Accepted in CVPR 2020
Python >= 3.5
PyTorch >= 1.0.0
Opencv2 ==3.4.1
Scipy == 1.1.0
Numpy == 1.14.3
Scikit-image (skimage) == 0.13.1
This is the environment for our experiments. Later versions of these packages might need a few modifications of the code.
Although our method is not limited to any specific CUDA and cudnn version, it's strongly encouraged that you use the latest version of these toolkits. It seems that the RFR-Net could run slowly in older CUDA version due to its recurrent design.
The link to the pretrained model. (Currently, Paris StreetView, CelebA datasets). We are expecting to release the Places2 weights before the end of January, we are sorry for the delay caused by the failure in our storage system.
https://drive.google.com/drive/folders/1EbRSL6SlJqeMliT9qU8V5g0idJqvirZr?usp=sharing
We strongly encourage the users to retrain the models if they are used for academic purpose, to ensure fair comparisons (which has been always desired). Achieving a good performance using the current version of code should not be difficult.
We've checked the reproducibilities of the results in the paper.
Reproducible | |
---|---|
Paris StreetView | True |
CelebA | True |
To perform training or testing, use
python run.py
There are several arguments that can be used, which are
--data_root +str #where to get the images for training/testing
--mask_root +str #where to get the masks for training/testing
--model_save_path +str #where to save the model during training
--result_save_path +str #where to save the inpainting results during testing
--model_path +str #the pretrained generator to use during training/testing
--target_size +int #the size of images and masks
--mask_mode +int #which kind of mask to be used, 0 for external masks with random order, 1 for randomly generated masks, 2 for external masks with fixed order
--batch_size +int #the size of mini-batch for training
--n_threads +int
--gpu_id +int #which gpu to use
--finetune #to finetune the model during training
--test #test the model
For example, to train the network using gpu 1, with pretrained models
python run.py --data_root data --mask_root mask --model_path checkpoints/g_10000.pth --batch_size 6 --gpu 1
to test the network
python run.py --data_root data/images --mask_root data/masks --model_path checkpoints/g_10000.pth --test --mask_mode 2
The RFR-Net for filling smaller holes is added. The only difference is the smaller number of pixels fixed in each iteration. If you are fixing small holes, you can use that version of code, to gain some speep-up.
To fully exploit the performance of the network, we suggest to use the following training procedure, in specific
- Train the network, i.e. use the command
python run.py
- Finetune the network, i.e. use the command
python run.py --finetune --model_path path-to-trained-generator
- Test the model
python run.py --test
All the descriptions below are under the assumption that the size of mini-batch is 6
For Paris Street View Dataset, train the model for 400,000 iterations and finetune for 200,000 iterations. (600,000 in total)
For CelebA Dataset, train the model for 350,000 iterations and finetune for 150,000 iterations. (500,000 in total)
For Places2 Challenge Dataset, train the model for 2,000,000 iterations and finetune for 1,000,000 iterations. (3,000,000 in total)
This part is for people who want to build their own methods based on this code.
The core of this code is the model.py
file. In specific, it defines the organization of the model, training procedures, loss functions and the parameter updating procedure.
Before we start training/testing, the model and its components are initialized by initialize_model(self, path=None, train=True)
method which builds a randomly initialized model and tries to load the pretrained parameters. The pipeline of the initialized model is provided in modules
(The RFR-Net in our case).
After the model is initialized, the method cuda(self, path=None, train=True)
is called, which moves the model to the gpu given there exists avaliable cuda devices.
When training the network, train(self, train_loader, save_path, finetune = False)
is called. This method requires an external dataloader that provides images and masks and the path to save the model. Given the dataloader correctly produces training data, the forward and backward propagation procedures are alternatively performed by calling forward(self, masked_image, mask, gt_image)
and update_parameters(self)
. The forward
method simply feeds the data to the generator network and saves the output results. The update_parameters(self)
updates the generator and discriminator separately (in our case, the discriminator doesn't exist). When updating the generator and discriminator, we calculate the loss functions and update the parameters.
After training, we can test the data. At this time, a dataloader that provides the test data are required and the path where you want to save the generated results should also be given.
To modify the method or build your own method based on this code, you can do this by changing the RFRNet.py
and model.py
files.
Some examples are given below:
To change the training targets for generator, you can modify the get_g_loss
method in model.py.
To change the architecture of the model, you might want to modify the RFRNet.py
file.
To add a discriminator for the RFR-Net, you need to 1.define the discirminator and its optimizer in initialize_model
and cuda
methods and 2.define the new loss functions for the discriminator and generator and 3. define parameter updating procedure in update_D
method.
This code will be improved constantly. More functions for visualization are still to be developed.
If you find the article or code useful for your project, please refer to
@InProceedings{Li_2020_CVPR,
author = {Li, Jingyuan and Wang, Ning and Zhang, Lefei and Du, Bo and Tao, Dacheng},
title = {Recurrent Feature Reasoning for Image Inpainting},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}
See the Paper folder