Version 0.5.0

[0.5.0] 2024-11-28

Major rework of torchoutil project folders.

Added

• Tensor typing classes for instance checks like Tensor2D, FloatTensor1D, etc.
• Function get_key_fn to customize sorting using patterns.
• Option on_error to load_yaml to ignore yaml scanner and parser exceptions.
• Function ranks to get the ranking values in a tensor.
• Function sort_with_patterns to sort values using string-based patterns.
• Function pad_and_crop_dim and class PadCropDim to pad then crop a tensor along a specified dimension.
• Function reimport_modules to re-import modules and functions in a notebook.
• Function checksum_module to compare module weights.
• Function nelement to get number of elements in Tensor-like object.
• Function flatten to get flat of elements in Tensors, numpy array, builtins scalars and lists.
• Added DatasetSlicer class in utils.

Modified

• Rename is_pickle_root to is_packed_root (old name was kept for backward compatibility).
• Split python functions into a separate directory called pyoutil.

Version 0.4.0

[0.4.0] 2024-05-27

Added

• Option subdir_size to pack_dataset.
• Classes CosDecayScheduler and CosDecayRule classes.
• Function sorted_dict to collections.
• Functions ndim and shape to search information in tensor-like objects.
• Function item to convert scalar-like objects to built-in scalar objects.
• Mixin classes to improve module features.

Modified

• Rename from_numpy to numpy_to_tensor to avoid confusion with torch.from_numpy.
• Rename save_to_yaml to to_yaml and save_to_csv to to_csv.
• Rename PickleDataset to PackedDataset.
• Rename pack_to_pickle to pack_dataset.
• Update EModule and ESequential classes with auto-config and auto-device detection.

Fixed

• Function can_be_converted_to_tensor now accepts numpy arrays and scalars.
• Wildcard imports, and module imports. (e.g. from torchoutil import nn)
• can_be_stacked now returns False with an empty sequence.

Version 0.3.1

[0.3.1] 2024-04-25

Added

• Method count_parameters to TModule.
• Option padding_idx to indices_to_onehot function.
• Functions dict_list_to_list_dict, flat_list, unflat_dict_of_dict to collections utils.
• Class PickleDataset and function pack_to_pickle to utils.
• Class ResampleNearest and function resample_nearest.
• Class TransformDrop and function transform_drop.

Fixed

• Remove invalid test file.
• Function is_scalar now returns True for numpy scalars when numpy package is installed.

Version 0.3.0

[0.3.0] 2024-04-17

Added

• PositionalEncoding layer for transformers networks.
• Property metadata to HDFDataset.
• Function create_params_groups_bias for avoid apply weight decay to networks bias.
• Functions is_numpy_scalar, is_torch_scalar and unzip.
• Options auto_open, numpy_to_torch and file_kwargs to customize loading in HDFDataset.

Modified

• Update keyword arguments for mask, pad, labels functions and modules.
• Getting added column using at from HDFDataset when return_added_columns is False.
• Renamed ModelCheckpointRegister to RegistryHub and add hash value check after download.

Fixed

• can_be_converted_to_tensor now returns True if input is a Tensor.

Version 0.2.2

[0.2.2] 2024-03-08

Fixed

• ModelCheckpointRegister now creates intermediate directories before download.
• MaskedMean and MaskedSum dim argument.

Version 0.2.1

[0.2.1] 2024-03-07

Added

• ModelCheckpointRegister class to make download and loading easier from checkpoint.

Modified

• pack_to_hdf now supports existing shape column.
• SizedDatasetLike is now compatible with Sequence-like objects.

Version 0.2.0

[0.2.0] 2024-03-04

Added

• Optional hdf datasets with HDFDataset class and pack_to_hdf function.
• Multiclass functions to convert labels.
• Arg diagonal to generate_square_subsequent_mask.
• Abs, Angle, Real and Imag modules.

Modified

• Use Literal for pad and crop functions typing.
• Minimal torch version is now 1.10.

Version 0.1.0

[0.1.0] 2024-01-29

Added

• 47 torch functions and 40 torch modules.