Our Project improves upon the quality and clarity of images created by autoencoders by training them in a novel recursive manner. The model is built out of smaller stacked models that share layers. The clarity of these images was also slightly improved using batch normalization, where batch normalization layers were shared between corresponding encoder and decoder layers. These training methods produced models that recreated clearer images with a slightly lower reconstruction error than an autoencoder trained end to end with the same architecture but failed to produce new face images when used on a variational autoencoder.
Need to do:
- Make recursive autoencoder variational
No Batch norm Shared Batch norm Non-shared Batch norm
Non-Variational Variational
Stacked Recursive
Things that already work:
- Stacked
- Recursive
- Shared batch norm recursive
- Shared batch norm stacked
- Stacked Variational
Easy
- Non-shared batch norm recursive
- Non-shared batch norm stacked
- Testing suite
Medium
- Recursive Variational
- Inception Score
Get data on:
- Stacked
- Stacked with non-shared batch norm
Data
- Parameter usage
- Train time
- Reconstruction error
- (Inception score eventually)
Namimg Schema:
- V - variational
- SB - Shared Batch Norm, NB - Normal Batch Norm
- S - Stacked, R - Recursive
- # - layers
- # - coding size