Danielle Pintz

Add usage logging APIs for the Runner's public entry points

1

Summary: Leverage the feature here: https://pytorch.org/docs/stable/notes/large_scale_deployments.html#api-usage-logging Differential Revision: D38750230

Add runner.predict

1

Reviewed By: ananthsub Differential Revision: D38692009

Add runner.evaluate

1

Reviewed By: ananthsub Differential Revision: D38599732

Remove __enter__ and __exit__ from MemorySnapshotProfiler

1

Summary: There is no need for MemorySnapshotProfiler to be a context manager since it conflicts with start_step and stop_step Differential Revision: D51049497

Add IProfiler protocol

5

Summary: Add a protocol to define a Profiler. Differential Revision: D50765964

Fix prepare_module docstring

1

Summary: Created from CodeHub with https://fburl.com/edit-in-codehub Reviewed By: ananthsub Differential Revision: D50369401

Set default value for max_steps in _is_epoch_done util

2

Summary: In evaluate and predict the user does not pass in a max_steps argument so it may seem strange that `state.max_steps` is used here: https://github.com/pytorch/tnt/blob/master/torchtnt/framework/evaluate.py#L143 Removing this for better readability...

Add annotations to `torchtnt/utils/flops.py`

2

Reviewed By: daniellepintz Differential Revision: D47868673

Consolidate meta callbacks and oss callbacks in docs

2

Differential Revision: D46847563

Separate grad scaler test out from test_app_state_mixin

1

Summary: Currently `test_app_state_mixin` is failing on OSS CI: ``` self = def test_app_state_mixin(self) -> None: """ Test that app_state, tracked_optimizers, tracked_lr_schedulers are set as expected with AutoUnit """ my_module =...