Deep Feature Consistent Variational Autoencoder

This repository contains the TensorFlow 2 implementation of the Variational Autoencoder described in Deep Feature Consistent Variational Autoencoder (Hou et al. 2016)

Requirements

• Python 3.7 or higher
• Tensorflow 2.2.0
• numpy
• opencv
• matplotlib

The other requirement is the CelebA dataset. The models from this repository were trained on CelebA Align&Cropped Images. For more details about the CelebA dataset, please refer to Large-scale CelebFaces Attributes (CelebA) Dataset.

Training

To train the model, set config parameters in config.py first and run the following command:

python main.py --model [name]

The model can be seleted from "VAE", "VAE_123" or "VAE_345", where:

• VAE: plain model trained with pixel-by-pixel loss
• VAE-123: model trained by using VGG19 layers relu1_1, relu2_1, relu3_1 for the feature perceptual loss
• VAE-345: model trained by using VGG19 layers relu3_1, relu4_1, relu5_1 for the feature perceptual loss

Results

1. Reconstruction

Original Images	Plain VAE	DFC VAE-123	DFC VAE-345

2. Sample

Plain VAE	DFC VAE-123	DFC VAE-345