U-GAT-IT — Unofficial PyTorch Implementation

Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization for Image-to-Image Translation

Authors: Junho Kim, Minjae Kim, Hyeonwoo Kang, Kwanghee Lee.

ICLR 2020

This folder provides a re-implementation of this paper in PyTorch, developed as part of the course METU CENG 796 - Deep Generative Models. The re-implementation is provided by:

• Onur Can Üner, onur.uner@metu.edu.tr

• Sinan Gençoğlu, sinan.gencoglu@metu.edu.tr

Please see the jupyter notebook file main.ipynb for a summary of paper, the implementation notes and our experimental results.

Dataset

You can download the dataset using following bash command or using the link below. It will create directory: data/selfie2anime/..

sh scripts/download_data.sh

Dataset

Training

After downloading the dataset you can train your own model using command:

python main --config ./config/selfie2anime

You can make lots of changes using config json file.

Testing

If you want to test model with pre-trained weights:

sh scripts/download_checkpoint.sh

Log files and Checkpoint

Import neccessary libraries

import json
import numpy as np
import matplotlib.pyplot as plt
from easydict import EasyDict as edict

import torch
import torchvision
from torchvision.utils import make_grid

import lib.models as models
import lib.data as datasets
from lib.utils import code_backup, load_checkpoint

Load the config file and model.

config = edict(json.load(open('configs/selfie2anime.json')))
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load model
checkpoint = './saved/UGATIT_selfie2anime/05-30_00-05/checkpoints/checkpoint-epoch42.pth'
model = getattr(models, config.arch.type)(**config.arch.args)
checkpoint = load_checkpoint(checkpoint)
model.load_pretrained_weights(checkpoint['state_dict'])
del checkpoint
model.to(device)
model.eval()

Create validation(test) set

val_set = getattr(datasets, config.datamanager.type)(
        config.datamanager.root, config.datamanager.dataset_dir,
        width=config.datamanager.width,
        height=config.datamanager.height,
        mean=config.datamanager.norm_mean,
        std=config.datamanager.norm_std,
        mode="test")

You can get sample images using G_AB generator:

images = []
for i, data in enumerate(val_set):
    if i == 10:
        break
    real_A = data['A'].to(device).unsqueeze(0)
    real_B = data['B'].to(device).unsqueeze(0)

    _, fake_B = model.forward(real_A, real_B)
    
    real_A = val_set.denormalize(real_A, device=device)
    fake_B = val_set.denormalize(fake_B, device=device)

    images.append(torch.cat((real_A.cpu().detach(), fake_B.cpu().detach()), dim=-1))
    
images = torch.cat(images, dim=-2)
images = make_grid(images, padding=100).numpy()
plt.figure(figsize = (128, 64))
plt.axis('off')
plt.imshow(np.transpose(images, (1,2,0)), interpolation='nearest')