v0.0.25

Performance

• Significantly improve the efficiency of delu.nn.NLinear for cases where batch size is greater than 1. The larger the input dimensions -- the larger the speedup. Since the computation algorithm is updated, the results can be slightly different with the new version (the underlying "math" is totally the same).

v0.0.23

This is a minor release.

• Various improvements in the documentation.
• delu.nn.named_sequential is deprecated.
• delu.utils.data.Enumerate is deprecated.
• delu.utils.data.IndexDataset is deprecated.

v0.0.22

This release improves the documentation website (both style and content).

v0.0.21

This is a relatively big release after v0.0.18.

Breaking changes

• delu.iter_batches: now, shuffle is a keyword-only argument
• delu.nn.Lambda
  • now, this module accepts only the functions from the torch module or methods of torch.Tensor
  • now, the passed callable is not accessible as a public attribute
• delu.random.seed: the algorithm computing the library- and device-specific seeds changed, so the result can change compared to the previous versions
• In the following functions, the first arguments are now positional-only:
  • delu.to
  • delu.cat
  • delu.iter_batches
  • delu.Timer.format
  • delu.data.Enumerate
  • delu.nn.Lambda
  • delu.random.seed
  • delu.random.set_state

New features

• Added delu.tools -- a new home for EarlyStopping, Timer and other general tools.

• Added delu.nn.NLinear -- a module representing N linear layers that are applied to N different inputs:
  (*B, *N, D1) -> (*B, *N, D2), where *B are the batch dimensions.

• Added delu.nn.named_sequential -- a shortcut for creating torch.nn.Sequential with named modules without OrderedDict:

  sequential = delu.nn.named_sequential(
      ('linear1', nn.Linear(10, 20)),
      ('activation', nn.ReLU()),
      ('linear2', nn.Linear(20, 1))
  )
  

• delu.nn.Lambda: now, the constructor accepts keyword arguments for the callable:

  m = delu.nn.Lambda(torch.squeeze, dim=1)
  

• delu.random.seed

  • the algorithm computing random seeds for all libraries was improved
  • now, None is allowed as base_seed; in this case, an unpredictable seed generated by OS will be used and returned:

  truly_random_seed = delu.random.seed(None)
  

• delu.random.set_state: now, omitting the 'torch.cuda' is allowed to avoid setting the states of CUDA RNGs

Deprecations & Renamings

• delu.data was renamed to delu.utils.data. The old name is now a deprecated alias.
• delu.Timer and delu.EarlyStopping were moved to the new delu.tools submodule. The old names are now deprecated aliases.

Dependencies

• Now, torch >=1.8,<3

Documentation

• Updated logo
• Simplified structure
• Removed the only (and not particularly representative) end-to-end example

Project

• Migrate from sphinx doctest to xdoctest

v0.0.18

Add support for PyTorch 2

v0.0.17

Breaking changes

• delu.cat: now, the input must be a list (before, any iterable was allowed)
• delu.Timer: now, print(timer), str(timer), f'{timer}' etc. return the full-precision representation (without rounding to seconds)

New features

• delu.EarlyStopping: a simpler replacement for delu.ProgressTracker, but now, without tracking the best score. The usage is very similar, see the documentation.
• delu.cat: now, supports nested collections (e.g. the input can be a list of tuple[Tensor, dict[str, tuple[Tensor, Tensor]]])

Deprecations

• delu.ProgressTracker: instead, use delu.EarlyStopping
• delu.data.FnDataset: no alternatives are provided

v0.0.15

Breaking changes

• delu.iter_batches is now powered by torch.arange/randperm, the interface was changed accordingly
• delu.Timer: the methods add and sub are removed

New features

• delu.to: like torch.Tensor.to, but for (nested) collections of tensors
• delu.cat: like torch.cat, but for collections of tensors
• delu.iter_batches is now faster and has a better interface

Deprecations

• delu.concat is deprecated in favor of delu.cat
• delu.hardware.free_memory is now a deprecated alias to delu.cuda.free_memory
• deprecate delu.data.Stream
• deprecate delu.data.collate
  • instead, use torch.utils.data.dataloader.default_collate
• deprecate delu.data.make_index_dataloader
  • instead, use delu.data.IndexDataset + torch.utils.data.DataLoader
• deprecate delu.evaluation
  • instead, use torch.nn.Module.eval + torch.inference_mode
• deprecate delu.hardware.get_gpus_info
  • instead, use the corresponding functions from torch.cuda
• deprecate delu.improve_reproducibility
  • instead, use delu.random.seed and manually set settings mention here

Documentation

• many improved explanations and examples

Dependencies

• require python>=3.8
• remove tqdm and pynvml from dependencies

Project

• switch from flake8 to ruff
• move tool settings from setup.cfg to pyproject.toml for coverage, isort, mypy
• freeze versions in requirements_dev.txt

v0.0.13

THE PROJECT WAS RENAMED FROM "Zero" TO "DeLU"

The changes since v0.0.8:

Deprecations

• delu.data.IndexLoader is deprecated in favor of the new delu.data.make_index_dataloader

New features

• delu.data.make_index_dataloader is a better replacement for the deprecated delu.data.IndexLoader

Documentation

• change theme to Furo (view)

Project

• move most tool settings to pyproject.toml

v0.0.8

API changes:

• zero.random.seed
  • the argument must be less than 2 ** 32 - 10000
  • the seed argument was renamed to base_seed

New features:

• zero.random.seed:
  • sets better seeds based on the given argument
  • the new argument one_cuda_seed (False by default) allows to choose if the same seed is set for all CUDA devices

Documentation

• style improvements

Dependencies

• numpy>=1.18

v0.0.7

This release includes several nice updates to the website:

• added the "copy button" to code blocks
• highlight Python signatures
• make section link buttons visible and usable