Skip to content

v0.4.0
Compare
Choose a tag to compare
@vmoens vmoens released this 25 Apr 16:14
· 364 commits to main since this release
v0.4.0
6dc8c68

What's Changed

This new version of tensordict comes with a great deal of new features:

  • You can now operate pointwise arithmetic operations on tensordict. For locked tensordicts and inplace operations such as += or data.mul_, fused cuda kernels will be used which will drastically improve the runtime.
    See

    • [Feature] Pointwise arithmetic operations using foreach by @vmoens in #722
    • [Feature] Mean, std, var, prod, sum by @vmoens in #751

  • Casting tensordicts to device is now much faster out-of-the box as data will be cast asynchronously (and it's safe too!)
    [BugFix,Feature] Optional non_blocking in set, to_module and update by @vmoens in #718
    [BugFix] consistent use of non_blocking in tensordict and torch.Tensor by @vmoens in #734
    [Feature] non_blocking=None by default by @vmoens in #748

  • The non-tensor data API has also been improved, see
    [BugFix] Allow inplace modification of non-tensor data in locked tds by @vmoens in #694
    [BugFix] Fix inheritance from non-tensor by @vmoens in #709
    [Feature] Allow non-tensordata to be shared across processes + memmap by @vmoens in #699
    [Feature] Better detection of non-tensor data by @vmoens in #685

  • @tensorclass now supports automatic type casting: annotating a value as a tensor or an int can ensure that the value will be cast to that type if the tensorclass decorator takes the autocast=True argument
    [Feature] Type casting for tensorclass by @vmoens in #735

  • TensorDict.map now supports the "fork" start method. Preallocated outputs are also a possibility.
    [Feature] mp_start_method in tensordict map by @vmoens in #695
    [Feature] map with preallocated output by @vmoens in #667

  • Miscellaneous performance improvements
    [Performance] Faster flatten_keys by @vmoens in #727
    [Performance] Faster update_ by @vmoens in #705
    [Performance] Minor efficiency improvements by @vmoens in #703
    [Performance] Random speedups by @albanD in #728
    [Feature] Faster to(device) by @vmoens in #740

  • Finally, we have opened a discord channel for tensordict!
    [Badge] Discord shield by @vmoens in #736

  • We cleaned up the API a bit, creating a save and a load methods, or adding some utils such as fromkeys. One can also check if a key belongs to a tensordict as it is done with a regular dictionary with key in tensordict!
    [Feature] contains, clear and fromkeys by @vmoens in #721

Thanks for all our contributors and community for the support!

Other PRs

  • [Benchmark] Benchmark to_module by @vmoens in #669
  • [Benchmark] Benchmark update_ by @vmoens in #704
  • [BugFIx] FIx tc.update_ by @vmoens in #750
  • [BugFix, Feature] pad_sequence refactoring by @dtsaras in #652
  • [BugFix, Feature] tensorclass.to_dict and from_dict by @vmoens in #707
  • [BugFix, Performance] Fewer imports at root by @vmoens in #682
  • [BugFix, Refactor] More reliable Sequential.get_dist by @vmoens in #678
  • [BugFix,Feature] filter_empty in apply by @vmoens in #661
  • [BugFix] Allow device overriding with None in apply by @vmoens in #720
  • [BugFix] Avoid lazy stacks in stack if not asked explicitly by @vmoens in #741
  • [BugFix] Dense stack lazy tds defaults to dense_stack_tds by @vmoens in #713
  • [BugFix] Faster to_module by @vmoens in #670
  • [BugFix] Fix colab in tutos by @vmoens in #757
  • [BugFix] Fix dense stack usage in torch.stack by @vmoens in #714
  • [BugFix] Fix empty(recurse) call in _apply_nest by @vmoens in #658
  • [BugFix] Fix indexing (lazy stack and names) by @vmoens in #657
  • [BugFix] Fix keys for nested lazy stacks by @vmoens in #745
  • [BugFix] Fix lazy params init by @vmoens in #681
  • [BugFix] Fix lazy stack keys by @vmoens in #744
  • [BugFix] Fix load_state_dict for TensorDictParams by @vmoens in #689
  • [BugFix] Fix name gathering with tensor indices by @vmoens in #690
  • [BugFix] Fix singleton dims in expand_as_right by @vmoens in #687
  • [BugFix] Fix to_module __exit__ update when td is locked by @vmoens in #671
  • [BugFix] Fix to_module batch-size mismatch by @vmoens in #688
  • [BugFix] Fix torch_function for uninit param by @vmoens in #683
  • [BugFix] Fix zipping in flatten_keys by @vmoens in #729
  • [BugFix] Improve update_ by @vmoens in #655
  • [BugFix] Keep dim names in transpose by @vmoens in #662
  • [BugFix] Loading phantom state-dicts by @vmoens in #650
  • [BugFix] Make functorch.dim optional by @vmoens in #737
  • [BugFix] Missing **kwargs in apply_ fallback by @vmoens in #664
  • [BugFix] Patch pad_sequence by @vmoens in #742
  • [BugFix] Remove monkey patching of uninit params by @vmoens in #684
  • [BugFix] Support empty tuple in lazy stack indexing by @vmoens in #696
  • [BugFix] Track sub-tds in memmap by @vmoens in #719
  • [BugFix] Unlock td during update in to_module by @vmoens in #686
  • [BugFix] module hook fixes by @vmoens in #673
  • [BugFix] tensorclass as a decorator by @vmoens in #691
  • [CI] Doc on release tag by @vmoens in #761
  • [CI] Fix wheels by @vmoens in #763
  • [CI] Pinning mpmath by @vmoens in #697
  • [CI] Remove OSX x86 jobs by @vmoens in #753
  • [CI] Remove all osx x86 workflows by @vmoens in #760
  • [CI] Remove snapshot from CI by @vmoens in #701
  • [CI] Schedule workflow for release branches by @vmoens in #759
  • [CI] Unpin mpmath by @vmoens in #702
  • [Doc,CI] Sanitize version by @vmoens in #762
  • [Doc] Fix EnsembleModule docstring by @BY571 in #712
  • [Doc] Fix probabilistic td module doc by @vmoens in #756
  • [Doc] Installation instructions in API ref by @vmoens in #660
  • [Doc] Per-release docs by @vmoens in #758
  • [Doc] fix typo by @husisy in #724
  • [Feature] Adds utils method isin by @albertbou92 in #654
  • [Feature] Adds utils method remove_duplicates by @albertbou92 in #653
  • [Feature] Inherit lock in shape / tensor ops by @vmoens in #730
  • [Feature] Store non tensor stacks in a single json by @vmoens in #711
  • [Feature] TensorDict logger by @vmoens in #710
  • [Feature] TensorDict.depth property by @vmoens in #732
  • [Feature] Use generator for map by @vmoens in #672
  • [Feature] Warn when reset_parameters_recursive is a no-op by @matteobettini in #693
  • [Feature] from_modules method for MOE / ensemble learning by @vmoens in #677
  • [Feature] td_flatten_with_keys by @vmoens in #675
  • [Feature] tensordict.to_padded_tensor by @vmoens in #723
  • [Feature] use return_mask as a string in pad_sequence by @dtsaras in #739
  • [Minor] Minor improvements to tensorclass by @vmoens in #743
  • [Minor] Remove double locks on data and grad by @vmoens in #752
  • [Refactor, Feature] Default to empty batch size by @vmoens in #674
  • [Refactor] Cleanup deprecation of empty td filtering in apply by @vmoens in #665
  • [Refactor] Refactor contiguous by @vmoens in #716
  • [Refactor] Set lazy_legacy to False by default by @vmoens in #680
  • [Test] Add proper tests for torch.stack with lazy stacks by @vmoens in #715
  • [Versioning] Remove deprecated features by @vmoens in #747
  • [Versioning] Torch version by @vmoens in #749
  • [Versioning] v0.4.0 by @vmoens in #649

New Contributors

Full Changelog: v0.3.0...v0.4.0

Contributors

albanD, dtsaras, and 5 other contributors
Assets 15
Loading