├── .flake8 ├── .github └── workflows │ ├── tests_full.yaml │ └── tests_reduced.yaml ├── .gitignore ├── CHANGELOG.md ├── LICENSE ├── README.md ├── noxfile.py ├── pyproject.toml ├── setup.py ├── src └── pytorch_fid │ ├── __init__.py │ ├── __main__.py │ ├── fid_score.py │ └── inception.py └── tests └── test_fid_score.py /.flake8: -------------------------------------------------------------------------------- 1 | [flake8] 2 | select = F,W,E,I,B,B9 3 | ignore = W503,E203,B950 4 | max-line-length = 88 5 | -------------------------------------------------------------------------------- /.github/workflows/tests_full.yaml: -------------------------------------------------------------------------------- 1 | name: "Full set of tests for PRs and master branch" 2 | on: 3 | push: 4 | branches: 5 | - "master" 6 | pull_request: 7 | 8 | concurrency: 9 | group: ${{ github.ref }} 10 | cancel-in-progress: true 11 | 12 | jobs: 13 | tests: 14 | runs-on: ubuntu-20.04 15 | strategy: 16 | matrix: 17 | python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"] 18 | name: Python ${{ matrix.python-version }} 19 | steps: 20 | - uses: actions/checkout@v4 21 | - name: Set up Python ${{ matrix.python-version }} 22 | uses: actions/setup-python@v5 23 | with: 24 | python-version: ${{ matrix.python-version }} 25 | - name: Install Nox 26 | run: pip install nox==2024.03.02 27 | - name: Run tests 28 | run: nox --non-interactive --error-on-missing-interpreter --session "tests-${{ matrix.python-version }}" 29 | 30 | lint: 31 | runs-on: ubuntu-latest 32 | steps: 33 | - uses: actions/checkout@v4 34 | - name: Set up Python 3.9 35 | uses: actions/setup-python@v5 36 | with: 37 | python-version: 3.9 38 | - name: Install Nox 39 | run: pip install nox==2024.03.02 40 | - name: Lint 41 | run: nox --non-interactive --error-on-missing-interpreter --session "lint" 42 | -------------------------------------------------------------------------------- /.github/workflows/tests_reduced.yaml: -------------------------------------------------------------------------------- 1 | name: "Reduced set of tests for push events" 2 | on: 3 | push: 4 | branches-ignore: 5 | - master 6 | 7 | concurrency: 8 | group: ${{ github.ref }} 9 | cancel-in-progress: true 10 | 11 | jobs: 12 | tests: 13 | runs-on: ubuntu-latest 14 | strategy: 15 | matrix: 16 | python-version: ["3.11"] 17 | name: Python ${{ matrix.python-version }} 18 | steps: 19 | - uses: actions/checkout@v4 20 | - name: Set up Python ${{ matrix.python-version }} 21 | uses: actions/setup-python@v5 22 | with: 23 | python-version: ${{ matrix.python-version }} 24 | - name: Install Nox 25 | run: pip install nox==2024.03.02 26 | - name: Run tests 27 | run: nox --non-interactive --error-on-missing-interpreter --session "tests-${{ matrix.python-version }}" 28 | 29 | lint: 30 | runs-on: ubuntu-latest 31 | steps: 32 | - uses: actions/checkout@v4 33 | - name: Set up Python 3.9 34 | uses: actions/setup-python@v5 35 | with: 36 | python-version: 3.9 37 | - name: Install Nox 38 | run: pip install nox==2024.03.02 39 | - name: Lint 40 | run: nox --non-interactive --error-on-missing-interpreter --session "lint" 41 | -------------------------------------------------------------------------------- /.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 | .hypothesis/ 51 | .pytest_cache/ 52 | 53 | # Translations 54 | *.mo 55 | *.pot 56 | 57 | # Django stuff: 58 | *.log 59 | local_settings.py 60 | db.sqlite3 61 | 62 | # Flask stuff: 63 | instance/ 64 | .webassets-cache 65 | 66 | # Scrapy stuff: 67 | .scrapy 68 | 69 | # Sphinx documentation 70 | docs/_build/ 71 | 72 | # PyBuilder 73 | target/ 74 | 75 | # Jupyter Notebook 76 | .ipynb_checkpoints 77 | 78 | # IPython 79 | profile_default/ 80 | ipython_config.py 81 | 82 | # pyenv 83 | .python-version 84 | 85 | # celery beat schedule file 86 | celerybeat-schedule 87 | 88 | # SageMath parsed files 89 | *.sage.py 90 | 91 | # Environments 92 | .env 93 | .venv 94 | env/ 95 | venv/ 96 | ENV/ 97 | env.bak/ 98 | venv.bak/ 99 | 100 | # Spyder project settings 101 | .spyderproject 102 | .spyproject 103 | 104 | # Rope project settings 105 | .ropeproject 106 | 107 | # mkdocs documentation 108 | /site 109 | 110 | # mypy 111 | .mypy_cache/ 112 | .dmypy.json 113 | dmypy.json 114 | 115 | # Pyre type checker 116 | .pyre/ -------------------------------------------------------------------------------- /CHANGELOG.md: -------------------------------------------------------------------------------- 1 | # Changelog 2 | 3 | ## [0.3.0] - 2023-01-05 4 | 5 | ### Added 6 | 7 | * Add argument `--save-stats` allowing to compute dataset statistics and save them as an `.npz` file ([#80](https://github.com/mseitzer/pytorch-fid/pull/80)). The `.npz` file can be used in subsequent FID computations instead of recomputing the dataset statistics. This option can be used in the following way: `python -m pytorch_fid --save-stats path/to/dataset path/to/outputfile`. 8 | 9 | ### Fixed 10 | 11 | * Do not use `os.sched_getaffinity` to get number of available CPUs on Windows, as it is not available there ([232b3b14](https://github.com/mseitzer/pytorch-fid/commit/232b3b1468800102fcceaf6f2bb8977811fc991a), [#84](https://github.com/mseitzer/pytorch-fid/issues/84)). 12 | * Do not use Inception model argument `pretrained`, as it was deprecated in torchvision 0.13 ([#88](https://github.com/mseitzer/pytorch-fid/pull/88)). 13 | 14 | ## [0.2.1] - 2021-10-10 15 | 16 | ### Added 17 | 18 | * Add argument `--num-workers` to select number of dataloader processes ([#66](https://github.com/mseitzer/pytorch-fid/pull/66)). Defaults to 8 or the number of available CPUs if less than 8 CPUs are available. 19 | 20 | ### Fixed 21 | 22 | * Fixed package setup to work under Windows ([#55](https://github.com/mseitzer/pytorch-fid/pull/55), [#72](https://github.com/mseitzer/pytorch-fid/issues/72)) 23 | 24 | ## [0.2.0] - 2020-11-30 25 | 26 | ### Added 27 | 28 | * Load images using a Pytorch dataloader, which should result in a speed-up. ([#47](https://github.com/mseitzer/pytorch-fid/pull/47)) 29 | * Support more image extensions ([#53](https://github.com/mseitzer/pytorch-fid/pull/53)) 30 | * Improve tooling by setting up Nox, add linting and test support ([#52](https://github.com/mseitzer/pytorch-fid/pull/52)) 31 | * Add some unit tests 32 | 33 | ## [0.1.1] - 2020-08-16 34 | 35 | ### Fixed 36 | 37 | * Fixed software license string in `setup.py` 38 | 39 | ## [0.1.0] - 2020-08-16 40 | 41 | Initial release as a pypi package. Use `pip install pytorch-fid` to install. 42 | -------------------------------------------------------------------------------- /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 | [![PyPI](https://img.shields.io/pypi/v/pytorch-fid.svg)](https://pypi.org/project/pytorch-fid/) 2 | 3 | # FID score for PyTorch 4 | 5 | This is a port of the official implementation of [Fréchet Inception Distance](https://arxiv.org/abs/1706.08500) to PyTorch. 6 | See [https://github.com/bioinf-jku/TTUR](https://github.com/bioinf-jku/TTUR) for the original implementation using Tensorflow. 7 | 8 | FID is a measure of similarity between two datasets of images. 9 | It was shown to correlate well with human judgement of visual quality and is most often used to evaluate the quality of samples of Generative Adversarial Networks. 10 | FID is calculated by computing the [Fréchet distance](https://en.wikipedia.org/wiki/Fr%C3%A9chet_distance) between two Gaussians fitted to feature representations of the Inception network. 11 | 12 | Further insights and an independent evaluation of the FID score can be found in [Are GANs Created Equal? A Large-Scale Study](https://arxiv.org/abs/1711.10337). 13 | 14 | The weights and the model are exactly the same as in [the official Tensorflow implementation](https://github.com/bioinf-jku/TTUR), and were tested to give very similar results (e.g. `.08` absolute error and `0.0009` relative error on LSUN, using ProGAN generated images). However, due to differences in the image interpolation implementation and library backends, FID results still differ slightly from the original implementation. So if you report FID scores in your paper, and you want them to be *exactly comparable* to FID scores reported in other papers, you should consider using [the official Tensorflow implementation](https://github.com/bioinf-jku/TTUR). 15 | 16 | ## Installation 17 | 18 | Install from [pip](https://pypi.org/project/pytorch-fid/): 19 | 20 | ``` 21 | pip install pytorch-fid 22 | ``` 23 | 24 | Requirements: 25 | - python3 26 | - pytorch 27 | - torchvision 28 | - pillow 29 | - numpy 30 | - scipy 31 | 32 | ## Usage 33 | 34 | To compute the FID score between two datasets, where images of each dataset are contained in an individual folder: 35 | ``` 36 | python -m pytorch_fid path/to/dataset1 path/to/dataset2 37 | ``` 38 | 39 | To run the evaluation on GPU, use the flag `--device cuda:N`, where `N` is the index of the GPU to use. 40 | 41 | ### Using different layers for feature maps 42 | 43 | In difference to the official implementation, you can choose to use a different feature layer of the Inception network instead of the default `pool3` layer. 44 | As the lower layer features still have spatial extent, the features are first global average pooled to a vector before estimating mean and covariance. 45 | 46 | This might be useful if the datasets you want to compare have less than the otherwise required 2048 images. 47 | Note that this changes the magnitude of the FID score and you can not compare them against scores calculated on another dimensionality. 48 | The resulting scores might also no longer correlate with visual quality. 49 | 50 | You can select the dimensionality of features to use with the flag `--dims N`, where N is the dimensionality of features. 51 | The choices are: 52 | - 64: first max pooling features 53 | - 192: second max pooling features 54 | - 768: pre-aux classifier features 55 | - 2048: final average pooling features (this is the default) 56 | 57 | ## Generating a compatible `.npz` archive from a dataset 58 | A frequent use case will be to compare multiple models against an original dataset. 59 | To save training multiple times on the original dataset, there is also the ability to generate a compatible `.npz` archive from a dataset. This is done using any combination of the previously mentioned arguments with the addition of the `--save-stats` flag. For example: 60 | ``` 61 | python -m pytorch_fid --save-stats path/to/dataset path/to/outputfile 62 | ``` 63 | 64 | The output file may then be used in place of the path to the original dataset for further comparisons. 65 | 66 | ## Citing 67 | 68 | If you use this repository in your research, consider citing it using the following Bibtex entry: 69 | 70 | ``` 71 | @misc{Seitzer2020FID, 72 | author={Maximilian Seitzer}, 73 | title={{pytorch-fid: FID Score for PyTorch}}, 74 | month={August}, 75 | year={2020}, 76 | note={Version 0.3.0}, 77 | howpublished={\url{https://github.com/mseitzer/pytorch-fid}}, 78 | } 79 | ``` 80 | 81 | ## License 82 | 83 | This implementation is licensed under the Apache License 2.0. 84 | 85 | FID was introduced by Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler and Sepp Hochreiter in "GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium", see [https://arxiv.org/abs/1706.08500](https://arxiv.org/abs/1706.08500) 86 | 87 | The original implementation is by the Institute of Bioinformatics, JKU Linz, licensed under the Apache License 2.0. 88 | See [https://github.com/bioinf-jku/TTUR](https://github.com/bioinf-jku/TTUR). 89 | -------------------------------------------------------------------------------- /noxfile.py: -------------------------------------------------------------------------------- 1 | import nox 2 | 3 | LOCATIONS = ("src/", "tests/", "noxfile.py", "setup.py") 4 | 5 | 6 | @nox.session 7 | def lint(session): 8 | session.install("flake8") 9 | session.install("flake8-bugbear") 10 | session.install("flake8-isort") 11 | session.install("black==24.3.0") 12 | 13 | args = session.posargs or LOCATIONS 14 | session.run("flake8", *args) 15 | session.run("black", "--check", "--diff", *args) 16 | 17 | 18 | @nox.session(python=["3.8", "3.9", "3.10", "3.11", "3.12"]) 19 | def tests(session): 20 | session.install( 21 | "torch==2.2.1", 22 | "torchvision", 23 | "--index-url", 24 | "https://download.pytorch.org/whl/cpu", 25 | ) 26 | session.install(".") 27 | session.install("pytest") 28 | session.install("pytest-mock") 29 | session.run("pytest", *session.posargs) 30 | -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [tool.black] 2 | target-version = ["py311"] 3 | 4 | [tool.isort] 5 | profile = "black" 6 | line_length = 88 7 | multi_line_output = 3 8 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | import setuptools 4 | 5 | 6 | def read(rel_path): 7 | base_path = os.path.abspath(os.path.dirname(__file__)) 8 | with open(os.path.join(base_path, rel_path), "r") as f: 9 | return f.read() 10 | 11 | 12 | def get_version(rel_path): 13 | for line in read(rel_path).splitlines(): 14 | if line.startswith("__version__"): 15 | # __version__ = "0.9" 16 | delim = '"' if '"' in line else "'" 17 | return line.split(delim)[1] 18 | 19 | raise RuntimeError("Unable to find version string.") 20 | 21 | 22 | if __name__ == "__main__": 23 | setuptools.setup( 24 | name="pytorch-fid", 25 | version=get_version(os.path.join("src", "pytorch_fid", "__init__.py")), 26 | author="Max Seitzer", 27 | description=( 28 | "Package for calculating Frechet Inception Distance (FID)" " using PyTorch" 29 | ), 30 | long_description=read("README.md"), 31 | long_description_content_type="text/markdown", 32 | url="https://github.com/mseitzer/pytorch-fid", 33 | package_dir={"": "src"}, 34 | packages=setuptools.find_packages(where="src"), 35 | classifiers=[ 36 | "Programming Language :: Python :: 3", 37 | "License :: OSI Approved :: Apache Software License", 38 | ], 39 | python_requires=">=3.5", 40 | entry_points={ 41 | "console_scripts": [ 42 | "pytorch-fid = pytorch_fid.fid_score:main", 43 | ], 44 | }, 45 | install_requires=[ 46 | "numpy", 47 | "pillow", 48 | "scipy", 49 | "torch>=1.0.1", 50 | "torchvision>=0.2.2", 51 | ], 52 | extras_require={ 53 | "dev": ["flake8", "flake8-bugbear", "flake8-isort", "black==24.3.0", "nox"] 54 | }, 55 | ) 56 | -------------------------------------------------------------------------------- /src/pytorch_fid/__init__.py: -------------------------------------------------------------------------------- 1 | __version__ = "0.3.0" 2 | -------------------------------------------------------------------------------- /src/pytorch_fid/__main__.py: -------------------------------------------------------------------------------- 1 | import pytorch_fid.fid_score 2 | 3 | pytorch_fid.fid_score.main() 4 | -------------------------------------------------------------------------------- /src/pytorch_fid/fid_score.py: -------------------------------------------------------------------------------- 1 | """Calculates the Frechet Inception Distance (FID) to evalulate GANs 2 | 3 | The FID metric calculates the distance between two distributions of images. 4 | Typically, we have summary statistics (mean & covariance matrix) of one 5 | of these distributions, while the 2nd distribution is given by a GAN. 6 | 7 | When run as a stand-alone program, it compares the distribution of 8 | images that are stored as PNG/JPEG at a specified location with a 9 | distribution given by summary statistics (in pickle format). 10 | 11 | The FID is calculated by assuming that X_1 and X_2 are the activations of 12 | the pool_3 layer of the inception net for generated samples and real world 13 | samples respectively. 14 | 15 | See --help to see further details. 16 | 17 | Code apapted from https://github.com/bioinf-jku/TTUR to use PyTorch instead 18 | of Tensorflow 19 | 20 | Copyright 2018 Institute of Bioinformatics, JKU Linz 21 | 22 | Licensed under the Apache License, Version 2.0 (the "License"); 23 | you may not use this file except in compliance with the License. 24 | You may obtain a copy of the License at 25 | 26 | http://www.apache.org/licenses/LICENSE-2.0 27 | 28 | Unless required by applicable law or agreed to in writing, software 29 | distributed under the License is distributed on an "AS IS" BASIS, 30 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 31 | See the License for the specific language governing permissions and 32 | limitations under the License. 33 | """ 34 | 35 | import os 36 | import pathlib 37 | from argparse import ArgumentDefaultsHelpFormatter, ArgumentParser 38 | 39 | import numpy as np 40 | import torch 41 | import torchvision.transforms as TF 42 | from PIL import Image 43 | from scipy import linalg 44 | from torch.nn.functional import adaptive_avg_pool2d 45 | 46 | try: 47 | from tqdm import tqdm 48 | except ImportError: 49 | # If tqdm is not available, provide a mock version of it 50 | def tqdm(x): 51 | return x 52 | 53 | 54 | from pytorch_fid.inception import InceptionV3 55 | 56 | parser = ArgumentParser(formatter_class=ArgumentDefaultsHelpFormatter) 57 | parser.add_argument("--batch-size", type=int, default=50, help="Batch size to use") 58 | parser.add_argument( 59 | "--num-workers", 60 | type=int, 61 | help=( 62 | "Number of processes to use for data loading. " "Defaults to `min(8, num_cpus)`" 63 | ), 64 | ) 65 | parser.add_argument( 66 | "--device", type=str, default=None, help="Device to use. Like cuda, cuda:0 or cpu" 67 | ) 68 | parser.add_argument( 69 | "--dims", 70 | type=int, 71 | default=2048, 72 | choices=list(InceptionV3.BLOCK_INDEX_BY_DIM), 73 | help=( 74 | "Dimensionality of Inception features to use. " 75 | "By default, uses pool3 features" 76 | ), 77 | ) 78 | parser.add_argument( 79 | "--save-stats", 80 | action="store_true", 81 | help=( 82 | "Generate an npz archive from a directory of " 83 | "samples. The first path is used as input and the " 84 | "second as output." 85 | ), 86 | ) 87 | parser.add_argument( 88 | "path", 89 | type=str, 90 | nargs=2, 91 | help=("Paths to the generated images or " "to .npz statistic files"), 92 | ) 93 | 94 | IMAGE_EXTENSIONS = {"bmp", "jpg", "jpeg", "pgm", "png", "ppm", "tif", "tiff", "webp"} 95 | 96 | 97 | class ImagePathDataset(torch.utils.data.Dataset): 98 | def __init__(self, files, transforms=None): 99 | self.files = files 100 | self.transforms = transforms 101 | 102 | def __len__(self): 103 | return len(self.files) 104 | 105 | def __getitem__(self, i): 106 | path = self.files[i] 107 | img = Image.open(path).convert("RGB") 108 | if self.transforms is not None: 109 | img = self.transforms(img) 110 | return img 111 | 112 | 113 | def get_activations( 114 | files, model, batch_size=50, dims=2048, device="cpu", num_workers=1 115 | ): 116 | """Calculates the activations of the pool_3 layer for all images. 117 | 118 | Params: 119 | -- files : List of image files paths 120 | -- model : Instance of inception model 121 | -- batch_size : Batch size of images for the model to process at once. 122 | Make sure that the number of samples is a multiple of 123 | the batch size, otherwise some samples are ignored. This 124 | behavior is retained to match the original FID score 125 | implementation. 126 | -- dims : Dimensionality of features returned by Inception 127 | -- device : Device to run calculations 128 | -- num_workers : Number of parallel dataloader workers 129 | 130 | Returns: 131 | -- A numpy array of dimension (num images, dims) that contains the 132 | activations of the given tensor when feeding inception with the 133 | query tensor. 134 | """ 135 | model.eval() 136 | 137 | if batch_size > len(files): 138 | print( 139 | ( 140 | "Warning: batch size is bigger than the data size. " 141 | "Setting batch size to data size" 142 | ) 143 | ) 144 | batch_size = len(files) 145 | 146 | dataset = ImagePathDataset(files, transforms=TF.ToTensor()) 147 | dataloader = torch.utils.data.DataLoader( 148 | dataset, 149 | batch_size=batch_size, 150 | shuffle=False, 151 | drop_last=False, 152 | num_workers=num_workers, 153 | ) 154 | 155 | pred_arr = np.empty((len(files), dims)) 156 | 157 | start_idx = 0 158 | 159 | for batch in tqdm(dataloader): 160 | batch = batch.to(device) 161 | 162 | with torch.no_grad(): 163 | pred = model(batch)[0] 164 | 165 | # If model output is not scalar, apply global spatial average pooling. 166 | # This happens if you choose a dimensionality not equal 2048. 167 | if pred.size(2) != 1 or pred.size(3) != 1: 168 | pred = adaptive_avg_pool2d(pred, output_size=(1, 1)) 169 | 170 | pred = pred.squeeze(3).squeeze(2).cpu().numpy() 171 | 172 | pred_arr[start_idx : start_idx + pred.shape[0]] = pred 173 | 174 | start_idx = start_idx + pred.shape[0] 175 | 176 | return pred_arr 177 | 178 | 179 | def calculate_frechet_distance(mu1, sigma1, mu2, sigma2, eps=1e-6): 180 | """Numpy implementation of the Frechet Distance. 181 | The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1) 182 | and X_2 ~ N(mu_2, C_2) is 183 | d^2 = ||mu_1 - mu_2||^2 + Tr(C_1 + C_2 - 2*sqrt(C_1*C_2)). 184 | 185 | Stable version by Dougal J. Sutherland. 186 | 187 | Params: 188 | -- mu1 : Numpy array containing the activations of a layer of the 189 | inception net (like returned by the function 'get_predictions') 190 | for generated samples. 191 | -- mu2 : The sample mean over activations, precalculated on an 192 | representative data set. 193 | -- sigma1: The covariance matrix over activations for generated samples. 194 | -- sigma2: The covariance matrix over activations, precalculated on an 195 | representative data set. 196 | 197 | Returns: 198 | -- : The Frechet Distance. 199 | """ 200 | 201 | mu1 = np.atleast_1d(mu1) 202 | mu2 = np.atleast_1d(mu2) 203 | 204 | sigma1 = np.atleast_2d(sigma1) 205 | sigma2 = np.atleast_2d(sigma2) 206 | 207 | assert ( 208 | mu1.shape == mu2.shape 209 | ), "Training and test mean vectors have different lengths" 210 | assert ( 211 | sigma1.shape == sigma2.shape 212 | ), "Training and test covariances have different dimensions" 213 | 214 | diff = mu1 - mu2 215 | 216 | # Product might be almost singular 217 | covmean, _ = linalg.sqrtm(sigma1.dot(sigma2), disp=False) 218 | if not np.isfinite(covmean).all(): 219 | msg = ( 220 | "fid calculation produces singular product; " 221 | "adding %s to diagonal of cov estimates" 222 | ) % eps 223 | print(msg) 224 | offset = np.eye(sigma1.shape[0]) * eps 225 | covmean = linalg.sqrtm((sigma1 + offset).dot(sigma2 + offset)) 226 | 227 | # Numerical error might give slight imaginary component 228 | if np.iscomplexobj(covmean): 229 | if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3): 230 | m = np.max(np.abs(covmean.imag)) 231 | raise ValueError("Imaginary component {}".format(m)) 232 | covmean = covmean.real 233 | 234 | tr_covmean = np.trace(covmean) 235 | 236 | return diff.dot(diff) + np.trace(sigma1) + np.trace(sigma2) - 2 * tr_covmean 237 | 238 | 239 | def calculate_activation_statistics( 240 | files, model, batch_size=50, dims=2048, device="cpu", num_workers=1 241 | ): 242 | """Calculation of the statistics used by the FID. 243 | Params: 244 | -- files : List of image files paths 245 | -- model : Instance of inception model 246 | -- batch_size : The images numpy array is split into batches with 247 | batch size batch_size. A reasonable batch size 248 | depends on the hardware. 249 | -- dims : Dimensionality of features returned by Inception 250 | -- device : Device to run calculations 251 | -- num_workers : Number of parallel dataloader workers 252 | 253 | Returns: 254 | -- mu : The mean over samples of the activations of the pool_3 layer of 255 | the inception model. 256 | -- sigma : The covariance matrix of the activations of the pool_3 layer of 257 | the inception model. 258 | """ 259 | act = get_activations(files, model, batch_size, dims, device, num_workers) 260 | mu = np.mean(act, axis=0) 261 | sigma = np.cov(act, rowvar=False) 262 | return mu, sigma 263 | 264 | 265 | def compute_statistics_of_path(path, model, batch_size, dims, device, num_workers=1): 266 | if path.endswith(".npz"): 267 | with np.load(path) as f: 268 | m, s = f["mu"][:], f["sigma"][:] 269 | else: 270 | path = pathlib.Path(path) 271 | files = sorted( 272 | [file for ext in IMAGE_EXTENSIONS for file in path.glob("*.{}".format(ext))] 273 | ) 274 | m, s = calculate_activation_statistics( 275 | files, model, batch_size, dims, device, num_workers 276 | ) 277 | 278 | return m, s 279 | 280 | 281 | def calculate_fid_given_paths(paths, batch_size, device, dims, num_workers=1): 282 | """Calculates the FID of two paths""" 283 | for p in paths: 284 | if not os.path.exists(p): 285 | raise RuntimeError("Invalid path: %s" % p) 286 | 287 | block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims] 288 | 289 | model = InceptionV3([block_idx]).to(device) 290 | 291 | m1, s1 = compute_statistics_of_path( 292 | paths[0], model, batch_size, dims, device, num_workers 293 | ) 294 | m2, s2 = compute_statistics_of_path( 295 | paths[1], model, batch_size, dims, device, num_workers 296 | ) 297 | fid_value = calculate_frechet_distance(m1, s1, m2, s2) 298 | 299 | return fid_value 300 | 301 | 302 | def save_fid_stats(paths, batch_size, device, dims, num_workers=1): 303 | """Saves FID statistics of one path""" 304 | if not os.path.exists(paths[0]): 305 | raise RuntimeError("Invalid path: %s" % paths[0]) 306 | 307 | if os.path.exists(paths[1]): 308 | raise RuntimeError("Existing output file: %s" % paths[1]) 309 | 310 | block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims] 311 | 312 | model = InceptionV3([block_idx]).to(device) 313 | 314 | print(f"Saving statistics for {paths[0]}") 315 | 316 | m1, s1 = compute_statistics_of_path( 317 | paths[0], model, batch_size, dims, device, num_workers 318 | ) 319 | 320 | np.savez_compressed(paths[1], mu=m1, sigma=s1) 321 | 322 | 323 | def main(): 324 | args = parser.parse_args() 325 | 326 | if args.device is None: 327 | device = torch.device("cuda" if (torch.cuda.is_available()) else "cpu") 328 | else: 329 | device = torch.device(args.device) 330 | 331 | if args.num_workers is None: 332 | try: 333 | num_cpus = len(os.sched_getaffinity(0)) 334 | except AttributeError: 335 | # os.sched_getaffinity is not available under Windows, use 336 | # os.cpu_count instead (which may not return the *available* number 337 | # of CPUs). 338 | num_cpus = os.cpu_count() 339 | 340 | num_workers = min(num_cpus, 8) if num_cpus is not None else 0 341 | else: 342 | num_workers = args.num_workers 343 | 344 | if args.save_stats: 345 | save_fid_stats(args.path, args.batch_size, device, args.dims, num_workers) 346 | return 347 | 348 | fid_value = calculate_fid_given_paths( 349 | args.path, args.batch_size, device, args.dims, num_workers 350 | ) 351 | print("FID: ", fid_value) 352 | 353 | 354 | if __name__ == "__main__": 355 | main() 356 | -------------------------------------------------------------------------------- /src/pytorch_fid/inception.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | import torchvision 5 | 6 | try: 7 | from torchvision.models.utils import load_state_dict_from_url 8 | except ImportError: 9 | from torch.utils.model_zoo import load_url as load_state_dict_from_url 10 | 11 | # Inception weights ported to Pytorch from 12 | # http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz 13 | FID_WEIGHTS_URL = "https://github.com/mseitzer/pytorch-fid/releases/download/fid_weights/pt_inception-2015-12-05-6726825d.pth" # noqa: E501 14 | 15 | 16 | class InceptionV3(nn.Module): 17 | """Pretrained InceptionV3 network returning feature maps""" 18 | 19 | # Index of default block of inception to return, 20 | # corresponds to output of final average pooling 21 | DEFAULT_BLOCK_INDEX = 3 22 | 23 | # Maps feature dimensionality to their output blocks indices 24 | BLOCK_INDEX_BY_DIM = { 25 | 64: 0, # First max pooling features 26 | 192: 1, # Second max pooling featurs 27 | 768: 2, # Pre-aux classifier features 28 | 2048: 3, # Final average pooling features 29 | } 30 | 31 | def __init__( 32 | self, 33 | output_blocks=(DEFAULT_BLOCK_INDEX,), 34 | resize_input=True, 35 | normalize_input=True, 36 | requires_grad=False, 37 | use_fid_inception=True, 38 | ): 39 | """Build pretrained InceptionV3 40 | 41 | Parameters 42 | ---------- 43 | output_blocks : list of int 44 | Indices of blocks to return features of. Possible values are: 45 | - 0: corresponds to output of first max pooling 46 | - 1: corresponds to output of second max pooling 47 | - 2: corresponds to output which is fed to aux classifier 48 | - 3: corresponds to output of final average pooling 49 | resize_input : bool 50 | If true, bilinearly resizes input to width and height 299 before 51 | feeding input to model. As the network without fully connected 52 | layers is fully convolutional, it should be able to handle inputs 53 | of arbitrary size, so resizing might not be strictly needed 54 | normalize_input : bool 55 | If true, scales the input from range (0, 1) to the range the 56 | pretrained Inception network expects, namely (-1, 1) 57 | requires_grad : bool 58 | If true, parameters of the model require gradients. Possibly useful 59 | for finetuning the network 60 | use_fid_inception : bool 61 | If true, uses the pretrained Inception model used in Tensorflow's 62 | FID implementation. If false, uses the pretrained Inception model 63 | available in torchvision. The FID Inception model has different 64 | weights and a slightly different structure from torchvision's 65 | Inception model. If you want to compute FID scores, you are 66 | strongly advised to set this parameter to true to get comparable 67 | results. 68 | """ 69 | super(InceptionV3, self).__init__() 70 | 71 | self.resize_input = resize_input 72 | self.normalize_input = normalize_input 73 | self.output_blocks = sorted(output_blocks) 74 | self.last_needed_block = max(output_blocks) 75 | 76 | assert self.last_needed_block <= 3, "Last possible output block index is 3" 77 | 78 | self.blocks = nn.ModuleList() 79 | 80 | if use_fid_inception: 81 | inception = fid_inception_v3() 82 | else: 83 | inception = _inception_v3(weights="DEFAULT") 84 | 85 | # Block 0: input to maxpool1 86 | block0 = [ 87 | inception.Conv2d_1a_3x3, 88 | inception.Conv2d_2a_3x3, 89 | inception.Conv2d_2b_3x3, 90 | nn.MaxPool2d(kernel_size=3, stride=2), 91 | ] 92 | self.blocks.append(nn.Sequential(*block0)) 93 | 94 | # Block 1: maxpool1 to maxpool2 95 | if self.last_needed_block >= 1: 96 | block1 = [ 97 | inception.Conv2d_3b_1x1, 98 | inception.Conv2d_4a_3x3, 99 | nn.MaxPool2d(kernel_size=3, stride=2), 100 | ] 101 | self.blocks.append(nn.Sequential(*block1)) 102 | 103 | # Block 2: maxpool2 to aux classifier 104 | if self.last_needed_block >= 2: 105 | block2 = [ 106 | inception.Mixed_5b, 107 | inception.Mixed_5c, 108 | inception.Mixed_5d, 109 | inception.Mixed_6a, 110 | inception.Mixed_6b, 111 | inception.Mixed_6c, 112 | inception.Mixed_6d, 113 | inception.Mixed_6e, 114 | ] 115 | self.blocks.append(nn.Sequential(*block2)) 116 | 117 | # Block 3: aux classifier to final avgpool 118 | if self.last_needed_block >= 3: 119 | block3 = [ 120 | inception.Mixed_7a, 121 | inception.Mixed_7b, 122 | inception.Mixed_7c, 123 | nn.AdaptiveAvgPool2d(output_size=(1, 1)), 124 | ] 125 | self.blocks.append(nn.Sequential(*block3)) 126 | 127 | for param in self.parameters(): 128 | param.requires_grad = requires_grad 129 | 130 | def forward(self, inp): 131 | """Get Inception feature maps 132 | 133 | Parameters 134 | ---------- 135 | inp : torch.autograd.Variable 136 | Input tensor of shape Bx3xHxW. Values are expected to be in 137 | range (0, 1) 138 | 139 | Returns 140 | ------- 141 | List of torch.autograd.Variable, corresponding to the selected output 142 | block, sorted ascending by index 143 | """ 144 | outp = [] 145 | x = inp 146 | 147 | if self.resize_input: 148 | x = F.interpolate(x, size=(299, 299), mode="bilinear", align_corners=False) 149 | 150 | if self.normalize_input: 151 | x = 2 * x - 1 # Scale from range (0, 1) to range (-1, 1) 152 | 153 | for idx, block in enumerate(self.blocks): 154 | x = block(x) 155 | if idx in self.output_blocks: 156 | outp.append(x) 157 | 158 | if idx == self.last_needed_block: 159 | break 160 | 161 | return outp 162 | 163 | 164 | def _inception_v3(*args, **kwargs): 165 | """Wraps `torchvision.models.inception_v3`""" 166 | try: 167 | version = tuple(map(int, torchvision.__version__.split(".")[:2])) 168 | except ValueError: 169 | # Just a caution against weird version strings 170 | version = (0,) 171 | 172 | # Skips default weight inititialization if supported by torchvision 173 | # version. See https://github.com/mseitzer/pytorch-fid/issues/28. 174 | if version >= (0, 6): 175 | kwargs["init_weights"] = False 176 | 177 | # Backwards compatibility: `weights` argument was handled by `pretrained` 178 | # argument prior to version 0.13. 179 | if version < (0, 13) and "weights" in kwargs: 180 | if kwargs["weights"] == "DEFAULT": 181 | kwargs["pretrained"] = True 182 | elif kwargs["weights"] is None: 183 | kwargs["pretrained"] = False 184 | else: 185 | raise ValueError( 186 | "weights=={} not supported in torchvision {}".format( 187 | kwargs["weights"], torchvision.__version__ 188 | ) 189 | ) 190 | del kwargs["weights"] 191 | 192 | return torchvision.models.inception_v3(*args, **kwargs) 193 | 194 | 195 | def fid_inception_v3(): 196 | """Build pretrained Inception model for FID computation 197 | 198 | The Inception model for FID computation uses a different set of weights 199 | and has a slightly different structure than torchvision's Inception. 200 | 201 | This method first constructs torchvision's Inception and then patches the 202 | necessary parts that are different in the FID Inception model. 203 | """ 204 | inception = _inception_v3(num_classes=1008, aux_logits=False, weights=None) 205 | inception.Mixed_5b = FIDInceptionA(192, pool_features=32) 206 | inception.Mixed_5c = FIDInceptionA(256, pool_features=64) 207 | inception.Mixed_5d = FIDInceptionA(288, pool_features=64) 208 | inception.Mixed_6b = FIDInceptionC(768, channels_7x7=128) 209 | inception.Mixed_6c = FIDInceptionC(768, channels_7x7=160) 210 | inception.Mixed_6d = FIDInceptionC(768, channels_7x7=160) 211 | inception.Mixed_6e = FIDInceptionC(768, channels_7x7=192) 212 | inception.Mixed_7b = FIDInceptionE_1(1280) 213 | inception.Mixed_7c = FIDInceptionE_2(2048) 214 | 215 | state_dict = load_state_dict_from_url(FID_WEIGHTS_URL, progress=True) 216 | inception.load_state_dict(state_dict) 217 | return inception 218 | 219 | 220 | class FIDInceptionA(torchvision.models.inception.InceptionA): 221 | """InceptionA block patched for FID computation""" 222 | 223 | def __init__(self, in_channels, pool_features): 224 | super(FIDInceptionA, self).__init__(in_channels, pool_features) 225 | 226 | def forward(self, x): 227 | branch1x1 = self.branch1x1(x) 228 | 229 | branch5x5 = self.branch5x5_1(x) 230 | branch5x5 = self.branch5x5_2(branch5x5) 231 | 232 | branch3x3dbl = self.branch3x3dbl_1(x) 233 | branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) 234 | branch3x3dbl = self.branch3x3dbl_3(branch3x3dbl) 235 | 236 | # Patch: Tensorflow's average pool does not use the padded zero's in 237 | # its average calculation 238 | branch_pool = F.avg_pool2d( 239 | x, kernel_size=3, stride=1, padding=1, count_include_pad=False 240 | ) 241 | branch_pool = self.branch_pool(branch_pool) 242 | 243 | outputs = [branch1x1, branch5x5, branch3x3dbl, branch_pool] 244 | return torch.cat(outputs, 1) 245 | 246 | 247 | class FIDInceptionC(torchvision.models.inception.InceptionC): 248 | """InceptionC block patched for FID computation""" 249 | 250 | def __init__(self, in_channels, channels_7x7): 251 | super(FIDInceptionC, self).__init__(in_channels, channels_7x7) 252 | 253 | def forward(self, x): 254 | branch1x1 = self.branch1x1(x) 255 | 256 | branch7x7 = self.branch7x7_1(x) 257 | branch7x7 = self.branch7x7_2(branch7x7) 258 | branch7x7 = self.branch7x7_3(branch7x7) 259 | 260 | branch7x7dbl = self.branch7x7dbl_1(x) 261 | branch7x7dbl = self.branch7x7dbl_2(branch7x7dbl) 262 | branch7x7dbl = self.branch7x7dbl_3(branch7x7dbl) 263 | branch7x7dbl = self.branch7x7dbl_4(branch7x7dbl) 264 | branch7x7dbl = self.branch7x7dbl_5(branch7x7dbl) 265 | 266 | # Patch: Tensorflow's average pool does not use the padded zero's in 267 | # its average calculation 268 | branch_pool = F.avg_pool2d( 269 | x, kernel_size=3, stride=1, padding=1, count_include_pad=False 270 | ) 271 | branch_pool = self.branch_pool(branch_pool) 272 | 273 | outputs = [branch1x1, branch7x7, branch7x7dbl, branch_pool] 274 | return torch.cat(outputs, 1) 275 | 276 | 277 | class FIDInceptionE_1(torchvision.models.inception.InceptionE): 278 | """First InceptionE block patched for FID computation""" 279 | 280 | def __init__(self, in_channels): 281 | super(FIDInceptionE_1, self).__init__(in_channels) 282 | 283 | def forward(self, x): 284 | branch1x1 = self.branch1x1(x) 285 | 286 | branch3x3 = self.branch3x3_1(x) 287 | branch3x3 = [ 288 | self.branch3x3_2a(branch3x3), 289 | self.branch3x3_2b(branch3x3), 290 | ] 291 | branch3x3 = torch.cat(branch3x3, 1) 292 | 293 | branch3x3dbl = self.branch3x3dbl_1(x) 294 | branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) 295 | branch3x3dbl = [ 296 | self.branch3x3dbl_3a(branch3x3dbl), 297 | self.branch3x3dbl_3b(branch3x3dbl), 298 | ] 299 | branch3x3dbl = torch.cat(branch3x3dbl, 1) 300 | 301 | # Patch: Tensorflow's average pool does not use the padded zero's in 302 | # its average calculation 303 | branch_pool = F.avg_pool2d( 304 | x, kernel_size=3, stride=1, padding=1, count_include_pad=False 305 | ) 306 | branch_pool = self.branch_pool(branch_pool) 307 | 308 | outputs = [branch1x1, branch3x3, branch3x3dbl, branch_pool] 309 | return torch.cat(outputs, 1) 310 | 311 | 312 | class FIDInceptionE_2(torchvision.models.inception.InceptionE): 313 | """Second InceptionE block patched for FID computation""" 314 | 315 | def __init__(self, in_channels): 316 | super(FIDInceptionE_2, self).__init__(in_channels) 317 | 318 | def forward(self, x): 319 | branch1x1 = self.branch1x1(x) 320 | 321 | branch3x3 = self.branch3x3_1(x) 322 | branch3x3 = [ 323 | self.branch3x3_2a(branch3x3), 324 | self.branch3x3_2b(branch3x3), 325 | ] 326 | branch3x3 = torch.cat(branch3x3, 1) 327 | 328 | branch3x3dbl = self.branch3x3dbl_1(x) 329 | branch3x3dbl = self.branch3x3dbl_2(branch3x3dbl) 330 | branch3x3dbl = [ 331 | self.branch3x3dbl_3a(branch3x3dbl), 332 | self.branch3x3dbl_3b(branch3x3dbl), 333 | ] 334 | branch3x3dbl = torch.cat(branch3x3dbl, 1) 335 | 336 | # Patch: The FID Inception model uses max pooling instead of average 337 | # pooling. This is likely an error in this specific Inception 338 | # implementation, as other Inception models use average pooling here 339 | # (which matches the description in the paper). 340 | branch_pool = F.max_pool2d(x, kernel_size=3, stride=1, padding=1) 341 | branch_pool = self.branch_pool(branch_pool) 342 | 343 | outputs = [branch1x1, branch3x3, branch3x3dbl, branch_pool] 344 | return torch.cat(outputs, 1) 345 | -------------------------------------------------------------------------------- /tests/test_fid_score.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import pytest 3 | import torch 4 | from PIL import Image 5 | 6 | from pytorch_fid import fid_score, inception 7 | 8 | 9 | @pytest.fixture 10 | def device(): 11 | return torch.device("cpu") 12 | 13 | 14 | def test_calculate_fid_given_statistics(mocker, tmp_path, device): 15 | dim = 2048 16 | m1, m2 = np.zeros((dim,)), np.ones((dim,)) 17 | sigma = np.eye(dim) 18 | 19 | def dummy_statistics(path, model, batch_size, dims, device, num_workers): 20 | if path.endswith("1"): 21 | return m1, sigma 22 | elif path.endswith("2"): 23 | return m2, sigma 24 | else: 25 | raise ValueError 26 | 27 | mocker.patch( 28 | "pytorch_fid.fid_score.compute_statistics_of_path", side_effect=dummy_statistics 29 | ) 30 | 31 | dir_names = ["1", "2"] 32 | paths = [] 33 | for name in dir_names: 34 | path = tmp_path / name 35 | path.mkdir() 36 | paths.append(str(path)) 37 | 38 | fid_value = fid_score.calculate_fid_given_paths( 39 | paths, batch_size=dim, device=device, dims=dim, num_workers=0 40 | ) 41 | 42 | # Given equal covariance, FID is just the squared norm of difference 43 | assert fid_value == np.sum((m1 - m2) ** 2) 44 | 45 | 46 | def test_compute_statistics_of_path(mocker, tmp_path, device): 47 | model = mocker.MagicMock(inception.InceptionV3)() 48 | model.side_effect = lambda inp: [inp.mean(dim=(2, 3), keepdim=True)] 49 | 50 | size = (4, 4, 3) 51 | arrays = [np.zeros(size), np.ones(size) * 0.5, np.ones(size)] 52 | images = [(arr * 255).astype(np.uint8) for arr in arrays] 53 | 54 | paths = [] 55 | for idx, image in enumerate(images): 56 | paths.append(str(tmp_path / "{}.png".format(idx))) 57 | Image.fromarray(image, mode="RGB").save(paths[-1]) 58 | 59 | stats = fid_score.compute_statistics_of_path( 60 | str(tmp_path), 61 | model, 62 | batch_size=len(images), 63 | dims=3, 64 | device=device, 65 | num_workers=0, 66 | ) 67 | 68 | assert np.allclose(stats[0], np.ones((3,)) * 0.5, atol=1e-3) 69 | assert np.allclose(stats[1], np.ones((3, 3)) * 0.25) 70 | 71 | 72 | def test_compute_statistics_of_path_from_file(mocker, tmp_path, device): 73 | model = mocker.MagicMock(inception.InceptionV3)() 74 | 75 | mu = np.random.randn(5) 76 | sigma = np.random.randn(5, 5) 77 | 78 | path = tmp_path / "stats.npz" 79 | with path.open("wb") as f: 80 | np.savez(f, mu=mu, sigma=sigma) 81 | 82 | stats = fid_score.compute_statistics_of_path( 83 | str(path), model, batch_size=1, dims=5, device=device, num_workers=0 84 | ) 85 | 86 | assert np.allclose(stats[0], mu) 87 | assert np.allclose(stats[1], sigma) 88 | 89 | 90 | def test_image_types(tmp_path): 91 | in_arr = np.ones((24, 24, 3), dtype=np.uint8) * 255 92 | in_image = Image.fromarray(in_arr, mode="RGB") 93 | 94 | paths = [] 95 | for ext in fid_score.IMAGE_EXTENSIONS: 96 | paths.append(str(tmp_path / "img.{}".format(ext))) 97 | in_image.save(paths[-1]) 98 | 99 | dataset = fid_score.ImagePathDataset(paths) 100 | 101 | for img in dataset: 102 | assert np.allclose(np.array(img), in_arr) 103 | --------------------------------------------------------------------------------