Summary

Enable timm-backed backbones for classification and switch to a config-driven multi-head classifier. This allows adding new attribute tasks by configuration without code changes, and exposes per-head metrics during training/evaluation.

Details

• Add timm to requirements.txt and document dependency in README.md
• Rewrite src/classification/models.py to instantiate backbones via timm.create_model (ResNet family) with options for pretrained weights, freezing, and feature extraction
• Implement config-driven multi-head classifier: parse per-attribute definitions in configs/classification_config.yaml (num_classes, loss, metrics) via resolve_classification_task_config and dynamic heads
• Generalize ClassificationTrainer to support arbitrary heads and log per-head metrics during train/validate
• Extend src/pipelines/pipeline.py to handle arbitrary number of heads/targets returned by timm models and log per-head metrics
• Update configs/classification_config.yaml with new schema and provide examples for vehicle_types and human_attributes
• Refresh docs/tutorials describing how to extend classification tasks via configuration
• Update README and docs to mention timm as dependency and new configuration-driven approach

Migration Notes

• Existing configuration files may require updates to align with the new classification schema. New tasks can be added by defining a new entry under classification.tasks.

Warning: Task VM test is not passing, cto.new will perform much better if you fix the setup