Manifold Autoencoder

This is code associated with Connor, Marissa, and Christopher Rozell. "Representing Closed Transformation Paths in Encoded Network Latent Space." AAAI. 2020.

Requirements

• torch
• scipy
• numpy
• cv2
• sklearn
• six
• glob

Code structure

Main Manifold Autoencoder Function

manifold_autoencoder_closedPaths.py is the main fule to run when training the manifold autoencoder mode. The argparser at the beginning of the file shows the possible parameters to select for a run. Two noteable parameters are the data_use parameter which specifies which dataset you'll be working with and the training_phase parameter which specifies which phase of training to run. The options for data_use are concen_circle, rotDigits, and gait corresponding to the tests from the paper. The training_phase options are AE_train, transOpt_train, and finetune corresponding to the three training phases described in Figure 2 in the paper.

Auxillary functions

• covNetModel.py - Defines the convolutional networks used in the MNIST experiments
• fullyConnectedLargeModel.py - Defines the larger fully connected network used in the gait experiments
• fullyConnectedModel.py - Defines the fully connected network used in the concentric circle experiments
• trans_opt_objectives.py - functions used to infer transport operator coefficients
• transOptModel.py - defines the transport operator neural network layer
• utils.py - contains functions for reading in and generating data as well as visualizing outputs

Guide to parameters for experiments

Concentric circle experiment:

• Autoencoder Training Phase: batch_size: 64, learning rate phi: 1e-4
• Transport Operator Training Phase: batch_size: 10, learning rate psi: 5, zeta: 1e-4, gamma: 0.005, M: 4
• Fine-tuning Phase: batch_size: 10, learning rate phi: 1e-4, learning rate psi: 5e-4, zeta: 0.0, gamma: 0.0, lambda: 1000, M: 1

Rotated MNIST experiment:

• Autoencoder Training Phase: batch_size: 64, learning rate phi: 1e-4
• Transport Operator Training Phase: batch_size: 32, learning rate psi: 0.01, zeta: 0.01, gamma: 8e-5, M: 10
• Fine-tuning Phase: batch_size: 32, learning rate phi: 0.005, learning rate psi: 1, zeta: 0.0, gamma: 0.0, lambda: 10, M: 1

Gait experiment:

• Autoencoder Training Phase: batch_size: 64, learning rate phi: 5e-4
• Transport Operator Training Phase: batch_size: 32, learning rate psi: 0.005, zeta: 0.05, gamma: 1e-4, M: 10