Torchoutil

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.

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.

Fixed
- `ModelCheckpointRegister` now creates intermediate directories before download.
- `MaskedMean` and `MaskedSum` dim argument.

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.

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.

Added
- 47 torch functions and 40 torch modules.