Resnet50/mseloss

[ghost]

Refactor: Dataset Structure, Transformations & Interface Improvements

This PR introduces several updates to the dataset pipeline, transformations, and model interface. Below are the main changes:

Dataset and Loader

• Dataset folder must now follow the structure:
  data/
  ├── abstrait/
  └── figuratif/

(Replaces previous folders like abstrait-v2, figuratif - aleat, etc.)

• PaintingsDataset:
• Now supports per-class augmentations via custom_augment_abstrait and custom_augment_figuratif.
• double_abstract=True duplicates abstract samples to include both original and transformed versions in training.
• Added n_transforms_augmented for dynamic random augmentation composition from a transform pool.
• Padding logic now supports zero, mirror, and replicate modes.
• Handles center crop and resizing for contrastive learning setups (transform=True).

Transformations

• Added transformations/ folder to organize custom transformations:
• SkewTransform
• RandomStretch
• These are used for abstract data augmentation and integrated into the default transform pool.

Dataset Split Script

• New script: get_new_data.py
• Splits a raw dataset into train/val/test.
• Saves images in the correct folder structure (new_dataset/train/abstrait, etc.).
• Applies default or custom augmentation to abstract images if provided.

Interface Enhancements

• Interface class updated with new arguments:
• load_model_path to resume training
• freeze_layers for partial fine-tuning of ResNet50
• weighted_loss to manually override automatic class weight calculation
• Integration with per-class augmentations and double_abstract

Example Notebook

• Added example usage in research/joel/20250617_weights.ipynb:
• Demonstrates how to use the updated PaintingsDataset with custom transforms
• Shows how to load the model and run embedding projections