0.38.4

:bug: Fix minor bugs

0.38.3

:bookmark: Release v0.38.3

0.38.2

:bookmark: Release v0.38.2

0.38.1

:bookmark: Release v0.38.1

0.38.0

Main changes

• Resuming from checkpoint moved to the TrainerConfig (setting resume_from_checkpoint in Trainer.train() is now deprecated and raises error)
• Resuming from checkpoints now supports inner-loop steps instead of only epochs
• Add data sampler for slicing data loaders (mainly used for training resumption)
• Re-order objects initialization in the Trainer's init function
• Add support for optimizer checkpointing in Trainer
• Add option to disable preprocess and post processing in Model.predict()
• Seperate generation config in Whisper's model config with a separate data class
• Drop support for Python 3.9

0.37.0

:bookmark: Release v0.37.0

0.36.1

:heavy_plus_sign: Add permissions to `version-release.yml`

0.36.0

v0.36.0

1. Add gradient accumulation support to the Trainer

Now, you can set gradient_accumulation_steps (defaults to 1 which is the same as regular training) in the TrainerConfig to enable this feature. This technique can mimic having larger batch sizes without changing the actual batch size! For example, having batch size of 16 and gradient accumulation steps of 4 equals to having batch size of 64! This can lead to faster convergance.

2. Implement tools for training speech recognition models

In this release we added SpeechRecognitionDataset, SpeechRecognitionDataCollator and SpeechRecognitionMetricsHandler so that you can easily train or finetune a Whisper model. Take a look at this example.

3. Split and refactor in Trainer for better subclassing

We split the training_step function of the Trainer in a way that now it only takes care of forward/backward pass and the optimization step is now moved to its own method called optimization_step. Also added lr_scheduler_step for customizing LR scheduling step.

4. Add support for more LR schedulers

5. Other bug fixes and improvements

0.35.1

:bookmark: Release `v0.35.1`

0.35.0

This is a big one! We made a lot of changes and improvements in Hezar.

Improvements

• Add support for accelerate for distributed training
• Add resume from checkpoint feature to Trainer
• Improve saving/logging capabilities in Trainer
• Improve print_info()
• Add ImageCaptioningDataset and ImageCaptioningDataCollator
• Enhance padding in tokenizers
• Rewrite contribution docs
• Add tests workflow to actions
• Add cache_dir parameter to all load() methods
• Improve OCRDataset and bug fixes
• Add training scripts for image captioning
• Add training script for CRNN training
• Clean registry.py
• Change license from MIT to Apache 2.0
• Some improvements and bug fixes in ViTRobertaImage2Text
• Bug fixes in tests
• Safe class var handling in configs
• Add return_scores to CRNNImage2Text
• Add get_state_dict_from_hub to support loading from any (non-Hezar) model on the Hub
• Set default LR scheduler (reduce on plateau) to Trainer

Bug fixes

• Fix image captioning decoding bug
• Fix mixed precision bug on CPU
• Fix embedding config bug

Deletions

• Delete empty models modules
• Remove all Union annotations and replace with |