About context-encoder

Context Encoders is an open-source implementation of the CVPR 2016 paper Unsupervised Feature Learning by Image Inpainting using GANs. This software trains deep neural networks to perform image inpainting, the task of filling in missing pixels within an image, to learn robust unsupervised visual features. The framework jointly optimizes an encoder-decoder architecture using both reconstruction loss and adversarial loss from a Generative Adversarial Network. It supports two primary training modes: center region inpainting for standard conditioning and arbitrary random region inpainting for more flexible scenarios. The distribution includes training scripts, testing utilities, and pre-trained models for inference on various datasets such as Paris Street View, ImageNet, and UC Berkeley cityscapes. Built initially on the Torch deep learning framework based on Soumith Chintala's DCGAN implementation, the codebase allows researchers to reproduce state-of-the-art semantic inpainting results, experiment with differen

p

Published by

pathak22

Visit View Profile

README.md

View on GitHub

Context Encoders: Feature Learning by Inpainting

CVPR 2016

[Project Website] [Imagenet Results]

Sample results on held-out images:

teaser

This is the training code for our CVPR 2016 paper on Context Encoders for learning deep feature representation in an unsupervised manner by image inpainting. Context Encoders are trained jointly with reconstruction and adversarial loss. This repo contains quick demo, training/testing code for center region inpainting and training/testing code for arbitray random region inpainting. This code is adapted from an initial fork of Soumith's DCGAN implementation. Scroll down to try out a quick demo or train your own inpainting models!

If you find Context Encoders useful in your research, please cite:

@inproceedings{pathakCVPR16context,
    Author = {Pathak, Deepak and Kr\"ahenb\"uhl, Philipp and Donahue, Jeff and Darrell, Trevor and Efros, Alexei},
    Title = {Context Encoders: Feature Learning by Inpainting},
    Booktitle = {Computer Vision and Pattern Recognition ({CVPR})},
    Year = {2016}
}

1) Semantic Inpainting Demo

Install Torch: http://torch.ch/docs/getting-started.html#_

Clone the repository

git clone https://github.com/pathak22/context-encoder.git

Demo


cd context-encoder
bash ./models/scripts/download_inpaintCenter_models.sh
# This will populate the `./models/` folder with trained models.

net=models/inpaintCenter/paris_inpaintCenter.t7 name=paris_result imDir=images/paris overlapPred=4 manualSeed=222 batchSize=21 gpu=1 th demo.lua net=models/inpaintCenter/imagenet_inpaintCenter.t7 name=imagenet_result imDir=images/imagenet overlapPred=4 manualSeed=222 batchSize=21 gpu=1 th demo.lua net=models/inpaintCenter/paris_inpaintCenter.t7 name=ucberkeley_result imDir=images/ucberkeley overlapPred=4 manualSeed=222 batchSize=4 gpu=1 th demo.lua

Note: If you are running on cpu, use gpu=0

Note: samples given in ./images/* are held-out images


### 2) Train Context Encoders

If you could successfully run the above demo, run following steps to train your own context encoder model for image inpainting.

0. [Optional] Install Display Package as follows. If you don't want to install it, then set `display=0` in `train.lua`.
  ```Shell
  luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec
  cd ~
  th -ldisplay.start 8000
  # if working on server machine create tunnel: ssh -f -L 8000:localhost:8000 -N server_address.com
  # on client side, open in browser: http://localhost:8000/

Make the dataset folder.

mkdir -p /path_to_wherever_you_want/mydataset/train/images/
# put all training images inside mydataset/train/images/
mkdir -p /path_to_wherever_you_want/mydataset/val/images/
# put all val images inside mydataset/val/images/
cd context-encoder/
ln -sf /path_to_wherever_you_want/mydataset dataset

Train the model


# For training center region inpainting model, run:
DATA_ROOT=dataset/train display_id=11 name=inpaintCenter overlapPred=4 wtl2=0.999 nBottleneck=4000 niter=500 loadSize=350 fineSize=128 gpu=1 th train.lua

For training random region inpainting model, run:

DATA_ROOT=dataset/train display_id=11 name=inpaintRandomNoOverlap useOverlapPred=0 wtl2=0.999 nBottleneck=4000 niter=500 loadSize=350 fineSize=128 gpu=1 th train_random.lua

or use fineSize=64 to train to generate 64x64 sized image (results are better):

DATA_ROOT=dataset/train display_id=11 name=inpaintRandomNoOverlap useOverlapPred=0 wtl2=0.999 nBottleneck=4000 niter=500 loadSize=350 fineSize=64 gpu=1 th train_random.lua


3. Test the model
  ```Shell
  # For training center region inpainting model, run:
  DATA_ROOT=dataset/val net=checkpoints/inpaintCenter_500_net_G.t7 name=test_patch overlapPred=4 manualSeed=222 batchSize=30 loadSize=350 gpu=1 th test.lua
  DATA_ROOT=dataset/val net=checkpoints/inpaintCenter_500_net_G.t7 name=test_full overlapPred=4 manualSeed=222 batchSize=30 loadSize=129 gpu=1 th test.lua

  # For testing random region inpainting model, run (with fineSize=64 or 124, same as training):
  DATA_ROOT=dataset/val net=checkpoints/inpaintRandomNoOverlap_500_net_G.t7 name=test_patch_random useOverlapPred=0 manualSeed=222 batchSize=30 loadSize=350 gpu=1 th test_random.lua
  DATA_ROOT=dataset/val net=checkpoints/inpaintRandomNoOverlap_500_net_G.t7 name=test_full_random useOverlapPred=0 manualSeed=222 batchSize=30 loadSize=129 gpu=1 th test_random.lua

3) Download Features Caffemodel

Features for context encoder trained with reconstruction loss.

4) TensorFlow Implementation

Checkout the TensorFlow implementation of our paper by Taeksoo here. However, it does not implement full functionalities of our paper.

5) Project Website

Click here.

6) Paris Street-View Dataset

Please email me if you need the dataset and I will share a private link with you. I can't post the public link to this dataset due to the policy restrictions from Google Street View.

context-encoder