PyTorch 2.5 support, model migrations, end of Python 3.8 support

Python version

• Update Python versions due to Python 3.8 EOL (PR #486)

Models

• Migrate all bottleneck-injected models to sc2bench (PR #490)
• Update with PyTorch Hub links (PR #489)
• Update yaml example (PR #487)

Misc

• Force-use pytorch hub when repo_or_dir is given (PR #488)
• Explicitly add evaluate to dependency (PR #483)
• Update versions (PRs #480, #491)

A new KD method, new benchmark results, and updated YAML constructors

New method

• Add KD with logits standardization (PR #460)

YAML configs

• Fix the official config for SRD (Issue #471, PR #473)
• Fix SRD config (Issue #471, PR #472)
• Add os.path YAML constructors (PR #454)

Logs

• Disable an auto-configuration for def_logger (Issue #465, PR #469)
• Use warning (PR #468)

Documentation

• Add a new benchmark (PR #464)
• Update Projects page (PRs #456, #475)

Misc

• Update README (PRs #461, #470)
• Update a URL (PR #459)
• Update GH Action vers (PRs #457, #458)
• Update CITATION (PR #455)
• Add a new DOI badge (PR #453)
• Update version (PRs #452, #479)

New KD methods, updated YAML constructors, and low-level loss support

New methods

• Add SRD method (PRs #436, #444, #446)
• Add Knowledge Distillation from A Stronger Teacher method (PR #433)
• Add Inter-Channel Correlation for Knowledge Distillation method (PR #432)

YAML constructor

• Update functions in yaml_util (PR #447)
• Fix docstrings and add import_call_method & yaml constructor (PR #442)

Distillation/Training boxes

• Enable auxiliary model wrapper builder to redesign input model (PR #437)

Registries

• Add low-level registry and get functions (PR #426)

Documentation

• Update benchmarks (PR #435)
• Fix a typo (PR #424)

Examples

• Replace dst with src (Issue roymiles/Simple-Recipe-Distillation#1, PR #445)
• Add Amazon SageMaker Studio Lab badges (PR #422)

Tests

• Add a test case for import_call_method (PR #443)
• Add import test (PR #441)

Misc

• Update citation info (PRs #438, #439, #440)
• Update publication links (PR #430)
• Update version (PR #425, #449, #451)
• Update README (PRs #423, #434, #450)
• Update image url (PR #421)

New generation with new features and documentation

torchdistill v1.0.0 Release Notes

This major release supports PyTorch 2.0 and contains a lot of new features, documentation support, and breaking changes.

PyYAML configurations and executable scripts with torchdistill <= v0.3.3 should be considered "legacy" and are no longer supported by torchdistill >= v1.0.0. New PyYAML configurations and executable scripts are provided for the major release.

This release adds support for Python 3.10 and 3.11, and Python 3.7 is no longer supported.

Documentation

• Update documents (PRs #400, #408)
• Add docstrings (PRs #392, #393, #394, #395, #396, #397)
• Add torchdistill logos (PRs #401, #402, #403)

Dependencies & Instantiation

• Add getattr constructor (PR #325)
• Make package arg optional (PR #322)
• Enable dynamic module import/get/call (PR #319)
• Add a function to import dependencies e.g., to register modules (PR #265)

Module registry

• Add *args (PR #345)
• Fix default value-related issues (PR #327)
• No longer use lowered keys (PR #326, #332)
• Disable lowering by default (PR #323)
• Rename type/name key (PR #312)
• Rename registry dicts and arguments for registry key (PR #269)
• Raise errors when requested module keys are not registered (PR #263)
• Enable naming modules to be registered (PR #262)

Distillation/Training boxes

• Remove default forward_proc for transparency (PR #417)
• Rename a forward_proc function (PR #414)
• Simplify (D)DP wrapper init (PR #410)
• Change the timing to print model setup info (PR #335)
• Add an option to specify find_unused_parameters for DDP (PR #334)
• Do not touch teacher model by default (PR #333)
• Training box does not have to inherit nn.Module class (PR #317)
• Add interfaces package to core (PR #310)
• Update forward interfaces (PR #307, #308)
• Rename post_process post_epoch_process for consistency (PR #306)
• Consider CosineAnnealingWarmRestarts in default post-epoch process functions (PR #305)
• Make some common procedures in training box registrable/replaceable (PR #304)
• Introduce {pre,post}-{epoch,forward} processes and registries (PR #274)
• Rename post_forward functions (PR #272)
• Make loss as kwarg (PR #273)

Forward hooks

• Fix initialization issues in IO dict for SELF_MODULE_PATH (PR #328)

Dataset modules

• Redesign split_dataset and remove unused functions (PR #360)
• Update CRD dataset wrapper (PR #352)
• Fix a bug (PR #351)
• Add default args and kwargs (PR #347)
• Add get_dataset (PR #324)

Loss modules

• Fix a typo (PR #413, #415)
• Add doc artifacts and an option to pass pre-instantiated loss module (PR #399)
• Add DictLossWrapper (PR #337)
• Rename an old function name PR #309)
• Rename single loss middle-level loss (PR #300)
• Explicitly define criterion wrapper (PR #298)
• Change concepts of OrgLoss and org_term (PR #296)
• Rename loss-related classes and functions (PR #294)
• Add default forward process function and KDLoss back as a single loss (PR #275)
• Remove org loss module and introduce self-module path (PR #271)

Model modules

• Support parameter operations (Discussion #387, PR #388)
• Replace pretrained with weights (PR #354)

Auxiliary model wrapper modules

• Add find_unused_parameters arg (PR #340)
• Rename special in configs to auxiliary_model_wrapper (PR #291)
• Rename special module for clarity (PR #276)

Optimizer/Scheduler modules

• Fix bugs around optimizer/scheduler (PR #358)
• epoch arg is deprecated for some LR schedulers (PR #338)

Examples

• Revert legacy file paths to non-legacy ones (PR #419)
• Update kwargs and scripts (PR #382)
• Update yaml util and sample configs (CIFAR-10, CIFAR-100) for the next major release (PR #361)
• Update sample script and configs (GLUE) for the next major release (PR #259)
• --log was replaced with --run_log (PR #350)
• dst_ckpt should be used when using -test_only (PR #349)
• Simplify the semantic segmentation script (PR #339)
• Move hardcoded-torchvision-specific code to local custom package (PR #331)
• Update world_size, cudnn configs, and checkpoint message (PR #330)
• Rename log argument due to the (abstract) conflict with torchrun (PR #329)
• Restructure examples and export some example-specific packages (PR #320)
• Add an option to disable torch.backend.cudnn.benchmark (PR #316)
• Support stage-wise loading/saving checkpoints (PR #315)
• Support src_ckpt and dst_ckpt for initialization and saving checkpoints respectively (PR #314)
• Use legacy configs and scripts tentatively (PR #292, #295)
• Add legacy examples and configs (PR #289)

Configs

• Declare forward_proc explicitly (PR #416)
• Add configs used in NLP-OSS 2023 paper (PR #407)
• Fix value based on log (PR #284)
• Update sample configs (ILSVRC 2012, COCO 2017, and PASCAL VOC 2012) for the next major release (PR #357)
• Update official configs for the next major release (PR #355)
• Merge single_/multi_stage directories (PR #346)
• Rename variables (PR #344)
• Rename "factor" "weight" (PR #302)
• Restructure criterion (PR #301)
• Consistently use "params" to indicate learnable parameters, not hyperparameters (PR #297)

Misc.

• Add Google Analytics ID (PR #406)
• Add sitemap.xml (PR #405)
• Update timm repo (PR #375)
• Add acknowledgments (PR #369)
• Update file paths (PR #356)
• Fix a typo and replace pretrained with weights (PR #353)
• Remove the dict option as it is not intuitive for building transform(s) (PR #303)
• Temporarily remove registry test (PR #293)
• Add an important notice (PR #286)
• Add read permission for content, following the new template (PR #284)
• Refactor (PRs #268, #270, #283, #343)
• Update README (PRs #252, #290, #299, #341, #342, #348, #364, #400, #409, #418)
• Update versions (PRs #251, #391, #420)

Workflows

• Add a GitHub Action for deploying Sphinx documentation (PR #404)

Updates, bug fixes, and end of apex support

Updates in APIs/scripts

• Add square-sized random crop option (PR #224)
• Replace torch.no_grad() with torch.inference_mode() (PR #245)
• Terminate apex support due to its maintenance mode (PRs #248, #249)

Bug fixes

• Add a default value (Discussion #229, PR #230)
• Fix a bug raised in torchvision (PR #231)
• Fix a default parameter (PR #235)

Misc.

• Fix a typo (PR #232)
• Update Travis (PR #236)
• Update README (PRs #228, #238, #240
• Update versions (PRs #223, #250)

Minor bug fix and updates

Bug fix

• Fix a potential bug in split_dataset (Issue #209, PR #210)

Misc.

• Update GitHub workflow (PR #217)
• Add local epoch for LambdaLR (PR #219)
• Update versions (PRs #208, #220)

Minor updates

Minor updates

• Freeze module before rebuild if applicable (PR #205)
• Refactor and improve result summary message (PR #206)
• Update version (PRs #204, #207)

Bug fix

Bug fix

• strict should not be used here (PR #202)

Minor update

• Update version (PRs #201, #203)

Example and minor updates

Example updates

• Restructure and make download=True (PR #190)
• Make log_freq configurable for test (PR #191)
• Refactor (PR #192)
• Probably torch.cuda.synchronize() is no longer needed (PR #194)
• Add an option to use teacher output (PR #195)
• Replace no_grad with inference_mode (PR #199)

Minor updates

• Add strict arg (PR #193)
• Add assert error message (PR #196)
• Check if ckpt file path is string (PR #197)
• Check if batch images are instance of Tensor (PR #198)
• Update version (PRs #189, #200)

Add new features, PASCAL examples and pretrained models

New features

• Add wrapped resize to enable specifying interpolation for resize (PR #182)
• Add wrapped random crop resize to enable specifying interpolation for random crop resize (PR #183)
• Enable to load ckpt containing only specific module and via URL (PR #187)

New examples and trained models

• Add examples for PASCAL VOC 2012 (PRs #184, #186)
• Update README (PR #185)
• Add model weights of DeepLabv3 with ResNet-50/101 fine-tuned on PASCAL VOC 2012 (Segmentation)

	mean IoU	global pixelwise acc
DeepLabv3 w/ ResNet-50	80.6	95.7
DeepLabv3 w/ ResNet-101	82.4	96.2

Model implementations are available in torchvision. These model weights are originally pretrained on COCO 2017 dataset (available in torchvision) and then fine-tuned on PASCAL VOC 2012 (Segmentation) dataset.

Minor updates

• Add a version constant (PR #175)
• Rename and add functions for ResNet-50 and ResNet-101 (PR #176)
• Add CITATION file (PR #178)
• Update version (PRs #174, #188)
• Update README (PRs #179, #180)