Finetuner 0.6.2 Update

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.

GitHub - jina-ai/finetuner: Task-oriented finetuning for better embeddings on neural search

:dart: Task-oriented finetuning for better embeddings on neural search - GitHub - jina-ai/finetuner: Task-oriented finetuning for better embeddings on neural search

GitHubjina-ai

This release covers Finetuner version 0.6.2, including dependencies finetuner-api 0.4.1 and finetuner-core 0.10.2.

It contains 3 new features and 1 bug fix.

🆕 Features

Finetuner can now produce PyTorch models

Previously, Finetuner only produced ONNX models. Users can now choose between ONNX and PyTorch models.

To select ONNX you must add the to_onnx flag to calls to finetuner.fit():

run = finetuner.fit(
    ...,
    to_onnx=True,
)

You must also add the flag to calls to finetuner.get_model() to use an ONNX model directly with DocArray:

 model = finetuner.get_model(..., is_onnx=True)

To use an ONNX model inside a Jina Flow:

f = Flow().add(
    uses='jinahub+docker://FinetunerExecutor/v0.10.2',
    uses_with={'is_onnx': True},
)

Resubmit jobs automatically

Previously, when submitting a request for Finetuner to use cloud computing resources, if the request failed, the job would fail and the user would have to resubmit it. Now, the job will be resubmitted automatically up to five times, before failing completely.

Concise and more readable log messages

We have improved the logging in Finetuner to provide fewer and more readable messages for users.

🐞 Bug Fixes

Require ONNX runtime version > 1.11.1

• This bug was causing version incompatibility errors for users of Python 3.10.
• The new version fixes the bug and makes Finetuner fully compatible with the latest Python releases.

🤟 Contributors

We would like to thank all contributors to this release:

• Michael Günther (@guenthermi)
• Zhaofeng Miao (@mapleeit)
• George Mastrapas (@gmastrapas)
• Wang Bo (@bwanglzu)