lindsayxx / image-synthesis-vae Goto Github PK

• Qualitative

• Quantitative, IS, FID, MS-SSIM, log-likelihood

• Training stability

• Computational complexity, convergence speed

General

• Data reading/processing (PyTorch has package to download some standard data)

• Argument parser

• KL divergence (slightly different between AGE and IntroVAE/VAE)

• Z-sampling

• Training batch loader

• Image generator

• Quantitative metrics

• L2 norm

AGE

• Training procedure

• 32x32 Encoder

• 32x32 Generator

• 128x128 Encoder

• 128x128 Generator

• Loss function (Based on KL divergence)

IntroVAE

• Training procedure

• 128x128 Encoder

• 128x128 Generator

• Loss function (Based on KL divergence)

Vanilla VAE

• Training procedure

• Loss function (Based on KL divergence)

• Study network architecture

• Training procedure

• Gradients and backprop procedure

• Parameter settings

• Optimization methods

• Data preprocess/augmentation

• Open accounts

• Learn how to use

• Ready for running code

• Read data
• Process data
• Decide which data to use for reproduction

• Study network architecture

• Training procedure

• Gradients and backprop procedure

• Parameter settings

• Optimization methods

• Data preprocess/augmentation