├── .dockerignore ├── .gitignore ├── .gitmodules ├── Dockerfile ├── LICENSE ├── README.md ├── app.py ├── config.json ├── evaluation ├── __init__.py ├── tsne_analysis_baseline │ ├── run_tsne_analysis.sh │ ├── test │ │ ├── assets │ │ │ ├── .gitignore │ │ │ ├── dataset │ │ │ │ ├── image_10.npz │ │ │ │ ├── image_10.png │ │ │ │ ├── image_5.npz │ │ │ │ ├── image_5.png │ │ │ │ ├── image_6.npz │ │ │ │ ├── image_6.png │ │ │ │ ├── image_7.npz │ │ │ │ ├── image_7.png │ │ │ │ ├── image_9.npz │ │ │ │ └── image_9.png │ │ │ ├── model_a │ │ │ │ ├── image_10.npz │ │ │ │ ├── image_10.png │ │ │ │ ├── image_5.npz │ │ │ │ ├── image_5.png │ │ │ │ ├── image_6.npz │ │ │ │ ├── image_6.png │ │ │ │ ├── image_7.npz │ │ │ │ ├── image_7.png │ │ │ │ ├── image_9.npz │ │ │ │ └── image_9.png │ │ │ └── model_b │ │ │ │ ├── image_0.npz │ │ │ │ ├── image_0.png │ │ │ │ ├── image_1.npz │ │ │ │ ├── image_1.png │ │ │ │ ├── image_2.npz │ │ │ │ ├── image_2.png │ │ │ │ ├── image_3.npz │ │ │ │ ├── image_3.png │ │ │ │ ├── image_5.npz │ │ │ │ ├── image_5.png │ │ │ │ └── image_6.png │ │ ├── test_gen_tsne.py │ │ └── test_grid.py │ ├── tmp_requirements.txt │ ├── tsne_analysis.py │ ├── tsne_evaluation_utils │ │ ├── grid.py │ │ └── metric.py │ └── tsne_output │ │ ├── data.csv │ │ ├── distances.csv │ │ ├── models_scatter.png │ │ ├── stats.csv │ │ ├── tsne_dataset.png │ │ ├── tsne_model_a.png │ │ └── tsne_model_b.png └── utils_tsne.py ├── images ├── Cartoons_example.jpeg └── Faces_example.jpeg ├── losses ├── __init__.py └── utils_loss.py ├── models ├── __init__.py ├── avatar_generator_model.py ├── cdann.py ├── decoder.py ├── denoiser.py ├── discriminator.py ├── encoder.py └── inception.py ├── requirements.txt ├── scripts ├── copyFiles.sh ├── download_faces.py ├── keepFiles.sh ├── plot_utils.py └── preprocessing_cartoons_data.py ├── sweeps ├── sweep-bs-1.yaml └── sweep-rs-1.yaml ├── train.py └── utils └── __init__.py /.dockerignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | pip-wheel-metadata/ 24 | share/python-wheels/ 25 | *.egg-info/ 26 | .installed.cfg 27 | *.egg 28 | MANIFEST 29 | 30 | # PyInstaller 31 | # Usually these files are written by a python script from a template 32 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 33 | *.manifest 34 | *.spec 35 | 36 | # Installer logs 37 | pip-log.txt 38 | pip-delete-this-directory.txt 39 | 40 | # Unit test / coverage reports 41 | htmlcov/ 42 | .tox/ 43 | .nox/ 44 | .coverage 45 | .coverage.* 46 | .cache 47 | nosetests.xml 48 | coverage.xml 49 | *.cover 50 | *.py,cover 51 | .hypothesis/ 52 | .pytest_cache/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | target/ 76 | 77 | # Jupyter Notebook 78 | .ipynb_checkpoints 79 | 80 | # IPython 81 | profile_default/ 82 | ipython_config.py 83 | 84 | # pyenv 85 | .python-version 86 | 87 | # pipenv 88 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 89 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 90 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 91 | # install all needed dependencies. 92 | #Pipfile.lock 93 | 94 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 95 | __pypackages__/ 96 | 97 | # Celery stuff 98 | celerybeat-schedule 99 | celerybeat.pid 100 | 101 | # SageMath parsed files 102 | *.sage.py 103 | 104 | # Environments 105 | .env 106 | .venv 107 | env/ 108 | venv/ 109 | ENV/ 110 | env.bak/ 111 | venv.bak/ 112 | 113 | # Spyder project settings 114 | .spyderproject 115 | .spyproject 116 | 117 | # Rope project settings 118 | .ropeproject 119 | 120 | # mkdocs documentation 121 | /site 122 | 123 | # mypy 124 | .mypy_cache/ 125 | .dmypy.json 126 | dmypy.json 127 | 128 | # Pyre type checker 129 | .pyre/ 130 | 131 | #jpg 132 | *.jpg 133 | 134 | 135 | #git 136 | .git 137 | .gitignore 138 | .gitmodules 139 | *.md 140 | LICENSE 141 | 142 | #docker 143 | Dockerfile 144 | .DS_Store 145 | 146 | #files 147 | __pycache__ 148 | avatar-image-generator-app/ 149 | images/ 150 | notebooks/ 151 | data/ 152 | datasets/ 153 | weights_trained/ 154 | wandb/ 155 | preprocessing/ 156 | 157 | #nohup 158 | *.out -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | pip-wheel-metadata/ 24 | share/python-wheels/ 25 | *.egg-info/ 26 | .installed.cfg 27 | *.egg 28 | MANIFEST 29 | 30 | # PyInstaller 31 | # Usually these files are written by a python script from a template 32 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 33 | *.manifest 34 | *.spec 35 | 36 | # Installer logs 37 | pip-log.txt 38 | pip-delete-this-directory.txt 39 | 40 | # Unit test / coverage reports 41 | htmlcov/ 42 | .tox/ 43 | .nox/ 44 | .coverage 45 | .coverage.* 46 | .cache 47 | nosetests.xml 48 | coverage.xml 49 | *.cover 50 | *.py,cover 51 | .hypothesis/ 52 | .pytest_cache/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | target/ 76 | 77 | # Jupyter Notebook 78 | .ipynb_checkpoints 79 | 80 | # IPython 81 | profile_default/ 82 | ipython_config.py 83 | 84 | # pyenv 85 | .python-version 86 | 87 | # pipenv 88 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 89 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 90 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 91 | # install all needed dependencies. 92 | #Pipfile.lock 93 | 94 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 95 | __pypackages__/ 96 | 97 | # Celery stuff 98 | celerybeat-schedule 99 | celerybeat.pid 100 | 101 | # SageMath parsed files 102 | *.sage.py 103 | 104 | # Environments 105 | .env 106 | .venv 107 | env/ 108 | venv/ 109 | ENV/ 110 | env.bak/ 111 | venv.bak/ 112 | 113 | # Spyder project settings 114 | .spyderproject 115 | .spyproject 116 | 117 | # Rope project settings 118 | .ropeproject 119 | 120 | # mkdocs documentation 121 | /site 122 | 123 | # mypy 124 | .mypy_cache/ 125 | .dmypy.json 126 | dmypy.json 127 | 128 | # Pyre type checker 129 | .pyre/ 130 | 131 | #jpg - data 132 | data/ 133 | *.jpg 134 | weights_trained/ 135 | weights/ 136 | wandb/ 137 | 138 | #nohup 139 | *.out -------------------------------------------------------------------------------- /.gitmodules: -------------------------------------------------------------------------------- 1 | [submodule "avatar-image-generator-app"] 2 | path = avatar-image-generator-app 3 | url = https://github.com/paper2code-pucp/avatar-image-generator-app.git 4 | -------------------------------------------------------------------------------- /Dockerfile: -------------------------------------------------------------------------------- 1 | FROM ubuntu:20.04 2 | ENV PATH="/root/miniconda3/bin:${PATH}" 3 | ARG PATH="/root/miniconda3/bin:${PATH}" 4 | 5 | RUN apt update \ 6 | && apt install -y python3-dev wget libgl1-mesa-dev libglib2.0-0 libsm6 libxext6 libxrender-dev 7 | 8 | RUN wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \ 9 | && mkdir root/.conda \ 10 | && sh Miniconda3-latest-Linux-x86_64.sh -b \ 11 | && rm -f Miniconda3-latest-Linux-x86_64.sh 12 | 13 | RUN conda create -y -n ml python=3.7 14 | 15 | COPY . src/ 16 | RUN /bin/bash -c "cd src \ 17 | && source activate ml \ 18 | && pip install -r requirements.txt" 19 | -------------------------------------------------------------------------------- /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 [yyyy] [name of copyright owner] 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 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Avatar Image Generator 2 | 3 | Faces Domain
4 | 5 | 6 | Generated Cartoons
7 | 8 | 9 | Based on the paper XGAN: https://arxiv.org/abs/1711.05139 10 | 11 | ## The problem 12 | 13 | This repo aims to contribute to the daunting problem of generating a cartoon given the picture of a face.

14 | This is an image-to-image translation problem, which involves many classic computer vision tasks, like style transfer, super-resolution, colorization and semnantic segmentation. Also, this is a many-to-many mapping, which means that for a given face there are multiple valid cartoons, and for a given cartoon there are multiple valid faces too.
15 | 16 | ## Dataset 17 | 18 | Faces dataset: we use the VggFace dataset (https://www.robots.ox.ac.uk/~vgg/data/vgg_face/) from the University of Oxford 19 | 20 | Cartoon dataset: we use the CartoonSet dataset from Google (https://google.github.io/cartoonset/), both the versions of 10000 and 100000 items. 21 | 22 | We filtered out the data just to keep realistic cartoons and faces images. This code is in `scripts`. To download the dataset: 23 | 24 | 1. `pip3 install gdown` 25 | 2. `gdown https://drive.google.com/uc?id=1tfMW5vZ0aUFnl-fSYpWexoGRKGSQsStL` 26 | 3. `unzip datasets.zip` 27 | 28 | ## Directory structure 29 | 30 | `config.json`: contains the model configuration to train the model and deploy the app 31 | 32 | `weights`: contains weights that we saved the last time we train the model. 33 | 34 | ``` 35 | ├── app.py 36 | ├── avatar-image-generator-app 37 | ├── config.json 38 | ├── Dockerfile 39 | ├── images 40 | │   ├── Cartoons_example.jpeg 41 | │   └── Faces_example.jpeg 42 | ├── LICENSE 43 | ├── losses 44 | │   └── __init__.py 45 | ├── models 46 | │   ├── avatar_generator_model.py 47 | │   ├── cdann.py 48 | │   ├── decoder.py 49 | │   ├── denoiser.py 50 | │   ├── discriminator.py 51 | │   ├── encoder.py 52 | │   └── __init__.py 53 | ├── README.md 54 | ├── requirements.txt 55 | ├── scripts 56 | │   ├── copyFiles.sh 57 | │   ├── download_faces.py 58 | │   ├── keepFiles.sh 59 | │   ├── plot_utils.py 60 | │   └── preprocessing_cartoons_data.py 61 | ├── sweeps 62 | │   ├── sweep-bs-1.yaml 63 | │   └── sweep-rs-1.yaml 64 | ├── train.py 65 | ├── utils 66 | │   └── __init__.py 67 | └── weights 68 | ├── c_dann.pth 69 | ├── d1.pth 70 | ├── d2.pth 71 | ├── denoiser.pth 72 | ├── disc1.pth 73 | ├── d_shared.pth 74 | ├── e1.pth 75 | ├── e2.pth 76 | └── e_shared.pth 77 | ``` 78 | ## The model 79 | Our codebase is in Python3. We suggest creating a new virtual environment. 80 | * The required packages can be installed by running `pip3 install -r requirements.txt` 81 | * Update `N_CUDA` by running `export N_CUDA=` if you want to specify the GPU to use 82 | 83 | It is based on the XGAN paper omitting the Teacher Loss and adding an autoencoder in the end. The latter was trained to learn well only the representation of the cartoons as to "denoise" the spots and wrong colorisation from the face-to-cartoon outputs of the XGAN. 84 | 85 | The model was trained using the hyperparameters located in `config.json`. Weights & Biases Sweep was used to find the best hyperparameters: 86 | 87 | 1. Change `root_path` in `config json`. It specifies where is `datasets` which contains the datasets. 88 | 2. Run `wandb login 17d2772d85cbda79162bd975e45fdfbf3bb18911` to use wandb to get the report 89 | 3. Run `python3 train.py --wandb --run_name --run_notes ` or `python3 train.py --no-wandb` 90 | 4. To launch an agent with a sweep configuration of wandb in bg from ssh `nohup wandb agent --count stevramos/avatar_image_generator/ &` 91 | 92 | You can see the Weights & Biases report here: https://wandb.ai/stevramos/avatar_image_generator 93 | 94 | This is the implementation of [our project](https://madewithml.com/projects/1233/generating-avatars-from-real-life-pictures/) created for the Made With ML Data Science Incubator (deprecated). 95 | 96 | 97 | ## Docker 98 | 1. Build the container: `sudo docker build -f Dockerfile -t avatar-image-generator .` 99 | * Run the container: `sudo docker run -ti avatar-image-generator /bin/bash` 100 | * Train the model: 101 | 102 | a. Create the folder: `mkdir weights_trained` 103 | 104 | b. Change the absolute path from which mount the volume. This is for both `weights_trained` and `datasets`. In this case: 105 | 106 | sudo docker run -v :/src/weights_trained/ -v :/src/datasets/ -ti avatar-image-generator /bin/bash -c "cd src/ && source activate ml && wandb login 17d2772d85cbda79162bd975e45fdfbf3bb18911 && python train.py --wandb --run_name --run_notes " 107 | 108 | * Run the app locally as a daemon in docker. `model_path` in `config.json` contains the weights to use in the app 109 | `sudo docker run -d -p 8000:9999 -ti avatar-image-generator /bin/bash -c "cd src/ && source activate ml && python app.py"` 110 | 111 | a. Local server: [http://0.0.0.0:8000/](http://0.0.0.0:8000/) 112 | -------------------------------------------------------------------------------- /app.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import flask 3 | import time 4 | from flask import Flask 5 | from flask import request 6 | from flask import Flask, render_template, Response, request, redirect, jsonify, send_from_directory, abort, send_file 7 | from flask_cors import CORS 8 | from models import Avatar_Generator_Model 9 | from utils import * 10 | import torch.nn as nn 11 | from PIL import Image 12 | import numpy as np 13 | import cv2 14 | import base64 15 | import os , io , sys 16 | 17 | ALLOWED_EXTENSIONS = set(['jpg', 'jpeg', 'png']) 18 | CONFIG_FILENAME = "config.json" 19 | DOWNLOAD_DIRECTORY = None 20 | 21 | def allowed_file(filename): 22 | return '.' in filename and filename.rsplit('.', 1)[1].lower() in ALLOWED_EXTENSIONS 23 | 24 | app = Flask(__name__) 25 | CORS(app) 26 | 27 | MODEL = None 28 | 29 | 30 | def face_to_cartoon(DOC_FILE, face): 31 | 32 | document_name = DOC_FILE.split('.')[0] 33 | extension = (DOC_FILE.split('.')[-1]).lower() 34 | document = Image.open(io.BytesIO(face)) 35 | 36 | if not os.path.exists(DOWNLOAD_DIRECTORY): 37 | os.makedirs(DOWNLOAD_DIRECTORY) 38 | 39 | if extension == "png": 40 | format_image = "PNG" 41 | else: 42 | extension = "jpg" 43 | format_image = "JPEG" 44 | 45 | filename_face = "{}.{}".format(document_name, extension) 46 | document.save(DOWNLOAD_DIRECTORY + filename_face, format_image, quality=80, optimize=True, progressive=True) 47 | 48 | 49 | filename_cartoon = "{}_cartoon.jpg".format(document_name) 50 | cartoon, cartoon_tensor = MODEL.generate(DOWNLOAD_DIRECTORY + filename_face, DOWNLOAD_DIRECTORY + filename_cartoon) 51 | 52 | return filename_cartoon 53 | 54 | 55 | @app.route('/send_image', methods=['POST']) 56 | def upload_file(): 57 | # check if the post request has the file part 58 | if 'face_image' not in request.files: 59 | resp = jsonify({'message' : 'No file part in the request'}) 60 | resp.status_code = 400 61 | return resp 62 | 63 | file = request.files['face_image'] 64 | 65 | errors = {} 66 | success = False 67 | 68 | if file and allowed_file(file.filename): 69 | filename_cartoon = face_to_cartoon(file.filename, file.read()) 70 | success = True 71 | else: 72 | errors[file.filename] = 'File type is not allowed' 73 | 74 | if success and errors: 75 | errors['message'] = 'File(s) successfully uploaded' 76 | resp = jsonify(errors) 77 | resp.status_code = 500 78 | return resp 79 | if success: 80 | resp = jsonify({'message' : 'Files successfully processed', 'filename_cartoon': filename_cartoon}) 81 | resp.status_code = 201 82 | resp.headers.add('Access-Control-Allow-Origin', '*') 83 | print('headers:: ', resp.headers) 84 | return resp 85 | else: 86 | resp = jsonify(errors) 87 | resp.status_code = 500 88 | return resp 89 | 90 | 91 | @app.route('/predict', methods=['POST']) 92 | def predict(): 93 | doc_name = request.form.get('filename_cartoon') 94 | try: 95 | 96 | return send_from_directory(DOWNLOAD_DIRECTORY, filename=doc_name, as_attachment=True) 97 | except FileNotFoundError: 98 | abort(404) 99 | 100 | 101 | if __name__ == "__main__": 102 | 103 | use_wandb = False 104 | config = configure_model(CONFIG_FILENAME,use_wandb=use_wandb) 105 | DOWNLOAD_DIRECTORY = config.download_directory 106 | 107 | MODEL = Avatar_Generator_Model(config, use_wandb=use_wandb) 108 | MODEL.load_weights(config.model_path) 109 | 110 | app.run(host="0.0.0.0", port="9999") -------------------------------------------------------------------------------- /config.json: -------------------------------------------------------------------------------- 1 | { 2 | "server_config":{ 3 | "model_path":"weights/", 4 | "download_directory":"data/" 5 | }, 6 | "train_dataset_params":{ 7 | "root_path":"/data/shuaman/xgan/", 8 | "dataset_path_faces":"datasets/face_datasets/face_images_wo_bg_permissive/", 9 | "dataset_path_cartoons":"datasets/cartoon_datasets/cartoonset100k_limited/", 10 | "dataset_path_test_faces":"datasets/test_faces/input_images/", 11 | "dataset_path_segmented_faces":"datasets/test_faces/segmented_faces/", 12 | "dataset_path_output_faces":"datasets/test_faces/generated_cartoon_images/", 13 | "loader_params":{ 14 | "batch_size":32 15 | }, 16 | "save_weights":true, 17 | "num_backups": 8, 18 | "save_path":"weights_trained/" 19 | }, 20 | "model_hparams":{ 21 | "dropout_rate_eshared":0.5, 22 | "use_critic_dann": true, 23 | "use_critic_disc": true, 24 | "use_spectral_norm": true, 25 | "use_denoiser": true, 26 | "use_disc_cartoon2face": false, 27 | "num_epochs": 200, 28 | "learning_rate_opTotal":1e-4, 29 | "learning_rate_opDisc":1e-3, 30 | "learning_rate_denoiser":1e-3, 31 | "learning_rate_opCdann":2e-4, 32 | "wRec_loss":0.9928583013837265, 33 | "wDann_loss":0.9252047602915646, 34 | "wSem_loss":0.44957120675107437, 35 | "wGan_loss":0.9790275245543392, 36 | "wTeach_loss":0.75, 37 | "use_gpu":true 38 | } 39 | } 40 | -------------------------------------------------------------------------------- /evaluation/__init__.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import pandas as pd 3 | from sklearn.preprocessing import StandardScaler 4 | 5 | from .utils_tsne import apply_tsne, generate_scatter 6 | 7 | def tsne_evaluation(ls_feature_arrays, ls_array_names, pca_components=None, perplexity=30, n_iter=1000, save_image=False, output_dir='./', save_wandb=False, plot_title='t-SNE evaluation'): 8 | assert len(ls_feature_arrays) == len(ls_array_names) 9 | 10 | feature_vectors = np.concatenate(ls_feature_arrays) 11 | 12 | # cancatenate names in a df with same length as feature_vectors 13 | feature_vector_names = [] 14 | list( map(feature_vector_names.extend, [[name]*ls_feature_arrays[i].shape[0] for i, name in enumerate(ls_array_names)]) ) 15 | df_feature_vector_info = pd.DataFrame({'name':feature_vector_names}) 16 | 17 | tsne_results, df_feature_vector_info = apply_tsne(df_feature_vector_info , feature_vectors, perplexity, n_iter, pca_components=pca_components) 18 | 19 | tsne_results_norm = StandardScaler().fit_transform(tsne_results) 20 | 21 | scatter_plot = None 22 | wandb_scatter_plot = None 23 | img_scatter_plot = None 24 | if save_image or save_wandb: 25 | wandb_scatter_plot, img_scatter_plot = generate_scatter(tsne_results_norm, df_feature_vector_info, save_image, output_dir, save_wandb, plot_title) 26 | 27 | 28 | return tsne_results_norm, df_feature_vector_info, wandb_scatter_plot, img_scatter_plot 29 | 30 | ############################### 31 | 32 | # distance_threshold, stats_df = calc_jaccard_index(df) 33 | # logger.info("stats %s", stats_df) 34 | # stats_df.to_csv(os.path.join(output_dir, "stats.csv"), index=False) 35 | # if enable_rmse: 36 | # df_distances = calc_rmse(df, image_shape) 37 | # df_distances.to_csv(os.path.join(output_dir, "distances.csv"), index=False) 38 | # return stats_df 39 | 40 | 41 | if __name__ == "__main__": 42 | output_dir = "./" 43 | # pca_components = 10 44 | pca_components = None 45 | with open('dataset.npy', 'rb') as f: 46 | np_a = np.load(f) 47 | with open('model_a.npy', 'rb') as f: 48 | np_b = np.load(f) 49 | with open('model_b.npy', 'rb') as f: 50 | np_c = np.load(f) 51 | 52 | tsne_evaluation([np_a, np_b, np_c],['dataset', 'm_a','m_b'], pca_components=pca_components, save_image=True, output_dir= output_dir, save_wandb = True) 53 | -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/run_tsne_analysis.sh: -------------------------------------------------------------------------------- 1 | python tsne_analysis.py -b test/assets/dataset -p test/assets/model_a -p test/assets/model_b -o tsne_output -f -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/.gitignore: -------------------------------------------------------------------------------- 1 | output 2 | -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/dataset/image_10.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/dataset/image_10.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/dataset/image_10.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/dataset/image_10.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/dataset/image_5.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/dataset/image_5.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/dataset/image_5.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/dataset/image_5.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/dataset/image_6.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/dataset/image_6.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/dataset/image_6.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/dataset/image_6.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/dataset/image_7.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/dataset/image_7.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/dataset/image_7.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/dataset/image_7.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/dataset/image_9.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/dataset/image_9.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/dataset/image_9.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/dataset/image_9.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_a/image_10.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_a/image_10.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_a/image_10.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_a/image_10.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_a/image_5.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_a/image_5.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_a/image_5.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_a/image_5.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_a/image_6.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_a/image_6.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_a/image_6.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_a/image_6.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_a/image_7.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_a/image_7.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_a/image_7.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_a/image_7.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_a/image_9.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_a/image_9.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_a/image_9.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_a/image_9.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_0.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_0.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_0.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_0.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_1.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_1.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_1.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_1.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_2.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_2.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_2.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_2.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_3.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_3.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_3.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_3.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_5.npz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_5.npz -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_5.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_5.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/assets/model_b/image_6.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/test/assets/model_b/image_6.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/test_gen_tsne.py: -------------------------------------------------------------------------------- 1 | import os 2 | import unittest 3 | 4 | from gen_tsne.gen_tsne import calculate 5 | 6 | ASSETS_DIR = os.path.join(os.path.dirname(__file__), "assets") 7 | 8 | 9 | class TestGenTsne(unittest.TestCase): 10 | 11 | def test_build(self): 12 | paths = [os.path.join(ASSETS_DIR, "dataset"), os.path.join(ASSETS_DIR, "model_a"), 13 | os.path.join(ASSETS_DIR, "model_b")] 14 | stats_df = calculate(paths, os.path.join(ASSETS_DIR, "output"), pca_components=None, frow=10, fcol=10) 15 | self.assertIn("jaccard_index", stats_df.columns) 16 | 17 | def test_build_with_features(self): 18 | paths = [os.path.join(ASSETS_DIR, "dataset"), os.path.join(ASSETS_DIR, "model_a")] 19 | stats_df = calculate(paths, os.path.join(ASSETS_DIR, "output"), pca_components=None, frow=10, fcol=10, 20 | use_features=True) 21 | self.assertIn("jaccard_index", stats_df.columns) 22 | -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/test/test_grid.py: -------------------------------------------------------------------------------- 1 | import os 2 | import unittest 3 | 4 | from gen_tsne import build_grid 5 | 6 | ASSETS_DIR = os.path.join(os.path.dirname(__file__), "assets") 7 | 8 | 9 | class TestGrid(unittest.TestCase): 10 | 11 | def test_build(self): 12 | paths = [os.path.join(ASSETS_DIR, "dataset"), os.path.join(ASSETS_DIR, "model_a"), 13 | os.path.join(ASSETS_DIR, "model_b")] 14 | df, _ = build_grid(paths, output_dir=os.path.join(ASSETS_DIR, "output"), pca_components=None, frow=10, fcol=10, 15 | save_scatter=True) 16 | self.assertIn("tsne_x", df.columns) 17 | self.assertIn("tsne_y", df.columns) 18 | 19 | def test_build_pca(self): 20 | paths = [os.path.join(ASSETS_DIR, "dataset"), os.path.join(ASSETS_DIR, "model_a"), 21 | os.path.join(ASSETS_DIR, "model_b")] 22 | df, _ = build_grid(paths, output_dir=os.path.join(ASSETS_DIR, "output"), pca_components=5, frow=10, fcol=10) 23 | self.assertIn("tsne_x", df.columns) 24 | self.assertIn("tsne_y", df.columns) 25 | 26 | def test_build_with_features(self): 27 | paths = [os.path.join(ASSETS_DIR, "dataset"), os.path.join(ASSETS_DIR, "model_a"), 28 | os.path.join(ASSETS_DIR, "model_b")] 29 | df, _ = build_grid(paths, output_dir=os.path.join(ASSETS_DIR, "output"), pca_components=None, frow=10, fcol=10, 30 | use_features=True) 31 | self.assertIn("tsne_x", df.columns) 32 | self.assertIn("tsne_y", df.columns) 33 | -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/tmp_requirements.txt: -------------------------------------------------------------------------------- 1 | scikit_learn==0.23.2 2 | numpy==1.19.4 3 | pandas==1.1.4 4 | pillow 5 | # pillow-simd==7.0.0.post3 6 | matplotlib==3.2.1 7 | seaborn==0.11.0 8 | MulticoreTSNE==0.1 -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/tsne_analysis.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import logging 3 | import os 4 | 5 | from tsne_evaluation_utils.grid import build as build_grid 6 | from tsne_evaluation_utils.metric import calc_jaccard_index, calc_rmse 7 | 8 | logging.basicConfig(level=logging.INFO) 9 | logger = logging.getLogger(__name__) 10 | 11 | # Code as in https://github.com/vfcosta/gen-tsne 12 | 13 | def calculate(paths, output_dir, enable_rmse=True, pca_components=None, frow=60, fcol=60, perplexity=30, 14 | n_iter=1000, save_data=True, use_features=False): 15 | df, image_shape = build_grid(paths, pca_components=pca_components, frow=frow, fcol=fcol, perplexity=perplexity, 16 | n_iter=n_iter, save_data=save_data, output_dir=output_dir, use_features=use_features) 17 | distance_threshold, stats_df = calc_jaccard_index(df) 18 | logger.info("stats %s", stats_df) 19 | stats_df.to_csv(os.path.join(output_dir, "stats.csv"), index=False) 20 | if enable_rmse: 21 | df_distances = calc_rmse(df, image_shape) 22 | df_distances.to_csv(os.path.join(output_dir, "distances.csv"), index=False) 23 | return stats_df 24 | 25 | 26 | if __name__ == "__main__": 27 | parser = argparse.ArgumentParser(description='Apply Gen t-SNE metric.') 28 | parser.add_argument('-b', '--baseline', help='Path to images from the dataset (baseline)', required=True) 29 | parser.add_argument('-p', '--paths', action='append', help='Paths to images from generative models', required=True) 30 | parser.add_argument('-o', '--output', help='Output dir', default="./output") 31 | parser.add_argument('-r', "--rows", type=int, help='rows', default=60) 32 | parser.add_argument('-c', "--cols", type=int, help='cols', default=60) 33 | parser.add_argument('-k', "--perplexity", type=int, help='perplexity', default=30) 34 | parser.add_argument('-n', "--iter", type=int, help='iterations', default=1000) 35 | parser.add_argument('-f', "--use-features", default=False, action='store_true', 36 | help='Use features to build the grid (.npy or .npz)') 37 | args = parser.parse_args() 38 | calculate([args.baseline] + args.paths, args.output, frow=args.rows, fcol=args.cols, perplexity=args.perplexity, 39 | n_iter=args.iter, use_features=args.use_features) 40 | -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/tsne_evaluation_utils/grid.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import os 3 | from glob import glob 4 | 5 | import matplotlib.pyplot as plt 6 | import numpy as np 7 | import pandas as pd 8 | import seaborn as sns 9 | from MulticoreTSNE import MulticoreTSNE as TSNE 10 | from PIL import Image 11 | from sklearn.decomposition import PCA 12 | from sklearn.preprocessing import StandardScaler 13 | 14 | logger = logging.getLogger(__name__) 15 | 16 | 17 | def build(paths, frow=60, fcol=60, perplexity=30, n_iter=1000, jitter_win=0, pca_components=50, 18 | output_dir="./output", save_data=True, save_scatter=True, use_features=False): 19 | os.makedirs(output_dir, exist_ok=True) 20 | df, image_shape, tsne_input = load_data(paths, use_features) 21 | tsne_results = apply_tsne(df, tsne_input, perplexity, n_iter, pca_components=pca_components) 22 | logger.info("tsne finished: %s", tsne_results.shape) 23 | df['tsne_x_raw'], df['tsne_y_raw'] = tsne_results[:, 0], tsne_results[:, 1] 24 | norm = StandardScaler().fit_transform(df[["tsne_x_raw", "tsne_y_raw"]]) 25 | df['tsne_x'], df['tsne_y'] = norm[:, 0], norm[:, 1] 26 | if save_scatter: 27 | generate_scatter(df, output_dir) 28 | df = generate_images(fcol, frow, image_shape, df, output_dir=output_dir, jitter_win=jitter_win) 29 | if save_data: 30 | logger.info("saving data.csv") 31 | df.to_csv(os.path.join(output_dir, "data.csv"), index=False) 32 | logger.info("finished") 33 | return df, image_shape 34 | 35 | 36 | def generate_images(fcol, frow, image_shape, df, output_dir=None, jitter_win=None): 37 | df["tsne_x_int"] = ((fcol - 1) * (df["tsne_x"] - np.min(df["tsne_x"])) / np.ptp(df["tsne_x"])).astype(int) 38 | df["tsne_y_int"] = ((frow - 1) * (df["tsne_y"] - np.min(df["tsne_y"])) / np.ptp(df["tsne_y"])).astype(int) 39 | all_possibilities = [] 40 | if jitter_win: 41 | yy, xx = np.mgrid[-jitter_win:jitter_win + 1, -jitter_win:jitter_win + 1] 42 | all_possibilities = np.vstack([xx.reshape(-1), yy.reshape(-1)]).T.tolist() 43 | all_possibilities.sort(key=lambda x: (max(abs(x[0]), abs(x[1])), abs(x[0]) + abs(x[1]))) 44 | all_possibilities.pop(0) 45 | 46 | for model_name, group in df.groupby(by="name"): 47 | ordered_images = np.zeros((frow, fcol, *image_shape)) 48 | overlap, show = 0, 0 49 | for i, row in group.iterrows(): 50 | x, y = row["tsne_x_int"], row["tsne_y_int"] 51 | possibilities = list(all_possibilities) 52 | while len(possibilities) and np.sum(ordered_images[x, y]) != 0: 53 | dx, dy = possibilities.pop(0) 54 | x, y = np.clip(x + dx, 0, fcol - 1), np.clip(y + dy, 0, frow - 1) 55 | if np.sum(ordered_images[x, y]) == 0: 56 | show += 1 57 | ordered_images[x, y] = row[get_features(image_shape)].values.reshape((-1, *image_shape)) 58 | else: 59 | overlap += 1 60 | logger.info("overlap for %s: %d, show: %d", model_name, overlap, show) 61 | ordered_images = np.flipud(np.transpose(ordered_images, (1, 0, 2, 3, 4))).reshape(frow * fcol, *image_shape) 62 | 63 | grid = (ordered_images.reshape(frow, fcol, *image_shape).swapaxes(1, 2) 64 | .reshape(image_shape[0] * frow, image_shape[1] * fcol, image_shape[2])) 65 | logger.info("tsne grid shape: %s", grid.shape) 66 | plt.figure(figsize=(20, 20)) 67 | plt.imsave(os.path.join(output_dir, f"tsne_{model_name}.png"), grid) 68 | return df 69 | 70 | 71 | def apply_tsne(df, data, perplexity, n_iter, learning_rate=200, pca_components=None, tsne_jobs=4): 72 | if pca_components: 73 | logger.info("shape before pca: %s", data.shape) 74 | pca = PCA(n_components=pca_components, svd_solver='randomized') 75 | data = pca.fit_transform(data) 76 | pca_cols = [f"pca_{c}" for c in range(pca.n_components)] 77 | df[pca_cols] = pd.DataFrame(data, index=df.index) 78 | logger.info("shape after pca: %s", data.shape) 79 | tsne = TSNE(n_components=2, verbose=1, perplexity=perplexity, n_iter=n_iter, learning_rate=learning_rate, 80 | n_jobs=tsne_jobs) 81 | return tsne.fit_transform(data) 82 | 83 | 84 | def load_features(image_path, extensions=("npz", "npy")): 85 | base_path = os.path.splitext(image_path)[0] 86 | for ext in extensions: 87 | f = f"{base_path}.{ext}" 88 | if os.path.exists(f): 89 | data = np.load(f) 90 | if ext == "npz": 91 | data = data["arr_0"] 92 | return data 93 | return None 94 | 95 | 96 | def load_data(paths, use_features): 97 | df = pd.DataFrame() 98 | image_shape = None 99 | all_features = [] 100 | for path in paths: 101 | logger.info("loading images from %s", path) 102 | name = os.path.basename(path) 103 | for f in glob(os.path.join(path, "*.png")): 104 | if use_features: 105 | features = load_features(f) 106 | if features is None: 107 | logger.warning("features not found for %s", f) 108 | continue 109 | all_features.append(features) 110 | image = np.array(Image.open(f))/255 111 | image_shape = image.shape 112 | df_new = pd.DataFrame(image.reshape((-1, np.prod(image_shape)))) 113 | df_new["name"] = name 114 | df_new["file"] = f 115 | df = df.append(df_new) 116 | # with open(os.path.basename(path).split('/')[-1]+'.npy', 'wb') as f: 117 | # np.save(f, all_features) 118 | # all_features = [] 119 | logger.info("loaded %d images with shape %s", len(df), image_shape) 120 | tsne_input = np.array(all_features) if use_features else get_image_data(df, image_shape) 121 | 122 | return df.reset_index(), image_shape, tsne_input 123 | 124 | 125 | def generate_scatter(df, output_dir): 126 | plt.figure(figsize=(10, 10)) 127 | sns.scatterplot(x="tsne_x", y="tsne_y", hue="name", data=df, legend="full", alpha=0.2) 128 | plt.savefig(os.path.join(output_dir, f"models_scatter.png")) 129 | 130 | 131 | def get_features(image_shape): 132 | return list(range(np.prod(image_shape))) 133 | 134 | 135 | def get_image_data(df, image_shape): 136 | return df[get_features(image_shape)].values 137 | -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/tsne_evaluation_utils/metric.py: -------------------------------------------------------------------------------- 1 | import logging 2 | 3 | import numpy as np 4 | import pandas as pd 5 | from sklearn.metrics import mean_squared_error 6 | 7 | logger = logging.getLogger(__name__) 8 | 9 | 10 | def calc_jaccard_index(df): 11 | df_dataset = df[df["name"] == df.iloc[0]["name"]] 12 | df_models = df[df["name"] != df.iloc[0]["name"]].reset_index() 13 | distance_matrix = _calc_distances(df_dataset, df_models) 14 | min_distances_matrix = distance_matrix.min(axis=1) 15 | logger.info("distance matrix percentile %f", np.percentile(min_distances_matrix, 50)) 16 | distance_threshold = np.percentile(min_distances_matrix, 50) 17 | logger.info("distance_threshold: %f", distance_threshold) 18 | model_names = df_models["name"].unique() 19 | cols = ["selected", "distance_threshold", "intersection", "jaccard_index"] 20 | stats_df = pd.DataFrame(index=model_names, columns=cols) 21 | for name in model_names: 22 | df_model = df_models[df_models["name"] == name] 23 | all_selected = set() 24 | min_distances, intersection_gen = [], [] 25 | for i, row in df_model.iterrows(): 26 | distances = distance_matrix[i] 27 | selected = np.where(distances < distance_threshold)[0] 28 | min_distances.append(np.min(distances)) 29 | if len(selected) > 0: 30 | all_selected = all_selected.union(selected) 31 | intersection_gen.append(i) 32 | logger.info("model %s selected: %d intersection: %d", name, len(all_selected), len(intersection_gen)) 33 | stats_df.loc[name]["selected"] = len(all_selected) 34 | stats_df.loc[name]["intersection"] = len(intersection_gen) 35 | stats_df.loc[name]["jaccard_index"] = len(intersection_gen)/(len(df_model) + len(df_dataset) - len(intersection_gen)) 36 | stats_df["distance_threshold"] = distance_threshold 37 | return distance_threshold, stats_df.reset_index() 38 | 39 | 40 | def _calc_distances(df_dataset, df_models): 41 | distance_matrix = np.empty((len(df_models), len(df_dataset))) 42 | for i, row in df_models.iterrows(): 43 | distances = np.sqrt(np.sum((df_dataset[["tsne_x", "tsne_y"]] - row[["tsne_x", "tsne_y"]]) ** 2, axis=1)) 44 | distance_matrix[i] = distances 45 | return distance_matrix 46 | 47 | 48 | def calc_rmse(df, shape): 49 | dataset_name = df.iloc[0]["name"] 50 | df_dataset = df[df["name"] == dataset_name] 51 | df_models = df[df["name"] != dataset_name] 52 | row_distances = [] 53 | for _, row in df_models.iterrows(): 54 | distances = np.sqrt(np.sum((df_dataset[["tsne_x", "tsne_y"]] - row[["tsne_x", "tsne_y"]])**2, axis=1)) 55 | min_index = np.argmin(distances) 56 | max_index = np.argmax(distances) 57 | cols = list(range(np.prod(shape))) 58 | cols_act = [c for c in df_dataset.columns if str(c).startswith("act_")] 59 | cols_pca = [c for c in df_dataset.columns if str(c).startswith("pca_")] 60 | 61 | values = {} 62 | for k, index in {"min": min_index, "max": max_index}.items(): 63 | row_dataset = df_dataset.iloc[index] 64 | rmse = np.sqrt(mean_squared_error(row[cols], row_dataset[cols])) 65 | rmse_act = np.sqrt(mean_squared_error(row[cols_act], row_dataset[cols_act])) if cols_act else None 66 | rmse_pca = np.sqrt(mean_squared_error(row[cols_pca], row_dataset[cols_pca])) if cols_pca else None 67 | values = {**values, f"rmse_{k}": rmse, f"rmse_act_{k}": rmse_act, f"rmse_pca_{k}": rmse_pca, 68 | f"distance_{k}": distances[index], f"index_{k}": index} 69 | row_distances.append(values) 70 | return pd.DataFrame(row_distances) 71 | -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/tsne_output/distances.csv: -------------------------------------------------------------------------------- 1 | rmse_min,rmse_act_min,rmse_pca_min,distance_min,index_min,rmse_max,rmse_act_max,rmse_pca_max,distance_max,index_max 2 | 0.0,,,0.5299099634151018,0,0.23874200354570316,,,3.4182638654331203,2 3 | 0.0,,,0.5511406028574707,1,0.4226597363877475,,,3.367744388489437,2 4 | 0.0,,,0.5721681224628774,2,0.23874200354570316,,,3.614981425335309,0 5 | 0.3992857884213457,,,1.494951438335612,1,0.35899513931559507,,,3.156267410458738,0 6 | 0.0,,,0.5883145291463263,4,0.6148720845015195,,,1.9763873709280828,1 7 | 0.321735112156867,,,0.8856215144732921,4,0.6108290221548718,,,2.595041016039699,1 8 | 0.27533660355687184,,,1.0606789858687125,2,0.3349378150364164,,,2.738964400532061,0 9 | 0.2966289252292293,,,0.6261604338228165,4,0.48272389643562397,,,2.776364194136647,1 10 | 0.0,,,0.7849563976991836,1,0.4933058730972799,,,3.078485495259059,0 11 | 0.31686550926405477,,,0.6659541170896194,4,0.3228451601943453,,,2.338557459202454,2 12 | -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/tsne_output/models_scatter.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/tsne_output/models_scatter.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/tsne_output/stats.csv: -------------------------------------------------------------------------------- 1 | index,selected,distance_threshold,intersection,jaccard_index 2 | model_a,4,0.6460572754562179,4,0.6666666666666666 3 | model_b,1,0.6460572754562179,1,0.1111111111111111 4 | -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/tsne_output/tsne_dataset.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/tsne_output/tsne_dataset.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/tsne_output/tsne_model_a.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/tsne_output/tsne_model_a.png -------------------------------------------------------------------------------- /evaluation/tsne_analysis_baseline/tsne_output/tsne_model_b.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/evaluation/tsne_analysis_baseline/tsne_output/tsne_model_b.png -------------------------------------------------------------------------------- /evaluation/utils_tsne.py: -------------------------------------------------------------------------------- 1 | import os 2 | import pandas as pd 3 | from sklearn.decomposition import PCA 4 | from MulticoreTSNE import MulticoreTSNE as TSNE 5 | import matplotlib.pyplot as plt 6 | import seaborn as sns 7 | import wandb 8 | import io 9 | import PIL 10 | 11 | def apply_tsne(df_feature_vector_info, feature_vectors, perplexity, n_iter, learning_rate=200, pca_components=None, tsne_jobs=4): 12 | if pca_components: 13 | print("shape before pca: %s", feature_vectors.shape) 14 | pca = PCA(n_components=pca_components, svd_solver='randomized') 15 | feature_vectors = pca.fit_transform(feature_vectors) 16 | pca_cols = [f"pca_{c}" for c in range(pca.n_components)] 17 | df_feature_vector_info[pca_cols] = pd.DataFrame(feature_vectors, index=df_feature_vector_info.index) 18 | print("shape after pca: %s", feature_vectors.shape) 19 | print();print('TSNE:') 20 | tsne = TSNE(n_components=2, verbose=1, perplexity=perplexity, n_iter=n_iter, learning_rate=learning_rate, 21 | n_jobs=tsne_jobs) 22 | print() 23 | return tsne.fit_transform(feature_vectors), df_feature_vector_info 24 | 25 | def generate_scatter(tsne_results, df_feature_vector_info, save_image, output_dir, save_wandb, plot_title): 26 | 27 | plt.figure(figsize=(10, 10)) 28 | plt.title(plot_title) 29 | plt.xlabel('tsne_x') 30 | plt.ylabel('tsne_y') 31 | sns.scatterplot(x=tsne_results[:,0], y=tsne_results[:,1], hue=df_feature_vector_info['name'], legend="full", alpha=0.8) 32 | 33 | if save_image: 34 | plt.savefig(os.path.join(output_dir, plot_title+"_scatter_plot.png")) 35 | 36 | wandb_scatter_plot =None 37 | img_scatter_plot = None 38 | if save_wandb: 39 | data = [[x,y, name] for (x, y, name) in zip(list(tsne_results[:,0]), list(tsne_results[:,1]), list(df_feature_vector_info['name']))] 40 | table = wandb.Table(data=data, columns = ["tsne_x", "tsne_y",'name']) 41 | wandb_scatter_plot = wandb.plot.scatter(table, "tsne_x", "tsne_y", title=plot_title) 42 | # wandb.log({"tsne evaluation" : wandb.plot.scatter(table, "tsne_x", "tsne_y", title="t-SNE evaluation")}) 43 | 44 | buf = io.BytesIO() 45 | plt.savefig(buf) 46 | buf.seek(0) 47 | img_scatter_plot = PIL.Image.open(buf) 48 | 49 | return wandb_scatter_plot, img_scatter_plot 50 | -------------------------------------------------------------------------------- /images/Cartoons_example.jpeg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/images/Cartoons_example.jpeg -------------------------------------------------------------------------------- /images/Faces_example.jpeg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/IAmigos/avatar-image-generator/9bf11125f4ea3090e217cf15866ec19ce944f9c6/images/Faces_example.jpeg -------------------------------------------------------------------------------- /losses/__init__.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from .utils_loss import calc_eval_stats, calc_fid 3 | 4 | def L2_norm(image_rec, image_orig): 5 | 6 | assert image_rec.shape == image_orig.shape, "Assertion error: shape of input should be as same as target" 7 | 8 | return torch.linalg.norm(image_rec.reshape(image_rec.shape[0], -1) - image_orig.reshape(image_orig.shape[0], -1), ord=2, dim=1).mean() 9 | 10 | 11 | 12 | def L1_norm(encoder,encoder_rec): 13 | 14 | assert encoder.shape == encoder_rec.shape, "Assertion error: shape of input should be as same as target" 15 | 16 | return torch.linalg.norm(encoder.reshape(encoder.shape[0], -1) - encoder_rec.reshape(encoder_rec.shape[0], -1), ord=1, dim=1).mean() 17 | 18 | def fid(real_img,fake_img): 19 | #Code obtained from: https://www.kaggle.com/ibtesama/gan-in-pytorch-with-fid 20 | mu_1,std_1=calc_eval_stats(real_img) 21 | mu_2,std_2=calc_eval_stats(fake_img) 22 | 23 | """get fretched distance""" 24 | fid_value = calc_fid(mu_1, std_1, mu_2, std_2) 25 | return fid_value 26 | 27 | def MMD(x, y, kernel, device): 28 | #Code obtained from: https://www.kaggle.com/onurtunali/maximum-mean-discrepancy 29 | """Emprical maximum mean discrepancy. The lower the result, the more evidence that distributions are the same. 30 | 31 | Args: 32 | x: first sample, distribution P 33 | y: second sample, distribution Q 34 | kernel: kernel type such as "multiscale" or "rbf" 35 | """ 36 | xx, yy, zz = torch.mm(x, x.t()), torch.mm(y, y.t()), torch.mm(x, y.t()) 37 | rx = (xx.diag().unsqueeze(0).expand_as(xx)) 38 | ry = (yy.diag().unsqueeze(0).expand_as(yy)) 39 | 40 | dxx = rx.t() + rx - 2. * xx # Used for A in (1) 41 | dyy = ry.t() + ry - 2. * yy # Used for B in (1) 42 | dxy = rx.t() + ry - 2. * zz # Used for C in (1) 43 | 44 | XX, YY, XY = (torch.zeros(xx.shape).to(device), 45 | torch.zeros(xx.shape).to(device), 46 | torch.zeros(xx.shape).to(device)) 47 | 48 | if kernel == "multiscale": 49 | 50 | bandwidth_range = [0.2, 0.5, 0.9, 1.3] 51 | for a in bandwidth_range: 52 | XX += a**2 * (a**2 + dxx)**-1 53 | YY += a**2 * (a**2 + dyy)**-1 54 | XY += a**2 * (a**2 + dxy)**-1 55 | 56 | if kernel == "rbf": 57 | 58 | bandwidth_range = [10, 15, 20, 50] 59 | for a in bandwidth_range: 60 | XX += torch.exp(-0.5*dxx/a) 61 | YY += torch.exp(-0.5*dyy/a) 62 | XY += torch.exp(-0.5*dxy/a) 63 | return torch.mean(XX + YY - 2. * XY).item() 64 | 65 | 66 | 67 | def get_gradient(crit, real, fake, epsilon): 68 | ''' 69 | Return the gradient of the critic's scores with respect to mixes of real and fake images. 70 | Parameters: 71 | crit: the critic model 72 | real: a batch of real images 73 | fake: a batch of fake images 74 | epsilon: a vector of the uniformly random proportions of real/fake per mixed image 75 | Returns: 76 | gradient: the gradient of the critic's scores, with respect to the mixed image 77 | ''' 78 | # Mix the images together 79 | 80 | 81 | mixed_images = real * epsilon + fake * (1 - epsilon) 82 | 83 | # Calculate the critic's scores on the mixed images 84 | mixed_scores = crit(mixed_images) 85 | 86 | # Take the gradient of the scores with respect to the images 87 | gradient = torch.autograd.grad( 88 | # Note: You need to take the gradient of outputs with respect to inputs. 89 | # This documentation may be useful, but it should not be necessary: 90 | # https://pytorch.org/docs/stable/autograd.html#torch.autograd.grad 91 | #### START CODE HERE #### 92 | inputs=mixed_images, 93 | outputs=mixed_scores, 94 | #### END CODE HERE #### 95 | # These other parameters have to do with the pytorch autograd engine works 96 | grad_outputs=torch.ones_like(mixed_scores), 97 | create_graph=True, 98 | retain_graph=True, 99 | )[0] 100 | return gradient 101 | 102 | 103 | def gradient_penalty(gradient): 104 | ''' 105 | Return the gradient penalty, given a gradient. 106 | Given a batch of image gradients, you calculate the magnitude of each image's gradient 107 | and penalize the mean quadratic distance of each magnitude to 1. 108 | Parameters: 109 | gradient: the gradient of the critic's scores, with respect to the mixed image 110 | Returns: 111 | penalty: the gradient penalty 112 | ''' 113 | # Flatten the gradients so that each row captures one image 114 | gradient = gradient.view(len(gradient), -1) 115 | 116 | # Calculate the magnitude of every row 117 | gradient_norm = gradient.norm(2, dim=1) 118 | 119 | # Penalize the mean squared distance of the gradient norms from 1 120 | #### START CODE HERE #### 121 | penalty = ((gradient_norm - 1)**2).mean() 122 | #### END CODE HERE #### 123 | return penalty 124 | 125 | 126 | def get_crit_loss(crit_fake_pred, crit_real_pred, gp, c_lambda): 127 | ''' 128 | Return the loss of a critic given the critic's scores for fake and real images, 129 | the gradient penalty, and gradient penalty weight. 130 | Parameters: 131 | crit_fake_pred: the critic's scores of the fake images 132 | crit_real_pred: the critic's scores of the real images 133 | gp: the unweighted gradient penalty 134 | c_lambda: the current weight of the gradient penalty 135 | Returns: 136 | crit_loss: a scalar for the critic's loss, accounting for the relevant factors 137 | ''' 138 | #### START CODE HERE #### 139 | crit_loss = -(crit_real_pred - crit_fake_pred - c_lambda*gp).mean() 140 | #### END CODE HERE #### 141 | return crit_loss 142 | 143 | 144 | def get_gen_loss(crit_fake_pred): 145 | ''' 146 | Return the loss of a generator given the critic's scores of the generator's fake images. 147 | Parameters: 148 | crit_fake_pred: the critic's scores of the fake images 149 | Returns: 150 | gen_loss: a scalar loss value for the current batch of the generator 151 | ''' 152 | #### START CODE HERE #### 153 | gen_loss = -1 * crit_fake_pred.mean() 154 | #### END CODE HERE #### 155 | return gen_loss -------------------------------------------------------------------------------- /losses/utils_loss.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | from scipy import linalg 4 | import torch.nn.functional as F 5 | 6 | def calc_fid(mu1, sigma1, mu2, sigma2, eps=1e-6): 7 | ##Code obtained from: https://www.kaggle.com/ibtesama/gan-in-pytorch-with-fid 8 | """Numpy implementation of the Frechet Distance. 9 | The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1) 10 | and X_2 ~ N(mu_2, C_2) is 11 | d^2 = ||mu_1 - mu_2||^2 + Tr(C_1 + C_2 - 2*sqrt(C_1*C_2)). 12 | """ 13 | 14 | mu1 = np.atleast_1d(mu1) 15 | mu2 = np.atleast_1d(mu2) 16 | 17 | sigma1 = np.atleast_2d(sigma1) 18 | sigma2 = np.atleast_2d(sigma2) 19 | 20 | assert mu1.shape == mu2.shape, \ 21 | 'Training and test mean vectors have different lengths' 22 | assert sigma1.shape == sigma2.shape, \ 23 | 'Training and test covariances have different dimensions' 24 | 25 | diff = mu1 - mu2 26 | 27 | 28 | covmean, _ = linalg.sqrtm(sigma1.dot(sigma2), disp=False) 29 | if not np.isfinite(covmean).all(): 30 | msg = ('fid calculation produces singular product; ' 31 | 'adding %s to diagonal of cov estimates') % eps 32 | print(msg) 33 | offset = np.eye(sigma1.shape[0]) * eps 34 | covmean = linalg.sqrtm((sigma1 + offset).dot(sigma2 + offset)) 35 | 36 | 37 | if np.iscomplexobj(covmean): 38 | if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3): 39 | m = np.max(np.abs(covmean.imag)) 40 | raise ValueError('Imaginary component {}'.format(m)) 41 | covmean = covmean.real 42 | 43 | tr_covmean = np.trace(covmean) 44 | 45 | return (diff.dot(diff) + np.trace(sigma1) + 46 | np.trace(sigma2) - 2 * tr_covmean) 47 | 48 | def calc_eval_stats(act): 49 | ##Code obtained from: https://www.kaggle.com/ibtesama/gan-in-pytorch-with-fid 50 | #model.eval() 51 | #act=np.empty((len(images), dims)) 52 | #batch=images.to(device) 53 | 54 | #pred = model(batch)[0] 55 | 56 | # If model output is not scalar, apply global spatial average pooling. 57 | # This happens if you choose a dimensionality not equal 2048. 58 | #if pred.size(2) != 1 or pred.size(3) != 1: 59 | # pred = F.adaptive_avg_pool2d(pred, output_size=(1, 1)) 60 | 61 | #act= pred.cpu().data.numpy().reshape(pred.size(0), -1) 62 | 63 | mu = np.mean(act, axis=0) 64 | sigma = np.cov(act, rowvar=False) 65 | return mu, sigma -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- 1 | from .encoder import * 2 | from .decoder import * 3 | from .discriminator import * 4 | from .denoiser import * 5 | from .cdann import * 6 | from .avatar_generator_model import * 7 | 8 | -------------------------------------------------------------------------------- /models/avatar_generator_model.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | import torchvision.transforms as transforms 5 | import torchvision 6 | from torch.autograd import Variable 7 | from PIL import Image 8 | from keras_segmentation.pretrained import pspnet_101_voc12 9 | import cv2 10 | import numpy as np 11 | 12 | from .encoder import * 13 | from .decoder import * 14 | from .discriminator import * 15 | from .denoiser import * 16 | from .cdann import * 17 | from .inception import * 18 | from utils import * 19 | from losses import * 20 | from evaluation import tsne_evaluation 21 | 22 | import wandb 23 | import os 24 | import sys 25 | from tqdm import tqdm 26 | from itertools import cycle 27 | 28 | 29 | class Avatar_Generator_Model(): 30 | """ 31 | # Methods 32 | __init__(dict_model): initializer 33 | dict_model: layers required to perform face-to-image generation (e1, e_shared, d_shared, d2, denoiser) 34 | generate(face_image, output_path=None): reutrn cartoon generated from given face image, saves it to output path if given 35 | load_weights(weights_path): loads weights from given path 36 | """ 37 | 38 | def __init__(self, config, use_wandb=True): 39 | self.use_wandb = use_wandb 40 | self.config = config 41 | self.device = torch.device("cuda:" + (os.getenv('N_CUDA')if os.getenv('N_CUDA') else "0") if self.config.use_gpu and torch.cuda.is_available() else "cpu") 42 | self.mmd_kernel_type = "multiscale" 43 | self.segmentation = pspnet_101_voc12() 44 | self.e1, self.e2, self.d1, self.d2, self.e_shared, self.d_shared, self.c_dann, self.discriminator1, self.denoiser, self.inception, self.discriminator2 = self.init_model(self.device, 45 | self.config.dropout_rate_eshared, 46 | self.config.use_critic_dann, 47 | self.config.use_critic_disc, 48 | self.config.use_spectral_norm, 49 | self.use_wandb) 50 | 51 | 52 | def init_model(self, device, 53 | dropout_rate_eshared, 54 | use_critic_dann, use_critic_disc, 55 | use_spectral_norm, use_wandb=True): 56 | 57 | e1 = Encoder() 58 | e2 = Encoder() 59 | e_shared = Eshared(dropout_rate_eshared) 60 | d_shared = Dshared() 61 | d1 = Decoder() 62 | d2 = Decoder() 63 | c_dann = Cdann(use_critic_dann=use_critic_dann, use_spectral_norm=use_spectral_norm) 64 | discriminator1 = Discriminator(use_critic_disc=use_critic_disc, use_spectral_norm=use_spectral_norm) 65 | discriminator2 = Discriminator(use_critic_disc=use_critic_disc, use_spectral_norm=use_spectral_norm) 66 | denoiser = Denoiser() 67 | inception = Inception([Inception.BLOCK_INDEX_BY_DIM[2048]]) #fid 68 | 69 | e1.to(device) 70 | e2.to(device) 71 | e_shared.to(device) 72 | d_shared.to(device) 73 | d1.to(device) 74 | d2.to(device) 75 | c_dann.to(device) 76 | discriminator1.to(device) 77 | denoiser = denoiser.to(device) 78 | inception = inception.to(device) 79 | discriminator2 = discriminator2.to(device) 80 | 81 | if use_wandb: 82 | wandb.watch(e1, log="all") 83 | wandb.watch(e2, log="all") 84 | wandb.watch(e_shared, log="all") 85 | wandb.watch(d_shared, log="all") 86 | wandb.watch(d1, log="all") 87 | wandb.watch(d2, log="all") 88 | wandb.watch(c_dann, log="all") 89 | wandb.watch(discriminator1, log="all") 90 | wandb.watch(denoiser, log="all") 91 | wandb.watch(discriminator2, log="all") 92 | #wandb.watch(inception, log="all") 93 | 94 | return (e1, e2, d1, d2, e_shared, d_shared, c_dann, discriminator1, denoiser, inception, discriminator2) 95 | 96 | 97 | def generate(self, path_filename, output_path): 98 | face = self.__extract_face(path_filename, output_path) 99 | return self.__to_cartoon(face, output_path) 100 | 101 | 102 | def load_weights(self, weights_path): 103 | 104 | self.e1.load_state_dict(torch.load( 105 | weights_path + 'e1.pth', map_location=torch.device(self.device))) 106 | 107 | self.e_shared.load_state_dict( 108 | torch.load(weights_path + 'e_shared.pth', map_location=torch.device(self.device))) 109 | 110 | self.e2.load_state_dict( 111 | torch.load(weights_path + 'e2.pth', map_location=torch.device(self.device))) 112 | 113 | self.d_shared.load_state_dict( 114 | torch.load(weights_path + 'd_shared.pth', map_location=torch.device(self.device))) 115 | 116 | self.d2.load_state_dict(torch.load( 117 | weights_path + 'd2.pth', map_location=torch.device(self.device))) 118 | 119 | self.d1.load_state_dict(torch.load( 120 | weights_path + 'd1.pth', map_location=torch.device(self.device))) 121 | 122 | self.denoiser.load_state_dict( 123 | torch.load(weights_path + 'denoiser.pth', map_location=torch.device(self.device))) 124 | 125 | self.discriminator1.load_state_dict( 126 | torch.load(weights_path + 'disc1.pth', map_location=torch.device(self.device))) 127 | 128 | self.c_dann.load_state_dict( 129 | torch.load(weights_path + 'c_dann.pth', map_location=torch.device(self.device))) 130 | 131 | self.discriminator2.load_state_dict( 132 | torch.load(weights_path + 'disc2.pth', map_location=torch.device(self.device))) 133 | 134 | 135 | def __extract_face(self, path_filename, output_path): 136 | out = self.segmentation.predict_segmentation( 137 | inp=path_filename, 138 | out_fname=output_path 139 | ) 140 | 141 | img_mask = cv2.imread(output_path) 142 | img1 = cv2.imread(path_filename) # READ BGR 143 | 144 | seg_gray = cv2.cvtColor(img_mask, cv2.COLOR_BGR2GRAY) 145 | _, bg_mask = cv2.threshold( 146 | seg_gray, 0, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU) 147 | 148 | bg_mask = cv2.cvtColor(bg_mask, cv2.COLOR_GRAY2BGR) 149 | 150 | bg = cv2.bitwise_or(img1, bg_mask) 151 | 152 | cv2.imwrite(output_path, bg) 153 | face_image = Image.open(output_path) 154 | 155 | return face_image 156 | 157 | 158 | def __to_cartoon(self, face, output_path): 159 | self.e1.eval() 160 | self.e_shared.eval() 161 | self.d_shared.eval() 162 | self.d2.eval() 163 | self.denoiser.eval() 164 | 165 | transform_list_faces = get_transforms_config_face() 166 | transform = transforms.Compose(transform_list_faces) 167 | face = transform(face).float() 168 | X = face.unsqueeze(0).to(self.device) 169 | 170 | with torch.no_grad(): 171 | output = self.e1(X) 172 | output = self.e_shared(output) 173 | output = self.d_shared(output) 174 | output = self.d2(output) 175 | if self.config.use_denoiser: 176 | output = self.denoiser(output) 177 | 178 | output = denorm(output) 179 | output = output[0] 180 | 181 | torchvision.utils.save_image(tensor=output, fp=output_path) 182 | 183 | return (torchvision.transforms.ToPILImage()(output), output) 184 | 185 | 186 | def get_feature_inception(self, images, dims=2048): 187 | self.inception = self.inception.eval() 188 | act = np.empty((len(images), dims)) 189 | batch = images.to(self.device) 190 | 191 | pred = self.inception(batch)[0] 192 | 193 | # If model output is not scalar, apply global spatial average pooling. 194 | # This happens if you choose a dimensionality not equal 2048. 195 | if pred.size(2) != 1 or pred.size(3) != 1: 196 | pred = F.adaptive_avg_pool2d(pred, output_size=(1, 1)) 197 | 198 | act= pred.cpu().data.numpy().reshape(pred.size(0), -1) 199 | return act 200 | 201 | def get_loss_test_set(self, test_loader_faces, test_loader_cartoons, criterion_bc): 202 | 203 | self.e1.eval() 204 | self.e2.eval() 205 | self.e_shared.eval() 206 | self.d_shared.eval() 207 | self.d1.eval() 208 | self.d2.eval() 209 | self.c_dann.eval() 210 | self.discriminator1.eval() 211 | self.denoiser.eval() 212 | self.inception.eval() 213 | self.discriminator2.eval() 214 | 215 | cartoons_batch_test = [] 216 | cartoons_construct_test = [] 217 | 218 | 219 | faces_encoder_test = [] 220 | cartoons_encoder_test = [] 221 | cartoons_construct_encoder_test = [] 222 | cartoon_inception_test = [] 223 | cartoons_construct_inception_test = [] 224 | 225 | with torch.no_grad(): 226 | for faces_batch, cartoons_batch in zip(cycle(test_loader_faces), test_loader_cartoons): 227 | 228 | faces_batch, _ = faces_batch 229 | faces_batch = Variable(faces_batch.type(torch.Tensor)) 230 | faces_batch = faces_batch.to(self.device) 231 | 232 | cartoons_batch, _ = cartoons_batch 233 | cartoons_batch = Variable(cartoons_batch.type(torch.Tensor)) 234 | cartoons_batch = cartoons_batch.to(self.device) 235 | 236 | if faces_batch.shape != cartoons_batch.shape: 237 | continue 238 | 239 | faces_enc1 = self.e1(faces_batch) 240 | faces_encoder = self.e_shared(faces_enc1) 241 | faces_decoder = self.d_shared(faces_encoder) 242 | faces_rec = self.d1(faces_decoder) 243 | cartoons_construct = self.d2(faces_decoder) 244 | cartoons_construct_enc2 = self.e2(cartoons_construct) 245 | cartoons_construct_encoder = self.e_shared(cartoons_construct_enc2) 246 | 247 | cartoons_enc2 = self.e2(cartoons_batch) 248 | cartoons_encoder = self.e_shared(cartoons_enc2) 249 | cartoons_decoder = self.d_shared(cartoons_encoder) 250 | cartoons_rec = self.d2(cartoons_decoder) 251 | faces_construct = self.d1(cartoons_decoder) 252 | faces_construct_enc1 = self.e1(faces_construct) 253 | faces_construct_encoder = self.e_shared(faces_construct_enc1) 254 | 255 | #inception 256 | cartoon_inception = self.get_feature_inception(cartoons_batch) 257 | cartoon_inception_test.append(cartoon_inception) 258 | #inception 259 | 260 | cartoons_batch_test.append(cartoons_batch) 261 | 262 | if self.config.use_denoiser: 263 | cartoons_construct = self.denoiser(cartoons_construct) 264 | 265 | #inception 266 | cartoons_construct_inception = self.get_feature_inception(cartoons_construct) 267 | cartoons_construct_inception_test.append(cartoons_construct_inception) 268 | #inception 269 | 270 | cartoons_construct_test.append(cartoons_construct) 271 | 272 | faces_encoder_test.append(faces_encoder) 273 | cartoons_encoder_test.append(cartoons_encoder) 274 | cartoons_construct_encoder_test.append(cartoons_construct_encoder) 275 | 276 | 277 | # return np.mean(loss_test) 278 | 279 | cartoons_batch_test = torch.cat(cartoons_batch_test) 280 | cartoons_construct_test = torch.cat(cartoons_construct_test) 281 | cartoons_construct_test = torch.unique(cartoons_construct_test, dim=0, sorted=False) 282 | 283 | cartoon_inception_test = np.concatenate(cartoon_inception_test) 284 | cartoons_construct_inception_test = np.concatenate(cartoons_construct_inception_test) 285 | cartoons_construct_inception_test = np.unique(cartoons_construct_inception_test, axis=0) 286 | 287 | cartoons_batch_feature_view = cartoons_batch_test.view(cartoons_batch_test.size()[0], -1) 288 | cartoons_construct_feature_view = cartoons_construct_test.view(cartoons_construct_test.size()[0], -1) 289 | cartoons_construct_feature_view = torch.unique(cartoons_construct_feature_view, dim=0, sorted=False) 290 | cartoons_batch_feature_view = cartoons_batch_feature_view[:cartoons_construct_feature_view.shape[0]] 291 | 292 | assert cartoons_construct_test.shape[0] == cartoons_construct_feature_view.shape[0], "torch unique cant get the same shape in constructed cartoons" 293 | 294 | fid_test = fid(cartoon_inception_test, cartoons_construct_inception_test) 295 | mmd_test = MMD(cartoons_batch_feature_view, cartoons_construct_feature_view, self.mmd_kernel_type, self.device) 296 | 297 | # tsne analysis 298 | faces_encoder_test = torch.cat(faces_encoder_test).cpu() 299 | faces_encoder_test = torch.unique(faces_encoder_test, dim=0, sorted=False) 300 | cartoons_encoder_test = torch.cat(cartoons_encoder_test).cpu() 301 | cartoons_construct_encoder_test = torch.cat(cartoons_construct_encoder_test).cpu() 302 | cartoons_construct_encoder_test = torch.unique(cartoons_construct_encoder_test, dim=0, sorted=False) 303 | 304 | assert faces_encoder_test.shape[0] == cartoons_construct_encoder_test.shape[0], "torch unique cant get the same shape in faces and constructed cartoons" 305 | 306 | # tsne of faces encoder and cartoons encoder 307 | tsne_results_norm, df_feature_vector_info, wandb_scatter_plot_1_fe_ce, img_scatter_plot_1_fe_ce = tsne_evaluation([faces_encoder_test, cartoons_encoder_test], ['faces encoder', 'cartoons encoder'], pca_components=None, perplexity=30, n_iter=1000, save_image=False, save_wandb=self.use_wandb, plot_title='t-SNE evaluation - FE and CE') 308 | 309 | # tsne of faces encoder and cartoons construct encoder 310 | tsne_results_norm, df_feature_vector_info, wandb_scatter_plot_2_fe_cce, img_scatter_plot_2_fe_cce = tsne_evaluation([faces_encoder_test, cartoons_construct_encoder_test], ['faces encoder', 'cartoons construct encoder'], pca_components=None, perplexity=30, n_iter=1000, save_image=False, save_wandb=self.use_wandb, plot_title='t-SNE evaluation - FE and CCE') 311 | 312 | return fid_test, mmd_test, wandb_scatter_plot_1_fe_ce, wandb_scatter_plot_2_fe_cce, img_scatter_plot_1_fe_ce, img_scatter_plot_2_fe_cce 313 | 314 | 315 | 316 | def train_crit_repeats(self, opt, fake, real, model, type_model, crit_repeats=5): 317 | 318 | if type_model=="discriminator": 319 | fake = fake.detach() 320 | loss_weight = self.config.wGan_loss 321 | elif type_model=="cdann": 322 | loss_weight = self.config.wDann_loss 323 | 324 | 325 | mean_iteration_critic_loss = torch.zeros(1).to(self.device) 326 | for i in range(crit_repeats): 327 | ### Update critic ### 328 | opt.zero_grad() 329 | lim_inf = i * int(len(fake)/crit_repeats) 330 | lim_sup = lim_inf + int(len(fake)/crit_repeats) if i < crit_repeats - 1 else 1000 331 | fake_sample = fake[lim_inf: lim_sup] 332 | real_sample = real[lim_inf: lim_sup] 333 | 334 | crit_fake_pred = model(fake_sample) 335 | crit_real_pred = model(real_sample) 336 | 337 | if type_model=="discriminator": 338 | epsilon = torch.rand(len(real_sample), 1, 1, 1, 339 | device=self.device, requires_grad=True) 340 | elif type_model=="cdann": 341 | epsilon = torch.rand(len(real_sample), 1, 342 | device=self.device, requires_grad=True) 343 | 344 | gradient = get_gradient( 345 | model, real_sample, fake_sample, epsilon) 346 | gp = gradient_penalty(gradient) 347 | crit_loss = get_crit_loss( 348 | crit_fake_pred.squeeze(), crit_real_pred.squeeze(), gp, 10) * self.config.wDann_loss 349 | 350 | # Keep track of the average critic loss in this batch 351 | mean_iteration_critic_loss += crit_loss / crit_repeats 352 | # Update gradients 353 | crit_loss.backward(retain_graph=True) 354 | # Update optimizer 355 | opt.step() 356 | loss = mean_iteration_critic_loss 357 | 358 | return loss 359 | 360 | def train_disc(self, disc, e, d, batch, real, opt): 361 | # discriminator face(1)->cartoon(2) 362 | # discriminator cartoon(2)->face(1) 363 | disc.zero_grad() 364 | 365 | enc = e(batch).detach() 366 | encoder = self.e_shared(enc).detach() 367 | decoder = self.d_shared(encoder).detach() 368 | construct = d(decoder).detach() 369 | 370 | if not self.config.use_critic_disc: 371 | # train discriminator with real cartoon images 372 | output_real = disc(real) 373 | loss_disc_real = self.config.wGan_loss * \ 374 | criterion_bc(output_real.squeeze(), torch.ones_like( 375 | output_real.squeeze(), device=self.device)) 376 | # loss_disc1_real_cartoons.backward() 377 | 378 | # train discriminator with fake cartoon images 379 | # class_faces.fill_(0) 380 | 381 | output_fake = disc(construct) 382 | loss_disc_fake = self.config.wGan_loss * \ 383 | criterion_bc(output_fake.squeeze(), torch.zeros_like( 384 | output_fake.squeeze(), device=self.device)) 385 | # loss_disc1_fake_cartoons.backward() 386 | 387 | loss_disc = loss_disc_real + loss_disc_fake 388 | loss_disc.backward() 389 | opt.step() 390 | else: 391 | loss_disc = self.train_crit_repeats(opt, construct, 392 | real, disc, 393 | "discriminator", crit_repeats=5) 394 | 395 | return loss_disc 396 | 397 | def train_step(self, train_loader_faces, train_loader_cartoons, optimizers, criterion_bc): 398 | 399 | optimizerDenoiser, optimizerDisc1, optimizerTotal, optimizerCdann, optimizerDisc2 = optimizers 400 | 401 | self.e1.train() 402 | self.e2.train() 403 | self.e_shared.train() 404 | self.d_shared.train() 405 | self.d1.train() 406 | self.d2.train() 407 | self.c_dann.train() 408 | self.discriminator1.train() 409 | self.denoiser.train() 410 | self.discriminator2.train() 411 | 412 | for faces_batch, cartoons_batch in zip(cycle(train_loader_faces), train_loader_cartoons): 413 | 414 | faces_batch, _ = faces_batch 415 | faces_batch = Variable(faces_batch.type(torch.Tensor)) 416 | faces_batch = faces_batch.to(self.device) 417 | 418 | cartoons_batch, _ = cartoons_batch 419 | cartoons_batch = Variable(cartoons_batch.type(torch.Tensor)) 420 | cartoons_batch = cartoons_batch.to(self.device) 421 | 422 | self.e1.zero_grad() 423 | self.e2.zero_grad() 424 | self.e_shared.zero_grad() 425 | self.d_shared.zero_grad() 426 | self.d1.zero_grad() 427 | self.d2.zero_grad() 428 | self.c_dann.zero_grad() 429 | 430 | if faces_batch.shape != cartoons_batch.shape: 431 | continue 432 | 433 | # architecture 434 | faces_enc1 = self.e1(faces_batch) 435 | faces_encoder = self.e_shared(faces_enc1) 436 | faces_decoder = self.d_shared(faces_encoder) 437 | faces_rec = self.d1(faces_decoder) 438 | cartoons_construct = self.d2(faces_decoder) 439 | cartoons_construct_enc2 = self.e2(cartoons_construct) 440 | cartoons_construct_encoder = self.e_shared(cartoons_construct_enc2) 441 | 442 | cartoons_enc2 = self.e2(cartoons_batch) 443 | cartoons_encoder = self.e_shared(cartoons_enc2) 444 | cartoons_decoder = self.d_shared(cartoons_encoder) 445 | cartoons_rec = self.d2(cartoons_decoder) 446 | faces_construct = self.d1(cartoons_decoder) 447 | faces_construct_enc1 = self.e1(faces_construct) 448 | faces_construct_encoder = self.e_shared(faces_construct_enc1) 449 | 450 | # train generator 451 | 452 | #training cdann 453 | if not self.config.use_critic_dann: 454 | label_output_face = self.c_dann(faces_encoder) 455 | label_output_cartoon = self.c_dann(cartoons_encoder) 456 | loss_dann = criterion_bc(label_output_face.squeeze(), torch.zeros_like(label_output_face.squeeze( 457 | ), device=self.device)) + criterion_bc(label_output_cartoon.squeeze(), torch.ones_like(label_output_cartoon.squeeze(), device=self.device)) 458 | loss_dann = self.config.wDann_loss * loss_dann 459 | loss_dann.backward(retain_graph=True) 460 | optimizerCdann.step() 461 | else: 462 | # train critic(cdann) 463 | loss_dann = self.train_crit_repeats(optimizerCdann, faces_encoder, 464 | cartoons_encoder, self.c_dann, 465 | "cdann", crit_repeats=5) 466 | 467 | loss_rec1 = L2_norm(faces_batch, faces_rec) 468 | loss_rec2 = L2_norm(cartoons_batch, cartoons_rec) 469 | loss_rec = loss_rec1 + loss_rec2 470 | 471 | loss_sem1 = L1_norm(faces_encoder.detach(), cartoons_construct_encoder) 472 | loss_sem2 = L1_norm(cartoons_encoder.detach(), faces_construct_encoder) 473 | loss_sem = loss_sem1 + loss_sem2 474 | 475 | # teach loss 476 | #faces_embedding = resnet(faces_batch.squeeze()) 477 | #loss_teach = L1_norm(faces_embedding.squeeze(), faces_encoder) 478 | # constant until train facenet 479 | loss_teach = torch.Tensor([0]).requires_grad_() 480 | loss_teach = loss_teach.to(self.device) 481 | 482 | # class_faces.fill_(1) 483 | output = self.discriminator1(cartoons_construct) 484 | if not self.config.use_critic_disc: 485 | loss_gen1 = criterion_bc(output.squeeze(), torch.ones_like( 486 | output.squeeze(), device=self.device)) 487 | else: 488 | loss_gen1 = get_gen_loss(output.squeeze()) 489 | 490 | 491 | if self.config.use_disc_cartoon2face: 492 | output2 = self.discriminator2(faces_construct) 493 | if not self.config.use_critic_disc: 494 | loss_gen2 = criterion_bc(output2.squeeze(), torch.ones_like( 495 | output2.squeeze(), device=self.device)) 496 | else: 497 | loss_gen2 = get_gen_loss(output2.squeeze()) 498 | else: 499 | loss_gen2 = torch.Tensor([0]).requires_grad_() 500 | loss_gen2 = loss_gen2.to(self.device) 501 | 502 | #it has been deleted config.wDann_loss*loss_dann 503 | loss_total = self.config.wRec_loss*loss_rec + \ 504 | self.config.wSem_loss*loss_sem + self.config.wGan_loss * \ 505 | loss_gen1 + self.config.wTeach_loss*loss_teach + \ 506 | self.config.wGan_loss * loss_gen2 507 | loss_total.backward() 508 | loss_total += loss_dann 509 | 510 | optimizerTotal.step() 511 | 512 | # discriminator face(1)->cartoon(2) 513 | self.discriminator1.zero_grad() 514 | loss_disc1 = self.train_disc(self.discriminator1, self.e1, 515 | self.d2, faces_batch, 516 | cartoons_batch, optimizerDisc1) 517 | 518 | # discriminator cartoon(2)->face(1) 519 | if self.config.use_disc_cartoon2face: 520 | self.discriminator2.zero_grad() 521 | loss_disc2 = self.train_disc(self.discriminator2, self.e2, 522 | self.d1, cartoons_batch, 523 | faces_batch, optimizerDisc2) 524 | else: 525 | loss_disc2 = torch.Tensor([0]).requires_grad_() 526 | 527 | # Denoiser 528 | if self.config.use_denoiser: 529 | self.denoiser.zero_grad() 530 | cartoons_denoised = self.denoiser(cartoons_rec.detach()) 531 | 532 | # Train Denoiser 533 | loss_denoiser = L2_norm(cartoons_batch, cartoons_denoised) 534 | loss_denoiser.backward() 535 | 536 | optimizerDenoiser.step() 537 | else: 538 | loss_denoiser = torch.Tensor([0]).requires_grad_() 539 | 540 | #break #Delete break 541 | 542 | 543 | return loss_rec1, loss_rec2, loss_dann, loss_sem1, loss_sem2, loss_disc1, loss_gen1, loss_disc2, loss_gen2, loss_total, loss_denoiser, loss_teach 544 | 545 | 546 | def train(self): 547 | 548 | if self.config.use_gpu and torch.cuda.is_available(): 549 | print("Training in " + torch.cuda.get_device_name(0)) 550 | else: 551 | print("Training in CPU") 552 | 553 | if self.config.save_weights: 554 | if self.use_wandb: 555 | path_save_weights = self.config.root_path + wandb.run.id + "_" + self.config.save_path 556 | else: 557 | path_save_weights = self.config.root_path + self.config.save_path 558 | try: 559 | os.mkdir(path_save_weights) 560 | except OSError: 561 | pass 562 | 563 | model = (self.e1, self.e2, self.d1, self.d2, self.e_shared, self.d_shared, self.c_dann, self.discriminator1, self.denoiser, self.discriminator2) 564 | 565 | train_loader_faces, test_loader_faces, train_loader_cartoons, test_loader_cartoons = get_datasets(self.config.root_path, self.config.dataset_path_faces, self.config.dataset_path_cartoons, self.config.batch_size) 566 | optimizers = init_optimizers(model, self.config.learning_rate_opDisc, self.config.learning_rate_opTotal, self.config.learning_rate_denoiser, self.config.learning_rate_opCdann) 567 | 568 | criterion_bc = nn.BCEWithLogitsLoss() 569 | criterion_bc.to(self.device) 570 | 571 | images_faces_to_test = get_test_images(self.segmentation, self.config.batch_size, self.config.root_path + self.config.dataset_path_test_faces, self.config.root_path + self.config.dataset_path_segmented_faces) 572 | 573 | for epoch in tqdm(range(self.config.num_epochs)): 574 | loss_rec1, loss_rec2, loss_dann, loss_sem1, loss_sem2, loss_disc1, loss_gen1, loss_disc2, loss_gen2, loss_total, loss_denoiser, loss_teach = self.train_step(train_loader_faces, train_loader_cartoons, optimizers, criterion_bc) 575 | 576 | metrics_log = {"train_epoch": epoch+1, 577 | "loss_rec1": loss_rec1.item(), 578 | "loss_rec2": loss_rec2.item(), 579 | "loss_dann": loss_dann.item(), 580 | "loss_semantic12": loss_sem1.item(), 581 | "loss_semantic21": loss_sem2.item(), 582 | "loss_disc1": loss_disc1.item(), 583 | "loss_gen1": loss_gen1.item(), 584 | "loss_disc2": loss_disc2.item(), 585 | "loss_gen2": loss_gen2.item(), 586 | "loss_teach": loss_teach.item(), 587 | "loss_total": loss_total.item()} 588 | 589 | if self.config.save_weights and ((epoch+1) % int(self.config.num_epochs/self.config.num_backups)) == 0: 590 | path_save_epoch = path_save_weights + 'epoch_{}'.format(epoch+1) 591 | try: 592 | os.mkdir(path_save_epoch) 593 | except OSError: 594 | pass 595 | save_weights(model, path_save_epoch, self.use_wandb) 596 | fid_test, mmd_test, wandb_scatter_plot_1_fe_ce, wandb_scatter_plot_2_fe_cce, img_scatter_plot_1_fe_ce, img_scatter_plot_2_fe_cce = self.get_loss_test_set(test_loader_faces, test_loader_cartoons, criterion_bc) 597 | generated_images = test_image(model, self.device, images_faces_to_test, self.config.use_denoiser) 598 | 599 | metrics_log["fid"] = fid_test 600 | metrics_log["mmd"] = mmd_test 601 | metrics_log["Generated images"] = [wandb.Image(img) for img in generated_images] 602 | metrics_log['t-SNE evaluation plot 1 - FE and CE'] = wandb_scatter_plot_1_fe_ce 603 | metrics_log['t-SNE evaluation plot 2 - FE and CCE'] = wandb_scatter_plot_2_fe_cce 604 | 605 | if self.use_wandb: 606 | metrics_log["t-SNE evaluation images"] = [wandb.Image(img) for img in [img_scatter_plot_1_fe_ce, img_scatter_plot_2_fe_cce]] 607 | 608 | if self.use_wandb: 609 | wandb.log(metrics_log) 610 | 611 | 612 | print("Losses") 613 | print('Epoch [{}/{}], Loss rec1: {:.4f}'.format(epoch + 614 | 1, self.config.num_epochs, loss_rec1.item())) 615 | print('Epoch [{}/{}], Loss rec2: {:.4f}'.format(epoch + 616 | 1, self.config.num_epochs, loss_rec2.item())) 617 | print('Epoch [{}/{}], Loss dann: {:.4f}'.format(epoch + 618 | 1, self.config.num_epochs, loss_dann.item())) 619 | print('Epoch [{}/{}], Loss semantic 1->2: {:.4f}'.format(epoch + 620 | 1, self.config.num_epochs, loss_sem1.item())) 621 | print('Epoch [{}/{}], Loss semantic 2->1: {:.4f}'.format(epoch + 622 | 1, self.config.num_epochs, loss_sem2.item())) 623 | print('Epoch [{}/{}], Loss disc1: {:.4f}'.format(epoch + 624 | 1, self.config.num_epochs, loss_disc1.item())) 625 | print('Epoch [{}/{}], Loss gen1: {:.4f}'.format(epoch + 626 | 1, self.config.num_epochs, loss_gen1.item())) 627 | print('Epoch [{}/{}], Loss disc2: {:.4f}'.format(epoch + 628 | 1, self.config.num_epochs, loss_disc2.item())) 629 | print('Epoch [{}/{}], Loss gen2: {:.4f}'.format(epoch + 630 | 1, self.config.num_epochs, loss_gen2.item())) 631 | print('Epoch [{}/{}], Loss teach: {:.4f}'.format(epoch + 632 | 1, self.config.num_epochs, loss_teach.item())) 633 | print('Epoch [{}/{}], Loss total: {:.4f}'.format(epoch + 634 | 1, self.config.num_epochs, loss_total.item())) 635 | print('Epoch [{}/{}], Loss denoiser: {:.4f}'.format(epoch + 636 | 1, self.config.num_epochs, loss_denoiser.item())) 637 | 638 | if self.use_wandb: 639 | wandb.finish() 640 | -------------------------------------------------------------------------------- /models/cdann.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from torch.autograd import Variable 5 | 6 | 7 | class GradReverse(torch.autograd.Function): 8 | @staticmethod 9 | def forward(ctx, x): 10 | return x.view_as(x) 11 | 12 | @staticmethod 13 | def backward(ctx, grad_output): 14 | return grad_output.neg() 15 | 16 | 17 | def grad_reverse(x): 18 | return GradReverse.apply(x) 19 | 20 | """ 21 | class Cdann(nn.Module): 22 | def __init__(self, dropout_rate): 23 | super(Cdann, self).__init__() 24 | self.fc1 = nn.Linear(in_features=1024, out_features=512) 25 | self.fc2 = nn.Linear(in_features=512, out_features=256) 26 | self.dropout2 = nn.Dropout(dropout_rate) 27 | self.fc3 = nn.Linear(in_features=256, out_features=128) 28 | self.fc4 = nn.Linear(in_features=128, out_features=64) 29 | self.dropout4 = nn.Dropout(dropout_rate) 30 | self.fc5 = nn.Linear(in_features=64, out_features=32) 31 | self.fc6 = nn.Linear(in_features=32, out_features=16) 32 | self.dropout6 = nn.Dropout(dropout_rate) 33 | self.fc7 = nn.Linear(in_features=16, out_features=1) 34 | 35 | nn.init.kaiming_normal_(self.fc1.weight) 36 | nn.init.kaiming_normal_(self.fc2.weight) 37 | nn.init.kaiming_normal_(self.fc3.weight) 38 | nn.init.kaiming_normal_(self.fc4.weight) 39 | nn.init.kaiming_normal_(self.fc5.weight) 40 | nn.init.kaiming_normal_(self.fc6.weight) 41 | nn.init.xavier_normal_(self.fc7.weight) 42 | 43 | def forward(self, x): 44 | x = grad_reverse(x) 45 | x = F.relu(self.fc1(x)) 46 | x = F.relu(self.fc2(x)) 47 | x = self.dropout2(x) 48 | x = F.relu(self.fc3(x)) 49 | x = F.relu(self.fc4(x)) 50 | x = self.dropout4(x) 51 | x = F.relu(self.fc5(x)) 52 | x = F.relu(self.fc6(x)) 53 | x = self.dropout6(x) 54 | x = torch.sigmoid(self.fc7(x)) 55 | 56 | return x 57 | """ 58 | 59 | 60 | 61 | class Cdann(nn.Module): 62 | ''' 63 | Taken from Coursera - GANNs 64 | ''' 65 | 66 | def __init__(self, use_critic_dann, im_chan=1024, hidden_dim=512, use_spectral_norm=False): 67 | super(Cdann, self).__init__() 68 | self.use_critic_dann = use_critic_dann 69 | self.cdan = nn.Sequential( 70 | self.make_cdan_block(im_chan, hidden_dim, use_spectral_norm), 71 | self.make_cdan_block(hidden_dim, hidden_dim // 2, use_spectral_norm), 72 | self.make_cdan_block(hidden_dim // 2, hidden_dim // 4, use_spectral_norm), 73 | self.make_cdan_block(hidden_dim // 4, 1, use_spectral_norm, final_layer=True), 74 | ) 75 | self.cdan = self.cdan.apply(self.weights_init) 76 | 77 | def make_cdan_block(self, input_channels, output_channels, use_spectral_norm, final_layer=False): 78 | ''' 79 | Function to return a sequence of operations corresponding to a critic block of DCGAN; 80 | a convolution, a batchnorm (except in the final layer), and an activation (except in the final layer). 81 | Parameters: 82 | input_channels: how many channels the input feature representation has 83 | output_channels: how many channels the output feature representation should have 84 | final_layer: a boolean, true if it is the final layer and false otherwise 85 | (affects activation and batchnorm) 86 | ''' 87 | if not final_layer: 88 | return nn.Sequential( 89 | self.make_linear_block(input_channels, output_channels, use_spectral_norm), 90 | nn.BatchNorm1d(output_channels), 91 | nn.LeakyReLU(0.2, inplace=True), 92 | ) 93 | else: 94 | return nn.Sequential( 95 | self.make_linear_block(input_channels, output_channels, use_spectral_norm) 96 | ) 97 | 98 | def make_linear_block(self, input_channels, output_channels, use_spectral_norm): 99 | if use_spectral_norm and self.use_critic_dann: 100 | return nn.Sequential( 101 | nn.utils.spectral_norm(nn.Linear(in_features=input_channels, 102 | out_features=output_channels)) 103 | ) 104 | 105 | else: 106 | return nn.Sequential( 107 | nn.Linear(in_features=input_channels, 108 | out_features=output_channels) 109 | ) 110 | 111 | def weights_init(self, m): 112 | if isinstance(m, nn.Linear): 113 | nn.init.kaiming_normal_(m.weight) 114 | if isinstance(m, nn.BatchNorm1d): 115 | torch.nn.init.normal_(m.weight, 0.0, 0.02) 116 | torch.nn.init.constant_(m.bias, 0) 117 | 118 | def forward(self, feature): 119 | ''' 120 | Function for completing a forward pass of the critic: Given an image tensor, 121 | returns a 1-dimension tensor representing fake/real. 122 | Parameters: 123 | feature: a tensor with dimension (im_chan) 124 | ''' 125 | feature = grad_reverse(feature) 126 | cdan_pred = self.cdan(feature) 127 | return cdan_pred.view(len(cdan_pred), -1) 128 | 129 | 130 | 131 | -------------------------------------------------------------------------------- /models/decoder.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | 6 | class Dshared(nn.Module): 7 | def __init__(self): 8 | super(Dshared, self).__init__() 9 | self.deconv1 = nn.ConvTranspose2d( 10 | in_channels=1024, out_channels=512, kernel_size=4, stride=2, bias=False) 11 | self.bd1 = nn.BatchNorm2d(512) 12 | self.deconv2 = nn.ConvTranspose2d( 13 | in_channels=512, out_channels=256, kernel_size=2, stride=2, bias=False) 14 | self.bd2 = nn.BatchNorm2d(256) 15 | 16 | nn.init.kaiming_normal_(self.deconv1.weight) 17 | nn.init.kaiming_normal_(self.deconv2.weight) 18 | 19 | def forward(self, x): 20 | x = x.view(-1, 1024, 1, 1) 21 | x = F.relu(self.deconv1(x)) 22 | x = self.bd1(x) 23 | x = F.relu(self.deconv2(x)) 24 | x = self.bd2(x) 25 | 26 | return x 27 | 28 | 29 | class Decoder(nn.Module): 30 | def __init__(self): 31 | super(Decoder, self).__init__() 32 | 33 | self.deconv3 = nn.ConvTranspose2d( 34 | in_channels=256, out_channels=128, kernel_size=2, stride=2, bias=False) 35 | self.bd3 = nn.BatchNorm2d(128) 36 | self.deconv4 = nn.ConvTranspose2d( 37 | in_channels=128, out_channels=64, kernel_size=2, stride=2, bias=False) 38 | self.bd4 = nn.BatchNorm2d(64) 39 | self.deconv5 = nn.ConvTranspose2d( 40 | in_channels=64, out_channels=3, kernel_size=2, stride=2) 41 | 42 | nn.init.kaiming_normal_(self.deconv3.weight) 43 | nn.init.kaiming_normal_(self.deconv4.weight) 44 | nn.init.xavier_normal_(self.deconv5.weight) 45 | 46 | def forward(self, x): 47 | x = F.relu(self.deconv3(x)) 48 | x = self.bd3(x) 49 | x = F.relu(self.deconv4(x)) 50 | x = self.bd4(x) 51 | x = torch.tanh(self.deconv5(x)) 52 | 53 | return x -------------------------------------------------------------------------------- /models/denoiser.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from torch.autograd import Variable 5 | 6 | 7 | class Denoiser(nn.Module): 8 | def __init__(self): 9 | super(Denoiser, self).__init__() 10 | 11 | self.encoder = nn.Sequential( 12 | nn.Conv2d(3, 64, kernel_size=3, padding=1), 13 | nn.ReLU(), 14 | nn.MaxPool2d(2, 2)) 15 | 16 | self.decoder = nn.Sequential( 17 | nn.Conv2d(64, 64, kernel_size=3, padding=1), 18 | nn.ReLU(), 19 | nn.Upsample(scale_factor=2), 20 | nn.Conv2d(64, 3, kernel_size=3, padding=1)) 21 | 22 | def forward(self, x): 23 | return torch.tanh(self.decoder(self.encoder(x))) -------------------------------------------------------------------------------- /models/discriminator.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from torch.autograd import Variable 5 | from torch.utils.data import DataLoader 6 | 7 | 8 | class Discriminator(nn.Module): 9 | def __init__(self, use_critic_disc, use_spectral_norm): 10 | super(Discriminator, self).__init__() 11 | self.use_critic_disc = use_critic_disc 12 | self.conv1 = self.make_block(in_channels=3, 13 | out_channels=16, 14 | kernel_size=3, 15 | stride=2, 16 | padding=1, 17 | bias=True, 18 | use_spectral_norm=use_spectral_norm) 19 | 20 | self.conv2 = self.make_block(in_channels=16, 21 | out_channels=32, 22 | kernel_size=3, 23 | stride=2, 24 | padding=1, 25 | bias=False, 26 | use_spectral_norm=use_spectral_norm) 27 | 28 | self.conv3 = self.make_block(in_channels=32, 29 | out_channels=32, 30 | kernel_size=3, 31 | stride=2, 32 | padding=1, 33 | bias=False, 34 | use_spectral_norm=use_spectral_norm) 35 | 36 | self.conv4 = self.make_block(in_channels=32, 37 | out_channels=32, 38 | kernel_size=3, 39 | stride=2, 40 | padding=1, 41 | bias=True, 42 | use_spectral_norm=use_spectral_norm) 43 | 44 | self.fc1 = self.make_block(in_channels=4*4*32, 45 | out_channels=1, 46 | use_spectral_norm=use_spectral_norm, 47 | final_layer=True) 48 | 49 | #self.conv1 = nn.Conv2d(in_channels=3, out_channels=16, 50 | # kernel_size=3, stride=2, padding=1) # out: 32 x 32 x 32 51 | #self.conv2 = nn.Conv2d(in_channels=16, out_channels=32, kernel_size=3, 52 | # stride=2, padding=1, bias=False) # out: 32 x 32 x 32 53 | self.b2 = nn.BatchNorm2d(32) 54 | #self.conv3 = nn.Conv2d(in_channels=32, out_channels=32, kernel_size=3, 55 | # stride=2, padding=1, bias=False) # out: 32 x 32 x 32 56 | self.b3 = nn.BatchNorm2d(32) 57 | #self.conv4 = nn.Conv2d(in_channels=32, out_channels=32, 58 | # kernel_size=3, stride=2, padding=1) # out: 32 x 32 x 32 59 | self.flatten = nn.Flatten() 60 | #self.fc1 = nn.Linear(in_features=4*4*32, out_features=1) 61 | 62 | nn.init.kaiming_normal_(self.conv1.weight) 63 | nn.init.kaiming_normal_(self.conv2.weight) 64 | nn.init.kaiming_normal_(self.conv3.weight) 65 | nn.init.kaiming_normal_(self.conv4.weight) 66 | nn.init.xavier_normal_(self.fc1.weight) 67 | 68 | def make_block(self, in_channels=None, out_channels=None, 69 | kernel_size=None, stride=None, padding=None, 70 | use_spectral_norm=None, bias=True, final_layer=False): 71 | if not final_layer: 72 | if use_spectral_norm and self.use_critic_disc: 73 | return nn.utils.spectral_norm( 74 | nn.Conv2d(in_channels=in_channels, out_channels=out_channels, 75 | kernel_size=kernel_size, stride=stride, 76 | padding=padding, bias=bias) 77 | ) 78 | else: 79 | return nn.Conv2d(in_channels=in_channels, out_channels=out_channels, 80 | kernel_size=kernel_size, stride=stride, 81 | padding=padding, bias=bias) 82 | else: 83 | if use_spectral_norm and self.use_critic_disc: 84 | return nn.utils.spectral_norm(nn.Linear(in_features=in_channels, 85 | out_features=out_channels)) 86 | else: 87 | return nn.Linear(in_features=in_channels, 88 | out_features=out_channels) 89 | 90 | 91 | 92 | def forward(self, x): 93 | x = F.leaky_relu(self.conv1(x), 0.2) 94 | x = F.leaky_relu(self.conv2(x), 0.2) 95 | x = self.b2(x) 96 | x = F.leaky_relu(self.conv3(x), 0.2) 97 | x = self.b3(x) 98 | x = F.leaky_relu(self.conv4(x), 0.2) 99 | x = self.flatten(x) 100 | x = self.fc1(x) 101 | 102 | return x 103 | -------------------------------------------------------------------------------- /models/encoder.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | 6 | class Encoder(nn.Module): 7 | def __init__(self): 8 | super(Encoder, self).__init__() 9 | 10 | self.conv1 = nn.Conv2d(in_channels=3, out_channels=32, kernel_size=3, 11 | stride=2, padding=1, bias=False) # out: 32 x 32 x 32 12 | self.b1 = nn.BatchNorm2d(32) 13 | self.conv2 = nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, 14 | stride=2, padding=1, bias=False) # out: 16 x 16 x 64 15 | self.b2 = nn.BatchNorm2d(64) 16 | 17 | nn.init.kaiming_normal_(self.conv1.weight) 18 | nn.init.kaiming_normal_(self.conv2.weight) 19 | 20 | def forward(self, x): 21 | x = F.relu(self.conv1(x)) 22 | x = self.b1(x) 23 | x = F.relu(self.conv2(x)) 24 | x = self.b2(x) 25 | 26 | return x 27 | 28 | 29 | class Eshared(nn.Module): 30 | def __init__(self, dropout_rate=0.5): 31 | super(Eshared, self).__init__() 32 | self.conv3 = nn.Conv2d(in_channels=64, out_channels=128, 33 | kernel_size=3, stride=2, padding=1, bias=False) 34 | self.b3 = nn.BatchNorm2d(128) 35 | self.conv4 = nn.Conv2d(in_channels=128, out_channels=256, 36 | kernel_size=3, stride=2, padding=1, bias=False) 37 | self.b4 = nn.BatchNorm2d(256) 38 | self.flatten = nn.Flatten() 39 | self.fc1 = nn.Linear(in_features=4*4*256, 40 | out_features=1024, bias=False) 41 | self.bfc1 = nn.BatchNorm1d(1024) 42 | self.dropout2 = nn.Dropout(dropout_rate) 43 | self.fc2 = nn.Linear(in_features=1024, out_features=1024, bias=False) 44 | self.bfc2 = nn.BatchNorm1d(1024) 45 | 46 | nn.init.kaiming_normal_(self.conv3.weight) 47 | nn.init.kaiming_normal_(self.conv4.weight) 48 | nn.init.kaiming_normal_(self.fc1.weight) 49 | nn.init.kaiming_normal_(self.fc2.weight) 50 | 51 | def forward(self, x): 52 | x = F.relu(self.conv3(x)) 53 | x = self.b3(x) 54 | x = F.relu(self.conv4(x)) 55 | x = self.b4(x) 56 | x = self.flatten(x) 57 | x = F.relu(self.fc1(x)) 58 | x = self.bfc1(x) 59 | x = self.dropout2(x) 60 | x = F.relu(self.fc2(x)) 61 | x = self.bfc2(x) 62 | 63 | return x -------------------------------------------------------------------------------- /models/inception.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | import torchvision.models as models 5 | 6 | #code as in https://www.kaggle.com/scratchpad/notebooke311656179/edit 7 | class Inception(nn.Module): 8 | """Pretrained InceptionV3 network returning feature maps""" 9 | 10 | # Index of default block of inception to return, 11 | # corresponds to output of final average pooling 12 | DEFAULT_BLOCK_INDEX = 3 13 | 14 | # Maps feature dimensionality to their output blocks indices 15 | BLOCK_INDEX_BY_DIM = { 16 | 64: 0, # First max pooling features 17 | 192: 1, # Second max pooling featurs 18 | 768: 2, # Pre-aux classifier features 19 | 2048: 3 # Final average pooling features 20 | } 21 | 22 | def __init__(self, 23 | output_blocks=[DEFAULT_BLOCK_INDEX], 24 | resize_input=True, 25 | normalize_input=True, 26 | requires_grad=False): 27 | 28 | super(Inception, self).__init__() 29 | 30 | self.resize_input = resize_input 31 | self.normalize_input = normalize_input 32 | self.output_blocks = sorted(output_blocks) 33 | self.last_needed_block = max(output_blocks) 34 | 35 | assert self.last_needed_block <= 3, \ 36 | 'Last possible output block index is 3' 37 | 38 | self.blocks = nn.ModuleList() 39 | 40 | 41 | inception = models.inception_v3(pretrained=True) 42 | 43 | # Block 0: input to maxpool1 44 | block0 = [ 45 | inception.Conv2d_1a_3x3, 46 | inception.Conv2d_2a_3x3, 47 | inception.Conv2d_2b_3x3, 48 | nn.MaxPool2d(kernel_size=3, stride=2) 49 | ] 50 | self.blocks.append(nn.Sequential(*block0)) 51 | 52 | # Block 1: maxpool1 to maxpool2 53 | if self.last_needed_block >= 1: 54 | block1 = [ 55 | inception.Conv2d_3b_1x1, 56 | inception.Conv2d_4a_3x3, 57 | nn.MaxPool2d(kernel_size=3, stride=2) 58 | ] 59 | self.blocks.append(nn.Sequential(*block1)) 60 | 61 | # Block 2: maxpool2 to aux classifier 62 | if self.last_needed_block >= 2: 63 | block2 = [ 64 | inception.Mixed_5b, 65 | inception.Mixed_5c, 66 | inception.Mixed_5d, 67 | inception.Mixed_6a, 68 | inception.Mixed_6b, 69 | inception.Mixed_6c, 70 | inception.Mixed_6d, 71 | inception.Mixed_6e, 72 | ] 73 | self.blocks.append(nn.Sequential(*block2)) 74 | 75 | # Block 3: aux classifier to final avgpool 76 | if self.last_needed_block >= 3: 77 | block3 = [ 78 | inception.Mixed_7a, 79 | inception.Mixed_7b, 80 | inception.Mixed_7c, 81 | nn.AdaptiveAvgPool2d(output_size=(1, 1)) 82 | ] 83 | self.blocks.append(nn.Sequential(*block3)) 84 | 85 | for param in self.parameters(): 86 | param.requires_grad = requires_grad 87 | 88 | def forward(self, inp): 89 | """Get Inception feature maps 90 | Parameters 91 | ---------- 92 | inp : torch.autograd.Variable 93 | Input tensor of shape Bx3xHxW. Values are expected to be in 94 | range (0, 1) 95 | Returns 96 | ------- 97 | List of torch.autograd.Variable, corresponding to the selected output 98 | block, sorted ascending by index 99 | """ 100 | outp = [] 101 | x = inp 102 | 103 | if self.resize_input: 104 | x = F.interpolate(x, 105 | size=(299, 299), 106 | mode='bilinear', 107 | align_corners=False) 108 | 109 | if self.normalize_input: 110 | x = 2 * x - 1 # Scale from range (0, 1) to range (-1, 1) 111 | 112 | for idx, block in enumerate(self.blocks): 113 | x = block(x) 114 | if idx in self.output_blocks: 115 | outp.append(x) 116 | 117 | if idx == self.last_needed_block: 118 | break 119 | 120 | return outp 121 | 122 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | Flask==1.1.1 2 | numpy==1.19.4 3 | matplotlib==3.2.2 4 | torch==1.7.0 5 | torchvision==0.8.1 6 | Pillow==7.1.0 7 | flask_cors==3.0.10 8 | opencv-python==4.5.1.48 9 | keras==2.4.3 10 | keras-segmentation==0.3.0 11 | tensorflow==2.4.1 12 | wandb==0.10.12 13 | tqdm==4.55.1 14 | helper==2.4.2 15 | cycler==0.11.0 16 | pandas==1.1.4 17 | scikit_learn==0.23.2 18 | seaborn==0.11.0 19 | MulticoreTSNE==0.1 20 | -------------------------------------------------------------------------------- /scripts/copyFiles.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #copying files from 0..9 directory 4 | #target directory 5 | mkdir $1 6 | 7 | for i in {0..9} 8 | do 9 | echo "Copying files from $i" 10 | cp $i/*.png $1 11 | done 12 | 13 | echo "Total files in $1:" 14 | ls $1 | wc -l 15 | -------------------------------------------------------------------------------- /scripts/download_faces.py: -------------------------------------------------------------------------------- 1 | import os 2 | import requests 3 | from PIL import Image 4 | import matplotlib.pyplot as plt 5 | from tqdm.notebook import tqdm as tqdm_n 6 | from tqdm import tqdm 7 | from time import time 8 | import cv2 9 | from keras_segmentation.pretrained import pspnet_101_voc12 10 | 11 | # !pip install keras-segmentation 12 | 13 | 14 | def read_file(path, min_pose, min_score, curation, formats_allowed): 15 | ''' Reads a text file and returns info of each image ''' 16 | 17 | with open(path, 'r') as in_file: 18 | for line in in_file: 19 | # url 20 | # bbox coordinates 21 | # pose: frontal/profile (pose>2 signifies a frontal face while pose<=2 represents left and right profile detection) 22 | # detection score: Score of a DPM detector 23 | # curation: Whether this image was a part of final curated dataset (1 or 0) 24 | _, url, x1, y1, x2, y2, pose, score, curated_dataset = line.split( 25 | ' ') 26 | if url.split('.')[-1] in formats_allowed and float(pose) >= min_pose: 27 | # Same as: not curation or (curation and int(curated_dataset)) 28 | if not curation or int(curated_dataset): 29 | yield url, [x1, y1, x2, y2] 30 | return [] 31 | 32 | 33 | def download_crop_image(item, offset_x, offset_top_percent, offset_bottom_percent): 34 | url, bbox = item 35 | x1, y1, x2, y2 = list(map(float, bbox)) 36 | 37 | im_cropped = None 38 | try: 39 | response = requests.get(url, stream=True) 40 | response.raw.decode_content = True 41 | im = Image.open(response.raw) 42 | w = (x2 - x1) 43 | h = (y2 - y1) 44 | v_offset_x = w*offset_x/100 45 | v_offset_top = h*offset_top_percent/100 46 | v_offset_bottom = h*offset_bottom_percent/100 47 | im_cropped = im.crop( 48 | (x1-v_offset_x, y1-v_offset_top, x2+v_offset_x, y2+v_offset_bottom)) 49 | except OSError: 50 | # tqdm.write('404: '+url) 51 | pass 52 | 53 | return im_cropped 54 | 55 | 56 | def get_image(person_name, file_w_path, num_images, target_path, offset_x, 57 | offset_top_percent, offset_bottom_percent, min_pose, min_score, 58 | curation, formats_allowed): 59 | ''' Downloads images from a given text file, crops them and saves them locally ''' 60 | 61 | dir_path = target_path # os.path.dirname(target_path) 62 | 63 | if not os.path.exists(dir_path): 64 | os.makedirs(dir_path) 65 | 66 | i_image = 0 67 | for item in read_file(file_w_path, min_pose, min_score, curation, formats_allowed): 68 | if i_image >= num_images: 69 | # tqdm.write('Finished Downloading '+person_name) 70 | return 71 | image = download_crop_image( 72 | item, offset_x, offset_top_percent, offset_bottom_percent) 73 | if image: 74 | try: 75 | image.save(os.path.join( 76 | dir_path, f'{person_name}_{i_image+1}.jpg')) 77 | except Exception: 78 | continue 79 | i_image += 1 80 | tqdm.write(f'There is no more images available for {person_name}') 81 | tqdm.write(f'({i_image} image(s) downloaded of {num_images})\n') 82 | 83 | 84 | def download_vgg_images(data_path, num_people, num_images, target_path, offset_x_percent, 85 | offset_top_percent, offset_bottom_percent, min_pose=3, min_score=0, 86 | curation=False, formats_allowed=['jpg', 'jpeg'], from_notebook=False): 87 | ''' 88 | Parses info of text files in a given folder and downloads the images 89 | in them. Each file contains info about images of the same person. 90 | ''' 91 | 92 | ini = time() 93 | 94 | target_path += '/' 95 | files = os.listdir(data_path) 96 | n = num_people if num_people < len(files) else len(files) 97 | pbar_class = tqdm_n if from_notebook else tqdm 98 | pbar = pbar_class(files[:num_people], total=n) 99 | for fi in pbar: 100 | # The file's name is the name of the person 101 | fi_splited = fi.split('.') 102 | person_name = '.'.join(fi_splited[:-1]) if len(fi_splited) > 1 else fi 103 | pbar.set_description( 104 | f'Processing {person_name} ({num_images} image(s))') 105 | file_w_path = data_path + '/' + fi 106 | get_image(person_name, file_w_path, num_images, 107 | target_path, offset_x_percent, offset_top_percent, 108 | offset_bottom_percent, min_pose, min_score, curation, 109 | formats_allowed) 110 | 111 | end = time() 112 | print(f'Downloading time: {round((end-ini)/60, 2)} min\n') 113 | 114 | 115 | def clean_corrupt_files(path, formats_allowed=['jpg', 'jpeg']): 116 | ''' Filters files that have only one channel, can't be opened or have a format not allowed ''' 117 | 118 | n_removed = 0 119 | for filename in os.listdir(path): 120 | filename_lower = filename # .lower() 121 | if filename_lower.split('.')[-1] in formats_allowed: 122 | try: 123 | img = Image.open( 124 | os.path.join(path, filename)) # open the image file 125 | img.verify() # verify that it is in fact an image 126 | if len(plt.imread(path + filename).shape) != 3: 127 | os.remove(os.path.join(path, filename)) 128 | print(f'Removing corrupt file:', filename) 129 | n_removed += 1 130 | except Exception: 131 | os.remove(os.path.join(path, filename)) 132 | print('Removing corrupt file:', filename) 133 | n_removed += 1 134 | else: 135 | os.remove(os.path.join(path, filename)) 136 | print('Removing file with different format:', filename) 137 | n_removed += 1 138 | print(f'\n{n_removed} file(s) removed') 139 | 140 | 141 | def remove_background(images_path, masks_path, output_path, from_notebook=False): 142 | ''' Removes background from images (paints it white) ''' 143 | 144 | if not os.path.exists(masks_path): 145 | os.makedirs(masks_path) 146 | if not os.path.exists(output_path): 147 | os.makedirs(output_path) 148 | 149 | # load the pretrained model trained on Pascal VOC 2012 dataset 150 | model = pspnet_101_voc12() 151 | 152 | pbar_class = tqdm_n if from_notebook else tqdm 153 | pbar = pbar_class(os.listdir(images_path)) 154 | start = time() 155 | for filename in pbar: 156 | pbar.set_description(f'Processing {filename}') 157 | mask_file = filename[:-4] + "_mask.png" 158 | output_file = filename[:-4] + "_wo_bg.jpg" 159 | 160 | model.predict_segmentation( 161 | inp=images_path + filename, 162 | out_fname=masks_path + mask_file 163 | ) 164 | 165 | img_mask = cv2.imread(masks_path + mask_file) 166 | img1 = cv2.imread(images_path + filename) # READ BGR 167 | 168 | seg_gray = cv2.cvtColor(img_mask, cv2.COLOR_BGR2GRAY) 169 | _, bg_mask = cv2.threshold( 170 | seg_gray, 0, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU) 171 | 172 | # convert mask to 3-channels 173 | bg_mask = cv2.cvtColor(bg_mask, cv2.COLOR_GRAY2BGR) 174 | 175 | # cv2.bitwise_or to extract the region 176 | bg = cv2.bitwise_or(img1, bg_mask) 177 | 178 | # save 179 | cv2.imwrite(output_path + output_file, bg) 180 | end = time() 181 | 182 | print('Background removed') 183 | print(f'Processing time: {round((end-start)/60, 2)} min\n') 184 | 185 | 186 | if __name__ == "__main__": 187 | data_path = r'C:\Users\Daniel Ibáñez\Documents\Proyectos\Avatar Project\vgg_face_dataset\files' 188 | target_path = r'C:\Users\Daniel Ibáñez\Documents\Proyectos\Avatar Project\prueba/' 189 | 190 | download_vgg_images(data_path, num_people=50, num_images=3, target_path=target_path, 191 | offset_x_percent=15, offset_top_percent=55, 192 | offset_bottom_percent=12, min_pose=3, min_score=0, 193 | curation=False, formats_allowed=['jpg', 'jpeg'], from_notebook=False) 194 | clean_corrupt_files(target_path, formats_allowed=['jpg', 'jpeg']) 195 | 196 | images_path = r'C:\Users\Daniel Ibáñez\Documents\Proyectos\Avatar Project\tmp\face_images/' 197 | masks_path = r'C:\Users\Daniel Ibáñez\Documents\Proyectos\Avatar Project\tmp\masks/' 198 | output_path = r'C:\Users\Daniel Ibáñez\Documents\Proyectos\Avatar Project\face_images_wo_bg/' 199 | 200 | remove_background(images_path, masks_path, 201 | output_path, from_notebook=False) 202 | -------------------------------------------------------------------------------- /scripts/keepFiles.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | #keep filename which are in $1b generated by data preprocessing 4 | #$2 source files 5 | #$3 target directory 6 | 7 | mkdir $3 8 | cat $1 | xargs -I % cp $2/% $3 9 | echo "Total kept files:" 10 | ls $3 | wc -l 11 | -------------------------------------------------------------------------------- /scripts/plot_utils.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import matplotlib.pyplot as plt 3 | 4 | def show(images, n_cols=None): 5 | n_cols = n_cols or len(images) 6 | n_rows = (len(images) - 1) // n_cols + 1 7 | if images.shape[-1] == 1: 8 | images = np.squeeze(images, axis=-1) 9 | plt.figure(figsize=(n_cols, n_rows)) 10 | for index, image in enumerate(images): 11 | plt.subplot(n_rows, n_cols, index + 1) 12 | plt.imshow(image, cmap="binary") 13 | plt.axis("off") 14 | -------------------------------------------------------------------------------- /scripts/preprocessing_cartoons_data.py: -------------------------------------------------------------------------------- 1 | import os 2 | import pandas as pd 3 | import matplotlib.pyplot as plt 4 | import matplotlib.image as mpimg 5 | #import seaborn as sns 6 | from tqdm import tqdm 7 | 8 | # Create image info dataset 9 | def read_cartoon_dataset(path_cartoons, list_sub_folders): 10 | 11 | df = pd.DataFrame() 12 | for index in tqdm(list_sub_folders): 13 | path_folder = os.path.join(path_cartoons,index) 14 | list_files = [file for file in os.listdir(path_folder) if '.csv' in file ] 15 | 16 | for file in list_files: 17 | path_csv = os.path.join(path_folder, file) 18 | df_cartoon = pd.read_csv(path_csv,header=None) 19 | df_cartoon.columns = ["attribute_name", "index_variant","total_num_variants"] 20 | df_cartoon["filename"] = index + "/" + file 21 | df_cartoon = df_cartoon.pivot(index="filename",columns="attribute_name", values="index_variant") 22 | 23 | df = pd.concat([df, df_cartoon]) 24 | 25 | return df 26 | 27 | def make_df_cartoon_dataset(path_cartoons, list_sub_folders): 28 | df_cartoon = read_cartoon_dataset(path_cartoons, list_sub_folders) 29 | df_cartoon = df_cartoon.reset_index() 30 | df_cartoon['subfolder'] = df_cartoon["filename"].apply(lambda x: x.split('/')[0]) 31 | df_cartoon['filename'] = df_cartoon["filename"].apply(lambda x: x.split('/')[-1]) 32 | df_cartoon.to_csv(path_cartoons + "/cartoon100k.csv.gz", header=True, compression="gzip", index=False) 33 | 34 | return df_cartoon 35 | 36 | 37 | # Show samples per feature 38 | def show_samples_feature(df_cartoon, col, sample_len): 39 | unique_col = df_cartoon[col].unique() 40 | 41 | 42 | print("unique values of {}: {}".format(col, len(unique_col))) 43 | 44 | for value in unique_col: 45 | idx = (df_cartoon[col] == value) 46 | file_names = df_cartoon.loc[idx,["filename","subfolder"]] 47 | 48 | print("total filenames with {} equal to {}: {}".format(col, value, len(file_names))) 49 | 50 | for file in file_names[:sample_len].values: 51 | print(file) 52 | print(os.path.join(path_cartoons,str(file[1]), file[0])) 53 | img=mpimg.imread(os.path.join(path_cartoons,str(file[1]), (file[0].split('.'))[0]+".png")) 54 | imgplot = plt.imshow(img) 55 | plt.show() 56 | 57 | 58 | def show_samples_idx(df_cartoon, idx, sample_len): 59 | file_names = df_cartoon.loc[idx,["filename","subfolder"]] 60 | 61 | print("unique values of index: {}".format(len(file_names))) 62 | 63 | for file in file_names[:sample_len].values: 64 | print(file) 65 | print(os.path.join(path_cartoons,str(file[1]), file[0])) 66 | img=mpimg.imread(os.path.join(path_cartoons,str(file[1]), (file[0].split('.'))[0]+".png")) 67 | imgplot = plt.imshow(img) 68 | plt.show() 69 | 70 | 71 | if __name__ == "__main__": 72 | path_cartoons = "/data/shuaman/xgan/cartoonset100k" 73 | list_sub_folders = ["0","1","2","3", 74 | "4","5","6","7","8","9"] 75 | 76 | df_cartoon = make_df_cartoon_dataset(path_cartoons, list_sub_folders) 77 | #df_cartoon = pd.read_csv(path_cartoons + "/cartoon100k.csv.gz") 78 | 79 | #analize cartoons with show_samples_feature and show_samples_idx to get the idx to drop 80 | facial_hair_active = [0,2,3,4,5,6,7,8,9,10,11,12,13] 81 | facial_no_hair = [14] 82 | facial_hair_delete = [1,12] 83 | 84 | hair_type_woman = [3, 4, 13, 25, 27, 28, 29, 33, 34, 35, 41, 48, 56, 58, 59, 60, 63, 64, 67, 69, 70, 75, 78, 85 | 79, 80, 82, 86, 87, 90, 95, 96, 102,6,7,9,36,40,42,47,55,57,65, 68, 71, 81, 84, 85, 93, 94, 97, 100, 86 | 101, 103, 26] 87 | hair_type_man = [1, 2, 8, 11, 12, 17, 22, 23, 31, 32, 39, 43, 44, 45, 46, 49, 50, 51, 53, 54, 61, 62, 66, 72, 88 | 73, 74, 76, 77, 83, 89, 92, 99, 107,10,14,18,19,20,21,30,38,52, 105] 89 | hair_type_mix = [0, 5, 15, 16, 24, 37, 88, 91, 98, 104, 106] 90 | hair_type_delete = [108, 109, 110] 91 | 92 | glass_type_delete = [10] 93 | glass_type_keep = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 11] 94 | 95 | face_color_dark = [0, 1] 96 | face_color_light = [6, 7, 8, 9, 10] 97 | face_color_mix = [2, 3, 4, 5] 98 | 99 | glass_color_delete = [2, 5] 100 | glass_color_keep = [0, 1, 3, 4, 6] 101 | 102 | hair_color_keep = [0, 1, 2, 3, 5, 6, 7, 9] 103 | hair_color_delete = [8, 4] 104 | hair_color_light = [0, 1, 2, 4] 105 | hair_color_mix = [3,5,6,7,8,9] 106 | 107 | 108 | 109 | # Drop images which have removed attributes 110 | idx = (df_cartoon.glasses.isin(glass_type_delete)) | (df_cartoon.hair.isin(hair_type_delete)) | (df_cartoon.glasses_color.isin(glass_color_delete)) | (df_cartoon.hair_color.isin(hair_color_delete)) | (df_cartoon.facial_hair.isin(facial_hair_delete)) 111 | df_cartoon_filter1 = df_cartoon.loc[-idx,:].reset_index(drop=True) 112 | 113 | # Analize images with realistic shapes 114 | idx_women = ((df_cartoon_filter1.hair.isin(hair_type_woman)) & (df_cartoon_filter1.facial_hair.isin(facial_no_hair))) 115 | df_cartoon_filter2 = df_cartoon_filter1.loc[(-df_cartoon_filter1.hair.isin(hair_type_woman))|(idx_women) ,:].reset_index(drop=True) 116 | 117 | # Analize images with realistic color of skin and hair 118 | idx_dark_color = ((df_cartoon_filter2.face_color.isin(face_color_dark)) & (-df_cartoon_filter2.hair_color.isin(hair_color_light))) 119 | df_cartoon_filter_final = df_cartoon_filter2.loc[(-df_cartoon_filter2.face_color.isin(face_color_dark)) | (idx_dark_color),:].reset_index(drop=True) 120 | 121 | #save cartoons ids to keep 122 | df_cartoon_filter_final.to_csv(path_cartoons + "/cartoon100k_limited.csv.gz", header=True, compression="gzip", index=False) 123 | 124 | #save the png files and then execute th shell scripts 125 | writePath = "filelist.txt" 126 | df_cartoon_filter_final["filename_png"] = df_cartoon_filter_final["filename"].apply(lambda x: x.split('.')[0])+ '.png' 127 | df_cartoon_filter_final["filename_png"].to_csv(writePath, header=None, index=None, sep=' ', mode='a') -------------------------------------------------------------------------------- /sweeps/sweep-bs-1.yaml: -------------------------------------------------------------------------------- 1 | command: 2 | - ${env} 3 | - python3 4 | - ${program} 5 | - ${args} 6 | early_terminate: 7 | eta: 2 8 | min_iter: 10 9 | type: hyperband 10 | method: bayes 11 | metric: 12 | goal: minimize 13 | name: loss_total 14 | parameters: 15 | wDann_loss: 16 | distribution: uniform 17 | max: 1 18 | min: 0.5 19 | wGan_loss: 20 | distribution: uniform 21 | max: 1 22 | min: 0.5 23 | wSem_loss: 24 | distribution: uniform 25 | max: 1 26 | min: 0.5 27 | program: train.py 28 | project: avatar-image-generator -------------------------------------------------------------------------------- /sweeps/sweep-rs-1.yaml: -------------------------------------------------------------------------------- 1 | project: avatar-image-generator 2 | program: train.py 3 | method: random 4 | metric: 5 | name: loss_total 6 | goal: minimize 7 | parameters: 8 | wRec_loss: 9 | distribution: uniform 10 | min: 0.75 11 | max: 1.0 12 | wDann_loss: 13 | distribution: uniform 14 | min: 0.25 15 | max: 1.0 16 | wSem_loss: 17 | distribution: uniform 18 | min: 0.25 19 | max: 1.0 20 | wGan_loss: 21 | distribution: uniform 22 | min: 0.25 23 | max: 1.0 24 | early_terminate: 25 | type: hyperband 26 | min_iter: 10 27 | eta: 2 28 | command: 29 | - ${env} 30 | - python3 31 | - ${program} 32 | - ${args} -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | from utils import parse_arguments, set_seed, configure_model 3 | from models import Avatar_Generator_Model 4 | import os 5 | import sys 6 | import wandb 7 | 8 | CONFIG_FILENAME = "config.json" 9 | PROJECT_WANDB = "avatar_image_generator" 10 | ENTITY = "iamigos" 11 | 12 | 13 | def is_there_arg(args, master_arg): 14 | if(master_arg in args): 15 | return True 16 | else: 17 | return False 18 | 19 | 20 | def train(config_file, use_wandb, run_name, run_notes): 21 | set_seed(32) 22 | config = configure_model(config_file, use_wandb) 23 | 24 | if use_wandb: 25 | wandb.init(project=PROJECT_WANDB, entity=ENTITY, config=config, name=run_name, notes=run_notes) 26 | config = wandb.config 27 | wandb.watch_called = False 28 | 29 | 30 | xgan = Avatar_Generator_Model(config, use_wandb) 31 | xgan.train() 32 | 33 | 34 | if __name__ == '__main__': 35 | use_sweep = is_there_arg(sys.argv, '--use_sweep') 36 | 37 | if not use_sweep: 38 | args = parse_arguments() 39 | use_wandb = args.wandb 40 | run_name = args.run_name 41 | run_notes = args.run_notes 42 | else: 43 | use_wandb = True 44 | run_name = None 45 | run_notes = None 46 | 47 | train(CONFIG_FILENAME, use_wandb=use_wandb, run_name=run_name, run_notes=run_notes) 48 | -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- 1 | import wandb 2 | import math 3 | import numpy as np 4 | import matplotlib.pyplot as plt 5 | import os 6 | import sys 7 | import argparse 8 | 9 | import torch 10 | import torch.nn as nn 11 | import torch.nn.functional as F 12 | 13 | import torchvision.transforms as transforms 14 | import torchvision 15 | from torch.autograd import Variable 16 | from torch.utils.data import DataLoader 17 | 18 | from tqdm import tqdm 19 | from PIL import Image 20 | 21 | import logging 22 | import random 23 | 24 | from keras_segmentation.pretrained import pspnet_50_ADE_20K, pspnet_101_cityscapes, pspnet_101_voc12 25 | import cv2 26 | import helper 27 | import json 28 | 29 | 30 | IMAGE_SIZE = 64 31 | MEAN = 0.5 32 | SD = 0.5 33 | STATS = (MEAN, MEAN, MEAN), (SD, SD, SD) 34 | 35 | def set_seed(seed): 36 | """Set seed""" 37 | random.seed(seed) 38 | np.random.seed(seed) 39 | torch.manual_seed(seed) 40 | if torch.cuda.is_available(): 41 | torch.cuda.manual_seed(seed) 42 | torch.cuda.manual_seed_all(seed) 43 | torch.backends.cudnn.deterministic = True 44 | torch.backends.cudnn.benchmark = False 45 | os.environ["PYTHONHASHSEED"] = str(seed) 46 | 47 | 48 | def parse_arguments(): 49 | ap = argparse.ArgumentParser() 50 | ap.add_argument('-w', '--wandb', default=False, action='store_true', 51 | help="use weights and biases") 52 | ap.add_argument('-nw ', '--no-wandb', dest='wandb', action='store_false', 53 | help="not use weights and biases") 54 | ap.add_argument('-n', '--run_name', required=False, type=str, default=None, 55 | help="name of the execution to save in wandb") 56 | ap.add_argument('-nt', '--run_notes', required=False, type=str, default=None, 57 | help="notes of the execution to save in wandb") 58 | 59 | args = ap.parse_args() 60 | 61 | return args 62 | 63 | 64 | def parse_configuration(config_file): 65 | """Loads config file if a string was passed 66 | and returns the input if a dictionary was passed. 67 | """ 68 | if isinstance(config_file, str): 69 | with open(config_file, 'r') as json_file: 70 | return json.load(json_file) 71 | else: 72 | return config_file 73 | 74 | 75 | def init_logger(log_file=None, log_dir=None): 76 | 77 | fmt = '%(asctime)s %(filename)s[line:%(lineno)d] %(levelname)s: %(message)s' 78 | 79 | if log_dir is None: 80 | log_dir = '~/temp/log/' 81 | 82 | if not os.path.exists(log_dir): 83 | print("Creating dir") 84 | os.makedirs(log_dir) 85 | 86 | log_file = os.path.join(log_dir, log_file) 87 | 88 | print('log file path:' + log_file) 89 | 90 | logging.basicConfig(level=logging.INFO, 91 | filename=log_file, 92 | format=fmt) 93 | 94 | return logging 95 | 96 | 97 | def configure_model(config_file, use_wandb): 98 | 99 | config_file = parse_configuration(config_file) 100 | 101 | config = dict( 102 | model_path=config_file["server_config"]["model_path"], 103 | download_directory=config_file["server_config"]["download_directory"], 104 | 105 | root_path=config_file["train_dataset_params"]["root_path"], 106 | dataset_path_faces=config_file["train_dataset_params"]["dataset_path_faces"], 107 | dataset_path_cartoons=config_file["train_dataset_params"]["dataset_path_cartoons"], 108 | dataset_path_test_faces=config_file["train_dataset_params"]["dataset_path_test_faces"], 109 | dataset_path_segmented_faces=config_file["train_dataset_params"]["dataset_path_segmented_faces"], 110 | dataset_path_output_faces=config_file["train_dataset_params"]["dataset_path_output_faces"], 111 | batch_size=config_file["train_dataset_params"]["loader_params"]["batch_size"], 112 | 113 | save_weights=config_file["train_dataset_params"]["save_weights"], 114 | num_backups=config_file["train_dataset_params"]["num_backups"], 115 | save_path=config_file["train_dataset_params"]["save_path"], 116 | 117 | dropout_rate_eshared=config_file["model_hparams"]["dropout_rate_eshared"], 118 | use_critic_dann=config_file["model_hparams"]["use_critic_dann"], 119 | use_critic_disc=config_file["model_hparams"]["use_critic_disc"], 120 | use_spectral_norm=config_file["model_hparams"]["use_spectral_norm"], 121 | use_denoiser=config_file["model_hparams"]["use_denoiser"], 122 | use_disc_cartoon2face=config_file["model_hparams"]["use_disc_cartoon2face"], 123 | num_epochs=config_file["model_hparams"]["num_epochs"], 124 | learning_rate_opTotal=config_file["model_hparams"]["learning_rate_opTotal"], 125 | learning_rate_opDisc=config_file["model_hparams"]["learning_rate_opDisc"], 126 | learning_rate_denoiser=config_file["model_hparams"]["learning_rate_denoiser"], 127 | learning_rate_opCdann=config_file["model_hparams"]["learning_rate_opCdann"], 128 | wRec_loss=config_file["model_hparams"]["wRec_loss"], 129 | wDann_loss=config_file["model_hparams"]["wDann_loss"], 130 | wSem_loss=config_file["model_hparams"]["wSem_loss"], 131 | wGan_loss=config_file["model_hparams"]["wGan_loss"], 132 | wTeach_loss=config_file["model_hparams"]["wTeach_loss"], 133 | use_gpu=config_file["model_hparams"]["use_gpu"] 134 | ) 135 | 136 | if not use_wandb: 137 | config = type("configuration", (object,), config) 138 | 139 | return config 140 | 141 | 142 | def weights_init(m): 143 | classname = m.__class__.__name__ 144 | if classname.find('Conv') != -1: 145 | nn.init.kaiming_uniform_(m.weight.data) 146 | 147 | elif classname.find('BatchNorm') != -1: 148 | nn.init.normal_(m.weight.data, 1.0, 0.02) 149 | nn.init.constant_(m.bias.data, 0) 150 | 151 | 152 | def save_weights(model, path_sub, use_wandb=True): 153 | e1, e2, d1, d2, e_shared, d_shared, c_dann, discriminator1, denoiser, discriminator2 = model 154 | 155 | torch.save(e1.state_dict(), os.path.join(path_sub, 'e1.pth')) 156 | torch.save(e2.state_dict(), os.path.join(path_sub, 'e2.pth')) 157 | torch.save(e_shared.state_dict(), os.path.join(path_sub, 'e_shared.pth')) 158 | torch.save(d_shared.state_dict(), os.path.join(path_sub, 'd_shared.pth')) 159 | torch.save(d1.state_dict(), os.path.join(path_sub, 'd1.pth')) 160 | torch.save(d2.state_dict(), os.path.join(path_sub, 'd2.pth')) 161 | torch.save(c_dann.state_dict(), os.path.join(path_sub, 'c_dann.pth')) 162 | torch.save(discriminator1.state_dict(), os.path.join(path_sub, 'disc1.pth')) 163 | torch.save(denoiser.state_dict(), os.path.join(path_sub, 'denoiser.pth')) 164 | torch.save(discriminator2.state_dict(), os.path.join(path_sub, 'disc2.pth')) 165 | 166 | if use_wandb: 167 | wandb.save(os.path.join(path_sub, '*.pth'), 168 | base_path='/'.join(path_sub.split('/')[:-2])) 169 | 170 | 171 | def get_transforms_config_face(): 172 | list_transforms = [ 173 | transforms.Resize((IMAGE_SIZE, IMAGE_SIZE)), 174 | transforms.ToTensor(), 175 | transforms.Normalize(*STATS) 176 | ] 177 | 178 | return list_transforms 179 | 180 | 181 | def get_transforms_config_cartoon(): 182 | list_transforms = [ 183 | transforms.CenterCrop(400), 184 | transforms.Resize((IMAGE_SIZE, IMAGE_SIZE)), 185 | transforms.ToTensor(), 186 | transforms.Normalize(*STATS) 187 | ] 188 | 189 | return list_transforms 190 | 191 | def get_datasets(root_path, dataset_path_faces, dataset_path_cartoons, batch_size): 192 | 193 | path_faces = root_path + dataset_path_faces 194 | path_cartoons = root_path + dataset_path_cartoons 195 | 196 | transform_list_faces = get_transforms_config_face() 197 | transform_list_cartoons = get_transforms_config_cartoon() 198 | 199 | transform_faces = transforms.Compose(transform_list_faces) 200 | 201 | transform_cartoons = transforms.Compose(transform_list_cartoons) 202 | 203 | dataset_faces = torchvision.datasets.ImageFolder( 204 | path_faces, transform=transform_faces) 205 | dataset_cartoons = torchvision.datasets.ImageFolder( 206 | path_cartoons, transform=transform_cartoons) 207 | 208 | train_dataset_faces, test_dataset_faces = torch.utils.data.random_split(dataset_faces, 209 | (int(len(dataset_faces)*0.9), len(dataset_faces) - int(len(dataset_faces)*0.9))) 210 | 211 | train_loader_faces = torch.utils.data.DataLoader( 212 | train_dataset_faces, 213 | batch_size=batch_size, 214 | shuffle=True, 215 | num_workers=4) 216 | 217 | test_loader_faces = torch.utils.data.DataLoader( 218 | test_dataset_faces, 219 | batch_size=batch_size, 220 | shuffle=True, 221 | num_workers=4) 222 | 223 | train_dataset_cartoons, test_dataset_cartoons = torch.utils.data.random_split(dataset_cartoons, 224 | (int(len(dataset_cartoons)*0.9), len(dataset_cartoons) - int(len(dataset_cartoons)*0.9))) 225 | 226 | train_loader_cartoons = torch.utils.data.DataLoader( 227 | train_dataset_cartoons, 228 | batch_size=batch_size, 229 | shuffle=True, 230 | num_workers=4) 231 | 232 | test_loader_cartoons = torch.utils.data.DataLoader( 233 | test_dataset_cartoons, 234 | batch_size=batch_size, 235 | shuffle=True, 236 | num_workers=4) 237 | 238 | return (train_loader_faces, test_loader_faces, train_loader_cartoons, test_loader_cartoons) 239 | 240 | 241 | def remove_background_image(model, path_filename, output_path): 242 | 243 | output_file = path_filename.split('/')[-1].split('.')[0] + "_wo_bg.jpg" 244 | 245 | out = model.predict_segmentation( 246 | inp=path_filename, 247 | out_fname=output_path + output_file 248 | ) 249 | 250 | img_mask = cv2.imread(output_path + output_file) 251 | img1 = cv2.imread(path_filename) # READ BGR 252 | 253 | seg_gray = cv2.cvtColor(img_mask, cv2.COLOR_BGR2GRAY) 254 | _, bg_mask = cv2.threshold( 255 | seg_gray, 0, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU) 256 | 257 | bg_mask = cv2.cvtColor(bg_mask, cv2.COLOR_GRAY2BGR) 258 | 259 | bg = cv2.bitwise_or(img1, bg_mask) 260 | 261 | cv2.imwrite(output_path + output_file, bg) 262 | 263 | 264 | def remove_background(model, path_test_faces, path_segmented_faces): 265 | 266 | path = path_test_faces + 'data/' 267 | output_path = path_segmented_faces + 'data/' 268 | 269 | dir_path = os.path.dirname(output_path) 270 | if not os.path.exists(dir_path): 271 | os.makedirs(dir_path) 272 | 273 | for filename in tqdm(os.listdir(path)): 274 | 275 | remove_background_image(model, path + filename, output_path) 276 | 277 | 278 | def get_test_images(model, batch_size, path_test_faces, path_segmented_faces): 279 | 280 | remove_background(model, path_test_faces, path_segmented_faces) 281 | 282 | path_test_images = path_segmented_faces 283 | 284 | transform_list_faces = get_transforms_config_face() 285 | transform_list_faces += [transforms.CenterCrop(IMAGE_SIZE)] 286 | 287 | transform = transforms.Compose(transform_list_faces) 288 | 289 | dataset_test_images = torchvision.datasets.ImageFolder( 290 | path_test_images, transform=transform) 291 | 292 | test_loader_images = torch.utils.data.DataLoader( 293 | dataset_test_images, 294 | batch_size=batch_size, 295 | num_workers=4) 296 | 297 | dataiter = iter(test_loader_images) 298 | test_images = dataiter.next() 299 | 300 | return test_images 301 | 302 | 303 | def denorm(img_tensors): 304 | 305 | return img_tensors * STATS[1][0] + STATS[0][0] 306 | 307 | 308 | def test_image(model, device, images_faces, use_denoiser): 309 | 310 | e1, e2, d1, d2, e_shared, d_shared, c_dann, discriminator1, denoiser, discriminator2 = model 311 | 312 | e1.eval() 313 | e2.eval() 314 | e_shared.eval() 315 | d_shared.eval() 316 | d1.eval() 317 | d2.eval() 318 | c_dann.eval() 319 | discriminator1.eval() 320 | discriminator2.eval() 321 | denoiser.eval() 322 | 323 | with torch.no_grad(): 324 | output = e1(images_faces[0].to(device)) 325 | output = e_shared(output) 326 | output = d_shared(output) 327 | output = d2(output) 328 | if use_denoiser: 329 | output = denoiser(output) 330 | 331 | output = denorm(output) 332 | 333 | return output.cpu() 334 | 335 | 336 | 337 | 338 | 339 | def init_optimizers(model, learning_rate_opDisc, learning_rate_opTotal, learning_rate_denoiser, learning_rate_opCdann): 340 | 341 | e1, e2, d1, d2, e_shared, d_shared, c_dann, discriminator1, denoiser, discriminator2 = model 342 | 343 | optimizerDisc1 = torch.optim.Adam( 344 | discriminator1.parameters(), lr=learning_rate_opDisc, betas=(0.5, 0.999)) 345 | 346 | optimizerDisc2 = torch.optim.Adam( 347 | discriminator2.parameters(), lr=learning_rate_opDisc, betas=(0.5, 0.999)) 348 | 349 | #listParameters = list(e1.parameters()) + list(e2.parameters()) + list(e_shared.parameters()) + list(d_shared.parameters()) + list(d1.parameters()) + list(d2.parameters()) + list(c_dann.parameters()) 350 | listParameters = list(e1.parameters()) + list(e2.parameters()) + list(e_shared.parameters()) + \ 351 | list(d_shared.parameters()) + list(d1.parameters()) + list(d2.parameters()) 352 | optimizerTotal = torch.optim.Adam( 353 | listParameters, lr=learning_rate_opTotal, betas=(0.5, 0.999)) 354 | 355 | optimizerDenoiser = torch.optim.Adam( 356 | denoiser.parameters(), lr=learning_rate_denoiser) 357 | 358 | optimizerCdann = torch.optim.Adam( 359 | c_dann.parameters(), lr=learning_rate_opCdann, betas=(0.5, 0.999)) 360 | 361 | return (optimizerDenoiser, optimizerDisc1, optimizerTotal, optimizerCdann, optimizerDisc2) 362 | 363 | 364 | --------------------------------------------------------------------------------