Pytorch-pfn-extras

Enhancements and bug-fixes

This release includes the following enhancements and bug fixes:

- Adds `create_distributed_subset_indices` to efficiently scatter datasets across distributed nodes.
- Adds `ppe.nn.Ensure` module to assert shape or dtypes in PyTorch code.
- Reduces the snapshot size when using `append=True` (output logs in JSON Lines or YAML) in LogReport.
- Makes `Trainer` and `Evaluator` public APIs so that users can add type annotations to user-defined `Logic` and `Handler` classes with these classes.
- Deprecates `async` option from the trainer and introduces a `DeferredResult` API instead.
- Stores `time_summary` in TLS to avoid bugs related to multiprocessing.

Backward-incompatible changes

- Process-wide `TimeSummary` singleton instance (`ppe.profiler.time_summary`) has been removed as it was not thread-safe and causing issues on multiprocessing/threading. In this release, you can obtain a thread-local `TimeSummary` instance using `ppe.profiler.get_time_summary()`.

See the list of [merged pull-requests](https://github.com/pfnet/pytorch-pfn-extras/pulls?q=is%3Apr+is%3Aclosed+milestone%3Av0.5.2) for the details.

Enhancements and bug-fixes

This release includes the following enhancements and bug fixes:

- Adds `ppe.onnx.grad` API to emit Gradient operator
- Improves tensor stripping in `ppe.onnx`
- Fixes `ProgressBar` extension always says `inf iters/sec`
- Fixes several type annotation issues

See the list of [merged pull-requests](https://github.com/pfnet/pytorch-pfn-extras/pulls?q=is%3Apr+is%3Aclosed+milestone%3Av0.5.1) for the details.

Enhancements and bug-fixes

This release includes the following enhancements and bug fixes:

- Adds a device-agnostic trainer and evaluator. Check the [documentation](https://pytorch-pfn-extras.readthedocs.io/en/latest/user_guide/index.html#trainer-techinical-preview).
- Adds typing support through the codebase.
- Adds online [documentation](https://pytorch-pfn-extras.readthedocs.io/).
- Shows evaluator name in the progress bar.

Backward-incompatible changes

- Drops support for PyTorch 1.7
- Deprecates the `LazyLinear` and `LazyConvNd` modules since they have been upstreamed to PyTorch.

See the list of [merged pull-requests](https://github.com/pfnet/pytorch-pfn-extras/milestone/12?closed=1) for the details.

This release includes the following enhancements and bug fixes:

- Fixes the Profiler API introduced in the v0.4.3 release to avoid initializing multiprocessing contexts and threads on `import pytorch_pfn_extras`.
- Profiler now reports min/max times.

See the list of [merged pull-requests](https://github.com/pfnet/pytorch-pfn-extras/pulls?q=is%3Apr+is%3Aclosed+milestone%3Av0.4.5) for the details.

This release includes the following enhancements and bug fixes:

- Restores support of Python 3.6 and 3.7, which was unexpectedly broken in the v0.4.3 release.
- Fixes broken styles on the documentation website.

See the list of [merged pull-requests](https://github.com/pfnet/pytorch-pfn-extras/pulls?q=is%3Apr+is%3Aclosed+milestone%3Av0.4.4) for the details.

This release includes the following enhancements and bug fixes:

- Adds support to generate documentation using [Read The Docs](https://pytorch-pfn-extras.readthedocs.io/)
- Adds logging support using the [`ppe.logging.get_logger`](https://pytorch-pfn-extras.readthedocs.io/en/latest/reference/generated/pytorch_pfn_extras.logging.get_logger.html) interface.
- Adds a profiler to measure timing in annotated code sections. ([`ppe.training.extensions.ProfileReport`](https://pytorch-pfn-extras.readthedocs.io/en/latest/reference/generated/pytorch_pfn_extras.training.extensions.ProfileReport.html) and [`ppe.profiler.TimeSummary.report`](https://pytorch-pfn-extras.readthedocs.io/en/latest/reference/generated/pytorch_pfn_extras.profiler.TimeSummary.report.html))
- Adds type annotations to several modules.

See the list of [merged pull-requests](https://github.com/pfnet/pytorch-pfn-extras/pulls?q=is%3Apr+is%3Aclosed+milestone%3Av0.4.3) for the details.