recong / Boundless-in-Pytorch

Boundless: Generative Adversarial Networks for Image Extension in Pytorch

Boundless: Generative Adversarial Networks for Image Extension in Pytorch

Unofficial pytorch implementation of Boundless: Generative Adversarial Networks for Image Extension. I used this code esrgan as reference.

Requirements

pytorch
torchvision
torchsummary
numpy
Pillow
random
glob

Prepare a dataset

1. Download a dataset wget http://data.csail.mit.edu/places/places365/train_256_places365standard.tar
2. Unpack a tar file tar -xvf train_256_places365standard.tar
3. Run the script using command python make_datasets.py

Train

1. Run the script using command python train.py

Test

1. Run the script using command python test.py

In this code, the input size is 512 x 512(in the original paper, 257 x 257). Due to this change, I intend to align the outputs' sizes of the layers and add an additional layer(layer 9) to the discriminator.

Please let me know if you have any problems.

2019/9/13 Update!

Having applied the input size 256 x 256 indicated in the paper, assuming that 257 x 257 is a typo, I noticed some problems as follows:

1. Inception_v3 in pytorch doesn’t support input size 256 x 256; thus, I implemented resnet152 instead. Details are here
2. In the original paper, the kernel size is 5 x 5 in layer 7. However, this is incorrect since the input size is 4 x 4 so I specified the kernel size 4 x 4 in layer 7.

2019/9/26 Update!

Following the author's advice, having applied the input size 257 x 257. If you want to test the 257 x 257 input, prepare your dataset whose size is 257 x 257 and select it using argparse command --dataset_name