Skip to content

PSUVisualCognitionLab/MLR-2.0_pairs

 
 

Repository files navigation

Directories:

MLR-src: source code for the MLR model (mVAE, label networks, SVMs...)
checkpoints: saved weight sets for models including autoencoder, SVM and label networks
data:  Datasets  (mnist, emnist, quickdraw, cifar, bengali characters)
simulation_src: source code for all specific  simulations
simulations: output of simulation runs
documentation: packages to install
training_samples:  diagnostic images during training are placed here


These are key files

Root directory:
 - Training.py:  Wrapper function for training a new model from scratch
 - plotting.py:  Wrapper function for running specific simulations


MLR-2.0 model is outlined in MLR_src/    
 - mVAE.py: defines the cropped encoder, cropped decoder, and retinal decoder NNs in the VAE_CNN class. 
        Training objective functions are defined below this class
 - BP_functions.py: defines the Binding Pool memory functions
 - classifiers.py: defines the SVM classifiers that operate on MLR-2.0's latent representations
 - label_network.py: defines the label network to project one-hot vectors into latent representations
 - dataset_builder.py: defines the modified dataset class for training
 - train_mVAE.py:  function to manage training of the mVAE

Simulation files are defined in simulation_src
 


 Training is currently a bit convoluted  
 Example:
 python Training.py --dataset quickdraw

Training.py calls train_mVAE.py, which in turn calls the train function in mVAE.py.   This is tech debt we should solve

About

working repo for the MLR 2.0 model with object pairs

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%