Finetuner makes neural network fine-tuning easier and faster by streamlining the workflow and handling all the complexity and infrastructure requirements in the cloud. With Finetuner, one can easily enhance the performance of pre-trained models and make them production-ready without expensive hardware.
jina-aiThis release covers Finetuner version 0.6.3, including dependencies finetuner-api 0.4.2 and finetuner-core 0.10.4.
It contains 2 new features, 2 bug fixes, and 1 documentation improvement.
🆕 Features
Allocate more GPU memory in GPU environments
Previously, the run scheduler was allocating 16GB of VRAM for GPU runs. Now, it allocates 24GB.
Users can now fine-tune significantly larger models and use larger batch sizes.
Add WiSE-FT to CLIP finetuning (#571)
WiSE-FT is a recent development that has proven to be an effective way to fine-tune models with a strong zero-shot capability, such as CLIP. We have added it to Finetuner along with documentation on its use.
Finetuner allows you to apply WiSE-FT easily using WiSEFTCallback. Finetuner will trigger the callback when the fine-tuning job is finished and merge the weights between the pre-trained model and the fine-tuned model:
from finetuner.callbacks import WiSEFTCallback
run = finetuner.fit(
model='ViT-B-32#openai',
...,
loss='CLIPLoss',
callbacks=[WiSEFTCallback(alpha=0.5)],
)
See the documentation for advice on how to set alpha.
🐞 Bug Fixes
Fix image normalization for CLIP models (#569)
- Finetuner's image processing was not identical to that used by OpenAI for training CLIP, potentially leading to inconsistent results.
- The new version fixes the bug and matches OpenAI's preprocessing.
Add open_clip to FinetunerExecutor requirements
The previous version of FinetunerExecutor failed to include the open_clip package in its requirements, forcing users to add it manually to their executors. This has now been repaired.
📗 Documentation Improvements
Add callbacks documentation (#564)
There is now full documentation for using callbacks with Finetuner.
🤟 Contributors
We would like to thank all contributors to this release:
- Wang Bo (@bwanglzu)
- Louis Milliken (@LMMilliken)
- Michael Günther (@guenthermi)
- George Mastrapas (@gmastrapas)
