Project | Arxiv | PyTorch

Torch implementation for learning a mapping from input images to output images, for example:

Image-to-Image Translation with Conditional Adversarial Networks
Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros
CVPR, 2017.

On some tasks, decent results can be obtained fairly quickly and on small datasets. For example, to learn to generate facades (example shown above), we trained on just 400 images for about 2 hours (on a single Pascal Titan X GPU). However, for harder problems it may be important to train on far larger datasets, and for many hours or even days.

Note: Please check out our PyTorch implementation for pix2pix and CycleGAN. The PyTorch version is under active development and can produce results comparable to or better than this Torch version.

• Linux or OSX
• NVIDIA GPU + CUDA CuDNN (CPU mode and CUDA without CuDNN may work with minimal modification, but untested)

• Install torch and dependencies from https://github.com/torch/distro
• Install torch packages nngraph and display

luarocks install nngraph
luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec

• Clone this repo:

git clone git@github.com:phillipi/pix2pix.git
cd pix2pix

• Download the dataset (e.g., CMP Facades):

bash ./datasets/download_dataset.sh facades

• Train the model

DATA_ROOT=./datasets/facades name=facades_generation which_direction=BtoA th train.lua

• (CPU only) The same training command without using a GPU or CUDNN. Setting the environment variables gpu=0 cudnn=0 forces CPU only

DATA_ROOT=./datasets/facades name=facades_generation which_direction=BtoA gpu=0 cudnn=0 batchSize=10 save_epoch_freq=5 th train.lua

• (Optionally) start the display server to view results as the model trains. ( See Display UI for more details):

th -ldisplay.start 8000 0.0.0.0

• Finally, test the model:

DATA_ROOT=./datasets/facades name=facades_generation which_direction=BtoA phase=val th test.lua

The test results will be saved to an html file here: ./results/facades_generation/latest_net_G_val/index.html.

DATA_ROOT=/path/to/data/ name=expt_name which_direction=AtoB th train.lua

Switch AtoB to BtoA to train translation in opposite direction.

Models are saved to ./checkpoints/expt_name (can be changed by passing checkpoint_dir=your_dir in train.lua).

See opt in train.lua for additional training options.

DATA_ROOT=/path/to/data/ name=expt_name which_direction=AtoB phase=val th test.lua

This will run the model named expt_name in direction AtoB on all images in /path/to/data/val.

Result images, and a webpage to view them, are saved to ./results/expt_name (can be changed by passing results_dir=your_dir in test.lua).

See opt in test.lua for additional testing options.

Download the datasets using the following script. Some of the datasets are collected by other researchers. Please cite their papers if you use the data.

bash ./datasets/download_dataset.sh dataset_name

• facades: 400 images from CMP Facades dataset. [Citation]
• cityscapes: 2975 images from the Cityscapes training set. [Citation]
• maps: 1096 training images scraped from Google Maps
• edges2shoes: 50k training images from UT Zappos50K dataset. Edges are computed by HED edge detector + post-processing. [Citation]
• edges2handbags: 137K Amazon Handbag images from iGAN project. Edges are computed by HED edge detector + post-processing. [Citation]
• night2day: around 20K natural scene images from Transient Attributes dataset [Citation]. To train a day2night pix2pix model, you need to add which_direction=BtoA.

Download the pre-trained models with the following script. You need to rename the model (e.g., facades_label2image to /checkpoints/facades/latest_net_G.t7) after the download has finished.

bash ./models/download_model.sh model_name

• facades_label2image (label -> facade): trained on the CMP Facades dataset.
• cityscapes_label2image (label -> street scene): trained on the Cityscapes dataset.
• cityscapes_image2label (street scene -> label): trained on the Cityscapes dataset.
• edges2shoes (edge -> photo): trained on UT Zappos50K dataset.
• edges2handbags (edge -> photo): trained on Amazon handbags images.
• day2night (daytime scene -> nighttime scene): trained on around 100 webcams.

We provide a python script to generate training data in the form of pairs of images {A,B}, where A and B are two different depictions of the same underlying scene. For example, these might be pairs {label map, photo} or {bw image, color image}. Then we can learn to translate A to B or B to A:

Create folder /path/to/data with subfolders A and B. A and B should each have their own subfolders train, val, test, etc. In /path/to/data/A/train, put training images in style A. In /path/to/data/B/train, put the corresponding images in style B. Repeat same for other data splits (val, test, etc).

Corresponding images in a pair {A,B} must be the same size and have the same filename, e.g., /path/to/data/A/train/1.jpg is considered to correspond to /path/to/data/B/train/1.jpg.

Once the data is formatted this way, call:

python scripts/combine_A_and_B.py --fold_A /path/to/data/A --fold_B /path/to/data/B --fold_AB /path/to/data

This will combine each pair of images (A,B) into a single image file, ready for training.

No need to run combine_A_and_B.py for colorization. Instead, you need to prepare some natural images and set preprocess=colorization in the script. The program will automatically convert each RGB image into Lab color space, and create L -> ab image pair during the training. Also set input_nc=1 and output_nc=2.

We provide python and Matlab scripts to extract coarse edges from photos. Run scripts/edges/batch_hed.py to compute HED edges. Run scripts/edges/PostprocessHED.m to simplify edges with additional post-processing steps. Check the code documentation for more details.

We provide scripts for running the evaluation of the Labels2Photos task on the Cityscapes validation set. We assume that you have installed caffe (and pycaffe) in your system. If not, see the official website for installation instructions. Once caffe is successfully installed, download the pre-trained FCN-8s semantic segmentation model (512MB) by running

bash ./scripts/eval_cityscapes/download_fcn8s.sh

Then make sure ./scripts/eval_cityscapes/ is in your system's python path. If not, run the following command to add it

export PYTHONPATH=${PYTHONPATH}:./scripts/eval_cityscapes/

Now you can run the following command to evaluate your predictions:

python ./scripts/eval_cityscapes/evaluate.py --cityscapes_dir /path/to/original/cityscapes/dataset/ --result_dir /path/to/your/predictions/ --output_dir /path/to/output/directory/

Images stored under --result_dir should contain your model predictions on the Cityscapes validation split, and have the original Cityscapes naming convention (e.g., frankfurt_000001_038418_leftImg8bit.png). The script will output a text file under --output_dir containing the metric.

Further notes: Our pre-trained FCN model is not supposed to work on Cityscapes in the original resolution (1024x2048) as it was trained on 256x256 images that are then upsampled to 1024x2048 during training. The purpose of the resizing during training was to 1) keep the label maps in the original high resolution untouched and 2) avoid the need to change the standard FCN training code and the architecture for Cityscapes. During test time, you need to synthesize 256x256 results. Our test code will automatically upsample your results to 1024x2048 before feeding them to the pre-trained FCN model. The output is at 1024x2048 resolution and will be compared to 1024x2048 ground truth labels. You do not need to resize the ground truth labels. The best way to verify whether everything is correct is to reproduce the numbers for real images in the paper first. To achieve it, you need to resize the original/real Cityscapes images (not labels) to 256x256 and feed them to the evaluation code.

Optionally, for displaying images during training and test, use the display package.

• Install it with: luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec
• Then start the server with: th -ldisplay.start
• Open this URL in your browser: http://localhost:8000

By default, the server listens on localhost. Pass 0.0.0.0 to allow external connections on any interface:

th -ldisplay.start 8000 0.0.0.0

Then open http://(hostname):(port)/ in your browser to load the remote desktop.

L1 error is plotted to the display by default. Set the environment variable display_plot to a comma-separated list of values errL1, errG and errD to visualize the L1, generator, and discriminator error respectively. For example, to plot only the generator and discriminator errors to the display instead of the default L1 error, set display_plot="errG,errD".

If you use this code for your research, please cite our paper Image-to-Image Translation Using Conditional Adversarial Networks:

@article{pix2pix2017,
  title={Image-to-Image Translation with Conditional Adversarial Networks},
  author={Isola, Phillip and Zhu, Jun-Yan and Zhou, Tinghui and Efros, Alexei A},
  journal={CVPR},
  year={2017}
}

If you love cats, and love reading cool graphics, vision, and learning papers, please check out the Cat Paper Collection:
[Github] [Webpage]

Code borrows heavily from DCGAN. The data loader is modified from DCGAN and Context-Encoder.