PytorchUtils

Some utilities for building models in pytorch.

Installation

Install using pip:
pip install pytorch-utils

OR

First, clone this repository using
git clone https://github.com/rajatvd/PytorchUtils/

Then, cd into the directory named package and run the following command:
pip install .

NBFigure

An extension of matplotlib figures to jupyter notebooks which are rendered using IPython Image displays. The plots are saved on disk and reloaded as images, allowing them to be easily updated dynamically. Useful for live loss plots.

RNN modules

Currently contains:

• WrappedLSTM : a pytorch nn Module which wraps an input and output module around an lstm. The whole module now works solely with packed sequences, and padding is not required. (DEPRACATED in favor of WrappedRNN)
• WrappedRNN : a module which wraps an input and output module around a general RNNBase instance.

Train utils

Contains a Trainer class. It can be used to call a train loop with a model, DataLoader, optimizer, a trainOnBatch function and an epoch callback function to train a model for a given number of epochs. Automatically saves the model, displays a live animated metric plot, and a progress bar for each epoch.

Integration with sacred

Also has a sacred_trainer module which can be used with a sacred experiment to log metrics and artifacts. Provides a loop function which takes a Run instance and uses it to log batch and callback metrics, and also saves model weights as checkpoints. Combine it with the visdom observer to easily visualize your deep learning experiments as well.

Todo

• Save best models to allow for easy reloading
• Make it easy to continue a run from a previous model checkpoint