{"author":"Gongfan Fang","author_email":"gongfan@u.nus.edu","bugtrack_url":null,"classifiers":["License :: OSI Approved :: MIT License","Operating System :: OS Independent","Programming Language :: Python :: 3"],"description":"# Pytorch MS-SSIM\n\n[![License](https://img.shields.io/badge/License-MIT-blue.svg)](https://opensource.org/licenses/MIT)\n[![Downloads](https://pepy.tech/badge/pytorch_msssim)](https://pepy.tech/project/pytorch_msssim)\n[![Latest Version](https://img.shields.io/badge/Latest%20Version-0.2.1-blue.svg)](https://github.com/VainF/pytorch_msssim/releases/latest)\n\nFast and differentiable MS-SSIM and SSIM for pytorch.\n\n

\n\n#### Structural Similarity (SSIM): \n

\n\n#### Multi-Scale Structural Similarity (MS-SSIM): \n

\n\n#### Why it is faster than other versions?\n\nGaussian kernels used in SSIM & MS-SSIM are seperable. A [separable filter](https://en.wikipedia.org/wiki/Separable_filter) in image processing can be written as product of two more simple filters. Typically a 2-dimensional convolution operation is separated into two 1-dimensional filters. This reduces the computational costs on an $N\\times M$ image with a $m\\times n$ filter from $\\mathcal{O}(M\\cdot N \\cdot m \\cdot n)$ down to $\\mathcal{O}(M\\cdot N \\cdot (m+n))$. More importantly, seperated kernels are more contiguous and thus cache-friendly than 2-D kernels, which effectively accelerates the computing of SSIM/MS-SSIM. \n\n# Update\n#### _2020.08.21_ (v0.2.1)\n\n3D image support from [@FynnBe](https://github.com/FynnBe)! \n\n#### _2020.04.30_ (v0.2)\n\nNow (v0.2), **ssim & ms-ssim can produce consistent results as tensorflow and skimage**. A benchmark (pytorch-msssim, tensorflow and skimage) can be found in the Tests section.\n\n# Installation\n\n```bash\npip install pytorch-msssim\n```\n\n# Usage\n\n### 1. Basic Usage \n\n```python\nfrom pytorch_msssim import ssim, ms_ssim, SSIM, MS_SSIM\n# X: (N,3,H,W) a batch of non-negative RGB images (0~255)\n# Y: (N,3,H,W) \n\n# calculate ssim & ms-ssim for each image\nssim_val = ssim( X, Y, data_range=255, size_average=False) # return (N,)\nms_ssim_val = ms_ssim( X, Y, data_range=255, size_average=False ) #(N,)\n\n# set 'size_average=True' to get a scalar value as loss. see tests/tests_loss.py for more details\nssim_loss = 1 - ssim( X, Y, data_range=255, size_average=True) # return a scalar\nms_ssim_loss = 1 - ms_ssim( X, Y, data_range=255, size_average=True )\n\n# reuse the gaussian kernel with SSIM & MS_SSIM. \nssim_module = SSIM(data_range=255, size_average=True, channel=3) # channel=1 for grayscale images\nms_ssim_module = MS_SSIM(data_range=255, size_average=True, channel=3)\n\nssim_loss = 1 - ssim_module(X, Y)\nms_ssim_loss = 1 - ms_ssim_module(X, Y)\n```\n### 2. Normalized input\nIf you need to calculate MS-SSIM/SSIM on normalized images, please denormalize them to the range of [0, 1] or [0, 255] first.\n\n```python\n# X: (N,3,H,W) a batch of normalized images (-1 ~ 1)\n# Y: (N,3,H,W) \nX = (X + 1) / 2 # [-1, 1] => [0, 1]\nY = (Y + 1) / 2 \nms_ssim_val = ms_ssim( X, Y, data_range=1, size_average=False ) #(N,)\n```\n\n### 3. Enable nonnegative_ssim\n\nFor ssim, it is recommended to set `nonnegative_ssim=True` to avoid negative results. However, this option is set to `False` by default to keep it consistent with tensorflow and skimage.\n\nFor ms-ssim, there is no nonnegative_ssim option and the ssim reponses is forced to be non-negative to avoid NaN results.\n\n\n# Tests and Examples\n\n```bash\ncd tests\n```\n### 1. Benchmark\n\n```bash\n# requires tf2\npython tests_comparisons_tf_skimage.py \n\n# or skimage only\n# python tests_comparisons_skimage.py \n```\n\nOutputs:\n\n```\nDownloading test image...\n===================================\n Test SSIM\n===================================\n====> Single Image\nRepeat 100 times\nsigma=0.0 ssim_skimage=1.000000 (147.2605 ms), ssim_tf=1.000000 (343.4146 ms), ssim_torch=1.000000 (92.9151 ms)\nsigma=10.0 ssim_skimage=0.932423 (147.5198 ms), ssim_tf=0.932661 (343.5191 ms), ssim_torch=0.932421 (95.6283 ms)\nsigma=20.0 ssim_skimage=0.785744 (152.6441 ms), ssim_tf=0.785733 (343.4085 ms), ssim_torch=0.785738 (87.5639 ms)\nsigma=30.0 ssim_skimage=0.636902 (145.5763 ms), ssim_tf=0.636902 (343.5312 ms), ssim_torch=0.636895 (90.4084 ms)\nsigma=40.0 ssim_skimage=0.515798 (147.3798 ms), ssim_tf=0.515801 (344.8978 ms), ssim_torch=0.515791 (96.4440 ms)\nsigma=50.0 ssim_skimage=0.422011 (148.2900 ms), ssim_tf=0.422007 (345.4076 ms), ssim_torch=0.422005 (86.3799 ms)\nsigma=60.0 ssim_skimage=0.351139 (146.2039 ms), ssim_tf=0.351139 (343.4428 ms), ssim_torch=0.351133 (93.3445 ms)\nsigma=70.0 ssim_skimage=0.296336 (145.5341 ms), ssim_tf=0.296337 (345.2255 ms), ssim_torch=0.296331 (92.6771 ms)\nsigma=80.0 ssim_skimage=0.253328 (147.6655 ms), ssim_tf=0.253328 (343.1386 ms), ssim_torch=0.253324 (82.5985 ms)\nsigma=90.0 ssim_skimage=0.219404 (142.6025 ms), ssim_tf=0.219405 (345.8275 ms), ssim_torch=0.219400 (100.9946 ms)\nsigma=100.0 ssim_skimage=0.192681 (144.5597 ms), ssim_tf=0.192682 (346.5489 ms), ssim_torch=0.192678 (85.0229 ms)\nPass!\n====> Batch\nPass!\n\n\n===================================\n Test MS-SSIM\n===================================\n====> Single Image\nRepeat 100 times\nsigma=0.0 msssim_tf=1.000000 (671.5363 ms), msssim_torch=1.000000 (125.1403 ms)\nsigma=10.0 msssim_tf=0.991137 (669.0296 ms), msssim_torch=0.991086 (113.4078 ms)\nsigma=20.0 msssim_tf=0.967292 (670.5530 ms), msssim_torch=0.967281 (107.6428 ms)\nsigma=30.0 msssim_tf=0.934875 (668.7717 ms), msssim_torch=0.934875 (111.3334 ms)\nsigma=40.0 msssim_tf=0.897660 (669.0801 ms), msssim_torch=0.897658 (107.3700 ms)\nsigma=50.0 msssim_tf=0.858956 (671.4629 ms), msssim_torch=0.858954 (100.9959 ms)\nsigma=60.0 msssim_tf=0.820477 (670.5424 ms), msssim_torch=0.820475 (103.4489 ms)\nsigma=70.0 msssim_tf=0.783511 (671.9357 ms), msssim_torch=0.783507 (113.9048 ms)\nsigma=80.0 msssim_tf=0.749522 (672.3925 ms), msssim_torch=0.749518 (120.3891 ms)\nsigma=90.0 msssim_tf=0.716221 (672.9066 ms), msssim_torch=0.716217 (118.3788 ms)\nsigma=100.0 msssim_tf=0.684958 (675.2075 ms), msssim_torch=0.684953 (117.9481 ms)\nPass\n====> Batch\nPass\n```\n\n

ssim=1.0000

ssim=0.4225

ssim=0.1924

\n\n### 2. MS_SSIM as loss function\n\nSee ['tests/tests_loss.py'](https://github.com/VainF/pytorch-msssim/tree/master/tests/tests_loss.py) for more details about how to use ssim or ms_ssim as loss functions\n\n### 3. AutoEncoder\n\nSee ['tests/ae_example'](https://github.com/VainF/pytorch-msssim/tree/master/tests/ae_example)\n\n![results](https://github.com/VainF/Images/blob/master/pytorch_msssim/ae_ms_ssim.jpg)\n*left: the original image, right: the reconstructed image*\n\n# References\n\n[https://github.com/jorge-pessoa/pytorch-msssim](https://github.com/jorge-pessoa/pytorch-msssim) \n[https://ece.uwaterloo.ca/~z70wang/research/ssim/](https://ece.uwaterloo.ca/~z70wang/research/ssim/) \n[https://ece.uwaterloo.ca/~z70wang/publications/msssim.pdf](https://ece.uwaterloo.ca/~z70wang/publications/msssim.pdf) \n[Matlab Code](https://ece.uwaterloo.ca/~z70wang/research/iwssim/) \n[ssim & ms-ssim from tensorflow](https://github.com/tensorflow/tensorflow/blob/v2.1.0/tensorflow/python/ops/image_ops_impl.py#L3314-L3438) \n","description_content_type":"text/markdown","docs_url":null,"download_url":"","downloads":{"last_day":-1,"last_month":-1,"last_week":-1},"dynamic":null,"home_page":"https://github.com/VainF/pytorch-msssim","keywords":"","license":"","license_expression":null,"license_files":null,"maintainer":"","maintainer_email":"","name":"pytorch-msssim","package_url":"https://pypi.org/project/pytorch-msssim/","platform":null,"project_url":"https://pypi.org/project/pytorch-msssim/","project_urls":{"Homepage":"https://github.com/VainF/pytorch-msssim"},"provides_extra":null,"release_url":"https://pypi.org/project/pytorch-msssim/1.0.0/","requires_dist":["torch"],"requires_python":"","summary":"Fast and differentiable MS-SSIM and SSIM for pytorch.","version":"1.0.0","yanked":false,"yanked_reason":null}

Package search

pytorch-msssim

project_urls

Dependencies

All package metadata

Releases