DeformingAutoencoders-pytorch

Pytorch code for DAE and IntrinsicDAE

Project:

http://www3.cs.stonybrook.edu/~cvl/dae.html

Usage:

Requirements: PyTorch

To train a DAE, run

python train_DAE_CelebA.py --dirDataroot=[path_to_root_of_training_data] --dirCheckpoints=[path_to_checkpoints] --dirImageoutput=[path_to_output directory for training] --dirTestingoutput=[path_to_output directory for testing]

To train an IntrinsicDAE, run

python train_IntrinsicDAE_CelebA.py --dirDataroot=[path_to_root_of_training_data] --dirCheckpoints=[path_to_checkpoints] --dirImageoutput=[path_to_output directory for training] --dirTestingoutput=[path_to_output directory for testing]

set --useDense=True (default) for DenseNet-like encoder/decoder (no skip connections over the bottleneck latent representations); --useDense=False for a smaller encoder-decoder architecture.

Dataset: CelebA (http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html) A google drive link to a cropped and resized version of CelebA: https://drive.google.com/open?id=1ueB8BJxid2rZbvh3RaoZ9lDdlKH4B-pL

Place the training images in [path_to_root_of_training_data]/celeba_split/img_00 (Split the dataset into multiple subsets if wanted.)

Checkpoints: Some example checkpoints can be found at: https://drive.google.com/drive/folders/1A2Qj1NhzVU5XSjeilKhjWwAgNWvlRyuA?usp=sharing

Three examples are provided:

1. DAE for CelebA with Dense encoder decoder, where opt.idim = 8 (./DAE_CelebA_idim8)
2. DAE for CelebA with Dense encoder decoder, where opt.idim = 16 (./DAE_CelebA_idim16)
3. IntrinsicDAE for CelebA with Dense encoder decoder. (./IntrinsicDAE_CelebA)

If using the code, please cite:

Deforming Autoencoders: Unsupervised Disentangling of Shape and Appearance, Zhixin Shu, Mihir Sahasrabudhe, Riza Alp Guler, Dimitris Samaras, Nikos Paragios, and Iasonas Kokkinos. European Conference on Computer Vision (ECCV), 2018.

Update 12-13-2018

1. Previous models with batch normalization suffer from data batch bias in testing. Replacing all nn.BatchNorm2d() layers (DAENet.py) with nn.InstanceNorm2d() layers (DAENet_InstanceNorm.py)would fix the problem in testing time.
2. Fixed bug on getBasedGrid(): previous code contains bug on integer->float convertion.