Akshita Bhagia

Better documentation for creating training configs

1

* The guide gives an example of a training config, but not an exhaustive list of what fields are possible to include. * The `Vocabulary` and `Trainer` fields are general,...

Models loaded using the `from_archive` method need to be saved with original config

1

When `allennlp train` is used to fine-tune a pretrained model (`model A`) using `from_archive(path_to_A)`, the finetuned model (`model B`) is saved with the config that contains `from_archive`. This means that...

Handle distributed case for metrics that do not currently support distributed setting

1

Not supported at all - [ ] Covariance - [ ] PearsonCorrelation (dependent on Covariance) - [ ] SpanBasedF1Measure Not supported for unequal batch sizes - [ ] SpearmanCorrelation -...

Allow for a way to have unimplemented functions in base class without requiring `check_signature=False`

7

In the latest `overrides` version (`6.1.0`), the following code snippet leads to errors: ``` class A: def generic_func(self, *args, **kwargs): raise NotImplementedError class B(A): @overrides def generic_func(self, actual_arg, some_kwarg=1): ....

pip installing tango on Google Colab works for imports. But the cli commands result in error: `!tango --version` Error: ```Traceback (most recent call last): File "/usr/local/bin/tango", line 5, in from...

Steps cannot take `Model` as input

1

Right now, steps cannot take `Model` object as inputs, unless `_to_params()` is implemented. As it's unreasonable to expect users to write `_to_params()` for every implementation, we can add any empty...

NOTE: Absolute mess, do not attempt to review. Fixes https://github.com/allenai/tango/issues/270 Changes proposed in this pull request: - ## Before submitting - [ ] I've read and followed all steps in...

Rather than wrapping the model before passing it to the trainer. We want to do this because the `checkpoint_wrapper` in `with_wrapped_modules` creates a `weakref` for the wrapped modules, rendering the...

https://pytorch.org/blog/introducing-pytorch-fully-sharded-data-parallel-api/