Multitask regression

[ElanaPearl]

What's in the PR:

• Adds in code for RobertaForSequenceClassification into repo
• Makes variant of this for RobertaForRegression which handles single or multi-task regression. Also normalizes/unnormalizes the values of each task so we can do multiple tasks on different scales.
• Added regression dataset class that reads in data from a csv, keeps track of all labels, and the mean/std.dev of each task so they can be normalized. This dataset starts by embedding all the smiles which isn't the most efficient.
• Also added a lazy regression dataset which uses hugging face's lazy csv dataset loader but currently has some issues
• Data collator that works for multitask (the vanilla one assumes each label is a single integer)
• Made it so each model just returns logits when called without labels (for eval)

Testing:

• Sequence Classification has been able to easily overfit to a dataset that always has 1 label
• Multitask Regression has been able to overfit to a dataset that has multiple consistent labels

Next steps:

• Switch over to new lazy dataloader
• Get this to work when some labels are NaN
• Reduce the repeated code- maybe merge all training scripts into one since they are quite similar
• Get multi-task regression to learn a simple molecular property (MW maybe?)
• Misc. code cleaning / reorganization (these probably shouldn't all be in a masked-lm folder but don't wanna do too much in 1 PR)

[seyonechithrananda]

Looks good so far! I'll try out the sequence classification script on my own and see how it does

[seyonechithrananda]

Could we perhaps split the multi-task reg code into a new subdirectory? In terms of code organization, it would be nice to have a directory for seq-classification/fine-tuning (MoleculeNet), multi-task reg (Pubchem), and masked-lm.

I'm really excited to try this out, looks awesome. :) Looking forward to seeing how masked-LM and multi-task reg compare.