AG News Experiments and Basic Client Refactor

[yuchongzhang]

PR Type

[Feature | Fix ]

Short Description

1. A refactor of the BasicClient class to support the model taking in a dictionary of tensors as its input. This is useful, for example, if one wants to train a language model on text data. In this case, in addition to the tensors that represent the tokenized sentence, we may also need to pass in the attention masks.
2. An extension of the label-based samplers to support a dictionary-based dataset.
3. A refactor of the dynamic layer exchange example to use hugging face models and datasets. The new example also shows how to use collate functions to dynamically pad the batches as opposed to padding everything to the maximum length, which is what we did before.
4. The code for the same AG news experiment using the sparse tensor exchanger.
5. Moved the original dynamic layer exchanger example to the research folder and replaced it with a typical example on non i.i.d. CIFAR-10.
6. Some small fixes here and there.

Tests Added

N/A

[emersodb]

Good to go from me