Overview

This is you first assignment for familiarising yourself with PyTorch.

You have to complete the missing parts in the code, with the goal of training a baseline RNN model for sentiment classification in Twitter messages. The functions for loading the raw data (utils/load_data.py) and the pretrained word embeddings (utils/load_embeddings.py) are given to you.

Key points

The key points of the first assignment are:

• Utilize the dataloading abstractions of PyTorch, namely torch.utils.data.Dataset and torch.utils.data.DataLoader. Don't use torchtext.
• Initialize the embedding layer of your model with pretrained word embeddings. I recommend using Glove's 50 dimensional vectors , as the performance of the model is irrelevant and using low-dimensional embeddings will speed things up.
• Implement a baseline RNN model. Than means using the RNNs output from the last timestep as feature representation of the input (no attention!). Remember, you have to account for the zero-padded timesteps!

Implementation details

The training pipeline (root) is in train.py. The classes for the model definition and dataloading are defined here:

• modules/dataloaders.py
• modules/models.py

but you have to implement the necessary methods.

Helpful links

• http://pytorch.org/tutorials/beginner/data_loading_tutorial.html
• https://www.sanyamkapoor.com/machine-learning/pytorch-data-loaders/
• https://github.com/hunkim/PyTorchZeroToAll
• https://www.youtube.com/playlist?list=PLlMkM4tgfjnJ3I-dbhO9JTw7gNty6o_2m&disable_polymer=true