This repository contains the (1) Learned Perceptual Image Patch Similarity (LPIPS) metric and (2) Berkeley-Adobe Perceptual Patch Similarity (BAPPS) dataset. It can also be used as an implementation of the "perceptual loss". This is a PyTorch implementation; a Tensorflow alternative is [here].

The Unreasonable Effectiveness of Deep Features as a Perceptual Metric
Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, Oliver Wang. In CVPR, 2018.

• Install PyTorch 0.4+ and torchvision fom http://pytorch.org

pip install -r requirements.txt

• Clone this repo:

git clone https://github.com/richzhang/PerceptualSimilarity
cd PerceptualSimilarity

Using this code, you can simply call model.forward(im0,im1) to evaluate the distance between two image patches.

Computing the distance between two images: python compute_dists.py --path0 imgs/ex_ref.png --path1 imgs/ex_p0.png --use_gpu

Computing the distance between pairs of images within two directories: python ./compute_dists_dirs.py --dir0 ./imgs/ex_dir0 --dir1 ./imgs/ex_dir1 --out ./imgs/example_dists.txt --use_gpu

Script test_network.py contains example usage. Run python test_network.py to take the distance between example reference image ex_ref.png to distorted images ex_p0.png and ex_p1.png. Before running it - which do you think should be closer? A more detailed explanation is below.

Load a model with the following commands.

from models import dist_model as dm
model = dm.DistModel()
model.initialize(model='net-lin',net='alex',use_gpu=True,version='0.1')

Variable net can be squeeze, alex, vgg. Network alex is fastest, performs the best, and is the default. Set variable model=net for an uncalibrated off-the-shelf network (taking cos distance).

Finally, to call the model, run

d = model.forward(im0,im1)

where im0, im1 are PyTorch tensors with shape Nx3xHxW (N patches of size HxW, RGB images scaled in [-1,+1]). Variable d will be a length N numpy array.

File perceptual_loss.py shows how to iteratively optimize using the metric. Run python perceptual_loss.py for a demo. The code can also be used to implement vanilla VGG loss, without our learned weights.

We found that deep network activations work surprisingly well as a perceptual similarity metric. This was true across network architectures (SqueezeNet [2.8 MB], AlexNet [9.1 MB], and VGG [58.9 MB] provided similar scores) and supervisory signals (unsupervised, self-supervised, and supervised all perform strongly). We slightly improved scores by linearly "calibrating" networks - adding a linear layer on top of off-the-shelf classification networks. We provide 3 variants, using linear layers on top of the SqueezeNet, AlexNet (default), and VGG networks.

If you use LPIPS in your publication, please specify which version you are using. The current version is 0.1. You can set version='0.0' for the initial release.

Run bash ./scripts/download_dataset.sh to download and unzip the dataset. Dataset will appear in directory ./dataset. Dataset takes [6.6 GB] total.

• 2AFC train [5.3 GB]
• 2AFC val [1.1 GB]
• JND val [0.2 GB]
  Alternatively, run bash ./scripts/get_dataset_valonly.sh to only download the validation set (no training set).

Script test_dataset_model.py evaluates a perceptual model on a subset of the dataset.

Dataset flags

• dataset_mode: 2afc or jnd, which type of perceptual judgment to evaluate
• datasets: list the datasets to evaluate
  • if dataset_mode was 2afc, choices are [train/traditional, train/cnn, val/traditional, val/cnn, val/superres, val/deblur, val/color, val/frameinterp]
  • if dataset_mode was jnd, choices are [val/traditional, val/cnn]

Perceptual similarity model flags

• model: perceptual similarity model to use
  • net-lin for our LPIPS learned similarity model (linear network on top of internal activations of pretrained network)
  • net for a classification network (uncalibrated with all layers averaged)
  • l2 for Euclidean distance
  • ssim for Structured Similarity Image Metric
• net: choices are [squeeze,alex,vgg] for the net-lin and net models (ignored for l2 and ssim models)
• colorspace: choices are [Lab,RGB], used for the l2 and ssim models (ignored for net-lin and net models)

Misc flags

• batch_size: evaluation batch size (will default to 1 )
• --use_gpu: turn on this flag for GPU usage

An example usage is as follows: python ./test_dataset_model.py --dataset_mode 2afc --datasets val/traditional val/cnn --model net-lin --net alex --use_gpu --batch_size 50. This would evaluate our model on the "traditional" and "cnn" validation datasets.

The dataset contains two types of perceptual judgements: Two Alternative Forced Choice (2AFC) and Just Noticeable Differences (JND).

(1) Two Alternative Forced Choice (2AFC) - Data is contained in the 2afc subdirectory. Evaluators were given a reference patch, along with two distorted patches, and were asked to select which of the distorted patches was "closer" to the reference patch.

Training sets contain 2 human judgments/triplet.

• train/traditional [56.6k triplets]
• train/cnn [38.1k triplets]
• train/mix [56.6k triplets]

Validation sets contain 5 judgments/triplet.

• val/traditional [4.7k triplets]
• val/cnn [4.7k triplets]
• val/superres [10.9k triplets]
• val/deblur [9.4k triplets]
• val/color [4.7k triplets]
• val/frameinterp [1.9k triplets]

Each 2AFC subdirectory contains the following folders:

• ref contains the original reference patches
• p0,p1 contain the two distorted patches
• judge contains what the human evaluators chose - 0 if all humans preferred p0, 1 if all humans preferred p1

(2) Just Noticeable Differences (JND) - Data is contained in the jnd subdirectory. Evaluators were presented with two patches - a reference patch and a distorted patch - for a limited time, and were asked if they thought the patches were the same (identically) or difference.

Each set contains 3 human evaluations/example.

• val/traditional [4.8k patch pairs]
• val/cnn [4.8k patch pairs]

Each JND subdirectory contains the following folders:

• p0,p1 contain the two patches
• same contains fraction of human evaluators who thought the patches were the same (0 if all humans thought patches were different, 1 if all humans thought patches were the same)

See script train_test_metric.sh for an example of training and testing the metric. The script will train a model on the full training set for 10 epochs, and then test the learned metric on all of the validation sets. The numbers should roughly match the Alex - lin row in Table 5 in the paper. The code supports training a linear layer on top of an existing representation. Training will add a subdirectory in the checkpoints directory.

You can also train "scratch" and "tune" versions by running train_test_metric_scratch.sh and train_test_metric_tune.sh, respectively.

Docker set up by SuperShinyEyes.

If you find this repository useful for your research, please use the following.

@inproceedings{zhang2018perceptual,
  title={The Unreasonable Effectiveness of Deep Features as a Perceptual Metric},
  author={Zhang, Richard and Isola, Phillip and Efros, Alexei A and Shechtman, Eli and Wang, Oliver},
  booktitle={CVPR},
  year={2018}
}

This repository borrows partially from the pytorch-CycleGAN-and-pix2pix repository. The average precision (AP) code is borrowed from the py-faster-rcnn repository. Backpropping through the metric was implemented by Angjoo Kanazawa.