- Add required wandb integration with hardcoded API key
- Log training metrics (loss, accuracy, per-class metrics) to wandb
- In finetune mode, automatically add old codes to training set to prevent forgetting
- Zero-pad epoch numbers in model filenames (ep001, ep002) for better sorting
- Change upload.py to display sizes and speeds in KB instead of bytes
- Update Makefile finetune target to use 100 epochs instead of 30
- Set default wandb project to 'euphon/themblem'
