Milestones before the release of Vulcan 2

[priyatharsan]

Date to complete: October 31th, 2018 @rfratila @nirtiac Feel free to add the what else been left to be done. Let's do this:

Priya's tasks

• [x] Change the criterion to object instance instead of name
  • [ ] Test manual custom criterion
• [x] Multi-input Network (https://arxiv.org/abs/1808.04456)
  • [x] Thoroughly test this once implemented
  • [x] Freeze network weights option
• [x] Other plot functionalities (@rfratila please add anything specific that we will need)
  • [x] t-sne
  • [x] pca
  • [x] receptive fields of a specific layer (most importantly the first layer)
  • [x] Metrics plot (e.g. Confusion Matrix, ROC curves - but these may already be implemented)
• [ ] Add more unittests to models (cnn, dnn, basenetwork - especially basenetwork)
• [x] Clean up examples folder (there are a few Jupiter notebooks that aren't super relevant like PyTorch_implementation.ipynb, PyTorch_train_function, and PyTorch_test_imports). Lets have different model configurations that show off the various ways to construct models.

Caitrin's tasks

• [ ] tests for dataset
• [x] implement unimplemented methods in dataset
• [x] cross-validation function
• [x] save function
  • [x] load function
• [x] tests for save and load
• [x] document metrics, layers and add tests at the same time

Robert's tasks

• [x] Saliency map plot
• [x] Learning rate scheduling
• [x] Snapshot ensemble
  • [x] train
  • [x] forward_pass
  • [x] save ensemble
  • [x] load ensemble

[priyatharsan]

@nirtiac the save and load are yet to be reconstructed, maybe we should add them under logger? what say @rfratila ?

[rfratila]

@priyatharsan, saving and loading should be placed where it is most intuitive. I would argue that they belong in basenetwork so that one can call model.save_model(path=//) and model.load_model(path=//). The latter would be a class method so you can call it before you need to instantiate a basenetwork type object and it would just build from the saved file. Does that make sense?

[nirtiac]

@rfratila @priyatharsan I agree, and with the classmethod!