycharm 126 | .idea 127 | 128 | # TF code 129 | tensorflow_code 130 | 131 | # Models 132 | proc_data 133 | 134 | # examples 135 | runs 136 | /runs_old 137 | /wandb 138 | /examples/runs 139 | /examples/**/*.args 140 | /examples/rag/sweep 141 | 142 | # data 143 | /data 144 | serialization_dir 145 | 146 | # emacs 147 | *.*~ 148 | debug.env 149 | 150 | # vim 151 | .*.swp 152 | 153 | #ctags 154 | tags 155 | 156 | # pre-commit 157 | .pre-commit* 158 | 159 | # .lock 160 | *.lock 161 | 162 | # DS_Store (MacOS) 163 | .DS_Store 164 | # RL pipelines may produce mp4 outputs 165 | *.mp4 166 | 167 | # dependencies 168 | /transformers 169 | 170 | # ruff 171 | .ruff_cache 172 | 173 | wandb 174 | -------------------------------------------------------------------------------- /INFERENCE.md: -------------------------------------------------------------------------------- 1 | # Inference tips 2 | 3 | Parler-TTS benefits from a number of optimizations that can make the model up to 4x faster. Add to this the ability to stream audio as it's being generated, and you can achieve time-to-first audio in under 500ms on a modern GPU. 4 | 5 | ## 📖 Quick Index 6 | * [Efficient Attention Implementation](#efficient-attention-implementations) 7 | * [Compilation](#compilation) 8 | * [Streaming](#streaming) 9 | * [Batch generation](#batch-generation) 10 | * [Speaker Consistency](#speaker-consistency) 11 | 12 | ## Efficient Attention implementations 13 | 14 | Parler-TTS supports [SDPA](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention.html) and [Flash Attention 2](https://github.com/Dao-AILab/flash-attention). 15 | 16 | SDPA is used by default and speeds up generation time by up to 1.4x compared with eager attention. 17 | 18 | To switch between attention implementations, simply specify `attn_implementation=attn_implementation` when loading the checkpoints: 19 | 20 | ```py 21 | from parler_tts import ParlerTTSForConditionalGeneration 22 | 23 | torch_device = "cuda:0" # use "mps" for Mac 24 | torch_dtype = torch.bfloat16 25 | model_name = "parler-tts/parler-tts-mini-v1" 26 | 27 | attn_implementation = "eager" # "sdpa" or "flash_attention_2" 28 | 29 | model = ParlerTTSForConditionalGeneration.from_pretrained( 30 | model_name, 31 | attn_implementation=attn_implementation 32 | ).to(torch_device, dtype=torch_dtype) 33 | ``` 34 | 35 | ## Compilation 36 | 37 | [Compiling](https://pytorch.org/docs/stable/generated/torch.compile.html) the forward method of Parler can speed up generation time by up to 4.5x. 38 | 39 | As an indication, `mode=default` brings a speed-up of 1.4 times compared to no compilation, while `mode="reduce-overhead"` brings much faster generation, at the cost of a longer compilation time and the need to generate twice to see the benefits of compilation. 40 | 41 | ```py 42 | import torch 43 | from parler_tts import ParlerTTSForConditionalGeneration 44 | from transformers import AutoTokenizer 45 | 46 | torch_device = "cuda:0" 47 | torch_dtype = torch.bfloat16 48 | model_name = "parler-tts/parler-tts-mini-v1" 49 | 50 | # need to set padding max length 51 | max_length = 50 52 | 53 | # load model and tokenizer 54 | tokenizer = AutoTokenizer.from_pretrained(model_name) 55 | model = ParlerTTSForConditionalGeneration.from_pretrained( 56 | model_name, 57 | attn_implementation="eager" 58 | ).to(torch_device, dtype=torch_dtype) 59 | 60 | # compile the forward pass 61 | compile_mode = "default" # chose "reduce-overhead" for 3 to 4x speed-up 62 | model.generation_config.cache_implementation = "static" 63 | model.forward = torch.compile(model.forward, mode=compile_mode) 64 | 65 | # warmup 66 | inputs = tokenizer("This is for compilation", return_tensors="pt", padding="max_length", max_length=max_length).to(torch_device) 67 | 68 | model_kwargs = {**inputs, "prompt_input_ids": inputs.input_ids, "prompt_attention_mask": inputs.attention_mask, } 69 | 70 | n_steps = 1 if compile_mode == "default" else 2 71 | for _ in range(n_steps): 72 | _ = model.generate(**model_kwargs) 73 | 74 | 75 | # now you can benefit from compilation speed-ups 76 | ... 77 | 78 | ``` 79 | 80 | 81 | ## Streaming 82 | 83 | ### How Does It Work? 84 | 85 | Parler-TTS is an auto-regressive transformer-based model, meaning generates audio codes (tokens) in a causal fashion. 86 | 87 | At each decoding step, the model generates a new set of audio codes, conditional on the text input and all previous audio codes. From the 88 | frame rate of the [DAC model](https://huggingface.co/parler-tts/dac_44khZ_8kbps) used to decode the generated codes to audio waveform, each set of generated audio codes corresponds to 0.011 seconds. This means we require a total of 1720 decoding steps to generate 20 seconds of audio. 89 | 90 | Rather than waiting for the entire audio sequence to be generated, which would require the full 1720 decoding steps, we can start playing the audio after a specified number of decoding steps have been reached, a techinque known as [*streaming*](https://huggingface.co/docs/transformers/main/en/generation_strategies#streaming). 91 | For example, after 86 steps we have the first second of audio ready, and so can play this without waiting for the remaining decoding steps to be complete. As we continue to generate with the Parler-TTS model, we append new chunks of generated audio to our output waveform on-the-fly. After the full 1720 decoding steps, the generated audio is complete, and is composed of 20 chunks of audio, each corresponding to 86 tokens. 92 | This method of playing incremental generations reduces the latency of the Parler-TTS model from the total time to generate 1720 tokens, to the time taken to play the first chunk of audio (86 tokens). This can result in significant improvements to perceived latency, particularly when the chunk size is chosen to be small. In practice, the chunk size should be tuned to your device: using a smaller chunk size will mean that the first chunk is ready faster, but should not be chosen so small that the model generates slower than the time it takes to play the audio. 93 | 94 | 95 | ### How Can I Use It? 96 | 97 | We've added [ParlerTTSStreamer](https://github.com/huggingface/parler-tts/blob/main/parler_tts/streamer.py) to the library. Don't hesitate to adapt it to your use-case. 98 | 99 | Here's how to create a generator out of the streamer. 100 | 101 | ```py 102 | import torch 103 | from parler_tts import ParlerTTSForConditionalGeneration, ParlerTTSStreamer 104 | from transformers import AutoTokenizer 105 | from threading import Thread 106 | 107 | torch_device = "cuda:0" # Use "mps" for Mac 108 | torch_dtype = torch.bfloat16 109 | model_name = "parler-tts/parler-tts-mini-v1" 110 | 111 | # need to set padding max length 112 | max_length = 50 113 | 114 | # load model and tokenizer 115 | tokenizer = AutoTokenizer.from_pretrained(model_name) 116 | model = ParlerTTSForConditionalGeneration.from_pretrained( 117 | model_name, 118 | ).to(torch_device, dtype=torch_dtype) 119 | 120 | sampling_rate = model.audio_encoder.config.sampling_rate 121 | frame_rate = model.audio_encoder.config.frame_rate 122 | 123 | def generate(text, description, play_steps_in_s=0.5): 124 | play_steps = int(frame_rate * play_steps_in_s) 125 | streamer = ParlerTTSStreamer(model, device=torch_device, play_steps=play_steps) 126 | # tokenization 127 | inputs = tokenizer(description, return_tensors="pt").to(torch_device) 128 | prompt = tokenizer(text, return_tensors="pt").to(torch_device) 129 | # create generation kwargs 130 | generation_kwargs = dict( 131 | input_ids=inputs.input_ids, 132 | prompt_input_ids=prompt.input_ids, 133 | attention_mask=inputs.attention_mask, 134 | prompt_attention_mask=prompt.attention_mask, 135 | streamer=streamer, 136 | do_sample=True, 137 | temperature=1.0, 138 | min_new_tokens=10, 139 | ) 140 | # initialize Thread 141 | thread = Thread(target=model.generate, kwargs=generation_kwargs) 142 | thread.start() 143 | # iterate over chunks of audio 144 | for new_audio in streamer: 145 | if new_audio.shape[0] == 0: 146 | break 147 | print(f"Sample of length: {round(new_audio.shape[0] / sampling_rate, 4)} seconds") 148 | yield sampling_rate, new_audio 149 | 150 | 151 | # now you can do 152 | text = "This is a test of the streamer class" 153 | description = "Jon's talking really fast." 154 | 155 | chunk_size_in_s = 0.5 156 | 157 | for (sampling_rate, audio_chunk) in generate(text, description, chunk_size_in_s): 158 | # You can do everything that you need with the chunk now 159 | # For example: stream it, save it, play it. 160 | print(audio_chunk.shape) 161 | ``` 162 | 163 | ## Batch generation 164 | 165 | Batching means combining operations for multiple samples to bring the overall time spent generating the samples lower than generating sample per sample. 166 | 167 | Here is a quick example of how you can use it: 168 | 169 | ```py 170 | from parler_tts import ParlerTTSForConditionalGeneration 171 | from transformers import AutoTokenizer, AutoFeatureExtractor, set_seed 172 | import scipy 173 | 174 | 175 | repo_id = "parler-tts/parler-tts-mini-v1" 176 | 177 | model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id).to("cuda") 178 | tokenizer = AutoTokenizer.from_pretrained(repo_id, padding_side="left") 179 | feature_extractor = AutoFeatureExtractor.from_pretrained(repo_id) 180 | 181 | input_text = ["Hey, how are you doing?", "I'm not sure how to feel about it."] 182 | description = 2 * ["A male speaker with a monotone and high-pitched voice is delivering his speech at a really low speed in a confined environment."] 183 | 184 | inputs = tokenizer(description, return_tensors="pt", padding=True).to("cuda") 185 | prompt = tokenizer(input_text, return_tensors="pt", padding=True).to("cuda") 186 | 187 | set_seed(0) 188 | generation = model.generate( 189 | input_ids=inputs.input_ids, 190 | attention_mask=inputs.attention_mask, 191 | prompt_input_ids=prompt.input_ids, 192 | prompt_attention_mask=prompt.attention_mask, 193 | do_sample=True, 194 | return_dict_in_generate=True, 195 | ) 196 | 197 | audio_1 = generation.sequences[0, :generation.audios_length[0]] 198 | audio_2 = generation.sequences[1, :generation.audios_length[1]] 199 | 200 | print(audio_1.shape, audio_2.shape) 201 | scipy.io.wavfile.write("sample_out.wav", rate=feature_extractor.sampling_rate, data=audio_1.cpu().numpy().squeeze()) 202 | scipy.io.wavfile.write("sample_out_2.wav", rate=feature_extractor.sampling_rate, data=audio_2.cpu().numpy().squeeze()) 203 | ``` 204 | 205 | ## Speaker Consistency 206 | 207 | The checkpoint was trained on 34 speakers. The full list of available speakers includes: 208 | Laura, Gary, Jon, Lea, Karen, Rick, Brenda, David, Eileen, Jordan, Mike, Yann, Joy, James, Eric, Lauren, Rose, Will, Jason, Aaron, Naomie, Alisa, Patrick, Jerry, Tina, Jenna, Bill, Tom, Carol, Barbara, Rebecca, Anna, Bruce, and Emily. 209 | 210 | However, the models performed better with certain speakers. Below are the top 20 speakers for each model variant, ranked by their average speaker similarity scores: 211 | 212 | ### Large Model - Top 20 Speakers 213 | 214 | | Speaker | Similarity Score | 215 | |---------|------------------| 216 | | Will | 0.906055 | 217 | | Eric | 0.887598 | 218 | | Laura | 0.877930 | 219 | | Alisa | 0.877393 | 220 | | Patrick | 0.873682 | 221 | | Rose | 0.873047 | 222 | | Jerry | 0.871582 | 223 | | Jordan | 0.870703 | 224 | | Lauren | 0.867432 | 225 | | Jenna | 0.866455 | 226 | | Karen | 0.866309 | 227 | | Rick | 0.863135 | 228 | | Bill | 0.862207 | 229 | | James | 0.856934 | 230 | | Yann | 0.856787 | 231 | | Emily | 0.856543 | 232 | | Anna | 0.848877 | 233 | | Jon | 0.848828 | 234 | | Brenda | 0.848291 | 235 | | Barbara | 0.847998 | 236 | 237 | ### Mini Model - Top 20 Speakers 238 | 239 | | Speaker | Similarity Score | 240 | |---------|------------------| 241 | | Jon | 0.908301 | 242 | | Lea | 0.904785 | 243 | | Gary | 0.903516 | 244 | | Jenna | 0.901807 | 245 | | Mike | 0.885742 | 246 | | Laura | 0.882666 | 247 | | Lauren | 0.878320 | 248 | | Eileen | 0.875635 | 249 | | Alisa | 0.874219 | 250 | | Karen | 0.872363 | 251 | | Barbara | 0.871509 | 252 | | Carol | 0.863623 | 253 | | Emily | 0.854932 | 254 | | Rose | 0.852246 | 255 | | Will | 0.851074 | 256 | | Patrick | 0.850977 | 257 | | Eric | 0.845459 | 258 | | Rick | 0.845020 | 259 | | Anna | 0.844922 | 260 | | Tina | 0.839160 | 261 | 262 | The numbers represent the average speaker similarity between a random snippet of the person speaking and a randomly Parler-generated snippet. Higher scores indicate better model performance in maintaining voice consistency. 263 | 264 | These scores are derived from [dataset for Mini](https://huggingface.co/datasets/ylacombe/parler-tts-mini-v1_speaker_similarity) and [dataset for Large](https://huggingface.co/datasets/ylacombe/parler-large-v1-og_speaker_similarity). -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [2024] [The HuggingFace Inc. team] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /Makefile: -------------------------------------------------------------------------------- 1 | check_dirs := . 2 | 3 | quality: 4 | black --check $(check_dirs) 5 | ruff $(check_dirs) 6 | 7 | style: 8 | black $(check_dirs) 9 | ruff $(check_dirs) --fix 10 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Parler-TTS 2 | 3 | Parler-TTS is a lightweight text-to-speech (TTS) model that can generate high-quality, natural sounding speech in the style of a given speaker (gender, pitch, speaking style, etc). It is a reproduction of work from the paper [Natural language guidance of high-fidelity text-to-speech with synthetic annotations](https://www.text-description-to-speech.com) by Dan Lyth and Simon King, from Stability AI and Edinburgh University respectively. 4 | 5 | Contrarily to other TTS models, Parler-TTS is a **fully open-source** release. All of the datasets, pre-processing, training code and weights are released publicly under permissive license, enabling the community to build on our work and develop their own powerful TTS models. 6 | 7 | This repository contains the inference and training code for Parler-TTS. It is designed to accompany the [Data-Speech](https://github.com/huggingface/dataspeech) repository for dataset annotation. 8 | 9 | > [!IMPORTANT] 10 | > **08/08/2024:** We are proud to release two new Parler-TTS checkpoints: 11 | > 1. [Parler-TTS Mini](https://huggingface.co/parler-tts/parler-tts-mini-v1), an 880M parameter model. 12 | > 2. [Parler-TTS Large](https://huggingface.co/parler-tts/parler-tts-large-v1), a 2.3B parameter model. 13 | > 14 | > These checkpoints have been trained on 45k hours of audiobook data. 15 | > 16 | > In addition, the code is optimized for much faster generation: we've added SDPA and Flash Attention 2 compatibility, as well as the ability to compile the model. 17 | 18 | ## 📖 Quick Index 19 | * [Installation](#installation) 20 | * [Usage](#usage) 21 | - [🎲 Using a random voice](#-random-voice) 22 | - [🎯 Using a specific speaker](#-using-a-specific-speaker) 23 | * [Training](#training) 24 | * [Demo](https://huggingface.co/spaces/parler-tts/parler_tts) 25 | * [Model weights and datasets](https://huggingface.co/parler-tts) 26 | * [Optimizing inference](#-optimizing-inference-speed) 27 | 28 | ## Installation 29 | 30 | Parler-TTS has light-weight dependencies and can be installed in one line: 31 | 32 | ```sh 33 | pip install git+https://github.com/huggingface/parler-tts.git 34 | ``` 35 | 36 | Apple Silicon users will need to run a follow-up command to make use the nightly PyTorch (2.4) build for bfloat16 support: 37 | 38 | ```sh 39 | pip3 install --pre torch torchaudio --index-url https://download.pytorch.org/whl/nightly/cpu 40 | ``` 41 | 42 | ## Usage 43 | 44 | > [!TIP] 45 | > You can directly try it out in an interactive demo [here](https://huggingface.co/spaces/parler-tts/parler_tts)! 46 | 47 | Using Parler-TTS is as simple as "bonjour". Simply install the library once: 48 | 49 | ```sh 50 | pip install git+https://github.com/huggingface/parler-tts.git 51 | ``` 52 | 53 | ### 🎲 Random voice 54 | 55 | 56 | **Parler-TTS** has been trained to generate speech with features that can be controlled with a simple text prompt, for example: 57 | 58 | ```py 59 | import torch 60 | from parler_tts import ParlerTTSForConditionalGeneration 61 | from transformers import AutoTokenizer 62 | import soundfile as sf 63 | 64 | device = "cuda:0" if torch.cuda.is_available() else "cpu" 65 | 66 | model = ParlerTTSForConditionalGeneration.from_pretrained("parler-tts/parler-tts-mini-v1").to(device) 67 | tokenizer = AutoTokenizer.from_pretrained("parler-tts/parler-tts-mini-v1") 68 | 69 | prompt = "Hey, how are you doing today?" 70 | description = "A female speaker delivers a slightly expressive and animated speech with a moderate speed and pitch. The recording is of very high quality, with the speaker's voice sounding clear and very close up." 71 | 72 | input_ids = tokenizer(description, return_tensors="pt").input_ids.to(device) 73 | prompt_input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device) 74 | 75 | generation = model.generate(input_ids=input_ids, prompt_input_ids=prompt_input_ids) 76 | audio_arr = generation.cpu().numpy().squeeze() 77 | sf.write("parler_tts_out.wav", audio_arr, model.config.sampling_rate) 78 | ``` 79 | 80 | ### 🎯 Using a specific speaker 81 | 82 | To ensure speaker consistency across generations, this checkpoint was also trained on 34 speakers, characterized by name. The full list of available speakers includes: 83 | Laura, Gary, Jon, Lea, Karen, Rick, Brenda, David, Eileen, Jordan, Mike, Yann, Joy, James, Eric, Lauren, Rose, Will, Jason, Aaron, Naomie, Alisa, Patrick, Jerry, Tina, Jenna, Bill, Tom, Carol, Barbara, Rebecca, Anna, Bruce, Emily. 84 | 85 | To take advantage of this, simply adapt your text description to specify which speaker to use: `Jon's voice is monotone yet slightly fast in delivery, with a very close recording that almost has no background noise.` 86 | 87 | You can replace "Jon" with any of the names from the list above to utilize different speaker characteristics. Each speaker has unique vocal qualities that can be leveraged to suit your specific needs. For more detailed information on speaker performance with voice consistency, please refer [inference guide](INFERENCE.md#speaker-consistency). 88 | 89 | ```py 90 | import torch 91 | from parler_tts import ParlerTTSForConditionalGeneration 92 | from transformers import AutoTokenizer 93 | import soundfile as sf 94 | 95 | device = "cuda:0" if torch.cuda.is_available() else "cpu" 96 | 97 | model = ParlerTTSForConditionalGeneration.from_pretrained("parler-tts/parler-tts-mini-v1").to(device) 98 | tokenizer = AutoTokenizer.from_pretrained("parler-tts/parler-tts-mini-v1") 99 | 100 | prompt = "Hey, how are you doing today?" 101 | description = "Jon's voice is monotone yet slightly fast in delivery, with a very close recording that almost has no background noise." 102 | 103 | input_ids = tokenizer(description, return_tensors="pt").input_ids.to(device) 104 | prompt_input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device) 105 | 106 | generation = model.generate(input_ids=input_ids, prompt_input_ids=prompt_input_ids) 107 | audio_arr = generation.cpu().numpy().squeeze() 108 | sf.write("parler_tts_out.wav", audio_arr, model.config.sampling_rate) 109 | ``` 110 | 111 | **Tips**: 112 | * Include the term "very clear audio" to generate the highest quality audio, and "very noisy audio" for high levels of background noise 113 | * Punctuation can be used to control the prosody of the generations, e.g. use commas to add small breaks in speech 114 | * The remaining speech features (gender, speaking rate, pitch and reverberation) can be controlled directly through the prompt 115 | 116 | ### ✨ Optimizing Inference Speed 117 | 118 | We've set up an [inference guide](INFERENCE.md) to make generation faster. Think SDPA, torch.compile and streaming! 119 | 120 | 121 | https://github.com/huggingface/parler-tts/assets/52246514/251e2488-fe6e-42c1-81cd-814c5b7795b0 122 | 123 | ## Training 124 | 125 | 126 | Open In Colab 127 | 128 | 129 | The [training folder](/training/) contains all the information to train or fine-tune your own Parler-TTS model. It consists of: 130 | - [1. An introduction to the Parler-TTS architecture](/training/README.md#1-architecture) 131 | - [2. The first steps to get started](/training/README.md#2-getting-started) 132 | - [3. A training guide](/training/README.md#3-training) 133 | 134 | > [!IMPORTANT] 135 | > **TL;DR:** After having followed the [installation steps](/training/README.md#requirements), you can reproduce the Parler-TTS Mini v1 training recipe with the following command line: 136 | 137 | ```sh 138 | accelerate launch ./training/run_parler_tts_training.py ./helpers/training_configs/starting_point_v1.json 139 | ``` 140 | 141 | > [!IMPORTANT] 142 | > You can also follow [this fine-tuning guide](https://github.com/ylacombe/scripts_and_notebooks/blob/main/Finetuning_Parler_TTS_v1_on_a_single_speaker_dataset.ipynb) on a mono-speaker dataset example. 143 | 144 | ## Acknowledgements 145 | 146 | This library builds on top of a number of open-source giants, to whom we'd like to extend our warmest thanks for providing these tools! 147 | 148 | Special thanks to: 149 | - Dan Lyth and Simon King, from Stability AI and Edinburgh University respectively, for publishing such a promising and clear research paper: [Natural language guidance of high-fidelity text-to-speech with synthetic annotations](https://arxiv.org/abs/2402.01912). 150 | - the many libraries used, namely [🤗 datasets](https://huggingface.co/docs/datasets/v2.17.0/en/index), [🤗 accelerate](https://huggingface.co/docs/accelerate/en/index), [jiwer](https://github.com/jitsi/jiwer), [wandb](https://wandb.ai/), and [🤗 transformers](https://huggingface.co/docs/transformers/index). 151 | - Descript for the [DAC codec model](https://github.com/descriptinc/descript-audio-codec) 152 | - Hugging Face 🤗 for providing compute resources and time to explore! 153 | 154 | 155 | ## Citation 156 | 157 | If you found this repository useful, please consider citing this work and also the original Stability AI paper: 158 | 159 | ``` 160 | @misc{lacombe-etal-2024-parler-tts, 161 | author = {Yoach Lacombe and Vaibhav Srivastav and Sanchit Gandhi}, 162 | title = {Parler-TTS}, 163 | year = {2024}, 164 | publisher = {GitHub}, 165 | journal = {GitHub repository}, 166 | howpublished = {\url{https://github.com/huggingface/parler-tts}} 167 | } 168 | ``` 169 | 170 | ``` 171 | @misc{lyth2024natural, 172 | title={Natural language guidance of high-fidelity text-to-speech with synthetic annotations}, 173 | author={Dan Lyth and Simon King}, 174 | year={2024}, 175 | eprint={2402.01912}, 176 | archivePrefix={arXiv}, 177 | primaryClass={cs.SD} 178 | } 179 | ``` 180 | 181 | ## Contribution 182 | 183 | Contributions are welcome, as the project offers many possibilities for improvement and exploration. 184 | 185 | Namely, we're looking at ways to improve both quality and speed: 186 | - Datasets: 187 | - Train on more data 188 | - Add more features such as accents 189 | - Training: 190 | - Add PEFT compatibility to do Lora fine-tuning. 191 | - Add possibility to train without description column. 192 | - Add notebook training. 193 | - Explore multilingual training. 194 | - Explore mono-speaker finetuning. 195 | - Explore more architectures. 196 | - Optimization: 197 | - Compilation and static cache 198 | - Support to FA2 and SDPA 199 | - Evaluation: 200 | - Add more evaluation metrics 201 | 202 | -------------------------------------------------------------------------------- /helpers/gradio_demo/app.py: -------------------------------------------------------------------------------- 1 | import gradio as gr 2 | import torch 3 | from transformers import AutoFeatureExtractor, AutoTokenizer, set_seed 4 | 5 | from parler_tts import ParlerTTSForConditionalGeneration 6 | 7 | 8 | device = "cuda:0" if torch.cuda.is_available() else "cpu" 9 | 10 | repo_id = "parler-tts/parler_tts_mini_v0.1" 11 | 12 | model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id).to(device) 13 | tokenizer = AutoTokenizer.from_pretrained(repo_id) 14 | feature_extractor = AutoFeatureExtractor.from_pretrained(repo_id) 15 | 16 | 17 | SAMPLE_RATE = feature_extractor.sampling_rate 18 | SEED = 41 19 | 20 | default_text = "Please surprise me and speak in whatever voice you enjoy." 21 | 22 | title = "# Parler-TTS " 23 | 24 | examples = [ 25 | [ 26 | "'This is the best time of my life, Bartley,' she said happily.", 27 | "A female speaker with a slightly low-pitched, quite monotone voice delivers her words at a slightly faster-than-average pace in a confined space with very clear audio.", 28 | ], 29 | [ 30 | "Montrose also, after having experienced still more variety of good and bad fortune, threw down his arms, and retired out of the kingdom. ", 31 | "A male speaker with a slightly high-pitched voice delivering his words at a slightly slow pace in a small, confined space with a touch of background noise and a quite monotone tone.", 32 | ], 33 | [ 34 | "montrose also after having experienced still more variety of good and bad fortune threw down his arms and retired out of the kingdom", 35 | "A male speaker with a low-pitched voice delivering his words at a fast pace in a small, confined space with a lot of background noise and an animated tone.", 36 | ], 37 | ] 38 | 39 | 40 | def gen_tts(text, description): 41 | inputs = tokenizer(description, return_tensors="pt").to(device) 42 | prompt = tokenizer(text, return_tensors="pt").to(device) 43 | 44 | set_seed(SEED) 45 | generation = model.generate( 46 | input_ids=inputs.input_ids, prompt_input_ids=prompt.input_ids, do_sample=True, temperature=1.0 47 | ) 48 | audio_arr = generation.cpu().numpy().squeeze() 49 | 50 | return (SAMPLE_RATE, audio_arr) 51 | 52 | 53 | css = """ 54 | #share-btn-container { 55 | display: flex; 56 | padding-left: 0.5rem !important; 57 | padding-right: 0.5rem !important; 58 | background-color: #000000; 59 | justify-content: center; 60 | align-items: center; 61 | border-radius: 9999px !important; 62 | width: 13rem; 63 | margin-top: 10px; 64 | margin-left: auto; 65 | flex: unset !important; 66 | } 67 | #share-btn { 68 | all: initial; 69 | color: #ffffff; 70 | font-weight: 600; 71 | cursor: pointer; 72 | font-family: 'IBM Plex Sans', sans-serif; 73 | margin-left: 0.5rem !important; 74 | padding-top: 0.25rem !important; 75 | padding-bottom: 0.25rem !important; 76 | right:0; 77 | } 78 | #share-btn * { 79 | all: unset !important; 80 | } 81 | #share-btn-container div:nth-child(-n+2){ 82 | width: auto !important; 83 | min-height: 0px !important; 84 | } 85 | #share-btn-container .wrap { 86 | display: none !important; 87 | } 88 | """ 89 | with gr.Blocks(css=css) as block: 90 | gr.Markdown(title) 91 | with gr.Row(): 92 | with gr.Column(): 93 | input_text = gr.Textbox(label="Input Text", lines=2, value=default_text, elem_id="input_text") 94 | description = gr.Textbox(label="Description", lines=2, value="", elem_id="input_description") 95 | run_button = gr.Button("Generate Audio", variant="primary") 96 | with gr.Column(): 97 | audio_out = gr.Audio(label="Parler-TTS generation", type="numpy", elem_id="audio_out") 98 | 99 | inputs = [input_text, description] 100 | outputs = [audio_out] 101 | gr.Examples(examples=examples, fn=gen_tts, inputs=inputs, outputs=outputs, cache_examples=True) 102 | run_button.click(fn=gen_tts, inputs=inputs, outputs=outputs, queue=True) 103 | 104 | block.queue() 105 | block.launch(share=True) 106 | -------------------------------------------------------------------------------- /helpers/model_init_scripts/init_dummy_model.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | 4 | from transformers import AutoConfig 5 | 6 | from parler_tts import ParlerTTSDecoderConfig, ParlerTTSForCausalLM, ParlerTTSForConditionalGeneration 7 | 8 | 9 | if __name__ == "__main__": 10 | parser = argparse.ArgumentParser() 11 | parser.add_argument("save_directory", type=str, help="Directory where to save the model and the decoder.") 12 | parser.add_argument("--text_model", type=str, help="Repository id or path to the text encoder.") 13 | parser.add_argument("--audio_model", type=str, help="Repository id or path to the audio encoder.") 14 | 15 | args = parser.parse_args() 16 | 17 | text_model = args.text_model 18 | encodec_version = args.audio_model 19 | 20 | t5 = AutoConfig.from_pretrained(text_model) 21 | encodec = AutoConfig.from_pretrained(encodec_version) 22 | 23 | encodec_vocab_size = encodec.codebook_size 24 | num_codebooks = encodec.num_codebooks 25 | print("num_codebooks", num_codebooks) 26 | 27 | decoder_config = ParlerTTSDecoderConfig( 28 | vocab_size=encodec_vocab_size + 1, 29 | max_position_embeddings=2048, 30 | num_hidden_layers=4, 31 | ffn_dim=512, 32 | num_attention_heads=8, 33 | layerdrop=0.0, 34 | use_cache=True, 35 | activation_function="gelu", 36 | hidden_size=512, 37 | dropout=0.0, 38 | attention_dropout=0.0, 39 | activation_dropout=0.0, 40 | pad_token_id=encodec_vocab_size, 41 | eos_token_id=encodec_vocab_size, 42 | bos_token_id=encodec_vocab_size + 1, 43 | num_codebooks=num_codebooks, 44 | ) 45 | 46 | decoder = ParlerTTSForCausalLM(decoder_config) 47 | decoder.save_pretrained(os.path.join(args.save_directory, "decoder")) 48 | 49 | model = ParlerTTSForConditionalGeneration.from_sub_models_pretrained( 50 | text_encoder_pretrained_model_name_or_path=text_model, 51 | audio_encoder_pretrained_model_name_or_path=encodec_version, 52 | decoder_pretrained_model_name_or_path=os.path.join(args.save_directory, "decoder"), 53 | vocab_size=t5.vocab_size, 54 | ) 55 | 56 | # set the appropriate bos/pad token ids 57 | model.generation_config.decoder_start_token_id = encodec_vocab_size + 1 58 | model.generation_config.pad_token_id = encodec_vocab_size 59 | model.generation_config.eos_token_id = encodec_vocab_size 60 | 61 | # set other default generation config params 62 | model.generation_config.max_length = int(30 * model.audio_encoder.config.frame_rate) 63 | model.generation_config.do_sample = True # True 64 | 65 | 66 | model.config.pad_token_id = encodec_vocab_size 67 | model.config.decoder_start_token_id = encodec_vocab_size + 1 68 | 69 | model.save_pretrained(os.path.join(args.save_directory, "tiny-model")) 70 | -------------------------------------------------------------------------------- /helpers/model_init_scripts/init_dummy_model_with_encodec.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | 4 | from transformers import AutoConfig 5 | 6 | from parler_tts import ParlerTTSDecoderConfig, ParlerTTSForCausalLM, ParlerTTSForConditionalGeneration 7 | 8 | 9 | if __name__ == "__main__": 10 | parser = argparse.ArgumentParser() 11 | parser.add_argument("save_directory", type=str, help="Directory where to save the model and the decoder.") 12 | args = parser.parse_args() 13 | 14 | text_model = "google-t5/t5-small" 15 | encodec_version = "facebook/encodec_24khz" 16 | 17 | t5 = AutoConfig.from_pretrained(text_model) 18 | encodec = AutoConfig.from_pretrained(encodec_version) 19 | 20 | encodec_vocab_size = encodec.codebook_size 21 | num_codebooks = 8 22 | print("num_codebooks", num_codebooks) 23 | 24 | decoder_config = ParlerTTSDecoderConfig( 25 | vocab_size=encodec_vocab_size + 1, 26 | max_position_embeddings=2048, 27 | num_hidden_layers=4, 28 | ffn_dim=512, 29 | num_attention_heads=8, 30 | layerdrop=0.0, 31 | use_cache=True, 32 | activation_function="gelu", 33 | hidden_size=512, 34 | dropout=0.0, 35 | attention_dropout=0.0, 36 | activation_dropout=0.0, 37 | pad_token_id=encodec_vocab_size, 38 | eos_token_id=encodec_vocab_size, 39 | bos_token_id=encodec_vocab_size + 1, 40 | num_codebooks=num_codebooks, 41 | ) 42 | 43 | decoder = ParlerTTSForCausalLM(decoder_config) 44 | 45 | decoder.save_pretrained(os.path.join(args.save_directory, "decoder")) 46 | 47 | model = ParlerTTSForConditionalGeneration.from_sub_models_pretrained( 48 | text_encoder_pretrained_model_name_or_path=text_model, 49 | audio_encoder_pretrained_model_name_or_path=encodec_version, 50 | decoder_pretrained_model_name_or_path=os.path.join(args.save_directory, "decoder"), 51 | vocab_size=t5.vocab_size, 52 | ) 53 | 54 | # set the appropriate bos/pad token ids 55 | model.generation_config.decoder_start_token_id = encodec_vocab_size + 1 56 | model.generation_config.pad_token_id = encodec_vocab_size 57 | model.generation_config.eos_token_id = encodec_vocab_size 58 | 59 | # set other default generation config params 60 | model.generation_config.max_length = int(30 * model.audio_encoder.config.frame_rate) 61 | model.generation_config.do_sample = True # True 62 | 63 | 64 | model.config.pad_token_id = encodec_vocab_size 65 | model.config.decoder_start_token_id = encodec_vocab_size + 1 66 | 67 | model.save_pretrained(os.path.join(args.save_directory, "tiny-model")) 68 | -------------------------------------------------------------------------------- /helpers/model_init_scripts/init_large_model.py: -------------------------------------------------------------------------------- 1 | from parler_tts import ParlerTTSForCausalLM, ParlerTTSForConditionalGeneration, ParlerTTSDecoderConfig 2 | from transformers import AutoConfig 3 | import os 4 | import argparse 5 | 6 | 7 | if __name__ == "__main__": 8 | parser = argparse.ArgumentParser() 9 | parser.add_argument("save_directory", type=str, help="Directory where to save the model and the decoder.") 10 | parser.add_argument("--text_model", type=str, help="Repository id or path to the text encoder.") 11 | parser.add_argument("--audio_model", type=str, help="Repository id or path to the audio encoder.") 12 | 13 | args = parser.parse_args() 14 | 15 | text_model = args.text_model 16 | encodec_version = args.audio_model 17 | 18 | t5 = AutoConfig.from_pretrained(text_model) 19 | encodec = AutoConfig.from_pretrained(encodec_version) 20 | 21 | encodec_vocab_size = encodec.codebook_size 22 | num_codebooks = encodec.num_codebooks 23 | print("num_codebooks", num_codebooks) 24 | 25 | decoder_config = ParlerTTSDecoderConfig( 26 | vocab_size=encodec_vocab_size + 64, # + 64 instead of +1 to have a multiple of 64 27 | max_position_embeddings=4096, # 30 s = 2580 28 | num_hidden_layers=30, 29 | ffn_dim=6144, 30 | num_attention_heads=24, 31 | num_key_value_heads=24, 32 | layerdrop=0.0, 33 | use_cache=True, 34 | activation_function="gelu", 35 | hidden_size=1536, 36 | dropout=0.1, 37 | attention_dropout=0.0, 38 | activation_dropout=0.0, 39 | pad_token_id=encodec_vocab_size, 40 | eos_token_id=encodec_vocab_size, 41 | bos_token_id=encodec_vocab_size + 1, 42 | num_codebooks=num_codebooks, 43 | ) 44 | 45 | decoder = ParlerTTSForCausalLM(decoder_config) 46 | decoder.save_pretrained(os.path.join(args.save_directory, "decoder")) 47 | 48 | model = ParlerTTSForConditionalGeneration.from_sub_models_pretrained( 49 | text_encoder_pretrained_model_name_or_path=text_model, 50 | audio_encoder_pretrained_model_name_or_path=encodec_version, 51 | decoder_pretrained_model_name_or_path=os.path.join(args.save_directory, "decoder"), 52 | vocab_size=t5.vocab_size, 53 | ) 54 | 55 | # set the appropriate bos/pad token ids 56 | model.generation_config.decoder_start_token_id = encodec_vocab_size + 1 57 | model.generation_config.pad_token_id = encodec_vocab_size 58 | model.generation_config.eos_token_id = encodec_vocab_size 59 | 60 | # set other default generation config params 61 | model.generation_config.max_length = int(30 * model.audio_encoder.config.frame_rate) 62 | model.generation_config.do_sample = True # True 63 | 64 | 65 | model.config.pad_token_id = encodec_vocab_size 66 | model.config.decoder_start_token_id = encodec_vocab_size + 1 67 | 68 | model.save_pretrained(os.path.join(args.save_directory, "parler-tts-untrained-larger/")) 69 | -------------------------------------------------------------------------------- /helpers/model_init_scripts/init_model_600M.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | 4 | from transformers import AutoConfig 5 | 6 | from parler_tts import ParlerTTSDecoderConfig, ParlerTTSForCausalLM, ParlerTTSForConditionalGeneration 7 | 8 | 9 | if __name__ == "__main__": 10 | parser = argparse.ArgumentParser() 11 | parser.add_argument("save_directory", type=str, help="Directory where to save the model and the decoder.") 12 | parser.add_argument("--text_model", type=str, help="Repository id or path to the text encoder.") 13 | parser.add_argument("--audio_model", type=str, help="Repository id or path to the audio encoder.") 14 | 15 | args = parser.parse_args() 16 | 17 | text_model = args.text_model 18 | encodec_version = args.audio_model 19 | 20 | t5 = AutoConfig.from_pretrained(text_model) 21 | encodec = AutoConfig.from_pretrained(encodec_version) 22 | 23 | encodec_vocab_size = encodec.codebook_size 24 | num_codebooks = encodec.num_codebooks 25 | print("num_codebooks", num_codebooks) 26 | 27 | decoder_config = ParlerTTSDecoderConfig( 28 | vocab_size=encodec_vocab_size + 64, # + 64 instead of +1 to have a multiple of 64 29 | max_position_embeddings=4096, # 30 s = 2580 30 | num_hidden_layers=24, 31 | ffn_dim=4096, 32 | num_attention_heads=16, 33 | layerdrop=0.0, 34 | use_cache=True, 35 | activation_function="gelu", 36 | hidden_size=1024, 37 | dropout=0.1, 38 | attention_dropout=0.0, 39 | activation_dropout=0.0, 40 | pad_token_id=encodec_vocab_size, 41 | eos_token_id=encodec_vocab_size, 42 | bos_token_id=encodec_vocab_size + 1, 43 | num_codebooks=num_codebooks, 44 | ) 45 | 46 | decoder = ParlerTTSForCausalLM(decoder_config) 47 | decoder.save_pretrained(os.path.join(args.save_directory, "decoder")) 48 | 49 | model = ParlerTTSForConditionalGeneration.from_sub_models_pretrained( 50 | text_encoder_pretrained_model_name_or_path=text_model, 51 | audio_encoder_pretrained_model_name_or_path=encodec_version, 52 | decoder_pretrained_model_name_or_path=os.path.join(args.save_directory, "decoder"), 53 | vocab_size=t5.vocab_size, 54 | ) 55 | 56 | # set the appropriate bos/pad token ids 57 | model.generation_config.decoder_start_token_id = encodec_vocab_size + 1 58 | model.generation_config.pad_token_id = encodec_vocab_size 59 | model.generation_config.eos_token_id = encodec_vocab_size 60 | 61 | # set other default generation config params 62 | model.generation_config.max_length = int(30 * model.audio_encoder.config.frame_rate) 63 | model.generation_config.do_sample = True # True 64 | 65 | model.config.pad_token_id = encodec_vocab_size 66 | model.config.decoder_start_token_id = encodec_vocab_size + 1 67 | 68 | model.save_pretrained(os.path.join(args.save_directory, "parler-tts-untrained-600M/")) 69 | -------------------------------------------------------------------------------- /helpers/push_to_hub_scripts/push_dac_to_hub.py: -------------------------------------------------------------------------------- 1 | import dac 2 | from transformers import AutoConfig, AutoModel, EncodecFeatureExtractor 3 | 4 | from parler_tts import DACConfig, DACModel 5 | from transformers import AutoConfig, AutoModel 6 | from transformers import EncodecFeatureExtractor 7 | 8 | from importlib.metadata import version 9 | from packaging.version import Version 10 | 11 | if Version(version("transformers"))<= Version("4.44.2dev"): 12 | AutoConfig.register("dac", DACConfig) 13 | else: 14 | AutoConfig.register("dac_on_the_hub", DACConfig) 15 | 16 | AutoModel.register(DACConfig, DACModel) 17 | 18 | # Download a model 19 | model_path = dac.utils.download(model_type="44khz") 20 | model = dac.DAC.load(model_path) 21 | 22 | hf_dac = DACModel(DACConfig()) 23 | hf_dac.model.load_state_dict(model.state_dict()) 24 | 25 | hf_dac.push_to_hub("parler-tts/dac_44khZ_8kbps") 26 | EncodecFeatureExtractor(sampling_rate=44100).push_to_hub("parler-tts/dac_44khZ_8kbps") 27 | -------------------------------------------------------------------------------- /helpers/push_to_hub_scripts/push_trained_parler_tts_to_hub.py: -------------------------------------------------------------------------------- 1 | from transformers import AutoFeatureExtractor, AutoTokenizer 2 | 3 | from parler_tts import ParlerTTSForConditionalGeneration 4 | 5 | 6 | path = "TODO" 7 | repo_id = "parler_tts_600M" 8 | 9 | 10 | AutoFeatureExtractor.from_pretrained("ylacombe/dac_44khZ_8kbps").push_to_hub(repo_id) 11 | AutoTokenizer.from_pretrained("google/t5-v1_1-base").push_to_hub(repo_id) 12 | 13 | ParlerTTSForConditionalGeneration.from_pretrained(path).push_to_hub(repo_id) 14 | -------------------------------------------------------------------------------- /helpers/training_configs/librispeech_tts_r_300M_dummy.json: -------------------------------------------------------------------------------- 1 | { 2 | "model_name_or_path": "./parler-tts-untrained-600M/parler-tts-untrained-600M/", 3 | "save_to_disk": "./tmp_dataset_audio/", 4 | "temporary_save_to_disk": "./audio_code_tmp/", 5 | 6 | 7 | "feature_extractor_name":"ylacombe/dac_44khZ_8kbps", 8 | "description_tokenizer_name":"google/flan-t5-base", 9 | "prompt_tokenizer_name":"google/flan-t5-base", 10 | 11 | "report_to": ["wandb"], 12 | "overwrite_output_dir": true, 13 | "output_dir": "./output_dir_training", 14 | 15 | "train_dataset_name": "blabble-io/libritts_r", 16 | "train_metadata_dataset_name": "parler-tts/libritts_r_tags_tagged_10k_generated", 17 | "train_dataset_config_name": "clean", 18 | "train_split_name": "test.clean", 19 | 20 | "eval_dataset_name": "blabble-io/libritts_r", 21 | "eval_metadata_dataset_name": "parler-tts/libritts_r_tags_tagged_10k_generated", 22 | "eval_dataset_config_name": "clean", 23 | "eval_split_name": "test.clean", 24 | 25 | "target_audio_column_name": "audio", 26 | "description_column_name": "text_description", 27 | "prompt_column_name": "text", 28 | 29 | "max_eval_samples": 48, 30 | "max_train_samples": 96, 31 | 32 | "max_duration_in_seconds": 20, 33 | "min_duration_in_seconds": 2.0, 34 | 35 | "add_audio_samples_to_wandb": true, 36 | "id_column_name": "id", 37 | 38 | "preprocessing_num_workers": 8, 39 | 40 | "do_train": true, 41 | "num_train_epochs": 50, 42 | "gradient_accumulation_steps": 1, 43 | "gradient_checkpointing": false, 44 | "per_device_train_batch_size": 4, 45 | "learning_rate": 1e-3, 46 | "adam_beta1": 0.9, 47 | "adam_beta2": 0.99, 48 | "weight_decay": 0.01, 49 | 50 | "lr_scheduler_type": "cosine", 51 | "warmup_steps": 40, 52 | 53 | 54 | "logging_steps": 2, 55 | "freeze_text_encoder": true, 56 | 57 | 58 | "do_eval": true, 59 | "predict_with_generate": true, 60 | "include_inputs_for_metrics": true, 61 | "evaluation_strategy": "steps", 62 | "eval_steps": 500, 63 | "save_steps": 5000, 64 | 65 | "per_device_eval_batch_size": 12, 66 | 67 | "audio_encoder_per_device_batch_size":24, 68 | "dtype": "bfloat16", 69 | "seed": 456, 70 | 71 | "dataloader_num_workers":8 72 | } 73 | -------------------------------------------------------------------------------- /helpers/training_configs/starting_point_0.01.json: -------------------------------------------------------------------------------- 1 | { 2 | "model_name_or_path": "./parler-tts-untrained-600M/parler-tts-untrained-600M/", 3 | "save_to_disk": "./tmp_dataset_audio/", 4 | "temporary_save_to_disk": "./audio_code_tmp/", 5 | 6 | 7 | "feature_extractor_name":"ylacombe/dac_44khZ_8kbps", 8 | "description_tokenizer_name":"google/flan-t5-base", 9 | "prompt_tokenizer_name":"google/flan-t5-base", 10 | 11 | "report_to": ["wandb"], 12 | "overwrite_output_dir": true, 13 | "output_dir": "./output_dir_training", 14 | 15 | "train_dataset_name": "blabble-io/libritts_r+blabble-io/libritts_r+blabble-io/libritts_r+parler-tts/mls_eng_10k", 16 | "train_metadata_dataset_name": "parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/mls-eng-10k-tags_tagged_10k_generated", 17 | "train_dataset_config_name": "clean+clean+other+default", 18 | "train_split_name": "train.clean.360+train.clean.100+train.other.500+train", 19 | 20 | "eval_dataset_name": "blabble-io/libritts_r+parler-tts/mls_eng_10k", 21 | "eval_metadata_dataset_name": "parler-tts/libritts_r_tags_tagged_10k_generated+parler-tts/mls-eng-10k-tags_tagged_10k_generated", 22 | "eval_dataset_config_name": "other+default", 23 | "eval_split_name": "test.other+test", 24 | 25 | "target_audio_column_name": "audio", 26 | "description_column_name": "text_description", 27 | "prompt_column_name": "text", 28 | 29 | "max_eval_samples": 96, 30 | 31 | "max_duration_in_seconds": 30, 32 | "min_duration_in_seconds": 2.0, 33 | "max_text_length": 400, 34 | 35 | "group_by_length": true, 36 | 37 | "add_audio_samples_to_wandb": true, 38 | "id_column_name": "id", 39 | 40 | "preprocessing_num_workers": 8, 41 | 42 | "do_train": true, 43 | "num_train_epochs": 40, 44 | "gradient_accumulation_steps": 8, 45 | "gradient_checkpointing": false, 46 | "per_device_train_batch_size": 3, 47 | "learning_rate": 0.00095, 48 | "adam_beta1": 0.9, 49 | "adam_beta2": 0.99, 50 | "weight_decay": 0.01, 51 | 52 | "lr_scheduler_type": "constant_with_warmup", 53 | "warmup_steps": 20000, 54 | 55 | 56 | "logging_steps": 1000, 57 | "freeze_text_encoder": true, 58 | 59 | 60 | "do_eval": true, 61 | "predict_with_generate": true, 62 | "include_inputs_for_metrics": true, 63 | "evaluation_strategy": "steps", 64 | "eval_steps": 10000, 65 | "save_steps": 10000, 66 | 67 | "per_device_eval_batch_size": 12, 68 | 69 | "audio_encoder_per_device_batch_size":20, 70 | "dtype": "bfloat16", 71 | "seed": 456, 72 | 73 | "dataloader_num_workers":8 74 | } 75 | -------------------------------------------------------------------------------- /helpers/training_configs/starting_point_v1.json: -------------------------------------------------------------------------------- 1 | { 2 | "model_name_or_path": "./parler-tts-untrained-600M/parler-tts-untrained-600M/", 3 | "save_to_disk": "./tmp_dataset_audio/", 4 | "temporary_save_to_disk": "./audio_code_tmp/", 5 | "wandb_project": "parler-tts-50k-hours", 6 | "wandb_run_name": "Mini", 7 | 8 | "feature_extractor_name":"ylacombe/dac_44khZ_8kbps", 9 | "description_tokenizer_name":"google/flan-t5-large", 10 | "prompt_tokenizer_name":"google/flan-t5-large", 11 | 12 | "report_to": ["wandb"], 13 | "overwrite_output_dir": true, 14 | "output_dir": "./output_dir_training", 15 | 16 | "train_dataset_name": "ylacombe/libritts_r_filtered+ylacombe/libritts_r_filtered+ylacombe/libritts_r_filtered+parler-tts/mls_eng", 17 | "train_metadata_dataset_name": "ylacombe/libritts-r-filtered-descriptions-10k-v5-without-accents+ylacombe/libritts-r-filtered-descriptions-10k-v5-without-accents+ylacombe/libritts-r-filtered-descriptions-10k-v5-without-accents+ylacombe/mls-eng-descriptions-v4", 18 | "train_dataset_config_name": "clean+clean+other+default", 19 | "train_split_name": "train.clean.360+train.clean.100+train.other.500+train", 20 | 21 | "eval_dataset_name": "ylacombe/libritts_r_filtered+parler-tts/mls_eng", 22 | "eval_metadata_dataset_name": "ylacombe/libritts-r-filtered-descriptions-10k-v5-without-accents+ylacombe/mls-eng-descriptions-v4", 23 | "eval_dataset_config_name": "other+default", 24 | "eval_split_name": "test.other+test", 25 | 26 | "target_audio_column_name": "audio", 27 | "description_column_name": "text_description", 28 | "prompt_column_name": "text", 29 | 30 | "max_eval_samples": 96, 31 | 32 | "max_duration_in_seconds": 30, 33 | "min_duration_in_seconds": 2.0, 34 | "max_text_length": 600, 35 | 36 | "group_by_length": true, 37 | 38 | "add_audio_samples_to_wandb": true, 39 | "id_column_name": "id", 40 | 41 | "preprocessing_num_workers": 8, 42 | 43 | "do_train": true, 44 | "num_train_epochs": 4, 45 | "gradient_accumulation_steps": 4, 46 | "gradient_checkpointing": false, 47 | "per_device_train_batch_size": 6, 48 | "learning_rate": 0.00095, 49 | "adam_beta1": 0.9, 50 | "adam_beta2": 0.99, 51 | "weight_decay": 0.01, 52 | 53 | "lr_scheduler_type": "constant_with_warmup", 54 | "warmup_steps": 20000, 55 | 56 | 57 | "logging_steps": 1000, 58 | "freeze_text_encoder": true, 59 | 60 | 61 | "do_eval": true, 62 | "predict_with_generate": true, 63 | "include_inputs_for_metrics": true, 64 | "evaluation_strategy": "steps", 65 | "eval_steps": 10000, 66 | "save_steps": 10000, 67 | 68 | "per_device_eval_batch_size": 4, 69 | 70 | "audio_encoder_per_device_batch_size":24, 71 | "dtype": "bfloat16", 72 | "seed": 456, 73 | 74 | "dataloader_num_workers":8, 75 | "attn_implementation": "sdpa" 76 | } -------------------------------------------------------------------------------- /helpers/training_configs/starting_point_v1_large.json: -------------------------------------------------------------------------------- 1 | { 2 | "model_name_or_path": "./parler-tts-untrained-large/parler-tts-untrained-large", 3 | "save_to_disk": "./tmp_dataset_audio/", 4 | "temporary_save_to_disk": "./audio_code_tmp/", 5 | "wandb_project": "parler-tts-50k-hours", 6 | "wandb_run_name": "Large", 7 | 8 | "feature_extractor_name":"ylacombe/dac_44khZ_8kbps", 9 | "description_tokenizer_name":"google/flan-t5-large", 10 | "prompt_tokenizer_name":"google/flan-t5-large", 11 | 12 | "report_to": ["wandb"], 13 | "overwrite_output_dir": true, 14 | "output_dir": "./output_dir_training", 15 | 16 | "train_dataset_name": "ylacombe/libritts_r_filtered+ylacombe/libritts_r_filtered+ylacombe/libritts_r_filtered+parler-tts/mls_eng", 17 | "train_metadata_dataset_name": "ylacombe/libritts-r-filtered-descriptions-10k-v5-without-accents+ylacombe/libritts-r-filtered-descriptions-10k-v5-without-accents+ylacombe/libritts-r-filtered-descriptions-10k-v5-without-accents+ylacombe/mls-eng-descriptions-v4", 18 | "train_dataset_config_name": "clean+clean+other+default", 19 | "train_split_name": "train.clean.360+train.clean.100+train.other.500+train", 20 | 21 | "eval_dataset_name": "ylacombe/libritts_r_filtered+parler-tts/mls_eng", 22 | "eval_metadata_dataset_name": "ylacombe/libritts-r-filtered-descriptions-10k-v5-without-accents+ylacombe/mls-eng-descriptions-v4", 23 | "eval_dataset_config_name": "other+default", 24 | "eval_split_name": "test.other+test", 25 | 26 | "target_audio_column_name": "audio", 27 | "description_column_name": "text_description", 28 | "prompt_column_name": "text", 29 | 30 | "max_eval_samples": 96, 31 | 32 | "max_duration_in_seconds": 30, 33 | "min_duration_in_seconds": 2.0, 34 | "max_text_length": 600, 35 | 36 | "group_by_length": true, 37 | 38 | "add_audio_samples_to_wandb": true, 39 | "id_column_name": "id", 40 | 41 | "preprocessing_num_workers": 8, 42 | 43 | "do_train": true, 44 | "num_train_epochs": 4, 45 | "gradient_accumulation_steps": 4, 46 | "gradient_checkpointing": false, 47 | "per_device_train_batch_size": 3, 48 | "learning_rate": 0.0015, 49 | "adam_beta1": 0.9, 50 | "adam_beta2": 0.99, 51 | "weight_decay": 0.01, 52 | 53 | "lr_scheduler_type": "constant_with_warmup", 54 | "warmup_steps": 10000, 55 | 56 | 57 | "logging_steps": 1000, 58 | "freeze_text_encoder": true, 59 | 60 | 61 | "do_eval": true, 62 | "predict_with_generate": true, 63 | "include_inputs_for_metrics": true, 64 | "evaluation_strategy": "steps", 65 | "eval_steps": 10000, 66 | "save_steps": 10000, 67 | "save_total_limit": 10, 68 | 69 | "per_device_eval_batch_size": 6, 70 | 71 | "audio_encoder_per_device_batch_size":24, 72 | "dtype": "bfloat16", 73 | "seed": 738, 74 | 75 | "dataloader_num_workers":8, 76 | "attn_implementation": "sdpa" 77 | } 78 | -------------------------------------------------------------------------------- /parler_tts/__init__.py: -------------------------------------------------------------------------------- 1 | __version__ = "0.2.2" 2 | 3 | 4 | from transformers import AutoConfig, AutoModel 5 | 6 | from .configuration_parler_tts import ParlerTTSConfig, ParlerTTSDecoderConfig 7 | from .dac_wrapper import DACConfig, DACModel 8 | from .modeling_parler_tts import ( 9 | ParlerTTSForCausalLM, 10 | ParlerTTSForConditionalGeneration, 11 | apply_delay_pattern_mask, 12 | build_delay_pattern_mask, 13 | ) 14 | 15 | from .streamer import ParlerTTSStreamer 16 | 17 | from importlib.metadata import version 18 | from packaging.version import Version 19 | 20 | if Version(version("transformers"))<= Version("4.44.2dev"): 21 | AutoConfig.register("dac", DACConfig) 22 | else: 23 | AutoConfig.register("dac_on_the_hub", DACConfig) 24 | 25 | AutoModel.register(DACConfig, DACModel) 26 | -------------------------------------------------------------------------------- /parler_tts/configuration_parler_tts.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2024 and The HuggingFace Inc. team. All rights reserved. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """ Parler-TTS model configuration""" 16 | 17 | from transformers import AutoConfig, logging 18 | from transformers.configuration_utils import PretrainedConfig 19 | 20 | from importlib.metadata import version 21 | from packaging.version import Version 22 | 23 | use_dac_on_the_hub = Version(version("transformers")) > Version("4.44.2dev") 24 | 25 | logger = logging.get_logger(__name__) 26 | 27 | PARLER_TTS_PRETRAINED_CONFIG_ARCHIVE_MAP = { 28 | "parler-tts/parler-tts-mini-v1": "https://huggingface.co/parler-tts/parler-tts-mini-v1/resolve/main/config.json", 29 | # See all ParlerTTS models at https://huggingface.co/models?filter=parler_tts 30 | } 31 | 32 | 33 | class ParlerTTSDecoderConfig(PretrainedConfig): 34 | r""" 35 | This is the configuration class to store the configuration of an [`ParlerTTSDecoder`]. It is used to instantiate a 36 | Parler-TTS decoder according to the specified arguments, defining the model architecture. Instantiating a 37 | configuration with the defaults will yield a similar configuration to that of the Parler-TTS 38 | [parler-tts/parler-tts-mini-v1](https://huggingface.co/parler-tts/parler-tts-mini-v1) architecture. 39 | 40 | Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the 41 | documentation from [`PretrainedConfig`] for more information. 42 | 43 | 44 | Args: 45 | vocab_size (`int`, *optional*, defaults to 2049): 46 | Vocabulary size of the ParlerTTSDecoder model. Defines the number of different tokens that can be 47 | represented by the `inputs_ids` passed when calling [`ParlerTTSDecoder`]. 48 | hidden_size (`int`, *optional*, defaults to 1024): 49 | Dimensionality of the layers and the pooler layer. 50 | num_hidden_layers (`int`, *optional*, defaults to 24): 51 | Number of decoder layers. 52 | num_attention_heads (`int`, *optional*, defaults to 16): 53 | Number of attention heads for each attention layer in the Transformer block. 54 | num_key_value_heads (`int`, *optional*): 55 | This is the number of key_value heads that should be used to implement Grouped Query Attention. If 56 | `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if 57 | `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When 58 | converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed 59 | by meanpooling all the original heads within that group. For more details checkout [this 60 | paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to 61 | `num_attention_heads`. 62 | num_cross_attention_key_value_heads (`int`, *optional*): 63 | This is the number of key_value heads that should be used to implement Grouped Query Attention in the cross-attention layers. 64 | If it is not specified, will default to `num_key_value_heads`. 65 | ffn_dim (`int`, *optional*, defaults to 4096): 66 | Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer block. 67 | activation_function (`str` or `function`, *optional*, defaults to `"gelu"`): 68 | The non-linear activation function (function or string) in the decoder and pooler. If string, `"gelu"`, 69 | `"relu"`, `"silu"` and `"gelu_new"` are supported. 70 | dropout (`float`, *optional*, defaults to 0.1): 71 | The dropout probability for all fully connected layers in the embeddings, text_encoder, and pooler. 72 | attention_dropout (`float`, *optional*, defaults to 0.0): 73 | The dropout ratio for the attention probabilities. 74 | activation_dropout (`float`, *optional*, defaults to 0.0): 75 | The dropout ratio for activations inside the fully connected layer. 76 | max_position_embeddings (`int`, *optional*, defaults to 2048): 77 | The maximum sequence length that this model might ever be used with. Typically, set this to something large 78 | just in case (e.g., 512 or 1024 or 2048). 79 | initializer_factor (`float`, *optional*, defaults to 0.02): 80 | The standard deviation of the truncated_normal_initializer for initializing all weight matrices. 81 | layerdrop (`float`, *optional*, defaults to 0.0): 82 | The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556) 83 | for more details. 84 | scale_embedding (`bool`, *optional*, defaults to `False`): 85 | Scale embeddings by diving by sqrt(hidden_size). 86 | use_cache (`bool`, *optional*, defaults to `True`): 87 | Whether the model should return the last key/values attentions (not used by all models) 88 | num_codebooks (`int`, *optional*, defaults to 4): 89 | The number of parallel codebooks forwarded to the model. 90 | tie_word_embeddings(`bool`, *optional*, defaults to `False`): 91 | Whether input and output word embeddings should be tied. 92 | rope_embeddings (`bool`, *optional*, defaults to `False`): 93 | Whether to use ROPE or absolute positional embeddings. 94 | rope_theta (`float`, *optional*, defaults to 100000.0): 95 | The base period of the RoPE embeddings. 96 | cross_attention_implementation_strategy (`str`, *optional*): 97 | If not specified, the cross-attention implementation will be the same as `_attn_implementation`. If `always_eager`, it will always be the eager implementation. If `always_sdpa`, it will always be the sdpa implementation. 98 | use_fused_lm_heads(`bool`, *optional*, defaults to `False`): 99 | Whether to fuse audio LM heads instead of applying them sequentially. 100 | codebook_weights(`List[int]`, *optional*): 101 | Weights applied to each codebook when computing the loss. 102 | """ 103 | 104 | model_type = "parler_tts_decoder" 105 | keys_to_ignore_at_inference = ["past_key_values"] 106 | 107 | def __init__( 108 | self, 109 | vocab_size=2049, # vocab size = 2048 (encodec vocab size) + 1 (eos) 110 | max_position_embeddings=2048, 111 | num_hidden_layers=24, 112 | ffn_dim=4096, 113 | num_attention_heads=16, 114 | num_key_value_heads=None, 115 | num_cross_attention_key_value_heads=None, 116 | layerdrop=0.0, 117 | use_cache=True, 118 | activation_function="gelu", 119 | hidden_size=1024, 120 | dropout=0.1, 121 | attention_dropout=0.0, 122 | activation_dropout=0.0, 123 | initializer_factor=0.02, 124 | scale_embedding=False, 125 | num_codebooks=4, 126 | pad_token_id=2048, 127 | bos_token_id=2049, 128 | eos_token_id=2048, 129 | tie_word_embeddings=False, 130 | rope_embeddings=False, 131 | rope_theta=10_000.0, 132 | cross_attention_implementation_strategy=None, 133 | use_fused_lm_heads=False, 134 | codebook_weights=None, 135 | **kwargs, 136 | ): 137 | self.vocab_size = vocab_size 138 | self.max_position_embeddings = max_position_embeddings 139 | self.hidden_size = hidden_size 140 | self.ffn_dim = ffn_dim 141 | self.num_hidden_layers = num_hidden_layers 142 | self.num_attention_heads = num_attention_heads 143 | if num_key_value_heads is None: 144 | num_key_value_heads = num_attention_heads 145 | self.num_key_value_heads = num_key_value_heads 146 | if num_cross_attention_key_value_heads is None: 147 | num_cross_attention_key_value_heads = num_key_value_heads 148 | self.num_cross_attention_key_value_heads = num_cross_attention_key_value_heads 149 | self.dropout = dropout 150 | self.attention_dropout = attention_dropout 151 | self.activation_dropout = activation_dropout 152 | self.activation_function = activation_function 153 | self.initializer_factor = initializer_factor 154 | self.layerdrop = layerdrop 155 | self.use_cache = use_cache 156 | self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True 157 | self.num_codebooks = num_codebooks 158 | self.rope_embeddings = rope_embeddings 159 | self.rope_theta = rope_theta 160 | self.cross_attention_implementation_strategy = cross_attention_implementation_strategy 161 | self.use_fused_lm_heads = use_fused_lm_heads 162 | self.codebook_weights = codebook_weights 163 | 164 | if codebook_weights is not None and len(codebook_weights) != num_codebooks: 165 | raise ValueError(f"`codebook_weights` has length {len(codebook_weights)} when it should be of length {num_codebooks}.") 166 | super().__init__( 167 | pad_token_id=pad_token_id, 168 | bos_token_id=bos_token_id, 169 | eos_token_id=eos_token_id, 170 | tie_word_embeddings=tie_word_embeddings, 171 | **kwargs, 172 | ) 173 | 174 | 175 | class ParlerTTSConfig(PretrainedConfig): 176 | r""" 177 | This is the configuration class to store the configuration of a [`ParlerTTSModel`]. It is used to instantiate a 178 | Parler-TTS model according to the specified arguments, defining the text encoder, audio encoder and Parler-TTS decoder 179 | configs. 180 | 181 | Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the 182 | documentation from [`PretrainedConfig`] for more information. 183 | 184 | Args: 185 | vocab_size (`int`, *optional*, defaults to 1024): 186 | Vocabulary size of the prompt token ids. Defines the number of different tokens that can be 187 | represented by the `prompt_inputs_ids`. 188 | prompt_cross_attention (`bool`, *optional*, defaults to `False`): 189 | Whether to use cross-attention conditioning for the prompt (as well as the description). 190 | kwargs (*optional*): 191 | Dictionary of keyword arguments. Notably: 192 | 193 | - **text_encoder** ([`PretrainedConfig`], *optional*) -- An instance of a configuration object that 194 | defines the text encoder config. 195 | - **audio_encoder** ([`PretrainedConfig`], *optional*) -- An instance of a configuration object that 196 | defines the audio encoder config. 197 | - **decoder** ([`PretrainedConfig`], *optional*) -- An instance of a configuration object that defines 198 | the decoder config. 199 | 200 | Example: 201 | 202 | ```python 203 | >>> from transformers import ( 204 | ... ParlerTTSConfig, 205 | ... ParlerTTSDecoderConfig, 206 | ... T5Config, 207 | ... EncodecConfig, 208 | ... ParlerTTSForConditionalGeneration, 209 | ... ) 210 | 211 | >>> # Initializing text encoder, audio encoder, and decoder model configurations 212 | >>> text_encoder_config = T5Config() 213 | >>> audio_encoder_config = EncodecConfig() 214 | >>> decoder_config = ParlerTTSDecoderConfig() 215 | 216 | >>> configuration = ParlerTTSConfig.from_sub_models_config( 217 | ... text_encoder_config, audio_encoder_config, decoder_config 218 | ... ) 219 | 220 | >>> # Initializing a ParlerTTSForConditionalGeneration (with random weights) from the parler-tts/parler-tts-mini-v1 style configuration 221 | >>> model = ParlerTTSForConditionalGeneration(configuration) 222 | 223 | >>> # Accessing the model configuration 224 | >>> configuration = model.config 225 | >>> config_text_encoder = model.config.text_encoder 226 | >>> config_audio_encoder = model.config.audio_encoder 227 | >>> config_decoder = model.config.decoder 228 | 229 | >>> # Saving the model, including its configuration 230 | >>> model.save_pretrained("parler_tts-model") 231 | 232 | >>> # loading model and config from pretrained folder 233 | >>> parler_tts_config = ParlerTTSConfig.from_pretrained("parler_tts-model") 234 | >>> model = ParlerTTSForConditionalGeneration.from_pretrained("parler_tts-model", config=parler_tts_config) 235 | ```""" 236 | 237 | model_type = "parler_tts" 238 | is_composition = True 239 | 240 | def __init__(self, vocab_size=1024, prompt_cross_attention=False, **kwargs): 241 | super().__init__(**kwargs) 242 | if "text_encoder" not in kwargs or "audio_encoder" not in kwargs or "decoder" not in kwargs: 243 | raise ValueError("Config has to be initialized with text_encoder, audio_encoder and decoder config") 244 | 245 | text_encoder_config = kwargs.pop("text_encoder") 246 | text_encoder_model_type = text_encoder_config.pop("model_type") 247 | 248 | audio_encoder_config = kwargs.pop("audio_encoder") 249 | audio_encoder_model_type = audio_encoder_config.pop("model_type") 250 | 251 | model_version = kwargs.get("transformers_version", None) 252 | if model_version is not None and Version(model_version) <= Version("4.44.2dev") and use_dac_on_the_hub and audio_encoder_model_type=="dac": 253 | # here we have to manually change model type if DAC based on transformers version 254 | audio_encoder_model_type = "dac_on_the_hub" 255 | 256 | decoder_config = kwargs.pop("decoder") 257 | 258 | self.vocab_size = vocab_size 259 | self.prompt_cross_attention = prompt_cross_attention 260 | self.text_encoder = AutoConfig.for_model(text_encoder_model_type, **text_encoder_config) 261 | self.audio_encoder = AutoConfig.for_model(audio_encoder_model_type, **audio_encoder_config) 262 | self.decoder = ParlerTTSDecoderConfig(**decoder_config) 263 | self.is_encoder_decoder = True 264 | 265 | @classmethod 266 | def from_sub_models_config( 267 | cls, 268 | text_encoder_config: PretrainedConfig, 269 | audio_encoder_config: PretrainedConfig, 270 | decoder_config: ParlerTTSDecoderConfig, 271 | **kwargs, 272 | ): 273 | r""" 274 | Instantiate a [`ParlerTTSConfig`] (or a derived class) from text encoder, audio encoder and decoder 275 | configurations. 276 | 277 | Returns: 278 | [`ParlerTTSConfig`]: An instance of a configuration object 279 | """ 280 | 281 | return cls( 282 | text_encoder=text_encoder_config.to_dict(), 283 | audio_encoder=audio_encoder_config.to_dict(), 284 | decoder=decoder_config.to_dict(), 285 | **kwargs, 286 | ) 287 | 288 | @property 289 | # This is a property because you might want to change the codec model on the fly 290 | def sampling_rate(self): 291 | return self.audio_encoder.sampling_rate -------------------------------------------------------------------------------- /parler_tts/dac_wrapper/__init__.py: -------------------------------------------------------------------------------- 1 | from .configuration_dac import DACConfig 2 | from .modeling_dac import DACModel 3 | -------------------------------------------------------------------------------- /parler_tts/dac_wrapper/configuration_dac.py: -------------------------------------------------------------------------------- 1 | 2 | from transformers import PretrainedConfig 3 | from importlib.metadata import version 4 | from packaging.version import Version 5 | 6 | 7 | class DACConfig(PretrainedConfig): 8 | model_type = "dac" if Version(version("transformers"))<= Version("4.44.2dev") else "dac_on_the_hub" 9 | 10 | def __init__( 11 | self, 12 | num_codebooks: int = 9, 13 | model_bitrate: int = 8, # kbps 14 | codebook_size: int = 1024, 15 | latent_dim: int = 1024, 16 | frame_rate: int = 86, 17 | sampling_rate: int = 44100, 18 | **kwargs, 19 | ): 20 | self.codebook_size = codebook_size 21 | self.model_bitrate = model_bitrate 22 | self.latent_dim = latent_dim 23 | self.num_codebooks = num_codebooks 24 | self.frame_rate = frame_rate 25 | self.sampling_rate = sampling_rate 26 | 27 | super().__init__(**kwargs) 28 | -------------------------------------------------------------------------------- /parler_tts/dac_wrapper/modeling_dac.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from dac.model import DAC 3 | from torch import nn 4 | 5 | from transformers import PreTrainedModel 6 | from transformers.models.encodec.modeling_encodec import EncodecDecoderOutput, EncodecEncoderOutput 7 | 8 | from .configuration_dac import DACConfig 9 | 10 | 11 | # model doesn't support batching yet 12 | 13 | 14 | class DACModel(PreTrainedModel): 15 | config_class = DACConfig 16 | main_input_name = "input_values" 17 | 18 | # Set main input to 'input_values' for voice steering 19 | main_input_name = "input_values" 20 | 21 | def __init__(self, config): 22 | super().__init__(config) 23 | 24 | self.model = DAC( 25 | n_codebooks=config.num_codebooks, 26 | latent_dim=config.latent_dim, 27 | codebook_size=config.codebook_size, 28 | ) 29 | 30 | self.remove_weight_norm() 31 | self.apply_weight_norm() 32 | 33 | def encode( 34 | self, input_values, padding_mask=None, bandwidth=None, return_dict=None, n_quantizers=None, sample_rate=None 35 | ): 36 | """ 37 | Encodes the input audio waveform into discrete codes. 38 | 39 | Args: 40 | input_values (`torch.Tensor` of shape `(batch_size, channels, sequence_length)`): 41 | Float values of the input audio waveform. 42 | padding_mask (`torch.Tensor` of shape `(batch_size, channels, sequence_length)`): 43 | Padding mask used to pad the `input_values`. 44 | bandwidth (`float`, *optional*): 45 | Not used, kept to have the same inferface as HF encodec. 46 | n_quantizers (`int`, *optional*) : 47 | Number of quantizers to use, by default None 48 | If None, all quantizers are used. 49 | sample_rate (`int`, *optional*) : 50 | Signal sampling_rate 51 | 52 | Returns: 53 | A list of frames containing the discrete encoded codes for the input audio waveform, along with rescaling 54 | factors for each chunk when `normalize` is True. Each frames is a tuple `(codebook, scale)`, with 55 | `codebook` of shape `[batch_size, num_codebooks, frames]`. 56 | Scale is not used here. 57 | 58 | """ 59 | _, channels, input_length = input_values.shape 60 | 61 | if channels < 1 or channels > 2: 62 | raise ValueError(f"Number of audio channels must be 1 or 2, but got {channels}") 63 | 64 | audio_data = self.model.preprocess(input_values, sample_rate) 65 | 66 | return_dict = return_dict if return_dict is not None else self.config.return_dict 67 | 68 | # TODO: for now, no chunk length 69 | 70 | chunk_length = None # self.config.chunk_length 71 | if chunk_length is None: 72 | chunk_length = input_length 73 | stride = input_length 74 | else: 75 | stride = self.config.chunk_stride 76 | 77 | if padding_mask is None: 78 | padding_mask = torch.ones_like(input_values).bool() 79 | 80 | encoded_frames = [] 81 | scales = [] 82 | 83 | step = chunk_length - stride 84 | if (input_length % stride) - step != 0: 85 | raise ValueError( 86 | "The input length is not properly padded for batched chunked decoding. Make sure to pad the input correctly." 87 | ) 88 | 89 | for offset in range(0, input_length - step, stride): 90 | mask = padding_mask[..., offset : offset + chunk_length].bool() 91 | frame = audio_data[:, :, offset : offset + chunk_length] 92 | 93 | scale = None 94 | 95 | _, encoded_frame, _, _, _ = self.model.encode(frame, n_quantizers=n_quantizers) 96 | encoded_frames.append(encoded_frame) 97 | scales.append(scale) 98 | 99 | encoded_frames = torch.stack(encoded_frames) 100 | 101 | if not return_dict: 102 | return (encoded_frames, scales) 103 | 104 | return EncodecEncoderOutput(encoded_frames, scales) 105 | 106 | def decode( 107 | self, 108 | audio_codes, 109 | audio_scales, 110 | padding_mask=None, 111 | return_dict=None, 112 | ): 113 | """ 114 | Decodes the given frames into an output audio waveform. 115 | 116 | Note that the output might be a bit bigger than the input. In that case, any extra steps at the end can be 117 | trimmed. 118 | 119 | Args: 120 | audio_codes (`torch.FloatTensor` of shape `(batch_size, nb_chunks, chunk_length)`, *optional*): 121 | Discret code embeddings computed using `model.encode`. 122 | audio_scales (`torch.Tensor` of shape `(batch_size, nb_chunks)`, *optional*): 123 | Not used, kept to have the same inferface as HF encodec. 124 | padding_mask (`torch.Tensor` of shape `(batch_size, channels, sequence_length)`): 125 | Padding mask used to pad the `input_values`. 126 | Not used yet, kept to have the same inferface as HF encodec. 127 | return_dict (`bool`, *optional*): 128 | Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. 129 | 130 | """ 131 | return_dict = return_dict or self.config.return_dict 132 | 133 | # TODO: for now, no chunk length 134 | 135 | if len(audio_codes) != 1: 136 | raise ValueError(f"Expected one frame, got {len(audio_codes)}") 137 | 138 | audio_values = self.model.quantizer.from_codes(audio_codes.squeeze(0))[0] 139 | audio_values = self.model.decode(audio_values) 140 | if not return_dict: 141 | return (audio_values,) 142 | return EncodecDecoderOutput(audio_values) 143 | 144 | def forward(self, tensor): 145 | raise ValueError("`DACModel.forward` not implemented yet") 146 | 147 | 148 | def apply_weight_norm(self): 149 | weight_norm = nn.utils.weight_norm 150 | if hasattr(nn.utils.parametrizations, "weight_norm"): 151 | weight_norm = nn.utils.parametrizations.weight_norm 152 | 153 | def _apply_weight_norm(module): 154 | if isinstance(module, nn.Conv1d) or isinstance(module, nn.ConvTranspose1d): 155 | weight_norm(module) 156 | 157 | self.apply(_apply_weight_norm) 158 | 159 | 160 | def remove_weight_norm(self): 161 | def _remove_weight_norm(module): 162 | if isinstance(module, nn.Conv1d) or isinstance(module, nn.ConvTranspose1d): 163 | nn.utils.remove_weight_norm(module) 164 | self.apply(_remove_weight_norm) 165 | -------------------------------------------------------------------------------- /parler_tts/logits_processors.py: -------------------------------------------------------------------------------- 1 | from transformers import LogitsProcessor, LogitsProcessorList 2 | from transformers.pytorch_utils import isin_mps_friendly 3 | import math 4 | import torch 5 | 6 | class ParlerTTSLogitsProcessor(LogitsProcessor): 7 | r"""This processor ensures that the delayed pattern mask constraints are respected. 8 | 9 | 10 | 11 | This logits processor is exclusively compatible with Parler-TTS. 12 | See the model documentation for examples. 13 | 14 | 15 | 16 | Args: 17 | eos_token_id (`Union[int, List[int], torch.Tensor]`): 18 | The id(s) of the *end-of-sequence* token. 19 | min_eos_p (`float`, *optional*): 20 | Minimum end of speech threshold. 21 | """ 22 | 23 | def __init__(self, eos_token_id, num_codebooks: int, batch_size: int, device: str = "cpu"): 24 | if not isinstance(eos_token_id, torch.Tensor): 25 | if isinstance(eos_token_id, int): 26 | eos_token_id = [eos_token_id] 27 | eos_token_id = torch.tensor(eos_token_id, device=device) 28 | self.eos_token_id = eos_token_id 29 | self.batch_size = batch_size 30 | 31 | if torch.is_floating_point(eos_token_id) or (eos_token_id < 0).any(): 32 | raise ValueError(f"`eos_token_id` has to be a list of positive integers, but is {eos_token_id}") 33 | 34 | self.num_codebooks = num_codebooks 35 | self.device = device 36 | 37 | 38 | self.codebook_idx = torch.arange(self.batch_size*self.num_codebooks, device=self.device) 39 | self.first_codebooks_unfinished = torch.arange(batch_size, device=device)*num_codebooks 40 | 41 | max_codebooks = torch.arange(self.batch_size, device=self.device)*self.num_codebooks + self.num_codebooks -1 42 | self.max_codebooks = max_codebooks 43 | 44 | def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor: 45 | 46 | is_eos = isin_mps_friendly(input_ids, self.eos_token_id).sum(1) 47 | 48 | self.first_codebooks_unfinished = torch.where((is_eos[self.first_codebooks_unfinished]>0) & (self.first_codebooks_unfinished self.first_codebooks_unfinished.repeat_interleave(self.num_codebooks) 52 | scores[eos_token_mask, self.eos_token_id] = -math.inf 53 | 54 | return scores -------------------------------------------------------------------------------- /parler_tts/streamer.py: -------------------------------------------------------------------------------- 1 | 2 | from .modeling_parler_tts import ParlerTTSForConditionalGeneration 3 | from transformers.generation.streamers import BaseStreamer 4 | from typing import Optional 5 | import torch 6 | import numpy as np 7 | import math 8 | from queue import Queue 9 | 10 | 11 | class ParlerTTSStreamer(BaseStreamer): 12 | def __init__( 13 | self, 14 | model: ParlerTTSForConditionalGeneration, 15 | device: Optional[str] = None, 16 | play_steps: Optional[int] = 10, 17 | stride: Optional[int] = None, 18 | timeout: Optional[float] = None, 19 | ): 20 | """ 21 | Streamer that stores playback-ready audio in a queue, to be used by a downstream application as an iterator. This is 22 | useful for applications that benefit from accessing the generated audio in a non-blocking way (e.g. in an interactive 23 | Gradio demo). 24 | Parameters: 25 | model (`ParlerTTSForConditionalGeneration`): 26 | The Parler-TTS model used to generate the audio waveform. 27 | device (`str`, *optional*): 28 | The torch device on which to run the computation. If `None`, will default to the device of the model. 29 | play_steps (`int`, *optional*, defaults to 10): 30 | The number of generation steps with which to return the generated audio array. Using fewer steps will 31 | mean the first chunk is ready faster, but will require more codec decoding steps overall. This value 32 | should be tuned to your device and latency requirements. 33 | stride (`int`, *optional*): 34 | The window (stride) between adjacent audio samples. Using a stride between adjacent audio samples reduces 35 | the hard boundary between them, giving smoother playback. If `None`, will default to a value equivalent to 36 | play_steps // 6 in the audio space. 37 | timeout (`int`, *optional*): 38 | The timeout for the audio queue. If `None`, the queue will block indefinitely. Useful to handle exceptions 39 | in `.generate()`, when it is called in a separate thread. 40 | """ 41 | self.decoder = model.decoder 42 | self.audio_encoder = model.audio_encoder 43 | self.generation_config = model.generation_config 44 | self.device = device if device is not None else model.device 45 | self.use_audio_scales = model.use_audio_scales 46 | self.use_4dim_audio_codes = model.use_4dim_audio_codes 47 | self.audio_kwargs = {} 48 | if self.use_audio_scales: 49 | self.audio_kwargs["audio_scales"] = [None] 50 | 51 | # variables used in the streaming process 52 | self.play_steps = play_steps 53 | if stride is not None: 54 | self.stride = stride 55 | else: 56 | hop_length = math.floor(self.audio_encoder.config.sampling_rate / self.audio_encoder.config.frame_rate) 57 | self.stride = hop_length * (play_steps - self.decoder.num_codebooks) // 6 58 | self.token_cache = None 59 | self.to_yield = 0 60 | 61 | # varibles used in the thread process 62 | self.audio_queue = Queue() 63 | self.stop_signal = None 64 | self.timeout = timeout 65 | 66 | def apply_delay_pattern_mask(self, input_ids): 67 | # build the delay pattern mask for offsetting each codebook prediction by 1 (this behaviour is specific to Parler) 68 | _, delay_pattern_mask = self.decoder.build_delay_pattern_mask( 69 | input_ids[:, :1], 70 | bos_token_id=self.generation_config.bos_token_id, 71 | pad_token_id=self.generation_config.decoder_start_token_id, 72 | max_length=input_ids.shape[-1], 73 | ) 74 | # apply the pattern mask to the input ids 75 | input_ids = self.decoder.apply_delay_pattern_mask(input_ids, delay_pattern_mask) 76 | 77 | # revert the pattern delay mask by filtering the pad token id 78 | mask = (delay_pattern_mask != self.generation_config.bos_token_id) & (delay_pattern_mask != self.generation_config.pad_token_id) 79 | input_ids = input_ids[mask].reshape(1, self.decoder.num_codebooks, -1) 80 | 81 | if self.use_4dim_audio_codes: 82 | # append the frame dimension back to the audio codes 83 | input_ids = input_ids[None, ...] 84 | 85 | # send the input_ids to the correct device 86 | input_ids = input_ids.to(self.audio_encoder.device) 87 | 88 | decode_sequentially = ( 89 | self.generation_config.bos_token_id in input_ids 90 | or self.generation_config.pad_token_id in input_ids 91 | or self.generation_config.eos_token_id in input_ids 92 | ) 93 | if not decode_sequentially: 94 | sample = self.audio_encoder.decode( 95 | audio_codes=input_ids, 96 | **self.audio_kwargs, 97 | ).audio_values 98 | output_values = sample if sample.ndim == 3 else sample.unsqueeze(0) 99 | else: 100 | sample = input_ids[:, 0] if self.use_4dim_audio_codes else input_ids[0] 101 | sample_mask = ((sample >= self.audio_encoder.config.codebook_size).sum(dim=(0, 1)) == 0) if self.use_4dim_audio_codes else ((sample >= self.audio_encoder.config.codebook_size).sum(dim=0) == 0) 102 | sample = sample[:, :, sample_mask] if self.use_4dim_audio_codes else sample[:, sample_mask] 103 | sample = self.audio_encoder.decode(audio_codes=sample[None, ...], **self.audio_kwargs).audio_values 104 | output_values = sample if sample.ndim == 3 else sample.unsqueeze(0) 105 | 106 | audio_values = output_values[0, 0] 107 | return audio_values.cpu().float().numpy() 108 | 109 | def put(self, value): 110 | batch_size = value.shape[0] // self.decoder.num_codebooks 111 | if batch_size > 1: 112 | raise ValueError("ParlerTTSStreamer only supports batch size 1") 113 | 114 | if self.token_cache is None: 115 | self.token_cache = value 116 | else: 117 | self.token_cache = torch.concatenate([self.token_cache, value[:, None]], dim=-1) 118 | 119 | if self.token_cache.shape[-1] % self.play_steps == 0: 120 | audio_values = self.apply_delay_pattern_mask(self.token_cache) 121 | self.on_finalized_audio(audio_values[self.to_yield : -self.stride]) 122 | self.to_yield += len(audio_values) - self.to_yield - self.stride 123 | 124 | def end(self): 125 | """Flushes any remaining cache and appends the stop symbol.""" 126 | if self.token_cache is not None: 127 | audio_values = self.apply_delay_pattern_mask(self.token_cache) 128 | else: 129 | audio_values = np.zeros(self.to_yield) 130 | 131 | self.on_finalized_audio(audio_values[self.to_yield :], stream_end=True) 132 | 133 | def on_finalized_audio(self, audio: np.ndarray, stream_end: bool = False): 134 | """Put the new audio in the queue. If the stream is ending, also put a stop signal in the queue.""" 135 | self.audio_queue.put(audio, timeout=self.timeout) 136 | if stream_end: 137 | self.audio_queue.put(self.stop_signal, timeout=self.timeout) 138 | 139 | def __iter__(self): 140 | return self 141 | 142 | def __next__(self): 143 | value = self.audio_queue.get(timeout=self.timeout) 144 | if not isinstance(value, np.ndarray) and value == self.stop_signal: 145 | raise StopIteration() 146 | else: 147 | return value -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [tool.black] 2 | line-length = 119 3 | target-version = ['py37'] 4 | 5 | [tool.ruff] 6 | # Never enforce `E501` (line length violations). 7 | ignore = ["C901", "E501", "E741", "W605"] 8 | select = ["C", "E", "F", "I", "W"] 9 | line-length = 119 10 | 11 | # Ignore import violations in all `__init__.py` files. 12 | [tool.ruff.per-file-ignores] 13 | "__init__.py" = ["E402", "F401", "F403", "F811"] 14 | 15 | [tool.ruff.isort] 16 | lines-after-imports = 2 17 | known-first-party = ["distil_whisper"] -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 The HuggingFace Team. All rights reserved. 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import os 16 | 17 | import setuptools 18 | 19 | 20 | _deps = [ 21 | "transformers>=4.46.1,<=4.46.1", 22 | "torch", 23 | "sentencepiece", 24 | "descript-audio-codec", 25 | "descript-audiotools @ git+https://github.com/descriptinc/audiotools", # temporary fix as long as 0.7.4 is not published 26 | "protobuf>=4.0.0" 27 | ] 28 | 29 | _extras_dev_deps = [ 30 | "black~=23.1", 31 | "isort>=5.5.4", 32 | "ruff>=0.0.241,<=0.0.259", 33 | ] 34 | 35 | _extras_training_deps = [ 36 | "jiwer", 37 | "wandb", 38 | "accelerate", 39 | "evaluate", 40 | "datasets[audio]>=2.14.5", 41 | ] 42 | 43 | here = os.path.abspath(os.path.dirname(__file__)) 44 | 45 | with open(os.path.join(here, "README.md"), encoding="utf-8") as f: 46 | long_description = f.read() 47 | 48 | # read version 49 | with open(os.path.join(here, "parler_tts", "__init__.py"), encoding="utf-8") as f: 50 | for line in f: 51 | if line.startswith("__version__"): 52 | version = line.split("=")[1].strip().strip('"') 53 | break 54 | else: 55 | raise RuntimeError("Unable to find version string.") 56 | 57 | setuptools.setup( 58 | name="parler_tts", 59 | version=version, 60 | description="Toolkit for using and training Parler-TTS, a high-quality text-to-speech model.", 61 | long_description=long_description, 62 | long_description_content_type="text/markdown", 63 | packages=setuptools.find_packages(), 64 | install_requires=_deps, 65 | extras_require={ 66 | "dev": _extras_dev_deps, 67 | "train": _extras_training_deps, 68 | }, 69 | ) 70 | -------------------------------------------------------------------------------- /training/README.md: -------------------------------------------------------------------------------- 1 | # Training Parler-TTS 2 | 3 | 4 | Open In Colab 5 | 6 | 7 | **TL;DR:** After having followed the [installation steps](#requirements), you can reproduce the [Parler-TTS Mini v1](https://huggingface.co/parler-tts/parler-tts-mini-v1) training recipe with the following command line: 8 | 9 | ```sh 10 | accelerate launch ./training/run_parler_tts_training.py ./helpers/training_configs/starting_point_v1.json 11 | ``` 12 | 13 | ------------- 14 | 15 | This sub-folder contains all the information to train or fine-tune your own Parler-TTS model. It consists of: 16 | - [1. An introduction to the Parler-TTS architecture](#a-architecture) 17 | - [2. First steps to get started](#b-getting-started) 18 | - [3. Training guide](#c-training) 19 | 20 | > [!IMPORTANT] 21 | > You can also follow [this fine-tuning guide](https://github.com/ylacombe/scripts_and_notebooks/blob/main/Finetuning_Parler_TTS_v1_on_a_single_speaker_dataset.ipynb) on a mono-speaker dataset example. 22 | 23 | ## 1. Architecture 24 | 25 | At the moment, Parler-TTS architecture is almost a carbon copy of the [MusicGen architecture](https://huggingface.co/docs/transformers/v4.39.3/en/model_doc/musicgen#model-structure) and can be decomposed into three distinct stages: 26 | 1. Text encoder: maps the text descriptions to a sequence of hidden-state representations. Parler-TTS uses a frozen text encoder initialised entirely from Flan-T5 27 | 2. Parler-TTS decoder: a language model (LM) that auto-regressively generates audio tokens (or codes) conditional on the encoder hidden-state representations 28 | 3. Audio codec: used to recover the audio waveform from the audio tokens predicted by the decoder. We use the [DAC model](https://github.com/descriptinc/descript-audio-codec) from Descript, although other codec models, such as [EnCodec](https://huggingface.co/facebook/encodec_48khz), can also be used. 29 | 30 | Parler-TTS however introduces some small tweaks: 31 | - The text **description** is passed through the text encoder and used in the cross-attention layers of the decoder. 32 | - The text **prompt** is simply passed through an embedding layer and concatenated to the decoder input hidden states. 33 | - The audio encoder used is [**DAC**](https://descript.notion.site/Descript-Audio-Codec-11389fce0ce2419891d6591a68f814d5) instead of [Encodec](https://github.com/facebookresearch/encodec), as it exhibits better quality. 34 | 35 | 36 | ## 2. Getting started 37 | 38 | To get started, you need to follow a few steps: 39 | 1. Install the requirements. 40 | 2. Find or initialize the model you'll train on. 41 | 3. Find and/or annotate the dataset you'll train your model on. 42 | 43 | ### Requirements 44 | 45 | The Parler-TTS code is written in [PyTorch](https://pytorch.org) and [Accelerate](https://huggingface.co/docs/accelerate/index). It uses some additional requirements, like [wandb](https://wandb.ai/), especially for logging and evaluation. 46 | 47 | To install the package for training, you need to clone the repository from source... 48 | 49 | ```bash 50 | git clone https://github.com/huggingface/parler-tts.git 51 | cd parler-tts 52 | ``` 53 | 54 | ... And then install the requirements: 55 | 56 | ```bash 57 | pip install -e .[train] 58 | ``` 59 | 60 | Optionally, you can create a wandb account and login to it by following [this guide](https://docs.wandb.ai/quickstart). [`wandb`](https://docs.wandb.ai/) allows for better tracking of the experiments metrics and losses. 61 | 62 | You also have the option to configure Accelerate by running the following command. Note that you should set the number of GPUs you wish to use for training, and also the data type (dtype) to your preferred dtype for training/inference (e.g. `bfloat16` on A100 GPUs, `float16` on V100 GPUs, etc.): 63 | 64 | ```bash 65 | accelerate config 66 | ``` 67 | 68 | Lastly, you can link you Hugging Face account so that you can push model repositories on the Hub. This will allow you to save your trained models on the Hub so that you can share them with the community. Run the command: 69 | 70 | ```bash 71 | git config --global credential.helper store 72 | huggingface-cli login 73 | ``` 74 | And then enter an authentication token from https://huggingface.co/settings/tokens. Create a new token if you do not have one already. You should make sure that this token has "write" privileges. 75 | 76 | ### Initialize a model from scratch or use a pre-trained one. 77 | 78 | Depending on your compute resources and your dataset, you need to choose between fine-tuning a pre-trained model and training a new model from scratch. 79 | 80 | In that sense, we released a 880M checkpoint trained on 45K hours of annotated data under the repository id: [`parler-tts/parler-tts-mini-v1`](https://huggingface.co/parler-tts/parler-tts-mini-v1), that you can fine-tune for your own use-case. 81 | 82 | You can also train you own model from scratch. You can find [here](/helpers/model_init_scripts/) examples on how to initialize a model from scratch. For example, you can initialize a dummy model with: 83 | 84 | ```sh 85 | python helpers/model_init_scripts/init_dummy_model.py ./parler-tts-untrained-dummy --text_model "google-t5/t5-small" --audio_model "parler-tts/dac_44khZ_8kbps" 86 | ``` 87 | 88 | In the rest of this guide, and to reproduce the Parler-TTS Mini v1 training recipe, we'll use a 880M parameters model that we'll initialize with: 89 | 90 | ```sh 91 | python helpers/model_init_scripts/init_model_600M.py ./parler-tts-untrained-600M --text_model "google/flan-t5-large" --audio_model "parler-tts/dac_44khZ_8kbps" 92 | ``` 93 | 94 | 95 | ### Create or find datasets 96 | 97 | To train your own Parler-TTS, you need datasets with 3 main features: 98 | - speech data 99 | - text transcription of the speech data 100 | - conditionning text description - that you can create using [Data-Speech](https://github.com/huggingface/dataspeech), a library that allows you to annotate the speaker and utterance characteristics with natural language description. 101 | 102 | Note that we made the choice to use description of the main speech characteristics (speaker pitch, speaking rate, level of noise, etc.) but that you are free to use any handmade or generated text description that makes sense. 103 | 104 | To train Parler-TTS Mini v1, we used: 105 | * A [filtered version](https://huggingface.co/datasets/parler-tts/libritts_r_filtered) of [LibriTTS-R dataset](https://huggingface.co/datasets/blabble-io/libritts_r), a 1K hours high-quality speech dataset. 106 | * The [English subset](https://huggingface.co/datasets/parler-tts/mls_eng) of [Multilingual LibriSpeech](https://huggingface.co/datasets/facebook/multilingual_librispeech). 107 | 108 | Both datasets have been annotated using the [Data-Speech](https://github.com/huggingface/dataspeech) recipe, respectively [here](https://huggingface.co/datasets/parler-tts/libritts-r-filtered-speaker-descriptions) and [here](https://huggingface.co/datasets/parler-tts/mls-eng-speaker-descriptions). 109 | 110 | 111 | ## 3. Training 112 | 113 | The script [`run_parler_tts_training.py`](/training/run_parler_tts_training.py) is an end-to-end script that: 114 | 1. load dataset(s) and merge them to the annotation dataset(s) if necessary 115 | 2. pre-compute audio tokens 116 | 3. train Parler-TTS 117 | 118 | To train Parler-TTS Mini v1, we roughly used: 119 | 120 | ```sh 121 | accelerate launch ./training/run_parler_tts_training.py \ 122 | --model_name_or_path "./parler-tts-untrained-600M/parler-tts-untrained-600M/" \ 123 | --feature_extractor_name "parler-tts/dac_44khZ_8kbps" \ 124 | --description_tokenizer_name "google/flan-t5-large" \ 125 | --prompt_tokenizer_name "google/flan-t5-large" \ 126 | --report_to "wandb" \ 127 | --overwrite_output_dir true \ 128 | --train_dataset_name "parler-tts/libritts_r_filtered+parler-tts/libritts_r_filtered+parler-tts/libritts_r_filtered+parler-tts/mls_eng" \ 129 | --train_metadata_dataset_name "parler-tts/libritts-r-filtered-speaker-descriptions+parler-tts/libritts-r-filtered-speaker-descriptions+parler-tts/libritts-r-filtered-speaker-descriptions+parler-tts/mls-eng-speaker-descriptions" \ 130 | --train_dataset_config_name "clean+clean+other+default" \ 131 | --train_split_name "train.clean.360+train.clean.100+train.other.500+train" \ 132 | --eval_dataset_name "parler-tts/libritts_r_filtered+parler-tts/mls_eng" \ 133 | --eval_metadata_dataset_name "parler-tts/libritts-r-filtered-speaker-descriptions+parler-tts/mls-eng-speaker-descriptions" \ 134 | --eval_dataset_config_name "other+default" \ 135 | --eval_split_name "test.other+test" \ 136 | --target_audio_column_name "audio" \ 137 | --description_column_name "text_description" \ 138 | --prompt_column_name "text" \ 139 | --max_duration_in_seconds 30 \ 140 | --min_duration_in_seconds 2.0 \ 141 | --max_text_length 600 \ 142 | --add_audio_samples_to_wandb true \ 143 | --id_column_name "id" \ 144 | --preprocessing_num_workers 8 \ 145 | --do_train true \ 146 | --num_train_epochs 4 \ 147 | --gradient_accumulation_steps 6 \ 148 | --gradient_checkpointing false \ 149 | --per_device_train_batch_size 4 \ 150 | --learning_rate 0.00095 \ 151 | --adam_beta1 0.9 \ 152 | --adam_beta2 0.99 \ 153 | --weight_decay 0.01 \ 154 | --lr_scheduler_type "constant_with_warmup" \ 155 | --warmup_steps 20000 \ 156 | --logging_steps 1000 \ 157 | --freeze_text_encoder true \ 158 | --do_eval true \ 159 | --predict_with_generate true \ 160 | --include_inputs_for_metrics true \ 161 | --evaluation_strategy steps \ 162 | --eval_steps 10000 \ 163 | --save_steps 10000 \ 164 | --per_device_eval_batch_size 4 \ 165 | --audio_encoder_per_device_batch_size 24 \ 166 | --dtype "bfloat16" \ 167 | --seed 456 \ 168 | --output_dir "./output_dir_training/" \ 169 | --temporary_save_to_disk "./audio_code_tmp/" \ 170 | --save_to_disk "./tmp_dataset_audio/" \ 171 | --max_eval_samples 96 \ 172 | --dataloader_num_workers 8 \ 173 | --group_by_length true \ 174 | --attn_implementation "sdpa" 175 | ``` 176 | 177 | In particular, note how multiple training datasets, metadataset, configurations and splits can be loaded by separating the dataset arguments by + symbols: 178 | ```sh 179 | "train_dataset_name": "parler-tts/libritts_r_filtered+parler-tts/libritts_r_filtered+parler-tts/libritts_r_filtered+parler-tts/mls_eng", 180 | "train_metadata_dataset_name": "parler-tts/libritts-r-filtered-speaker-descriptions+parler-tts/libritts-r-filtered-speaker-descriptions+parler-tts/libritts-r-filtered-speaker-descriptions+parler-tts/mls-eng-speaker-descriptions", 181 | "train_dataset_config_name": "clean+clean+other+default", 182 | "train_split_name": "train.clean.360+train.clean.100+train.other.500+train", 183 | ``` 184 | 185 | 186 | Additionally, you can also write a JSON config file. Here, [starting_point_v1.json](helpers/training_configs/starting_point_v1.json) contains the exact same hyper-parameters than above and can be launched like that: 187 | ```sh 188 | accelerate launch ./training/run_parler_tts_training.py ./helpers/training_configs/starting_point_v1.json 189 | ``` 190 | 191 | Training logs will be reported to wandb, provided that you passed `--report_to "wandb"` to the arguments. 192 | 193 | > [!TIP] 194 | > Starting training a new model from scratch can easily be overwhelming, so here's what training looked like for v1: [logs](https://api.wandb.ai/links/ylacombe/j7g8isjn) 195 | 196 | Scaling to multiple GPUs using [distributed data parallelism (DDP)](https://pytorch.org/tutorials/beginner/ddp_series_theory.html) is trivial: simply run `accelerate config` and select the multi-GPU option, specifying the IDs of the GPUs you wish to use. The above script can then be run using DDP with no code changes. In our case, we used 4 nodes of 8 H100 80GB to train Parler-TTS Mini for around 1.5 days. 197 | 198 | 199 | There are a few other noteworthy arguments: 200 | 1. `train_metadata_dataset_name` and `eval_metadata_dataset_name` specify, if necessary, the names of the dataset(s) that contain(s) the conditionning text descriptions. For example, this [dataset resulting from the Data-Speech annotation process](https://huggingface.co/datasets/parler-tts/libritts-r-filtered-speaker-descriptions) is saved without the audio column, as it's costly to write and push audio data, so it needs to be concatenated back to the original LibriTTS-R dataset. 201 | 2. As noted above, the script pre-computes audio tokens as computing audio codes is costly and only needs to be done once, since we're freezing the audio encoder. `audio_encoder_per_device_batch_size` is used to precise the per devie batch size for this pre-processing step. 202 | 3. Additionnally, when scaling up the training data and iterating on the hyper-parameters or the model architecture, we might want to avoid recomputing the audio tokens at each training run. That's why we introduced two additional parameters, `save_to_disk` and `temporary_save_to_disk` that serves as temporary buffers to save intermediary datasets. Note that processed data is made of text and audio tokens which are much more memory efficient, so the additional required space is negligible. 203 | 4. `predict_with_generate` and `add_audio_samples_to_wandb` are required to store generated audios and to compute WER and CLAP similarity. 204 | 5. `freeze_text_encoder`: which allows to freeze the text encoder, to save compute resources. 205 | 206 | And finally, two additional comments: 207 | 1. `lr_scheduler_stype`: defines the learning rate schedule, one of `constant_with_warmup` or `cosine`. When experimenting with a training set-up or training for very few epochs, using `constant_with_warmup` is typically beneficial, since the learning rate remains high over the short training run. When performing longer training runs, using a `cosine` schedule shoud give better results. 208 | 2. `dtype`: data type (dtype) in which the model computation should be performed. Note that this only controls the dtype of the computations (forward and backward pass), and not the dtype of the parameters or optimiser states. 209 | 210 | > [!TIP] 211 | > Fine-tuning is as easy as modifying `model_name_or_path` to a pre-trained model. 212 | > For example: `--model_name_or_path parler-tts/parler-tts-mini-v1`. 213 | -------------------------------------------------------------------------------- /training/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/huggingface/parler-tts/d108732cd57788ec86bc857d99a6cabd66663d68/training/__init__.py -------------------------------------------------------------------------------- /training/arguments.py: -------------------------------------------------------------------------------- 1 | from dataclasses import dataclass, field 2 | from typing import Optional, List 3 | 4 | from transformers import Seq2SeqTrainingArguments 5 | 6 | 7 | @dataclass 8 | class ModelArguments: 9 | """ 10 | Arguments pertaining to which model/config/tokenizer we are going to fine-tune from. 11 | """ 12 | 13 | model_name_or_path: str = field( 14 | metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"} 15 | ) 16 | config_name: Optional[str] = field( 17 | default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"} 18 | ) 19 | feature_extractor_name: Optional[str] = field( 20 | default=None, metadata={"help": "Pretrained feature extractor name or path if not the same as model_name"} 21 | ) 22 | description_tokenizer_name: Optional[str] = field( 23 | default=None, metadata={"help": "Pretrained description tokenizer name or path if not the same as model_name"} 24 | ) 25 | prompt_tokenizer_name: Optional[str] = field( 26 | default=None, 27 | metadata={"help": "Pretrained prompt tokenizer name or path if not the same as description_tokenizer_name"}, 28 | ) 29 | cache_dir: Optional[str] = field( 30 | default=None, 31 | metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"}, 32 | ) 33 | use_fast_tokenizer: bool = field( 34 | default=True, 35 | metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."}, 36 | ) 37 | model_revision: str = field( 38 | default="main", 39 | metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."}, 40 | ) 41 | pad_token_id: int = field( 42 | default=None, 43 | metadata={"help": "If specified, change the model pad token id."}, 44 | ) 45 | decoder_start_token_id: int = field( 46 | default=None, 47 | metadata={"help": "If specified, change the model decoder start token id."}, 48 | ) 49 | freeze_text_encoder: bool = field( 50 | default=False, 51 | metadata={"help": "Whether to freeze the text encoder."}, 52 | ) 53 | do_sample: bool = field( 54 | default=True, 55 | metadata={"help": "Whether to do sampling or greedy decoding."}, 56 | ) 57 | temperature: float = field( 58 | default=1.0, 59 | metadata={"help": "Temperature if sampling."}, 60 | ) 61 | max_length: int = field( 62 | default=2580, 63 | metadata={"help": "Generation max length."}, 64 | ) 65 | bandwidth: float = field( 66 | default=6, 67 | metadata={"help": "Audio encoder bandwidth."}, 68 | ) 69 | asr_model_name_or_path: str = field( 70 | default="distil-whisper/distil-large-v2", 71 | metadata={ 72 | "help": "Used to compute WER during evaluation. Path to pretrained model or model identifier from huggingface.co/models" 73 | }, 74 | ) 75 | clap_model_name_or_path: str = field( 76 | default="laion/larger_clap_music_and_speech", 77 | metadata={ 78 | "help": "Used to compute audio similarity during evaluation. Path to pretrained model or model identifier from huggingface.co/models" 79 | }, 80 | ) 81 | attn_implementation: str = field( 82 | default="eager", 83 | metadata={"help": "Attention implementation used. One of `eager`, `sdpa`, `flash_attention_2`"}, 84 | ) 85 | cross_attention_implementation_strategy: str = field( 86 | default=None, 87 | metadata={ 88 | "help": "If not specified, the cross-attention implementation will be the same as `_attn_implementation`. If `always_eager`, it will always be the eager implementation. If `always_sdpa`, it will always be the sdpa implementation." 89 | }, 90 | ) 91 | prompt_padding_side: Optional[str] = field( 92 | default="left", 93 | metadata={ 94 | "help": "Prompt tokenizer padding side. Defaults to `left`. If the prompt is pre-pended to the codebooks hidden states, it should be padded on the left." 95 | }, 96 | ) 97 | 98 | 99 | @dataclass 100 | class DataTrainingArguments: 101 | """ 102 | Arguments pertaining to what data we are going to input our model for training and eval. 103 | 104 | Using `HfArgumentParser` we can turn this class 105 | into argparse arguments to be able to specify them on 106 | the command line. 107 | """ 108 | 109 | train_dataset_name: str = field( 110 | default=None, 111 | metadata={ 112 | "help": "The name of the training dataset to use (via the datasets library). Load and combine " 113 | "multiple datasets by separating dataset ids by a '+' symbol. For example, to load and combine " 114 | " librispeech and common voice, set `train_dataset_name='librispeech_asr+common_voice'`." 115 | }, 116 | ) 117 | train_dataset_config_name: Optional[str] = field( 118 | default=None, 119 | metadata={ 120 | "help": "The configuration name of the training dataset to use (via the datasets library). Load and combine " 121 | "multiple datasets by separating dataset configs by a '+' symbol." 122 | }, 123 | ) 124 | train_split_name: str = field( 125 | default="train", 126 | metadata={ 127 | "help": ("The name of the training data set split to use (via the datasets library). Defaults to 'train'") 128 | }, 129 | ) 130 | train_dataset_samples: str = field( 131 | default=None, 132 | metadata={ 133 | "help": "Number of samples in the training data. Load and combine " 134 | "multiple datasets by separating dataset samples by a '+' symbol." 135 | }, 136 | ) 137 | train_metadata_dataset_name: str = field( 138 | default=None, 139 | metadata={ 140 | "help": "The name of the metadata training dataset to use (via the datasets library). Load and combine " 141 | "multiple datasets by separating dataset ids by a '+' symbol. For example, to load and combine " 142 | " librispeech and common voice, set `train_dataset_name='librispeech_asr+common_voice'`." 143 | }, 144 | ) 145 | eval_dataset_name: str = field( 146 | default=None, 147 | metadata={ 148 | "help": "The name of the evaluation dataset to use (via the datasets library). Defaults to the training dataset name if unspecified." 149 | }, 150 | ) 151 | eval_dataset_config_name: Optional[str] = field( 152 | default=None, 153 | metadata={ 154 | "help": "The configuration name of the evaluation dataset to use (via the datasets library). Defaults to the training dataset config name if unspecified" 155 | }, 156 | ) 157 | eval_split_name: str = field( 158 | default="test", 159 | metadata={ 160 | "help": "The name of the evaluation data set split to use (via the datasets library). Defaults to 'test'" 161 | }, 162 | ) 163 | eval_metadata_dataset_name: str = field( 164 | default=None, 165 | metadata={ 166 | "help": "The name of the metadata training dataset to use (via the datasets library). Load and combine " 167 | "multiple datasets by separating dataset ids by a '+' symbol. For example, to load and combine " 168 | " librispeech and common voice, set `train_dataset_name='librispeech_asr+common_voice'`." 169 | }, 170 | ) 171 | target_audio_column_name: str = field( 172 | default="audio", 173 | metadata={"help": "The name of the dataset column containing the target audio data. Defaults to 'audio'"}, 174 | ) 175 | description_column_name: str = field( 176 | default=None, 177 | metadata={"help": "The name of the dataset column containing the description text data. Defaults to 'None'."}, 178 | ) 179 | prompt_column_name: str = field( 180 | default=None, 181 | metadata={"help": "The name of the dataset column containing the prompt text data. Defaults to 'None'."}, 182 | ) 183 | overwrite_cache: bool = field( 184 | default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."} 185 | ) 186 | preprocessing_num_workers: Optional[int] = field( 187 | default=None, 188 | metadata={"help": "The number of processes to use for the preprocessing."}, 189 | ) 190 | max_train_samples: Optional[int] = field( 191 | default=None, 192 | metadata={ 193 | "help": ( 194 | "For debugging purposes or quicker training, truncate the number of training examples to this " 195 | "value if set." 196 | ) 197 | }, 198 | ) 199 | max_eval_samples: Optional[int] = field( 200 | default=None, 201 | metadata={ 202 | "help": ( 203 | "For debugging purposes or quicker training, truncate the number of validation examples to this " 204 | "value if set." 205 | ) 206 | }, 207 | ) 208 | max_duration_in_seconds: float = field( 209 | default=35.0, 210 | metadata={ 211 | "help": ( 212 | "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`." 213 | "Also, used to set maximum audio length if `pad_to_max_length=True`." 214 | ) 215 | }, 216 | ) 217 | min_duration_in_seconds: float = field( 218 | default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"} 219 | ) 220 | max_text_length: int = field( 221 | default=500, metadata={"help": "If set, max description lengths in number of characters."} 222 | ) 223 | max_prompt_token_length: int = field( 224 | default=None, 225 | metadata={ 226 | "help": ( 227 | "If set, filter samples with prompts that are longer than `max_prompt_token_length` tokens." 228 | "Also, used to set maximum prompt token length if `pad_to_max_length=True`." 229 | ) 230 | }, 231 | ) 232 | max_description_token_length: int = field( 233 | default=None, 234 | metadata={ 235 | "help": ( 236 | "If set, filter samples with descriptions that are longer than `max_description_token_length` tokens." 237 | "Also, used to set maximum description token length if `pad_to_max_length=True`." 238 | ) 239 | }, 240 | ) 241 | pad_to_max_length: bool = field( 242 | default=False, 243 | metadata={ 244 | "help": ( 245 | "If `True`, pad audio, prompt and description to a maximum length set with respectively " 246 | "`max_duration_in_seconds`, `max_prompt_token_length`, `max_description_token_length`." 247 | ) 248 | }, 249 | ) 250 | preprocessing_only: bool = field( 251 | default=False, 252 | metadata={ 253 | "help": ( 254 | "Whether to only do data preprocessing and skip training. This is especially useful when data" 255 | " preprocessing errors out in distributed training due to timeout. In this case, one should run the" 256 | " preprocessing in a non-distributed setup with `preprocessing_only=True` so that the cached datasets" 257 | " can consequently be loaded in distributed training." 258 | " In this training script, `save_to_disk` must be set to the path in which the dataset should be saved. " 259 | ) 260 | }, 261 | ) 262 | token: str = field( 263 | default=None, 264 | metadata={ 265 | "help": ( 266 | "The token to use as HTTP bearer authorization for remote files. If not specified, will use the token " 267 | "generated when running `huggingface-cli login` (stored in `~/.huggingface`)." 268 | ) 269 | }, 270 | ) 271 | use_auth_token: bool = field( 272 | default=None, 273 | metadata={ 274 | "help": "The `use_auth_token` argument is deprecated and will be removed in v4.34. Please use `token` instead." 275 | }, 276 | ) 277 | trust_remote_code: bool = field( 278 | default=False, 279 | metadata={ 280 | "help": ( 281 | "Whether or not to allow for custom models defined on the Hub in their own modeling files. This option " 282 | "should only be set to `True` for repositories you trust and in which you have read the code, as it will " 283 | "execute code present on the Hub on your local machine." 284 | ) 285 | }, 286 | ) 287 | add_audio_samples_to_wandb: bool = field( 288 | default=False, 289 | metadata={"help": "If set and if `wandb` in args.report_to, will add generated audio samples to wandb logs."}, 290 | ) 291 | id_column_name: str = field(default=None, metadata={"help": "id column name."}) 292 | wandb_project: str = field( 293 | default="parler-speech", 294 | metadata={"help": "The name of the wandb project."}, 295 | ) 296 | wandb_run_name: str = field( 297 | default=None, 298 | metadata={ 299 | "help": "If specified, the name of the run. If not specified, wandb will give a random name to this run." 300 | }, 301 | ) 302 | save_to_disk: str = field( 303 | default=None, 304 | metadata={ 305 | "help": "If set, will save the dataset to this path if this is an empyt folder. If not empty, will load the datasets from it." 306 | }, 307 | ) 308 | temporary_save_to_disk: str = field(default=None, metadata={"help": "Temporarily save audio labels here."}) 309 | save_codec_steps: Optional[int] = field( 310 | default=500, 311 | metadata={"help": "Temporarily save the audio labels every `save_steps`."}, 312 | ) 313 | pad_to_multiple_of: Optional[int] = field( 314 | default=2, 315 | metadata={"help": ("Pad to multiple of for tokenizers.")}, 316 | ) 317 | 318 | 319 | @dataclass 320 | class ParlerTTSTrainingArguments(Seq2SeqTrainingArguments): 321 | dtype: Optional[str] = field( 322 | default="float32", 323 | metadata={ 324 | "help": ( 325 | "The data type (dtype) in which to run training. One of `float32` (full-precision), " 326 | "`float16` or `bfloat16` (both half-precision)." 327 | ) 328 | }, 329 | ) 330 | audio_encoder_per_device_batch_size: int = field( 331 | default=8, 332 | metadata={"help": ("Specify the batch size of the audio encoding pre-processing steps.")}, 333 | ) 334 | eval_dataloader_num_workers: Optional[int] = field( 335 | default=0, 336 | metadata={ 337 | "help": ( 338 | "Number of subprocesses to use for evaluation data loading (PyTorch only). 0 means that the data will be loaded in the main process." 339 | ) 340 | }, 341 | ) 342 | compute_clap_similarity_metric: bool = field( 343 | default=True, 344 | metadata={ 345 | "help": ( 346 | "Whether or not to compute the clap similarity metric between the description and the generation during evalution." 347 | ) 348 | }, 349 | ) 350 | compute_noise_level_metric: bool = field( 351 | default=True, 352 | metadata={"help": ("Whether or not to compute the squim si-sdr measure of the generations.")}, 353 | ) 354 | noise_level_to_compute_clean_wer: float = field( 355 | default=25, 356 | metadata={ 357 | "help": ( 358 | "if `compute_noise_level_metric=True`, will compute a 'clean' WER on samples with generated noise higher than `noise_level_to_compute_clean_wer`." 359 | "This is a proxy measure to compute WER on clean audios, provided that the model learn to generate clean audios." 360 | ) 361 | }, 362 | ) 363 | eval_generation_steps: Optional[int] = field( 364 | default=None, 365 | metadata={ 366 | "help": ( 367 | "Number of update steps between two generation evaluation. Will default to the same" 368 | "value as `eval_steps` if not set. Should be an integer and a multiple of `eval_steps`." 369 | ) 370 | }, 371 | ) 372 | codebook_weights: Optional[List[float]] = field( 373 | default=None, 374 | metadata={"help": "Weights applied to each codebook."}, 375 | ) -------------------------------------------------------------------------------- /training/data.py: -------------------------------------------------------------------------------- 1 | import logging 2 | from dataclasses import dataclass 3 | from typing import Dict, List, Optional, Set, Union 4 | 5 | import datasets 6 | import numpy as np 7 | import torch 8 | from accelerate import Accelerator 9 | from datasets import Dataset, IterableDataset, concatenate_datasets, interleave_datasets, load_dataset 10 | from tqdm import tqdm 11 | from transformers import AutoFeatureExtractor, AutoTokenizer 12 | 13 | 14 | @dataclass 15 | class DataCollatorEncodecWithPadding: 16 | """ 17 | Data collator that will dynamically pad the inputs received to the longest sequence in the batch or 18 | to `max_length` if `max_length` is set and `padding=max_length`. 19 | """ 20 | 21 | feature_extractor: AutoFeatureExtractor 22 | audio_column_name: str 23 | feature_extractor_input_name: Optional[str] = "input_values" 24 | max_length: Optional[int] = None 25 | padding: Optional[str] = "longest" 26 | 27 | def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]: 28 | # split inputs and labels since they have to be of different lengths and need 29 | # different padding methods 30 | audios = [feature[self.audio_column_name]["array"] for feature in features] 31 | len_audio = [len(audio) for audio in audios] 32 | if self.max_length is not None: 33 | audios = [audio[: min(l, self.max_length)] for audio, l in zip(audios, len_audio)] 34 | 35 | # since resampling has already been performed in the 'load_multiple_datasets' function, 36 | # a fixed sampling_rate(44100hz) is passed to the feature_extractor. 37 | sampling_rate = self.feature_extractor.sampling_rate 38 | batch = self.feature_extractor( 39 | audios, sampling_rate=sampling_rate, return_tensors="pt", padding=self.padding, max_length=self.max_length 40 | ) 41 | batch["len_audio"] = torch.tensor(len_audio).unsqueeze(1) 42 | return batch 43 | 44 | 45 | @dataclass 46 | class DataCollatorParlerTTSWithPadding: 47 | """ 48 | Data collator that will dynamically pad the inputs received. 49 | Args: 50 | prompt_tokenizer (:class:`~transformers.AutoTokenizer`) 51 | The prompt_tokenizer used for proccessing the data. 52 | description_tokenizer (:class:`~transformers.AutoTokenizer`) 53 | The description_tokenizer used for proccessing the data. 54 | padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`): 55 | Select a strategy to pad the returned sequences (according to the model's padding side and padding index) 56 | among: 57 | * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single 58 | sequence if provided). 59 | * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the 60 | maximum acceptable input length for the model if that argument is not provided. 61 | * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of 62 | different lengths). 63 | pad_to_multiple_of (:obj:`int`, `optional`): 64 | If set will pad the sequence to a multiple of the provided value. 65 | This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 66 | 7.5 (Volta). 67 | """ 68 | 69 | prompt_tokenizer: AutoTokenizer 70 | description_tokenizer: AutoTokenizer 71 | padding: Union[bool, str] = "longest" 72 | pad_to_multiple_of: Optional[int] = None 73 | prompt_max_length: Optional[int] = None 74 | description_max_length: Optional[int] = None 75 | audio_max_length: Optional[int] = None 76 | 77 | def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]: 78 | # split inputs and labels since they have to be of different lengths and need 79 | # different padding methods 80 | 81 | labels = [torch.tensor(feature["labels"]).transpose(0, 1) for feature in features] 82 | # (bsz, seq_len, num_codebooks) 83 | labels = torch.nn.utils.rnn.pad_sequence(labels, batch_first=True, padding_value=-100) 84 | if self.audio_max_length is not None and self.padding == "max_length": 85 | labels = torch.nn.functional.pad( 86 | labels, pad=(0, 0, 0, max(self.audio_max_length - labels.shape[1], 0)), value=-100 87 | ) 88 | 89 | input_ids = [{"input_ids": feature["input_ids"]} for feature in features] 90 | 91 | input_ids = self.description_tokenizer.pad( 92 | input_ids, 93 | return_tensors="pt", 94 | padding=self.padding, 95 | pad_to_multiple_of=self.pad_to_multiple_of, 96 | max_length=self.description_max_length, 97 | ) 98 | 99 | batch = {"labels": labels, **input_ids} 100 | 101 | prompt_input_ids = [{"input_ids": feature["prompt_input_ids"]} for feature in features] 102 | prompt_input_ids = self.prompt_tokenizer.pad( 103 | prompt_input_ids, 104 | return_tensors="pt", 105 | padding=self.padding, 106 | pad_to_multiple_of=self.pad_to_multiple_of, 107 | max_length=self.prompt_max_length, 108 | ) 109 | 110 | batch["prompt_input_ids"] = prompt_input_ids["input_ids"] 111 | if "attention_mask" in prompt_input_ids: 112 | batch["prompt_attention_mask"] = prompt_input_ids["attention_mask"] 113 | 114 | return batch 115 | 116 | 117 | def convert_dataset_str_to_list( 118 | dataset_names, 119 | dataset_config_names, 120 | metadata_dataset_names=None, 121 | splits=None, 122 | dataset_samples=None, 123 | default_split="train", 124 | ): 125 | if isinstance(dataset_names, str): 126 | dataset_names = dataset_names.split("+") 127 | dataset_config_names = dataset_config_names.split("+") 128 | splits = splits.split("+") if splits is not None else None 129 | dataset_samples = dataset_samples.split("+") if dataset_samples is not None else None 130 | metadata_dataset_names = metadata_dataset_names.split("+") if metadata_dataset_names is not None else None 131 | 132 | # basic checks to ensure we've got the right number of datasets/configs/splits/columns/probs 133 | if len(dataset_names) != len(dataset_config_names): 134 | raise ValueError( 135 | f"Ensure one config is passed for each dataset, got {len(dataset_names)} datasets and" 136 | f" {len(dataset_config_names)} configs." 137 | ) 138 | 139 | if splits is not None and len(splits) != len(dataset_names): 140 | raise ValueError( 141 | f"Ensure one split is passed for each dataset, got {len(dataset_names)} datasets and {len(splits)} splits." 142 | ) 143 | 144 | if metadata_dataset_names is not None and len(metadata_dataset_names) != len(dataset_names): 145 | raise ValueError( 146 | f"Ensure one metadata dataset is passed for each dataset, got {len(dataset_names)} datasets and {len(metadata_dataset_names)} metadata datasets." 147 | ) 148 | 149 | if dataset_samples is not None: 150 | if len(dataset_samples) != len(dataset_names): 151 | raise ValueError( 152 | f"Ensure one sample is passed for each dataset, got {len(dataset_names)} datasets and " 153 | f"{len(dataset_samples)} samples." 154 | ) 155 | dataset_samples = [float(ds_sample) for ds_sample in dataset_samples] 156 | else: 157 | dataset_samples = [None] * len(dataset_names) 158 | 159 | splits = splits if splits is not None else [default_split for _ in range(len(dataset_names))] 160 | 161 | dataset_names_dict = [] 162 | for i, ds_name in enumerate(dataset_names): 163 | dataset_names_dict.append( 164 | { 165 | "name": ds_name, 166 | "config": dataset_config_names[i], 167 | "split": splits[i], 168 | "metadata_dataset_name": metadata_dataset_names[i], 169 | "samples": dataset_samples[i], 170 | } 171 | ) 172 | return dataset_names_dict 173 | 174 | 175 | def load_multiple_datasets( 176 | accelerator: Accelerator, 177 | dataset_names: Union[List, str], 178 | dataset_config_names: Union[List, str], 179 | metadata_dataset_names: Optional[str] = None, 180 | splits: Optional[Union[List, str]] = None, 181 | label_column_names: Optional[List] = None, 182 | stopping_strategy: Optional[str] = "first_exhausted", 183 | dataset_samples: Optional[Union[List, np.array]] = None, 184 | streaming: Optional[bool] = False, 185 | seed: Optional[int] = None, 186 | id_column_name: Optional[str] = None, 187 | columns_to_keep: Optional[Set[str]] = None, 188 | prompt_column_name: Optional[str] = None, 189 | sampling_rate: Optional[int] = None, 190 | audio_column_name: Optional[str] = None, 191 | logger: Optional[logging.Logger] = None, 192 | **kwargs, 193 | ) -> Union[Dataset, IterableDataset]: 194 | dataset_names_dict = convert_dataset_str_to_list( 195 | dataset_names, dataset_config_names, metadata_dataset_names, splits, label_column_names, dataset_samples 196 | ) 197 | 198 | if dataset_samples is not None: 199 | dataset_samples = [ds_dict["samples"] for ds_dict in dataset_names_dict] 200 | probabilities = np.array(dataset_samples) / np.sum(dataset_samples) 201 | else: 202 | probabilities = None 203 | 204 | all_datasets = [] 205 | # iterate over the datasets we want to interleave 206 | for dataset_dict in tqdm(dataset_names_dict, desc="Combining datasets..."): 207 | with accelerator.local_main_process_first(): 208 | dataset = load_dataset( 209 | dataset_dict["name"], 210 | dataset_dict["config"], 211 | split=dataset_dict["split"], 212 | streaming=streaming, 213 | **kwargs, 214 | ) 215 | dataset_features = dataset.features.keys() 216 | 217 | if sampling_rate is not None and audio_column_name is not None: 218 | # resample target audio 219 | dataset = dataset.cast_column(audio_column_name, datasets.features.Audio(sampling_rate=sampling_rate)) 220 | 221 | metadata_dataset_name = dataset_dict["metadata_dataset_name"] 222 | if metadata_dataset_name is not None: 223 | logger.info( 224 | f'Merging {dataset_dict["name"]} - {dataset_dict["split"]} with {metadata_dataset_name} - {dataset_dict["split"]}' 225 | ) 226 | metadata_dataset = load_dataset( 227 | metadata_dataset_name, 228 | dataset_dict["config"], 229 | split=dataset_dict["split"], 230 | streaming=streaming, 231 | **kwargs, 232 | ) 233 | 234 | # TODO(YL): I forgot to create unique ids for MLS english. 235 | # To iterate faster, I bypass the original id check and do another one. - Done once because assuming it won't change next time 236 | # if dataset_dict["name"] == "parler-tts/mls_eng_10k": 237 | # def concat_ids(book_id, speaker_id, begin_time): 238 | # return {"id": f"{book_id}_{speaker_id}_{str(begin_time).replace('.', '_')}"} 239 | # dataset = dataset.map(concat_ids, input_columns=["book_id", "speaker_id", "begin_time"], num_proc=24) 240 | # metadata_dataset = metadata_dataset.map(concat_ids, input_columns=["book_id", "speaker_id", "begin_time"], num_proc=24) 241 | # metadata_dataset = metadata_dataset.rename_column(id_column_name, f"metadata_{id_column_name}") 242 | 243 | if dataset_dict["name"] not in {"parler-tts/mls_eng_10k", "parler-tts/mls_eng"}: 244 | if id_column_name is not None and id_column_name not in dataset.column_names: 245 | raise ValueError( 246 | f"id_column_name={id_column_name} but has not been found in the dataset columns" 247 | f"- one of {', '.join(list(dataset.column_names))}." 248 | ) 249 | if id_column_name is not None and id_column_name not in metadata_dataset.column_names: 250 | raise ValueError( 251 | f"id_column_name={id_column_name} but has not been found in the metadata dataset columns" 252 | f"- one of {', '.join(list(metadata_dataset.column_names))}." 253 | ) 254 | elif id_column_name is not None: 255 | metadata_dataset = metadata_dataset.rename_column(id_column_name, f"metadata_{id_column_name}") 256 | 257 | metadata_columns_to_remove = set(metadata_dataset.column_names).intersection(set(dataset.column_names)) 258 | 259 | if prompt_column_name is not None: 260 | # We might have applied some transformations to the prompts (e.g punctuation restoration) 261 | # so we make sure to remove it from the original dataset 262 | if prompt_column_name in dataset.column_names: 263 | logger.info( 264 | f"REMOVE {prompt_column_name} from dataset {dataset_dict['name']} - dataset_dict['split']" 265 | ) 266 | dataset.remove_columns(prompt_column_name) 267 | 268 | metadata_columns_to_remove = set(metadata_dataset.column_names).intersection(set(dataset.column_names)) 269 | metadata_dataset = metadata_dataset.remove_columns(metadata_columns_to_remove) 270 | 271 | dataset = concatenate_datasets([dataset, metadata_dataset], axis=1) 272 | 273 | if id_column_name is not None and dataset_dict["name"] not in { 274 | "parler-tts/mls_eng_10k", 275 | "parler-tts/mls_eng", 276 | }: 277 | if ( 278 | len( 279 | dataset.filter( 280 | lambda id1, id2: id1 != id2, 281 | input_columns=[id_column_name, f"metadata_{id_column_name}"], 282 | ) 283 | ) 284 | != 0 285 | ): 286 | raise ValueError( 287 | f"Concatenate didn't work. Some ids don't correspond on dataset {dataset_dict['name']}" 288 | ) 289 | 290 | dataset_features = dataset.features.keys() 291 | 292 | if columns_to_keep is not None: 293 | dataset = dataset.remove_columns(set(dataset_features - columns_to_keep)) 294 | all_datasets.append(dataset) 295 | 296 | if len(all_datasets) == 1: 297 | # we have a single dataset so just return it as is 298 | return all_datasets[0] 299 | 300 | if streaming: 301 | interleaved_dataset = interleave_datasets( 302 | all_datasets, 303 | stopping_strategy=stopping_strategy, 304 | probabilities=probabilities, 305 | seed=seed, 306 | ) 307 | else: 308 | with accelerator.local_main_process_first(): 309 | interleaved_dataset = concatenate_datasets(all_datasets) 310 | 311 | return interleaved_dataset 312 | -------------------------------------------------------------------------------- /training/eval.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from torchaudio.pipelines import SQUIM_OBJECTIVE 3 | import torchaudio 4 | import evaluate 5 | from transformers import ( 6 | AutoModel, 7 | AutoProcessor, 8 | pipeline, 9 | WhisperForConditionalGeneration, 10 | WhisperTokenizer, 11 | WhisperTokenizerFast, 12 | ) 13 | from accelerate.utils.memory import release_memory 14 | import numpy as np 15 | 16 | 17 | def clap_similarity(clap_model_name_or_path, texts, audios, device, input_sampling_rate=44100): 18 | clap = AutoModel.from_pretrained(clap_model_name_or_path) 19 | clap_processor = AutoProcessor.from_pretrained(clap_model_name_or_path) 20 | output_sampling_rate = clap_processor.feature_extractor.sampling_rate 21 | if input_sampling_rate != output_sampling_rate: 22 | audios = [ 23 | torchaudio.functional.resample(torch.from_numpy(audio), input_sampling_rate, output_sampling_rate).numpy() 24 | for audio in audios 25 | ] 26 | clap_inputs = clap_processor( 27 | text=texts, audios=audios, padding=True, return_tensors="pt", sampling_rate=output_sampling_rate 28 | ).to(device) 29 | 30 | clap.to(device) 31 | with torch.no_grad(): 32 | text_features = clap.get_text_features( 33 | clap_inputs["input_ids"], attention_mask=clap_inputs.get("attention_mask", None) 34 | ) 35 | audio_features = clap.get_audio_features(clap_inputs["input_features"]) 36 | 37 | cosine_sim = torch.nn.functional.cosine_similarity(audio_features, text_features, dim=1, eps=1e-8).mean() 38 | 39 | cosine_sim = cosine_sim.to("cpu") 40 | 41 | clap.to("cpu") 42 | clap, clap_inputs, audio_features, text_features = release_memory(clap, clap_inputs, audio_features, text_features) 43 | return cosine_sim 44 | 45 | 46 | def si_sdr(audios, device, input_sampling_rate=44100): 47 | max_audio_length = 15 * SQUIM_OBJECTIVE.sample_rate 48 | model = SQUIM_OBJECTIVE.get_model().to((device)) 49 | 50 | output_sampling_rate = SQUIM_OBJECTIVE.sample_rate 51 | if input_sampling_rate != output_sampling_rate: 52 | audios = [ 53 | torchaudio.functional.resample( 54 | torch.tensor(audio)[None, :].to(device).float(), input_sampling_rate, output_sampling_rate 55 | ) 56 | for audio in audios 57 | ] 58 | 59 | def apply_squim(waveform): 60 | with torch.no_grad(): 61 | waveform = waveform[:, : min(max_audio_length, waveform.shape[1])] 62 | _, _, sdr_sample = model(waveform) 63 | sdr_sample = sdr_sample.cpu()[0] 64 | return sdr_sample 65 | 66 | si_sdrs = [apply_squim(audio) for audio in audios] 67 | audios, model = release_memory(audios, model) 68 | return si_sdrs 69 | 70 | 71 | def wer( 72 | asr_model_name_or_path, 73 | prompts, 74 | audios, 75 | device, 76 | per_device_eval_batch_size, 77 | sampling_rate, 78 | noise_level_to_compute_clean_wer, 79 | si_sdr_measures, 80 | ): 81 | metric = evaluate.load("wer") 82 | asr_pipeline = pipeline(model=asr_model_name_or_path, device=device, chunk_length_s=25.0) 83 | 84 | return_language = None 85 | if isinstance(asr_pipeline.model, WhisperForConditionalGeneration): 86 | return_language = True 87 | 88 | transcriptions = asr_pipeline( 89 | [{"raw": audio, "sampling_rate": sampling_rate} for audio in audios], 90 | batch_size=int(per_device_eval_batch_size), 91 | return_language=return_language, 92 | ) 93 | 94 | if isinstance(asr_pipeline.tokenizer, (WhisperTokenizer, WhisperTokenizerFast)): 95 | tokenizer = asr_pipeline.tokenizer 96 | else: 97 | tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-large-v3") 98 | 99 | english_normalizer = tokenizer.normalize 100 | basic_normalizer = tokenizer.basic_normalize 101 | 102 | normalized_predictions = [] 103 | normalized_references = [] 104 | 105 | for pred, ref in zip(transcriptions, prompts): 106 | normalizer = ( 107 | english_normalizer 108 | if isinstance(pred.get("chunks", None), list) and pred["chunks"][0].get("language", None) == "english" 109 | else basic_normalizer 110 | ) 111 | norm_ref = normalizer(ref) 112 | if len(norm_ref) > 0: 113 | norm_pred = normalizer(pred["text"]) 114 | normalized_predictions.append(norm_pred) 115 | normalized_references.append(norm_ref) 116 | 117 | word_error = 100 118 | clean_word_error = None 119 | noisy_word_error = None 120 | percent_clean_samples = 0 121 | if len(normalized_references) > 0: 122 | word_error = 100 * metric.compute(predictions=normalized_predictions, references=normalized_references) 123 | 124 | 125 | if noise_level_to_compute_clean_wer and si_sdr_measures: 126 | si_sdr_measures = np.array(si_sdr_measures) 127 | mask = si_sdr_measures >= noise_level_to_compute_clean_wer 128 | if mask.any(): 129 | clean_word_error = 100 * metric.compute( 130 | predictions=np.array(normalized_predictions)[mask], references=np.array(normalized_references)[mask] 131 | ) 132 | if not mask.all(): 133 | noisy_word_error = 100 * metric.compute( 134 | predictions=np.array(normalized_predictions)[~mask], references=np.array(normalized_references)[~mask] 135 | ) 136 | else: 137 | noisy_word_error = 0 138 | percent_clean_samples = mask.sum() / len(mask) 139 | 140 | asr_pipeline.model.to("cpu") 141 | asr_pipeline = release_memory(asr_pipeline) 142 | return word_error, [t["text"] for t in transcriptions], clean_word_error, noisy_word_error, percent_clean_samples 143 | -------------------------------------------------------------------------------- /training/run_parler_tts_training.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # coding=utf-8 3 | # Copyright 2024 The HuggingFace Inc. team. All rights reserved. 4 | # 5 | # Licensed under the Apache License, Version 2.0 (the "License"); 6 | # you may not use this file except in compliance with the License. 7 | # You may obtain a copy of the License at 8 | # 9 | # http://www.apache.org/licenses/LICENSE-2.0 10 | # 11 | # Unless required by applicable law or agreed to in writing, software 12 | # distributed under the License is distributed on an "AS IS" BASIS, 13 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 14 | # See the License for the specific language governing permissions and 15 | # limitations under the License. 16 | 17 | """ Train Parler-TTS using 🤗 Accelerate""" 18 | 19 | import logging 20 | import os 21 | import re 22 | import sys 23 | import time 24 | import math 25 | import contextlib 26 | from multiprocess import set_start_method 27 | from datetime import timedelta 28 | import inspect 29 | from tqdm import tqdm 30 | from pathlib import Path 31 | 32 | import torch 33 | from torch.utils.data import DataLoader 34 | 35 | import datasets 36 | from datasets import DatasetDict, Dataset, IterableDataset, concatenate_datasets 37 | 38 | from huggingface_hub import HfApi 39 | 40 | import transformers 41 | from transformers import AutoFeatureExtractor, AutoTokenizer, HfArgumentParser 42 | from transformers.trainer_pt_utils import LengthGroupedSampler 43 | from transformers.optimization import get_scheduler 44 | from transformers.utils import send_example_telemetry 45 | 46 | 47 | from accelerate import Accelerator, skip_first_batches 48 | from accelerate.utils import set_seed, AutocastKwargs, InitProcessGroupKwargs, TorchDynamoPlugin, DistributedDataParallelKwargs 49 | from accelerate.utils.memory import release_memory 50 | 51 | from parler_tts import ( 52 | ParlerTTSConfig, 53 | ParlerTTSForConditionalGeneration, 54 | build_delay_pattern_mask, 55 | ) 56 | 57 | from training.utils import ( 58 | get_last_checkpoint, 59 | rotate_checkpoints, 60 | log_pred, 61 | log_metric, 62 | load_all_codec_checkpoints, 63 | save_codec_checkpoint, 64 | get_last_codec_checkpoint_step, 65 | ) 66 | from training.arguments import ModelArguments, DataTrainingArguments, ParlerTTSTrainingArguments 67 | from training.data import load_multiple_datasets, DataCollatorParlerTTSWithPadding, DataCollatorEncodecWithPadding 68 | from training.eval import clap_similarity, wer, si_sdr 69 | 70 | logger = logging.getLogger(__name__) 71 | 72 | 73 | def main(): 74 | # See all possible arguments in src/transformers/training_args.py 75 | # or by passing the --help flag to this script. 76 | # We now keep distinct sets of args, for a cleaner separation of concerns. 77 | 78 | parser = HfArgumentParser((ModelArguments, DataTrainingArguments, ParlerTTSTrainingArguments)) 79 | if len(sys.argv) == 2 and sys.argv[1].endswith(".json"): 80 | # If we pass only one argument to the script and it's the path to a json file, 81 | # let's parse it to get our arguments. 82 | model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1])) 83 | else: 84 | model_args, data_args, training_args = parser.parse_args_into_dataclasses() 85 | 86 | # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The 87 | # information sent is the one passed as arguments along with your Python/PyTorch versions. 88 | send_example_telemetry("run_parler_tts", model_args, data_args) 89 | 90 | if training_args.dtype == "float16": 91 | mixed_precision = "fp16" 92 | torch_dtype = torch.float16 93 | elif training_args.dtype == "bfloat16": 94 | mixed_precision = "bf16" 95 | torch_dtype = torch.bfloat16 96 | else: 97 | mixed_precision = "no" 98 | torch_dtype = torch.float32 99 | 100 | if data_args.pad_to_max_length and ( 101 | data_args.max_duration_in_seconds is None 102 | or data_args.max_prompt_token_length is None 103 | or data_args.max_description_token_length is None 104 | ): 105 | raise ValueError( 106 | "`pad_to_max_length` is `True` but one of the following parameters has not been set: `max_duration_in_seconds`, `max_prompt_token_length`, `max_description_token_length`" 107 | ) 108 | 109 | padding = "max_length" if data_args.pad_to_max_length else "longest" 110 | 111 | ####### A. Preparation 112 | kwargs_handlers = [InitProcessGroupKwargs(timeout=timedelta(minutes=120)), DistributedDataParallelKwargs(find_unused_parameters=False)] 113 | 114 | accelerator = Accelerator( 115 | gradient_accumulation_steps=training_args.gradient_accumulation_steps, 116 | mixed_precision=mixed_precision, 117 | log_with=training_args.report_to, 118 | project_dir=training_args.output_dir, 119 | kwargs_handlers=kwargs_handlers, 120 | ) 121 | 122 | accelerator.init_trackers( 123 | project_name=data_args.wandb_project, 124 | config={ 125 | "learning_rate": training_args.learning_rate, 126 | "model_name_or_path": model_args.model_name_or_path, 127 | "num_train_epochs": training_args.num_train_epochs, 128 | "gradient_accumulation_steps": training_args.gradient_accumulation_steps, 129 | "per_device_train_batch_size": training_args.per_device_train_batch_size, 130 | "global_batch_size": training_args.per_device_train_batch_size * accelerator.num_processes, 131 | "mixed_precision": mixed_precision, 132 | "lr_scheduler_type": training_args.lr_scheduler_type, 133 | "warmup_steps": training_args.warmup_steps, 134 | "freeze_text_encoder": model_args.freeze_text_encoder, 135 | "max_duration_in_seconds": data_args.max_duration_in_seconds, 136 | "weight_decay": training_args.weight_decay, 137 | "adam_beta1": training_args.adam_beta1, 138 | "adam_beta2": training_args.adam_beta2, 139 | "temperature": model_args.temperature, 140 | }, 141 | init_kwargs={"wandb": {"name": data_args.wandb_run_name}} if data_args.wandb_run_name else {}, 142 | ) 143 | 144 | # Detecting last checkpoint and eventually continue from last checkpoint 145 | last_checkpoint = None 146 | if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir: 147 | last_checkpoint = get_last_checkpoint(training_args.output_dir) 148 | if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0: 149 | raise ValueError( 150 | f"Output directory ({training_args.output_dir}) already exists and is not empty. " 151 | "Use --overwrite_output_dir to overcome." 152 | ) 153 | elif last_checkpoint is not None and training_args.resume_from_checkpoint is None: 154 | logger.info( 155 | f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change " 156 | "the `--output_dir` or add `--overwrite_output_dir` to train from scratch." 157 | ) 158 | 159 | # Setup logging 160 | logging.basicConfig( 161 | format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", 162 | datefmt="%m/%d/%Y %H:%M:%S", 163 | handlers=[logging.StreamHandler(sys.stdout)], 164 | ) 165 | logger.setLevel(logging.INFO if accelerator.is_main_process else logging.WARN) 166 | 167 | # Log a small summary on each proces 168 | logger.warning( 169 | f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, " 170 | f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}" 171 | ) 172 | 173 | # Set the verbosity to info of the Transformers logger (on main process only) 174 | if accelerator.is_local_main_process: 175 | datasets.utils.logging.set_verbosity_warning() 176 | transformers.utils.logging.set_verbosity_info() 177 | else: 178 | datasets.utils.logging.set_verbosity_error() 179 | transformers.utils.logging.set_verbosity_error() 180 | 181 | logger.info("Training/evaluation parameters %s", training_args) 182 | 183 | # Set seed before initializing model. 184 | set_seed(training_args.seed) 185 | num_workers = data_args.preprocessing_num_workers 186 | 187 | # 1. First, lett's instantiate the feature extractor, tokenizers and model 188 | # Note for distributed training, the .from_pretrained methods guarantee that only 189 | # one local process can concurrently download model & vocab. 190 | 191 | # load feature extractor 192 | feature_extractor = AutoFeatureExtractor.from_pretrained( 193 | model_args.feature_extractor_name or model_args.model_name_or_path, 194 | cache_dir=model_args.cache_dir, 195 | token=data_args.token, 196 | trust_remote_code=data_args.trust_remote_code, 197 | ) 198 | sampling_rate = feature_extractor.sampling_rate 199 | 200 | # load prompt tokenizer 201 | prompt_tokenizer = AutoTokenizer.from_pretrained( 202 | model_args.prompt_tokenizer_name or model_args.description_tokenizer_name or model_args.model_name_or_path, 203 | cache_dir=model_args.cache_dir, 204 | token=data_args.token, 205 | trust_remote_code=data_args.trust_remote_code, 206 | use_fast=model_args.use_fast_tokenizer, 207 | padding_side=model_args.prompt_padding_side, 208 | ) 209 | 210 | # load description tokenizer 211 | description_tokenizer = AutoTokenizer.from_pretrained( 212 | model_args.description_tokenizer_name or model_args.model_name_or_path, 213 | cache_dir=model_args.cache_dir, 214 | token=data_args.token, 215 | trust_remote_code=data_args.trust_remote_code, 216 | use_fast=model_args.use_fast_tokenizer, 217 | ) 218 | 219 | if model_args.use_fast_tokenizer: 220 | logger.warning( 221 | "Disabling fast tokenizer warning: https://github.com/huggingface/transformers/blob/main/src/transformers/tokenization_utils_base.py#L3231-L3235" 222 | ) 223 | prompt_tokenizer.deprecation_warnings["Asking-to-pad-a-fast-tokenizer"] = True 224 | description_tokenizer.deprecation_warnings["Asking-to-pad-a-fast-tokenizer"] = True 225 | 226 | # 2. Now, let's load the dataset 227 | 228 | if data_args.save_to_disk is not None: 229 | os.makedirs(data_args.save_to_disk, exist_ok=True) 230 | 231 | # assume that the dataset has been saved to `save_to_disk` if the latter is not empty 232 | dataset_was_precomputed = len(os.listdir(data_args.save_to_disk)) > 0 233 | if dataset_was_precomputed: 234 | with accelerator.local_main_process_first(): 235 | vectorized_datasets = datasets.load_from_disk(data_args.save_to_disk) 236 | else: 237 | raw_datasets = DatasetDict() 238 | 239 | columns_to_keep = { 240 | "target_audio_column_name": data_args.target_audio_column_name, 241 | "prompt_column_name": data_args.prompt_column_name, 242 | } 243 | if data_args.description_column_name is not None: 244 | columns_to_keep["description_column_name"] = data_args.description_column_name 245 | 246 | if training_args.do_train: 247 | raw_datasets["train"] = load_multiple_datasets( 248 | accelerator, 249 | data_args.train_dataset_name, 250 | data_args.train_dataset_config_name, 251 | metadata_dataset_names=data_args.train_metadata_dataset_name, 252 | splits=data_args.train_split_name, 253 | dataset_samples=data_args.train_dataset_samples, 254 | seed=training_args.seed, 255 | cache_dir=model_args.cache_dir, 256 | num_proc=data_args.preprocessing_num_workers, 257 | id_column_name=data_args.id_column_name, 258 | columns_to_keep=columns_to_keep.values(), 259 | prompt_column_name=data_args.prompt_column_name, 260 | audio_column_name=data_args.target_audio_column_name, 261 | sampling_rate=sampling_rate, 262 | logger=logger, 263 | # streaming=data_args.streaming, TODO(SG): optionally enable streaming mode 264 | ) 265 | 266 | for key in columns_to_keep: 267 | if columns_to_keep[key] not in raw_datasets["train"].column_names: 268 | raise ValueError( 269 | f"--{key} '{columns_to_keep[key]}' not found in dataset '{data_args.train_dataset_name}'." 270 | f" Make sure to set `--{key}` to the correct audio column - one of" 271 | f" {', '.join(raw_datasets['train'].column_names)}." 272 | ) 273 | 274 | if data_args.max_train_samples is not None: 275 | raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples)) 276 | 277 | if training_args.do_eval: 278 | raw_datasets["eval"] = load_multiple_datasets( 279 | accelerator, 280 | data_args.eval_dataset_name if data_args.eval_dataset_name else data_args.train_dataset_name, 281 | data_args.eval_dataset_config_name 282 | if data_args.eval_dataset_config_name 283 | else data_args.train_dataset_config_name, 284 | metadata_dataset_names=data_args.eval_metadata_dataset_name, 285 | splits=data_args.eval_split_name, 286 | cache_dir=model_args.cache_dir, 287 | num_proc=data_args.preprocessing_num_workers, 288 | id_column_name=data_args.id_column_name, 289 | columns_to_keep=columns_to_keep.values(), 290 | prompt_column_name=data_args.prompt_column_name, 291 | audio_column_name=data_args.target_audio_column_name, 292 | sampling_rate=sampling_rate, 293 | logger=logger, 294 | # streaming=data_args.streaming, TODO(SG): optionally enable streaming mode 295 | ) 296 | 297 | if data_args.max_eval_samples is not None: 298 | with accelerator.local_main_process_first(): 299 | raw_datasets["eval"] = ( 300 | raw_datasets["eval"].shuffle(seed=training_args.seed).select(range(data_args.max_eval_samples)) 301 | ) 302 | 303 | # 3. Next, let's load the config. 304 | config = ParlerTTSConfig.from_pretrained( 305 | model_args.model_name_or_path, 306 | cache_dir=model_args.cache_dir, 307 | token=data_args.token, 308 | trust_remote_code=data_args.trust_remote_code, 309 | ) 310 | 311 | if training_args.codebook_weights is not None and len(training_args.codebook_weights) != config.decoder.num_codebooks: 312 | raise ValueError(f"`codebook_weights` has length {len(training_args.codebook_weights)} when it should be of length {config.decoder.num_codebooks}.") 313 | 314 | # update pad token id and decoder_start_token_id 315 | config.decoder.update( 316 | { 317 | "cross_attention_implementation_strategy": model_args.cross_attention_implementation_strategy 318 | if model_args.cross_attention_implementation_strategy is not None 319 | else None, 320 | "codebook_weights": training_args.codebook_weights if training_args.codebook_weights is not None else config.decoder.codebook_weights 321 | } 322 | ) 323 | config.update( 324 | { 325 | "pad_token_id": model_args.pad_token_id if model_args.pad_token_id is not None else config.pad_token_id, 326 | "decoder_start_token_id": model_args.decoder_start_token_id 327 | if model_args.decoder_start_token_id is not None 328 | else config.decoder_start_token_id, 329 | } 330 | ) 331 | 332 | # create model 333 | model = ParlerTTSForConditionalGeneration.from_pretrained( 334 | model_args.model_name_or_path, 335 | cache_dir=model_args.cache_dir, 336 | config=config, 337 | token=data_args.token, 338 | trust_remote_code=data_args.trust_remote_code, 339 | attn_implementation={"decoder": model_args.attn_implementation, "text_encoder": "eager"}, 340 | ) 341 | 342 | # enable gradient checkpointing if necessary 343 | if training_args.gradient_checkpointing: 344 | model.gradient_checkpointing_enable() 345 | 346 | # 4. Now we preprocess the datasets including loading the audio, resampling and normalization 347 | # Thankfully, `datasets` takes care of automatically loading and resampling the audio, 348 | # so that we just need to set the correct target sampling rate and normalize the input 349 | # via the `feature_extractor` 350 | 351 | # derive max & min input length for sample rate & max duration 352 | sampling_rate = feature_extractor.sampling_rate 353 | max_target_length = int(data_args.max_duration_in_seconds * sampling_rate) 354 | min_target_length = int(data_args.min_duration_in_seconds * sampling_rate) 355 | target_audio_column_name = data_args.target_audio_column_name 356 | description_column_name = data_args.description_column_name 357 | prompt_column_name = data_args.prompt_column_name 358 | feature_extractor_input_name = feature_extractor.model_input_names[0] 359 | audio_encoder_pad_token_id = config.decoder.pad_token_id 360 | audio_encoder_eos_token_id = config.decoder.eos_token_id 361 | audio_encoder_bos_token_id = model.generation_config.decoder_start_token_id 362 | max_length = model.generation_config.max_length 363 | num_codebooks = model.decoder.config.num_codebooks 364 | bandwidth = model_args.bandwidth 365 | attn_implementation = model_args.attn_implementation 366 | 367 | # Freeze Encoders 368 | model.freeze_encoders(model_args.freeze_text_encoder) 369 | 370 | # Test all gather - used for warmout and avoiding timeout 371 | logger.debug(str(accelerator.process_index), main_process_only=False, in_order=True) 372 | test_tensor = torch.tensor([accelerator.process_index], device=accelerator.device) 373 | gathered_tensor = accelerator.gather(test_tensor) 374 | print("gathered_tensor", gathered_tensor) 375 | accelerator.wait_for_everyone() 376 | 377 | if not dataset_was_precomputed: 378 | # Filter on text length 379 | if description_column_name is not None and data_args.max_text_length is not None: 380 | with accelerator.local_main_process_first(): 381 | # filter description that is shorter than max_text_length 382 | raw_datasets = raw_datasets.filter( 383 | lambda x: len(x) < data_args.max_text_length, 384 | num_proc=num_workers, 385 | input_columns=[description_column_name], 386 | ) 387 | 388 | # Preprocessing the dataset. 389 | # We need to tokenize the texts. 390 | def pass_through_processors(description, prompt): 391 | batch = {} 392 | 393 | batch["input_ids"] = description_tokenizer(description.strip())["input_ids"] 394 | batch["prompt_input_ids"] = prompt_tokenizer(prompt.strip())["input_ids"] 395 | 396 | return batch 397 | 398 | with accelerator.local_main_process_first(): 399 | # this is a trick to avoid to rewrite the entire audio column which takes ages 400 | vectorized_datasets = raw_datasets.map( 401 | pass_through_processors, 402 | remove_columns=next(iter(raw_datasets.values())).column_names, 403 | input_columns=[description_column_name, prompt_column_name], 404 | num_proc=num_workers, 405 | desc="preprocess datasets", 406 | ) 407 | 408 | # We use Accelerate to perform distributed inference 409 | # T5 doesn't support fp16 410 | autocast_kwargs = AutocastKwargs(enabled=(mixed_precision != "fp16")) 411 | 412 | # Now we encode the audio labels with encodec. 413 | ####### B. Encode audio 414 | 415 | logger.info("*** Encode target audio with encodec ***") 416 | 417 | # no need to prepare audio_decoder because used for inference without mixed precision 418 | # see: https://huggingface.co/docs/accelerate/main/en/package_reference/accelerator#accelerate.Accelerator.prepare 419 | if training_args.torch_compile: 420 | audio_decoder = accelerator.prepare_model(model.audio_encoder, evaluation_mode=True) 421 | else: 422 | audio_decoder = model.audio_encoder 423 | 424 | encoder_data_collator = DataCollatorEncodecWithPadding( 425 | feature_extractor, 426 | audio_column_name=target_audio_column_name, 427 | feature_extractor_input_name=feature_extractor_input_name, 428 | max_length=max_target_length, 429 | padding=padding, 430 | ) 431 | encoder_signature = set(inspect.signature(audio_decoder.forward).parameters) 432 | 433 | def apply_audio_decoder(batch): 434 | len_audio = batch.pop("len_audio") 435 | audio_decoder.to(batch["input_values"].device).eval() 436 | if bandwidth is not None: 437 | batch["bandwidth"] = bandwidth 438 | elif "num_quantizers" in encoder_signature: 439 | batch["num_quantizers"] = num_codebooks 440 | elif "num_codebooks" in encoder_signature: 441 | batch["num_codebooks"] = num_codebooks 442 | elif "n_quantizers" in encoder_signature: 443 | batch["n_quantizers"] = num_codebooks 444 | 445 | with torch.no_grad(): 446 | labels = audio_decoder.encode(**batch)["audio_codes"] 447 | output = {} 448 | output["len_audio"] = len_audio 449 | # (1, bsz, codebooks, seq_len) -> (bsz, seq_len, codebooks) 450 | output["labels"] = labels.squeeze(0).transpose(1, 2) 451 | 452 | # if `pad_to_max_length`, the maximum corresponding audio length of the current batch is max_duration*sampling_rate 453 | max_length = len_audio.max() if padding != "max_length" else max_target_length 454 | output["ratio"] = torch.ones_like(len_audio) * labels.shape[-1] / max_length 455 | return output 456 | 457 | # (1, codebooks, seq_len) where seq_len=1 458 | bos_labels = torch.ones((1, num_codebooks, 1)) * audio_encoder_bos_token_id 459 | 460 | def postprocess_dataset(labels): 461 | # (1, codebooks, seq_len) 462 | labels = torch.tensor(labels).unsqueeze(0) 463 | # add bos 464 | labels = torch.cat([bos_labels, labels], dim=-1) 465 | 466 | labels, delay_pattern_mask = build_delay_pattern_mask( 467 | labels, 468 | bos_token_id=audio_encoder_bos_token_id, 469 | pad_token_id=audio_encoder_eos_token_id, 470 | max_length=labels.shape[-1] + num_codebooks, 471 | num_codebooks=num_codebooks, 472 | ) 473 | 474 | # the first ids of the delay pattern mask are precisely labels, we use the rest of the labels mask 475 | # to take care of EOS 476 | # we want labels to look like this: 477 | # - [B, a, b, E, E, E, E] 478 | # - [B, B, c, d, E, E, E] 479 | # - [B, B, B, e, f, E, E] 480 | # - [B, B, B, B, g, h, E] 481 | labels = torch.where(delay_pattern_mask == -1, audio_encoder_eos_token_id, delay_pattern_mask) 482 | 483 | # the first timestamp is associated to a row full of BOS, let's get rid of it 484 | # we also remove the last timestampts (full of PAD) 485 | output = {"labels": labels[:, 1:]} 486 | return output 487 | 488 | for split in vectorized_datasets: 489 | data_loader = DataLoader( 490 | raw_datasets[split], 491 | batch_size=training_args.audio_encoder_per_device_batch_size, 492 | collate_fn=encoder_data_collator, 493 | num_workers=training_args.dataloader_num_workers, 494 | pin_memory=True, 495 | ) 496 | data_loader = accelerator.prepare(data_loader) 497 | total_inference_steps = len(data_loader) 498 | 499 | start_step = get_last_codec_checkpoint_step(os.path.join(data_args.temporary_save_to_disk, split)) 500 | accelerator.wait_for_everyone() 501 | if start_step > 0: 502 | logger.info(f"Resuming {split} from step {start_step}") 503 | # efficiently skip the first n batches 504 | start_step += 1 505 | data_loader = skip_first_batches(data_loader, start_step) 506 | 507 | all_generated_labels = [] 508 | all_lens = [] 509 | if start_step < total_inference_steps: 510 | for i, batch in enumerate(tqdm(data_loader, disable=not accelerator.is_local_main_process)): 511 | cur_step = start_step + i 512 | generate_labels = apply_audio_decoder(batch) 513 | generate_labels = accelerator.pad_across_processes(generate_labels, dim=1, pad_index=0) 514 | generate_labels = accelerator.gather_for_metrics(generate_labels) 515 | 516 | if accelerator.is_main_process: 517 | lab = generate_labels["labels"].cpu().transpose(1, 2).to(torch.int16) 518 | rat = generate_labels["ratio"].cpu().squeeze(1) 519 | lens = generate_labels["len_audio"].cpu().squeeze(1) 520 | lab = [l[:, : int(ratio * length)] for (l, ratio, length) in zip(lab, rat, lens)] 521 | 522 | all_generated_labels.extend(lab) 523 | all_lens.extend(lens) 524 | 525 | if ((cur_step + 1) % data_args.save_codec_steps == 0) or ( 526 | cur_step == total_inference_steps - 1 527 | ): 528 | tmp_labels = Dataset.from_dict({"labels": all_generated_labels, "target_length": all_lens}) 529 | tmp_labels = tmp_labels.map( 530 | postprocess_dataset, 531 | num_proc=data_args.preprocessing_num_workers, # this one is resource consuming if many processor. 532 | input_columns=["labels"], 533 | desc="Postprocessing labeling", 534 | ) 535 | save_codec_checkpoint( 536 | os.path.join(data_args.temporary_save_to_disk, split), tmp_labels, cur_step 537 | ) 538 | all_generated_labels = [] 539 | all_lens = [] 540 | 541 | accelerator.wait_for_everyone() 542 | 543 | if accelerator.is_main_process and len(all_generated_labels) > 0: 544 | tmp_labels = Dataset.from_dict({"labels": all_generated_labels, "target_length": all_lens}) 545 | tmp_labels = tmp_labels.map( 546 | postprocess_dataset, 547 | num_proc=data_args.preprocessing_num_workers, # this one is resource consuming if many processor. 548 | input_columns=["labels"], 549 | desc="Postprocessing labeling", 550 | ) 551 | save_codec_checkpoint(os.path.join(data_args.temporary_save_to_disk, split), tmp_labels, cur_step) 552 | all_generated_labels = [] 553 | all_lens = [] 554 | accelerator.wait_for_everyone() 555 | 556 | del all_generated_labels 557 | accelerator.wait_for_everyone() 558 | 559 | with accelerator.local_main_process_first(): 560 | tmp_labels = load_all_codec_checkpoints(os.path.join(data_args.temporary_save_to_disk, split)).select( 561 | range(len(vectorized_datasets[split])) 562 | ) 563 | logger.info(f"Concatenating {split}: {tmp_labels} with {vectorized_datasets[split]}") 564 | vectorized_datasets[split] = concatenate_datasets([vectorized_datasets[split], tmp_labels], axis=1) 565 | 566 | accelerator.free_memory() 567 | del generate_labels, all_lens 568 | 569 | with accelerator.local_main_process_first(): 570 | # NOTE: filtering is done at the end because in the `datasets` library, caching audio files is done after most operations 571 | # caching audio files is time and disk-space consuming, so we want to avoid it at all costs, especially for large (>1Kh) audio datasets. 572 | # That's also why we avoid to concat the processed datasets (vectorized_datasets) with the audio column present in raw_datasets. 573 | 574 | def is_audio_in_length_range(length): 575 | return length > min_target_length and length < max_target_length 576 | 577 | # filter data that is shorter than min_target_length 578 | vectorized_datasets = vectorized_datasets.filter( 579 | is_audio_in_length_range, 580 | num_proc=num_workers, 581 | input_columns=["target_length"], 582 | ) 583 | 584 | if description_column_name is not None and data_args.max_description_token_length is not None: 585 | with accelerator.local_main_process_first(): 586 | # filter description that is shorter than max_text_length 587 | vectorized_datasets = vectorized_datasets.filter( 588 | lambda x: len(x) < data_args.max_description_token_length, 589 | num_proc=num_workers, 590 | input_columns=["input_ids"], 591 | ) 592 | 593 | if data_args.max_prompt_token_length is not None: 594 | with accelerator.local_main_process_first(): 595 | # filter description that is shorter than max_text_length 596 | vectorized_datasets = vectorized_datasets.filter( 597 | lambda x: len(x) < data_args.max_prompt_token_length, 598 | num_proc=num_workers, 599 | input_columns=["prompt_input_ids"], 600 | ) 601 | 602 | if data_args.save_to_disk is not None and not dataset_was_precomputed: 603 | if accelerator.is_main_process: 604 | vectorized_datasets.save_to_disk( 605 | data_args.save_to_disk, 606 | num_proc=min(data_args.preprocessing_num_workers, len(vectorized_datasets["eval"]) - 1), 607 | ) 608 | accelerator.wait_for_everyone() 609 | logger.info(f"Dataset saved at {data_args.save_to_disk}") 610 | 611 | audio_max_length = None 612 | if padding == "max_length": 613 | audio_max_length = max(vectorized_datasets["train"]["target_length"]) 614 | with accelerator.local_main_process_first(): 615 | max_sample = vectorized_datasets["train"].filter( 616 | lambda x: x == audio_max_length, 617 | num_proc=num_workers, 618 | input_columns=["target_length"], 619 | ) 620 | audio_max_length = max([len(l[0]) for l in max_sample["labels"]]) 621 | 622 | if description_column_name is not None and data_args.max_description_token_length is not None: 623 | with accelerator.local_main_process_first(): 624 | # filter description that is shorter than max_text_length 625 | vectorized_datasets = vectorized_datasets.filter( 626 | lambda x: len(x) < data_args.max_description_token_length, 627 | num_proc=num_workers, 628 | input_columns=["input_ids"], 629 | ) 630 | 631 | if data_args.max_prompt_token_length is not None: 632 | with accelerator.local_main_process_first(): 633 | # filter description that is shorter than max_text_length 634 | vectorized_datasets = vectorized_datasets.filter( 635 | lambda x: len(x) < data_args.max_prompt_token_length, 636 | num_proc=num_workers, 637 | input_columns=["prompt_input_ids"], 638 | ) 639 | 640 | if training_args.group_by_length: 641 | # apply a simple heuristic to take into account audio and text lengths 642 | def add_target_lengths(target_length, prompt, description): 643 | return {"target_length": target_length + len(prompt) + len(description)} 644 | 645 | with accelerator.local_main_process_first(): 646 | vectorized_datasets = vectorized_datasets.map( 647 | add_target_lengths, 648 | num_proc=num_workers, 649 | input_columns=["target_length", "prompt_input_ids", "input_ids"], 650 | ) 651 | 652 | # for large datasets it is advised to run the preprocessing on a 653 | # single machine first with ``args.preprocessing_only`` since there will mostly likely 654 | # be a timeout when running the script in distributed mode. 655 | # In a second step ``args.preprocessing_only`` can then be set to `False` to load the 656 | # cached dataset 657 | if data_args.preprocessing_only and data_args.save_to_disk is None: 658 | raise ValueError( 659 | "`preprocessing_only=True` but `save_to_disk` is not set. The latter should indicates where to save the dataset locally." 660 | ) 661 | elif data_args.preprocessing_only: 662 | logger.info(f"Data preprocessing finished. Files save at {data_args.save_to_disk}") 663 | return 664 | 665 | # 6. Next, we can prepare the training. 666 | 667 | # Let's use word CLAP similary and WER metrics as our evaluation metrics, 668 | def compute_metrics( 669 | audios, 670 | descriptions, 671 | prompts, 672 | device="cpu", 673 | compute_clap_similarity_metric=False, 674 | compute_noise_level_metric=False, 675 | noise_level_to_compute_clean_wer=None, 676 | ): 677 | results = {} 678 | input_ids = descriptions 679 | texts = description_tokenizer.batch_decode(input_ids, skip_special_tokens=True) 680 | prompts = prompt_tokenizer.batch_decode(prompts, skip_special_tokens=True) 681 | audios = [a.float().cpu().numpy() for a in audios] 682 | 683 | if compute_clap_similarity_metric: 684 | clap_score = clap_similarity( 685 | model_args.clap_model_name_or_path, texts, audios, device, input_sampling_rate=sampling_rate 686 | ) 687 | results["clap"] = clap_score 688 | 689 | si_sdr_measures = None 690 | if compute_noise_level_metric: 691 | si_sdr_measures = si_sdr(audios, device, input_sampling_rate=sampling_rate) 692 | 693 | word_error, transcriptions, clean_word_error, noisy_word_error, percent_clean_samples = wer( 694 | model_args.asr_model_name_or_path, 695 | prompts, 696 | audios, 697 | device, 698 | training_args.per_device_eval_batch_size, 699 | sampling_rate, 700 | noise_level_to_compute_clean_wer, 701 | si_sdr_measures, 702 | ) 703 | results["wer"] = word_error 704 | if clean_word_error is not None: 705 | results["clean_wer"] = clean_word_error 706 | results["noisy_word_error"] = noisy_word_error 707 | results["percent_clean_samples"] = percent_clean_samples 708 | 709 | return results, texts, prompts, audios, transcriptions, si_sdr_measures 710 | 711 | # Define Training Schedule 712 | # Store some constants 713 | per_device_train_batch_size = int(training_args.per_device_train_batch_size) 714 | train_batch_size = per_device_train_batch_size * accelerator.num_processes 715 | gradient_accumulation_steps = int(training_args.gradient_accumulation_steps) 716 | per_device_eval_batch_size = int(training_args.per_device_eval_batch_size) 717 | 718 | if training_args.max_steps < 0: 719 | num_epochs = int(training_args.num_train_epochs) 720 | steps_per_epoch = len(vectorized_datasets["train"]) // (train_batch_size * gradient_accumulation_steps) 721 | total_train_steps = steps_per_epoch * num_epochs 722 | elif training_args.max_steps > 0: 723 | logger.info("max_steps is given, it will override any value given in num_train_epochs") 724 | total_train_steps = int(training_args.max_steps) 725 | # Setting a very large number of epochs so we go as many times as necessary over the iterator. 726 | num_epochs = sys.maxsize 727 | steps_per_epoch = total_train_steps 728 | 729 | if training_args.eval_steps is None: 730 | logger.info(f"eval_steps is not set, evaluating at the end of each epoch") 731 | eval_steps = steps_per_epoch 732 | else: 733 | eval_steps = training_args.eval_steps 734 | 735 | if training_args.eval_generation_steps is None: 736 | eval_generation_steps = eval_steps 737 | else: 738 | eval_generation_steps = training_args.eval_generation_steps 739 | 740 | # T5 doesn't support fp16 741 | autocast_kwargs = AutocastKwargs(enabled=(mixed_precision != "fp16")) 742 | 743 | # Define optimizer, LR scheduler, collator 744 | optimizer = torch.optim.AdamW( 745 | params=model.parameters(), 746 | lr=training_args.learning_rate, 747 | betas=(training_args.adam_beta1, training_args.adam_beta2), 748 | eps=training_args.adam_epsilon, 749 | weight_decay=training_args.weight_decay, 750 | ) 751 | 752 | # LR scheduler gets stepped by `num_processes` each time -> account for this in warmup / total steps 753 | lr_scheduler = get_scheduler( 754 | name=training_args.lr_scheduler_type, 755 | optimizer=optimizer, 756 | num_warmup_steps=training_args.get_warmup_steps(total_train_steps) * accelerator.num_processes, 757 | num_training_steps=total_train_steps * accelerator.num_processes, 758 | ) 759 | 760 | # Instantiate custom data collator 761 | data_collator = DataCollatorParlerTTSWithPadding( 762 | prompt_tokenizer=prompt_tokenizer, 763 | description_tokenizer=description_tokenizer, 764 | pad_to_multiple_of=data_args.pad_to_multiple_of, 765 | padding=padding, 766 | prompt_max_length=data_args.max_prompt_token_length, 767 | description_max_length=data_args.max_description_token_length, 768 | audio_max_length=audio_max_length, 769 | ) 770 | 771 | # Prepare everything with accelerate 772 | model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler) 773 | 774 | num_examples = total_train_steps * train_batch_size * gradient_accumulation_steps 775 | logger.info("***** Running training *****") 776 | logger.info(f" Num examples = {num_examples}") 777 | logger.info(" Instantaneous batch size per device =" f" {per_device_train_batch_size}") 778 | logger.info(" Gradient accumulation steps =" f" {gradient_accumulation_steps}") 779 | logger.info( 780 | f" Total train batch size (w. parallel & distributed) = {train_batch_size * gradient_accumulation_steps}" 781 | ) 782 | logger.info(f" Total optimization steps = {total_train_steps}") 783 | 784 | # ======================== Training ================================ 785 | train_time = 0 786 | train_start = time.time() 787 | steps_trained_progress_bar = tqdm( 788 | range(total_train_steps), desc="Train steps ... ", position=0, disable=not accelerator.is_local_main_process 789 | ) 790 | continue_training = True 791 | epochs_trained = 0 792 | cur_step = 0 793 | 794 | checkpoint = None 795 | if training_args.resume_from_checkpoint is not None: 796 | checkpoint = training_args.resume_from_checkpoint 797 | elif last_checkpoint is not None: 798 | checkpoint = last_checkpoint 799 | 800 | if accelerator.is_main_process: 801 | if training_args.push_to_hub: 802 | api = HfApi(token=training_args.hub_token) 803 | 804 | # Create repo (repo_name from args or inferred) 805 | repo_name = training_args.hub_model_id 806 | if repo_name is None: 807 | repo_name = Path(training_args.output_dir).absolute().name 808 | repo_id = api.create_repo(repo_name, exist_ok=True).repo_id 809 | 810 | with open(os.path.join(training_args.output_dir, ".gitignore"), "w+") as gitignore: 811 | if "wandb" not in gitignore: 812 | gitignore.write("wandb\n") 813 | elif training_args.output_dir is not None: 814 | os.makedirs(training_args.output_dir, exist_ok=True) 815 | accelerator.wait_for_everyone() 816 | 817 | # Now save everything to be able to create a single processor later 818 | # make sure all processes wait until data is saved 819 | # only the main process saves them 820 | if accelerator.is_main_process: 821 | # save feature extractor, tokenizer and config 822 | if ( 823 | model_args.prompt_tokenizer_name is None 824 | and model_args.description_tokenizer_name 825 | or (model_args.prompt_tokenizer_name == model_args.description_tokenizer_name) 826 | ): 827 | prompt_tokenizer.save_pretrained(training_args.output_dir) 828 | else: 829 | logger.warning( 830 | f"Prompt tokenizer ('{model_args.prompt_tokenizer_name}') and description tokenizer ('{model_args.description_tokenizer_name}') are not the same. Saving only the prompt tokenizer." 831 | ) 832 | prompt_tokenizer.save_pretrained(training_args.output_dir) 833 | 834 | feature_extractor.save_pretrained(training_args.output_dir) 835 | config.save_pretrained(training_args.output_dir) 836 | accelerator.wait_for_everyone() 837 | 838 | if checkpoint is not None: 839 | accelerator.load_state(checkpoint) 840 | # Find num steps and epoch from saved state string pattern 841 | pattern = r"checkpoint-(\d+)-epoch-(\d+)" 842 | match = re.search(pattern, checkpoint) 843 | cur_step = int(match.group(1)) 844 | epochs_trained = int(match.group(2)) 845 | 846 | logger.info(" Continuing training from checkpoint, will skip to saved global_step") 847 | logger.info(f" Continuing training from epoch {epochs_trained}") 848 | logger.info(f" Continuing training from global step {cur_step}") 849 | 850 | steps_trained_progress_bar.update(cur_step) 851 | 852 | for epoch in range(0, epochs_trained): 853 | with accelerator.local_main_process_first(): 854 | vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed) 855 | 856 | if training_args.max_steps < 0: 857 | # we know exactly the number of steps per epoch, so can skip through the required number of batches 858 | resume_step = (cur_step - epochs_trained * steps_per_epoch) * gradient_accumulation_steps 859 | else: 860 | # Currently we don't know how many steps we've taken in the current epoch 861 | # So we just shuffle the dataset one extra time and start from a fresh epoch 862 | # This is "good enough" for our purposes but not fully correct 863 | resume_step = None 864 | with accelerator.local_main_process_first(): 865 | vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed) 866 | else: 867 | resume_step = None 868 | 869 | gen_kwargs = { 870 | "do_sample": model_args.do_sample, 871 | "temperature": model_args.temperature, 872 | "max_length": model_args.max_length, 873 | # Because of the delayed pattern mask, generation might stop earlier because of unexpected behaviour 874 | # on the first tokens of the codebooks that are delayed. 875 | # This fix the issue. 876 | "min_new_tokens": num_codebooks + 1, 877 | } 878 | 879 | # Define gradient update step fn 880 | def train_step( 881 | batch, 882 | accelerator, 883 | autocast_kwargs, 884 | num_items_in_batch, 885 | gradient_accumulation_steps, 886 | ): 887 | if mixed_precision == "fp16": 888 | # fp16 doesn't work with T5-like models 889 | with accelerator.autocast(autocast_handler=autocast_kwargs): 890 | if training_args.parallel_mode.value != "distributed": 891 | encoder_outputs = model.text_encoder( 892 | input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None) 893 | ) 894 | else: 895 | encoder_outputs = model.module.text_encoder( 896 | input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None) 897 | ) 898 | # we optionnally project last_hidden_state to avoid recomputing every time 899 | encoder_hidden_states = encoder_outputs.last_hidden_state 900 | if ( 901 | config.text_encoder.hidden_size != config.decoder.hidden_size 902 | and config.decoder.cross_attention_hidden_size is None 903 | ): 904 | encoder_hidden_states = ( 905 | model.enc_to_dec_proj(encoder_hidden_states) 906 | if training_args.parallel_mode.value != "distributed" 907 | else model.module.enc_to_dec_proj(encoder_hidden_states) 908 | ) 909 | 910 | if batch.get("attention_mask", None) is not None: 911 | encoder_hidden_states = encoder_hidden_states * batch.get("attention_mask", None)[..., None] 912 | 913 | encoder_outputs.last_hidden_state = encoder_hidden_states 914 | batch["encoder_outputs"] = encoder_outputs 915 | 916 | outputs = model(**batch, loss_reduction="sum") 917 | # CE (data) loss 918 | ce_loss = (outputs.loss * gradient_accumulation_steps * accelerator.num_processes) / num_items_in_batch 919 | 920 | metrics = {"loss": ce_loss} 921 | 922 | # per CE loss 923 | per_codebook_losses = outputs.per_codebook_losses 924 | metrics.update({f"codebook_{i}_loss": ((l * gradient_accumulation_steps * accelerator.num_processes) / num_items_in_batch) for (i,l) in enumerate(per_codebook_losses)}) 925 | return ce_loss, metrics 926 | 927 | # Define eval fn 928 | def eval_step( 929 | batch, 930 | accelerator, 931 | autocast_kwargs, 932 | ): 933 | eval_model = model if not training_args.torch_compile else model._orig_mod 934 | 935 | if mixed_precision == "fp16": 936 | # fp16 doesn't work with T5-like models 937 | with accelerator.autocast(autocast_handler=autocast_kwargs): 938 | if training_args.parallel_mode.value != "distributed": 939 | encoder_outputs = model.text_encoder( 940 | input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None) 941 | ) 942 | else: 943 | encoder_outputs = model.module.text_encoder( 944 | input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None) 945 | ) 946 | # we optionnally project last_hidden_state to avoid recomputing every time 947 | encoder_hidden_states = encoder_outputs.last_hidden_state 948 | if ( 949 | config.text_encoder.hidden_size != config.decoder.hidden_size 950 | and config.decoder.cross_attention_hidden_size is None 951 | ): 952 | encoder_hidden_states = ( 953 | model.enc_to_dec_proj(encoder_hidden_states) 954 | if training_args.parallel_mode.value != "distributed" 955 | else model.module.enc_to_dec_proj(encoder_hidden_states) 956 | ) 957 | 958 | if batch.get("attention_mask", None) is not None: 959 | encoder_hidden_states = encoder_hidden_states * batch.get("attention_mask", None)[..., None] 960 | 961 | encoder_outputs.last_hidden_state = encoder_hidden_states 962 | batch["encoder_outputs"] = encoder_outputs 963 | 964 | with torch.no_grad(): 965 | outputs = eval_model(**batch) 966 | # CE (data) loss 967 | ce_loss = outputs.loss 968 | metrics = {"loss": ce_loss} 969 | 970 | # per CE loss 971 | per_codebook_losses = outputs.per_codebook_losses 972 | metrics.update({f"codebook_{i}_loss": l for (i,l) in enumerate(per_codebook_losses)}) 973 | return metrics 974 | 975 | def generate_step(batch, accelerator): 976 | batch.pop("decoder_attention_mask", None) 977 | eval_model = accelerator.unwrap_model(model, keep_fp32_wrapper=True) 978 | if training_args.torch_compile: 979 | # if the model is compiled, we use the original model bc compile is not compatible with .generate 980 | eval_model = model._orig_mod 981 | 982 | # since we've might have loaded the weights in fp32, we have to autocast to ensure FA2 weights are in half-precision. 983 | # with accelerator.autocast(autocast_handler=AutocastKwargs(enabled=(attn_implementation=="flash_attention_2"))): 984 | output_audios = eval_model.generate(**batch, **gen_kwargs) 985 | output_audios = accelerator.pad_across_processes(output_audios, dim=1, pad_index=0) 986 | return output_audios 987 | 988 | model.train() 989 | 990 | total_batched_samples = resume_step if resume_step is not None else 0 991 | for epoch in range(epochs_trained, num_epochs): 992 | with accelerator.local_main_process_first(): 993 | vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed) 994 | sampler = None 995 | if training_args.group_by_length: 996 | sampler = LengthGroupedSampler(train_batch_size, lengths=vectorized_datasets["train"]["target_length"]) 997 | train_dataloader = DataLoader( 998 | vectorized_datasets["train"], 999 | collate_fn=data_collator, 1000 | batch_size=per_device_train_batch_size, 1001 | sampler=sampler, 1002 | shuffle=not training_args.group_by_length, 1003 | num_workers=training_args.dataloader_num_workers, 1004 | pin_memory=training_args.dataloader_pin_memory, 1005 | ) 1006 | train_dataloader = accelerator.prepare(train_dataloader) 1007 | if hasattr(train_dataloader, "dataset") and isinstance(train_dataloader.dataset, IterableDataset): 1008 | train_dataloader.dataset.set_epoch(epoch) 1009 | 1010 | if resume_step is not None: 1011 | # Skip the first N batches in the dataloader when resuming from a checkpoint 1012 | logger.info(f" Skip first {resume_step} batches") 1013 | train_dataloader = accelerator.skip_first_batches(train_dataloader, resume_step) 1014 | resume_step = None 1015 | accelerator.wait_for_everyone() 1016 | 1017 | # We chunkify the epoch iterator into gradient accumulation steps `n` batches 1018 | train_iterator = iter(train_dataloader) 1019 | num_steps_in_epoch = len(train_dataloader) 1020 | remainder = num_steps_in_epoch % gradient_accumulation_steps 1021 | remainder = remainder if remainder != 0 else gradient_accumulation_steps 1022 | total_updates = math.ceil(num_steps_in_epoch / gradient_accumulation_steps) 1023 | 1024 | update_step = -1 1025 | for _ in range(total_updates): 1026 | update_step += 1 1027 | 1028 | # preload the total batch per step 1029 | batch_samples = [] 1030 | num_batches_in_step = gradient_accumulation_steps if update_step != (total_updates - 1) else remainder 1031 | for _ in range(num_batches_in_step): 1032 | batch_samples += [next(train_iterator)] 1033 | 1034 | # get num items in batch - if different than BOS and than -100 1035 | num_items_in_batch = sum([(batch["labels"].ne(audio_encoder_bos_token_id) | batch["labels"].ne(-100) | batch["labels"].ne(audio_encoder_eos_token_id)).sum((0,1))[0] for batch in batch_samples]) 1036 | num_items_in_batch = accelerator.gather(num_items_in_batch).sum().item() 1037 | 1038 | # losses = [] 1039 | for i,batch in enumerate(batch_samples): 1040 | total_batched_samples += 1 1041 | ctx = model.no_sync if (i < len(batch_samples) - 1 and accelerator.num_processes > 1) else contextlib.nullcontext 1042 | 1043 | with ctx(): 1044 | loss, train_metric = train_step(batch, accelerator, autocast_kwargs, num_items_in_batch, gradient_accumulation_steps) 1045 | accelerator.backward(loss) 1046 | # losses.append(loss.detach()) 1047 | 1048 | grad_norm = accelerator.clip_grad_norm_(model.parameters(), training_args.max_grad_norm) 1049 | optimizer.step() 1050 | lr_scheduler.step() 1051 | optimizer.zero_grad() 1052 | 1053 | # The accelerator has performed an optimization step behind the scenes 1054 | steps_trained_progress_bar.update(1) 1055 | cur_step += 1 1056 | 1057 | # losses = accelerator.gather(sum(losses)).sum().item() / (accelerator.num_processes * gradient_accumulation_steps) 1058 | 1059 | if cur_step % training_args.logging_steps == 0: 1060 | steps_trained_progress_bar.write( 1061 | f"Step... ({cur_step} / {total_train_steps} | Loss:" 1062 | f" {train_metric['loss']}, Learning Rate:" 1063 | f" {lr_scheduler.get_last_lr()[0]})" 1064 | ) 1065 | train_metric["grad_norm"] = grad_norm.detach().item() if isinstance(grad_norm, torch.Tensor) else grad_norm 1066 | log_metric( 1067 | accelerator, 1068 | metrics=train_metric, 1069 | learning_rate=lr_scheduler.get_last_lr()[0], 1070 | train_time=train_time + time.time() - train_start, 1071 | step=cur_step, 1072 | epoch=epoch, 1073 | prefix="train", 1074 | ) 1075 | 1076 | # save checkpoint and weights after each save_steps and at the end of training 1077 | if (cur_step % training_args.save_steps == 0) or cur_step == total_train_steps: 1078 | intermediate_dir = os.path.join(training_args.output_dir, f"checkpoint-{cur_step}-epoch-{epoch}") 1079 | # safe_serialization=False to avoid shared tensors saving issue (TODO(YL): it's a temporary fix) 1080 | # https://github.com/huggingface/transformers/issues/27293#issuecomment-1872560074 1081 | accelerator.save_state(output_dir=intermediate_dir, safe_serialization=False) 1082 | accelerator.wait_for_everyone() 1083 | if accelerator.is_main_process: 1084 | rotate_checkpoints( 1085 | training_args.save_total_limit, output_dir=training_args.output_dir, logger=logger 1086 | ) 1087 | 1088 | if cur_step == total_train_steps: 1089 | # un-wrap student model for save 1090 | unwrapped_model = accelerator.unwrap_model(model) 1091 | unwrapped_model.save_pretrained(training_args.output_dir) 1092 | 1093 | if training_args.push_to_hub: 1094 | api.upload_folder( 1095 | repo_id=repo_id, 1096 | folder_path=training_args.output_dir, 1097 | commit_message=f"Saving train state of step {cur_step}", 1098 | run_as_future=True, 1099 | ) 1100 | accelerator.wait_for_everyone() 1101 | 1102 | if training_args.do_eval and (cur_step % eval_steps == 0 or cur_step == total_train_steps): 1103 | train_time += time.time() - train_start 1104 | # ======================== Evaluating ============================== 1105 | model.eval() 1106 | eval_metrics = [] 1107 | eval_preds = [] 1108 | eval_descriptions = [] 1109 | eval_prompts = [] 1110 | eval_start = time.time() 1111 | 1112 | # release training input batch 1113 | batch = release_memory(batch) 1114 | 1115 | validation_dataloader = DataLoader( 1116 | vectorized_datasets["eval"], 1117 | collate_fn=data_collator, 1118 | batch_size=per_device_eval_batch_size, 1119 | drop_last=False, 1120 | num_workers=training_args.eval_dataloader_num_workers, 1121 | pin_memory=training_args.dataloader_pin_memory, 1122 | ) 1123 | validation_dataloader = accelerator.prepare(validation_dataloader) 1124 | 1125 | for batch in tqdm( 1126 | validation_dataloader, 1127 | desc=f"Evaluating - Inference ...", 1128 | position=2, 1129 | disable=not accelerator.is_local_main_process, 1130 | ): 1131 | # Model forward 1132 | eval_metric = eval_step(batch, accelerator, autocast_kwargs) 1133 | eval_metric = accelerator.gather_for_metrics(eval_metric) 1134 | eval_metric = {key: val.unsqueeze(0) if val.ndim == 0 else val for (key,val) in eval_metric.items()} 1135 | eval_metrics.append(eval_metric) 1136 | 1137 | if training_args.predict_with_generate and (cur_step % eval_generation_steps == 0 or cur_step == total_train_steps): 1138 | validation_dataloader = DataLoader( 1139 | vectorized_datasets["eval"], 1140 | collate_fn=data_collator, 1141 | batch_size=per_device_eval_batch_size, 1142 | drop_last=False, 1143 | num_workers=training_args.eval_dataloader_num_workers, 1144 | pin_memory=training_args.dataloader_pin_memory, 1145 | ) 1146 | validation_dataloader = accelerator.prepare(validation_dataloader) 1147 | # generation 1148 | for batch in tqdm( 1149 | validation_dataloader, 1150 | desc=f"Evaluating - Generation ...", 1151 | position=2, 1152 | disable=not accelerator.is_local_main_process, 1153 | ): 1154 | generated_audios = generate_step(batch, accelerator) 1155 | # Gather all predictions and targets 1156 | generated_audios, input_ids, prompts = accelerator.pad_across_processes( 1157 | (generated_audios, batch["input_ids"], batch["prompt_input_ids"]), dim=1, pad_index=0 1158 | ) 1159 | generated_audios, input_ids, prompts = accelerator.gather_for_metrics( 1160 | (generated_audios, input_ids, prompts) 1161 | ) 1162 | eval_preds.extend(generated_audios.to("cpu")) 1163 | eval_descriptions.extend(input_ids.to("cpu")) 1164 | eval_prompts.extend(prompts.to("cpu")) 1165 | 1166 | eval_time = time.time() - eval_start 1167 | # normalize eval metrics 1168 | eval_metrics = { 1169 | key: torch.mean(torch.cat([d[key] for d in eval_metrics])).to("cpu") for key in eval_metrics[0] 1170 | } 1171 | 1172 | # compute metrics 1173 | metrics_desc = "" 1174 | if training_args.predict_with_generate and (cur_step % eval_generation_steps == 0 or cur_step == total_train_steps): 1175 | if accelerator.is_local_main_process: 1176 | ( 1177 | metric_values, 1178 | pred_descriptions, 1179 | pred_prompts, 1180 | audios, 1181 | transcriptions, 1182 | si_sdr_measures, 1183 | ) = compute_metrics( 1184 | eval_preds, 1185 | eval_descriptions, 1186 | eval_prompts, 1187 | accelerator.device, 1188 | training_args.compute_clap_similarity_metric, 1189 | training_args.compute_noise_level_metric, 1190 | training_args.noise_level_to_compute_clean_wer, 1191 | ) 1192 | eval_metrics.update(metric_values) 1193 | metrics_desc = " ".join([f"Eval {key}: {value} |" for key, value in metric_values.items()]) 1194 | if "wandb" in training_args.report_to: 1195 | log_pred( 1196 | accelerator, 1197 | pred_descriptions, 1198 | pred_prompts, 1199 | transcriptions, 1200 | audios, 1201 | si_sdr_measures, 1202 | sampling_rate=sampling_rate, 1203 | step=cur_step, 1204 | prefix="eval", 1205 | ) 1206 | accelerator.wait_for_everyone() 1207 | 1208 | # Print metrics and update progress bar 1209 | if accelerator.is_local_main_process: 1210 | steps_trained_progress_bar.write( 1211 | f"Eval results for step ({cur_step} / {total_train_steps} | Eval Loss: {eval_metrics['loss']} |" 1212 | f" {metrics_desc})" 1213 | ) 1214 | 1215 | log_metric( 1216 | accelerator, 1217 | metrics=eval_metrics, 1218 | train_time=eval_time, 1219 | step=cur_step, 1220 | epoch=epoch, 1221 | prefix="eval", 1222 | ) 1223 | 1224 | # release eval batch and relax metrics 1225 | eval_metrics, eval_preds, eval_descriptions, eval_prompts, batch, eval_metric = release_memory( 1226 | eval_metrics, eval_preds, eval_descriptions, eval_prompts, batch, eval_metric 1227 | ) 1228 | if training_args.predict_with_generate and (cur_step % eval_generation_steps == 0 or cur_step == total_train_steps): 1229 | generated_audios, input_ids, prompts = release_memory(generated_audios, input_ids, prompts) 1230 | 1231 | # train mode 1232 | model.train() 1233 | 1234 | # flush the train metrics 1235 | train_start = time.time() 1236 | 1237 | # break condition 1238 | if cur_step == total_train_steps: 1239 | continue_training = False 1240 | break 1241 | 1242 | if not continue_training: 1243 | break 1244 | 1245 | accelerator.end_training() 1246 | 1247 | 1248 | if __name__ == "__main__": 1249 | main() 1250 | -------------------------------------------------------------------------------- /training/utils.py: -------------------------------------------------------------------------------- 1 | import os 2 | import re 3 | import shutil 4 | from dataclasses import field 5 | from pathlib import Path 6 | from typing import Dict, List 7 | 8 | import torch 9 | from datasets import concatenate_datasets, load_from_disk 10 | from wandb import Audio 11 | from datasets import load_from_disk, concatenate_datasets 12 | 13 | 14 | def list_field(default=None, metadata=None): 15 | return field(default_factory=lambda: default, metadata=metadata) 16 | 17 | 18 | _RE_CHECKPOINT = re.compile(r"^checkpoint-(\d+)-epoch-(\d+)$") 19 | CHECKPOINT_CODEC_PREFIX = "checkpoint" 20 | _RE_CODEC_CHECKPOINT = re.compile(r"^checkpoint-(\d+)$") 21 | 22 | 23 | def get_last_checkpoint(folder): 24 | content = os.listdir(folder) 25 | checkpoints = [ 26 | path 27 | for path in content 28 | if _RE_CHECKPOINT.search(path) is not None and os.path.isdir(os.path.join(folder, path)) 29 | ] 30 | if len(checkpoints) == 0: 31 | return 32 | return os.path.join(folder, max(checkpoints, key=lambda x: int(_RE_CHECKPOINT.search(x).groups()[0]))) 33 | 34 | 35 | def sorted_checkpoints(output_dir=None, checkpoint_prefix="checkpoint") -> List[str]: 36 | """Helper function to sort saved checkpoints from oldest to newest.""" 37 | ordering_and_checkpoint_path = [] 38 | 39 | glob_checkpoints = [str(x) for x in Path(output_dir).glob(f"{checkpoint_prefix}-*") if os.path.isdir(x)] 40 | 41 | for path in glob_checkpoints: 42 | regex_match = re.match(f".*{checkpoint_prefix}-([0-9]+)", path) 43 | if regex_match is not None and regex_match.groups() is not None: 44 | ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path)) 45 | 46 | checkpoints_sorted = sorted(ordering_and_checkpoint_path) 47 | checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted] 48 | return checkpoints_sorted 49 | 50 | 51 | def rotate_checkpoints(save_total_limit=None, output_dir=None, checkpoint_prefix="checkpoint", logger=None) -> None: 52 | """Helper function to delete old checkpoints.""" 53 | if save_total_limit is None or save_total_limit <= 0: 54 | return 55 | # Check if we should delete older checkpoint(s) 56 | checkpoints_sorted = sorted_checkpoints(output_dir=output_dir, checkpoint_prefix=checkpoint_prefix) 57 | if len(checkpoints_sorted) <= save_total_limit: 58 | return 59 | 60 | number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - save_total_limit) 61 | checkpoints_to_be_deleted = checkpoints_sorted[:number_of_checkpoints_to_delete] 62 | for checkpoint in checkpoints_to_be_deleted: 63 | logger.info(f"Deleting older checkpoint [{checkpoint}] due to args.save_total_limit") 64 | shutil.rmtree(checkpoint, ignore_errors=True) 65 | 66 | 67 | def save_codec_checkpoint(output_dir, dataset, step): 68 | checkpoint_path = f"{CHECKPOINT_CODEC_PREFIX}-{step}" 69 | output_path = os.path.join(output_dir, checkpoint_path) 70 | dataset.save_to_disk(output_path) 71 | 72 | 73 | def load_codec_checkpoint(checkpoint_path): 74 | dataset = load_from_disk(checkpoint_path) 75 | return dataset 76 | 77 | 78 | def sorted_codec_checkpoints(output_dir=None) -> List[str]: 79 | """Helper function to sort saved checkpoints from oldest to newest.""" 80 | ordering_and_checkpoint_path = [] 81 | 82 | glob_checkpoints = [str(x) for x in Path(output_dir).glob(f"{CHECKPOINT_CODEC_PREFIX}-*")] 83 | 84 | for path in glob_checkpoints: 85 | regex_match = re.match(f".*{CHECKPOINT_CODEC_PREFIX}-([0-9]+)", path) 86 | if regex_match is not None and regex_match.groups() is not None: 87 | ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path)) 88 | 89 | checkpoints_sorted = sorted(ordering_and_checkpoint_path) 90 | checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted] 91 | return checkpoints_sorted 92 | 93 | 94 | def load_all_codec_checkpoints(output_dir=None) -> List[str]: 95 | """Helper function to load and concat all checkpoints.""" 96 | checkpoints_sorted = sorted_codec_checkpoints(output_dir=output_dir) 97 | datasets = [load_from_disk(checkpoint) for checkpoint in checkpoints_sorted] 98 | datasets = concatenate_datasets(datasets, axis=0) 99 | return datasets 100 | 101 | 102 | def get_last_codec_checkpoint_step(folder) -> int: 103 | if not os.path.exists(folder) or not os.path.isdir(folder): 104 | os.makedirs(folder, exist_ok=True) 105 | return 0 106 | content = os.listdir(folder) 107 | checkpoints = [path for path in content if _RE_CODEC_CHECKPOINT.search(path) is not None] 108 | if len(checkpoints) == 0: 109 | return 0 110 | last_checkpoint = os.path.join( 111 | folder, max(checkpoints, key=lambda x: int(_RE_CODEC_CHECKPOINT.search(x).groups()[0])) 112 | ) 113 | # Find num steps saved state string pattern 114 | pattern = r"checkpoint-(\d+)" 115 | match = re.search(pattern, last_checkpoint) 116 | cur_step = int(match.group(1)) 117 | return cur_step 118 | 119 | 120 | def log_metric( 121 | accelerator, 122 | metrics: Dict, 123 | train_time: float, 124 | step: int, 125 | epoch: int, 126 | learning_rate: float = None, 127 | prefix: str = "train", 128 | ): 129 | """Helper function to log all training/evaluation metrics with the correct prefixes and styling.""" 130 | log_metrics = {} 131 | for k, v in metrics.items(): 132 | if "codebook" in k: 133 | log_metrics[f"codebook_{prefix}/{k}"] = v 134 | else: 135 | log_metrics[f"{prefix}/{k}"] = v 136 | log_metrics[f"{prefix}/time"] = train_time 137 | log_metrics[f"{prefix}/epoch"] = epoch 138 | if learning_rate is not None: 139 | log_metrics[f"{prefix}/learning_rate"] = learning_rate 140 | accelerator.log(log_metrics, step=step) 141 | 142 | 143 | def log_pred( 144 | accelerator, 145 | pred_descriptions: List[str], 146 | pred_prompts: List[str], 147 | transcriptions: List[str], 148 | audios: List[torch.Tensor], 149 | si_sdr_measures: List[float], 150 | sampling_rate: int, 151 | step: int, 152 | prefix: str = "eval", 153 | num_lines: int = 200000, 154 | ): 155 | """Helper function to log target/predicted transcriptions to weights and biases (wandb).""" 156 | if accelerator.is_main_process: 157 | wandb_tracker = accelerator.get_tracker("wandb") 158 | # pretty name for current step: step 50000 -> step 50k 159 | cur_step_pretty = f"{int(step // 1000)}k" if step > 1000 else step 160 | prefix_pretty = prefix.replace("/", "-") 161 | 162 | if si_sdr_measures is None: 163 | # convert str data to a wandb compatible format 164 | str_data = [ 165 | [pred_descriptions[i], pred_prompts[i], transcriptions[i]] for i in range(len(pred_descriptions)) 166 | ] 167 | # log as a table with the appropriate headers 168 | wandb_tracker.log_table( 169 | table_name=f"predictions/{prefix_pretty}-step-{cur_step_pretty}", 170 | columns=["Target descriptions", "Target prompts", "Predicted transcriptions"], 171 | data=str_data[:num_lines], 172 | step=step, 173 | commit=False, 174 | ) 175 | else: 176 | # convert str data to a wandb compatible format 177 | str_data = [ 178 | [pred_descriptions[i], pred_prompts[i], transcriptions[i], si_sdr_measures[i]] 179 | for i in range(len(pred_descriptions)) 180 | ] 181 | # log as a table with the appropriate headers 182 | wandb_tracker.log_table( 183 | table_name=f"predictions/{prefix_pretty}-step-{cur_step_pretty}", 184 | columns=["Target descriptions", "Target prompts", "Predicted transcriptions", "Noise estimation"], 185 | data=str_data[:num_lines], 186 | step=step, 187 | commit=False, 188 | ) 189 | 190 | # wandb can only loads 100 audios per step 191 | wandb_tracker.log( 192 | { 193 | "Speech samples": [ 194 | Audio( 195 | audio, 196 | caption=f"{pred_prompts[i]} --- DESCRIPTION: {pred_descriptions[i]}", 197 | sample_rate=sampling_rate, 198 | ) 199 | for (i, audio) in enumerate(audios[: min(len(audios), 100)]) 200 | ] 201 | }, 202 | step=step, 203 | ) 204 | --------------------------------------------------------------------------------