[ECCV 2018] A PyTorch implementation of Patch2Image

• This is a pytorch implementation of the paper, Unsupervised Holistic Image Generation from Key Local Patches. (ECCV 2018).
• Paper link: https://arxiv.org/abs/1703.10730 (warning: this is an old version. Latest version will be uploaded!)

Concept

Examples (Faces)

Examples (Cars)

Requirements

• Python2 or 3
• Cuda device (NVIDIA GTX1080Ti was used to test)
• Pytorch
• Visdom (optional)
• Tensorflow & Tensorboard (optional)

Preparing dataset

Download dataset via visiting celebA or CompCar.

For celebA dataset,

You can download using download.py

> python download.py celebA

For compcar dataset, Download the entire compcar dataset and some pre-processing is required.

You should crop the car patches using the ground truth bounding boxes, resize them 128*128 resolution, and save them in a single directory.

Key-patches

We already extracted key patches from celebA and compcar dataset and save the bounding box coordinates to celebA_allbbs.mat and compcar_allbbs.mat.

You can extract key patches and use your own key patches.

Training celebA dataset

Run

python main.py --db_name=celebA --dataset_root=YOUR_DATA_ROOT --is_crop=True --image_size=108 --output_size=64 --model_structure=unet

The resolution of output image can be enlarged by --output_size=128 or --output_size=256 options.

Training compcar dataset

Run

python main.py --db_name=compcar --dataset_root=YOUR_DATA_ROOT --is_crop=False --image_size=128 --output_size=128 --conv_dim=64  --batch_size=32 --model_structure=unet

Misc.

Modify the options output_size, conv_dim, or batch_size to prevent out-of-memory error.