What's new

Added 🎉

• Added support for PEP 604 in FromParams, i.e. writing union types as "X | Y" instead of "Union[X, Y]".
• [internals] Added a spot for miscellaneous end-to-end integration tests (not to be confused with "tests of integrations") in tests/end_to_end/.
• [internals] Core tests now run on all officially supported Python versions.

Fixed ✅

• Fixed a bug in FromParams where non-FromParams class parameters were not instantiated
  properly (or at all).
• Fixed a bug in FromParams where kwargs were not passed on from a wrapper class to the wrapped class.
• Fixed small bug where some errors from git would be printed when executor metadata is created
  outside of a git repository.

Commits

ea6d2e5 another FromParams fix (#66)
eeb1560 Update datasets requirement from <1.15,>=1.12 to >=1.12,<1.16 (#60)
fe6dbe0 Bump isort from 5.9.3 to 5.10.0 (#61)
9f302b3 PEP 604 support (#59)
634bd71 Merge pull request #63 from allenai/StepGraphTests
9f513d6 Rename step_graph.py to step_graph_test.py
604001f tweak checkpointing/validate step
0920349 fix bug with git metadata (#56)
af1b438 fix another FromParams bug, add spot for miscellaneous end-to-end tests (#58)

What's new

Fixed ✅

• Fixed a bug with FromParams that caused .from_params() to fail when the params contained
  an object that was already instantiated.
• tango command no longer installs a SIGTERM handler, which fixes some bugs with integrations that use multiprocessing.

Commits

ab47e21 Merge pull request #55 from allenai/no-sigterm-handler
b759c46 remove sigterm handler
c4e96cb fix example
a1f9ec5 fix FromParams bug (#54)
d5ef0ae Bump black from 21.9b0 to 21.10b0 (#53)
b20f5b2 fix typo

What's new

Changed ⚠️

• Updated the LightningTrainStep to optionally take in a LightningDataModule as input.

Commits

4d77160 Merge pull request #52 from allenai/fix-typo
92874c2 fix release docs
26e36e4 Merge pull request #51 from allenai/lightning-data-module
b10d256 update changelog
6d068a4 fix import order
1744f55 adding option for data module

What's new

Added 🎉

• Added IterableDatasetDict, a version of DatasetDict for streaming-like datasets.
• Added a PyTorch Lightning integration with LightningTrainStep.

Fixed ✅

• Fixed bug with FromParams and Lazy where extra arguments would sometimes be passed down through
  to a Lazy class when they shouldn't.

Commits

a95dbae fix bugs with initializing lightning loggers and plugins (#50)
7e9c354 fix bug w/ DataLoader and PTL
d634ea8 add isort (#49)
277b0e2 add IterableDatasetDict (#46)
db06e70 Merge pull request #45 from allenai/pytorch-lightning
6174381 fix CHANGELOG
501bf73 fix failing test
686be74 Merge branch 'main' into pytorch-lightning
4f3f328 add torch:: to torch integrations
293fac1 doc and general fixes
2b2a6a3 Update tango/integrations/pytorch_lightning/init.py
d46a0ba Update docs/source/api/integrations/pytorch_lightning.rst
3aeb825 fix docs
375ff58 update docs
c6a1e38 update ci
8110242 PyTorch Lightning Integration
60121c0 update docs, print unicode characters by name (#44)
7beab21 clean up docs
2f6871b only print ascii characters (#43)

What's new

Added 🎉

• Added support for torch 1.10.0.

Changed ⚠️

• --file-friendly-logging flag is now an option to the main tango command, so needs
  to be passed before run, e.g. tango --file-friendly-logging run ....

Fixed ✅

• Fixed bug with Step.from_params.
• Ensure logging is initialized is spawn processes during distributed training with TorchTrainStep.

Commits

d497e7a Update torch requirement from <1.10.0,>=1.9.0 to >=1.9.0,<1.11.0 (#42)
fbad9b2 fix failing test
409b50a ensure logging initialize in spawn distributed workers
1714390 move file_friendly_logging flag back to main command
af8bc69 fix bug in Step.from_params
4e5b406 add TangoMetadata to docs

What's new

Added 🎉

• Added support for global settings file, tango.yml.
• Added 'include_package' (array of string) param to config spec.
• Added a custom error StopEarly that a TrainCallback can raise within the TorchTrainStep
  to stop training early without crashing.
• Added step config, tango command, and tango version to executor metadata.
• Executor now also saves pip dependencies and conda environment files to the run directory
  for each step.

Fixed ✅

• Ensured **kwargs arguments are logged in FromParams.

Commits

0094888 save pip and conda files to run directory, add step config to metadata (#41)
d588886 Early stopping via callbacks in torch train (#40)
c437887 ensure '**kwargs' are logged in FromParams
7de1837 add support for global settings file (#39)
4026df1 Update datasets requirement from <1.14,>=1.12 to >=1.12,<1.15 (#38)
7ce3493 add 'include_package' param to config spec

What's new

Added 🎉

• Added new steps to datasets integration: ConcatenateDatasets ("datasets::concatenate") and InterleaveDatasets (datasets::interleave).
• Added __contains__ and __iter__ methods on DatasetDict so that it is now a Mapping class.
• Added tango info command that - among other things - displays which integrations are installed.

Commits

d47399d add 'tango info' command (#34)
80f8457 add interleave and concatenate dataset steps (#33)
10f56f7 add test for generics from std lib (#32)
f77837a make DatasetDict an actual Mapping (#30)

What's new

Added 🎉

• Added convert_to_tango_dataset_dict() function in the datasets integration.
  It's important for step caching purposes to use this to convert a HF DatasetDict
  to a native Tango DatasetDict when that DatasetDict is part of the input to another
  step. Otherwise the HF DatasetDict will have to be pickled to determine its hash.

Changed ⚠️

• Format.checksum() is now an abstract method. Subclasses should only compute checksum
  on the serialized artifact and nothing else in the directory.
• [internals] Changed the relationship between Executor, StepCache, and Step.
Executor now owns the StepCache, and Step never interacts with StepCache directly.

Commits

bdc8486 make Format.checksum abstract
5514ce5 Refactor Executor, StepCache, and Step, improve hashing of DatasetDict (#29)

What's new

Added 🎉

• Added a Weights & Biases integration with a training callback ("wandb::log")
  for TorchTrainStep ("torch::train") that logs training and validation metrics to W&B.

Fixed ✅

• Fixed Format.checksum() when there is a symlink to a directory in the cache folder.

Commits

374f1ad Add W&B integration (#28)

What's new

Added 🎉

• Added the ability to track a metric other than "loss" for validation in TorchTrainStep ("torch::train").

Fixed ✅

• Final model returned from TorchTrainStep ("torch::train") will have best weights loaded.
• Checkpoints are saved from TorchTrainStep ("torch::train") even when there is no validation loop.
• Fixed TorchTrainStep ("torch::train") when validation_split is None.
• Fixed distributed training with TorchTrainStep ("torch::train") on GPU devices.

Commits

ba05e79 Torch train updates and distributed fixes (#27)