torchinfo

What's Changed

• Enable np.ndarrays as in- and outputs. by @snimu in https://github.com/TylerYep/torchinfo/pull/222
• Don't change device if input_data is given by @snimu in https://github.com/TylerYep/torchinfo/pull/236
• Move to using tensor.untyped_storage() for Pytorch 2.0 by @TylerYep in https://github.com/TylerYep/torchinfo/commit/7f2bed3

Full Changelog: https://github.com/TylerYep/torchinfo/compare/v1.7.2...v1.8.0

What's Changed

• Simplify layers_to_str by @mert-kurttutan in https://github.com/TylerYep/torchinfo/pull/173
• Add hide_recursive_layers option to ColumnSettings by @mert-kurttutan in https://github.com/TylerYep/torchinfo/pull/174
• Add contribution of recursive calls to total_output_bytes by @mert-kurttutan in https://github.com/TylerYep/torchinfo/pull/181
• Add params and MACs units specifiers by @richardtml in https://github.com/TylerYep/torchinfo/pull/188
• Add the column 'Params %' by @luke396 in https://github.com/TylerYep/torchinfo/pull/199
• Automatically get device from first model parameter if device parameter to summary is None by @snimu in https://github.com/TylerYep/torchinfo/pull/211
• Enable analyzing nested input- and output-dicts by @snimu in https://github.com/TylerYep/torchinfo/pull/212

Cleanup Work

• Add additional versions to GitHub Actions tests by @TylerYep in https://github.com/TylerYep/torchinfo/pull/189
• Fix torchvision deprecation warnings in test cases by @TylerYep in https://github.com/TylerYep/torchinfo/pull/191
• Separate nested_list_size function, add some documentation, improve mypy for setuptools by @TylerYep in https://github.com/TylerYep/torchinfo/pull/220

New Contributors

• @richardtml made their first contribution in https://github.com/TylerYep/torchinfo/pull/188
• @luke396 made their first contribution in https://github.com/TylerYep/torchinfo/pull/199
• @snimu made their first contribution in https://github.com/TylerYep/torchinfo/pull/211

Full Changelog: https://github.com/TylerYep/torchinfo/compare/v1.7.1...v1.7.2

What's Changed

• Update half precision test cases to support Pytorch v1.12 by @mert-kurttutan in https://github.com/TylerYep/torchinfo/pull/165
• Use layer_id instead of class_name in add_missing_layers by @mert-kurttutan in https://github.com/TylerYep/torchinfo/pull/163
• Replace add_missing_layers with add_missing_container_layers by @mert-kurttutan in https://github.com/TylerYep/torchinfo/pull/169

New Contributors

• @mert-kurttutan made their first contribution in https://github.com/TylerYep/torchinfo/pull/165

Full Changelog: https://github.com/TylerYep/torchinfo/compare/v1.7.0...v1.7.1

What's Changed

• Calculate parameter counts for nn.Parameter
• Add parameter counts for nn.UninitializedParameters
• Show output shapes for the overall model
• Internal optimizations (hooks applied iteratively instead of recursively, add profiling code)

Full Changelog: https://github.com/TylerYep/torchinfo/compare/v1.6.6...v1.7.0

What's Changed

• Add "Trainable" column by @bsridatta in https://github.com/TylerYep/torchinfo/pull/128
• Do not error if there are None values in ModuleLists
• Obtain kernel_size using the actual attribute name.

Fix a regression in which torchinfo would crash on PyTorch versions < 1.9

What's Changed

• Support half-precision dtypes when calculating model size by @jedi2610 in https://github.com/TylerYep/torchinfo/pull/109
• Uninitialized tensor fix by @jedi2610 in https://github.com/TylerYep/torchinfo/pull/120
• Allow user to set the training/eval mode of the model before calling summary.

New Contributors

• @jedi2610 made their first contribution in https://github.com/TylerYep/torchinfo/pull/109
• @bsridatta made their first contribution in https://github.com/TylerYep/torchinfo/pull/116

Full Changelog: https://github.com/TylerYep/torchinfo/compare/v1.6.3...v1.64

Fixed bug with layer reuse in the same variable, with and without existing hooks.

• Support for pruned models
• Support for "ascii_only" as a row setting to disable the fancy branch logging.

Deprecate Python 3.6 support. Install v1.5.4 if you want to use Python 3.6.

LayerInfo's trainable field is now trainable_params

• Display layers that share the same variable
  • e.g. activation layers that are defined once and then reused throughout the model
• README updates.

• Use sys.getsizeof for calculating input size. In the future, we will use this instead of the tensor shape to calculate the size.
• Rework the input_data correction to allow nested dicts and other data structure combinations.
• Add missing basic summary test.
• Refactor the main summary function to use a common traversal helper function.

• Fix bug causing inconsistent Mult-Add totals that do not sum correctly.
• Overhaul output testing to work automatically for all tests
• Add cache_forward_pass to make it easier to iterate on depths in Jupyter Notebooks

Upgrade the version number past v1.4.5 in order for pip to resolve the version correctly across older versions of pip.

• Fix issues with torch.jit scripted modules
• Add support for ParameterLists
• Display bias layers in verbose=2 mode

Add py.typed to surface type annotations.

Fix bug with inconsistent calculations for total_params using different depths.

Fix MACs calculation and differences in output when using different depths.

depth parameter now only affects formatting, calculated values do not change.