14. Colab Training

Notebook Location

• colab/cattle_buffalo_trainer.ipynb — main Colab notebook
• colab/cattle_buffalo_trainer.py — same content as percent-format script
• colab/README.md — setup instructions

Project Setup Options (in Colab)

1. GitHub clone (recommended):

   !rm -rf /content/project   # force-refresh to pick up latest code
   !git clone https://github.com/Bharaths31/ML-CB-B-identifier /content/project
   

   Important: Always rm -rf /content/project before cloning so that updated code (bug fixes, new features) is pulled correctly. Simply re-running the clone cell without deleting first silently keeps the old cached copy.

2. Upload colab_project.zip: created by python scripts/create_colab_project_zip.py
3. Google Drive mount: copy colab_project.zip from My Drive/ML-CB-B-identifier/

Dataset Options (in Colab)

1. Kaggle API / Direct Download: auto-downloads unified algsoch/breed-cattle-buffalo dataset containing both cattle and buffalo breeds
2. Upload archive.zip: created by python scripts/create_colab_archive.py
3. Google Drive: copy archive.zip from Drive

T4 GPU Optimizations

Setting	Value	Reason
batch_size	64	Maximizes T4 utilization (15 GB VRAM)
grad_accum	2	Effective batch = 128
num_workers	2	Colab has 2 CPU cores
prefetch_factor	4	Keeps GPU fed
pin_memory	True	Faster CPU→GPU transfer
AMP	phases 1-2 only	Disabled for QAT phase 3
Dataset location	/content/data/raw/	Local SSD, not Drive

Known Colab Gotchas

Symptom	Root Cause	Fix
AttributeError: 'torch._C._CudaDeviceProperties' object has no attribute 'total_mem'	PyTorch attribute is total_memory, not total_mem	Fixed in src/train.py:49 and colab/cattle_buffalo_trainer.py:29,900
Re-running GitHub clone cell doesn't pick up new code	Colab re-uses cached /content/project/ directory	Add !rm -rf /content/project before git clone in §1
Runtime restarts wipe all files	Colab free tier has ephemeral storage	Re-run all cells from §0 top-to-bottom after any restart

---