feat: add early stopping functionality to Maestro models and examples

[AshAnand34]

Description

This pull request introduces early stopping functionality to the Multimodal Maestro framework, enabling more efficient training by halting when a monitored metric (e.g., validation loss) stops improving. The changes include updates to model configurations, the addition of an EarlyStoppingCallback, and examples demonstrating how to use this feature.

Early stopping functionality:

• Callback addition:

  • Added EarlyStoppingCallback class in maestro/trainer/common/callbacks.py to integrate early stopping into PyTorch Lightning trainers. This monitors a specified metric and stops training when improvements cease. [1] [2]

• Configuration updates:

  • Added early stopping parameters (early_stopping, early_stopping_patience, early_stopping_threshold, early_stopping_monitor) to the configurations for Florence2Configuration, PaliGemma2Configuration, and Qwen25VLConfiguration. These enable fine-tuning of early stopping behavior. [1] [2] [3] [4] [5] [6]

• Training logic:

  • Updated the train methods for Florence-2, PaliGemma-2, and Qwen2.5-VL models to include the EarlyStoppingCallback when early stopping is enabled in the configuration. [1] [2] [3]

Related issue: #198

Type of change

Please delete options that are not relevant.

• New feature (non-breaking change which adds functionality)

How has this change been tested, please provide a testcase or example of how you tested the change?

• Example script:
  • Added a Python script (cookbooks/maestro_early_stopping_example.py) with examples of enabling early stopping for Florence-2, PaliGemma-2, and Qwen2.5-VL models.

Docs

• Docs updated? What were the changes:

• Notebook:
  • Added a Jupyter notebook (cookbooks/maestro_early_stopping.ipynb) that demonstrates how to use early stopping with the Florence-2 model. It includes setup, training, and visualization of training metrics.