code

code:

This folder contains various Complex Shearlet Network (CoShNet) trained on the Fashion MNIST + MNIST Dataset. Two most prominent file being test_fashion which is the main script for CoShCVNN to build our CVNN and test_resnet.py which is used to train our baselines ResNet(18|50). convolution etc.

Contents:

1. logs/: This folder contains logs of various runs related to ShearNet on Fashion. These logs are extremely useful to see stuff like:

   • Response of various layers (wstats ([t1.]) and histograms)
   • Input to the network (raw input and wstats)
   • Weight Matrices of different layers. (wstats).
   • Testing Results of different types of runs.

   Different naming conventions like [t2.], [t3.] etc. denote the machine on which the tests were conducted. There are a few other self explanatory naming conventions used in the context of the folder.

2. modelling/: This folder contains model architectures for models of different types and some related useful utilities for the same.

3. pipeline/: This folder contains useful utilities for pipeline of the model.

4. snapshots/: This folder is supposed to contain saved snapshots of models and their relevant optimizers (saving and restoring optimizers does in practice affect performance). They can be easily identified by their naming convention along with number of epochs. (Note: you may look at loadmodel.py on how to load a sample snapshot. Note it won't run unless you have a snapshot with that name.)

5. projconfig: Utility for getting specific folders for the dataset.

6. test_fashion: Train and test Fashion with CoShCVNN().

7. test_resnet: Trains and tests ResNet(18|50) in our training pipeline.

Terminology:

[t1.] wstats: Mean, Min, Max of an array of values. In exactly that order.

[t2.] mck: genesis in Manny Ko's machine.

[t3.] uj: genesis in Ujjawal K. Panchal's machine.

[t4.] ARD: Automatic Relevance Determination.

References:

[1.] Fashion MNIST Dataset from torchvision