Colab Notebook Development Pattern (v3.7.1)

CRITICAL RULE

ALWAYS copy the structure from notebooks/training.ipynb when creating new notebooks.

DO NOT:

• Improvise cell structure
• Combine setup cells
• Skip drive mount
• Use different extraction methods
• Assume paths work differently

Canonical Reference

The authoritative template is notebooks/training.ipynb. Every new notebook MUST follow its pattern for:

1. Part 1: Setup (Cells 1-6)

- GPU verification - Drive mount (separate cell) - Repository extraction (%%bash cell) - Path setup - Dependencies - API keys

1. Part 2+: Your Content

- Configuration cells - Data loading (use prefetchalldata) - Training/experiment loops - Results saving

Cell-by-Cell Template

Setup Cells (MANDATORY - Copy Exactly)

# Cell 1: GPU Verification
!nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv
import torch
assert torch.cuda.is_available(), "CUDA not available!"
# ... (see training.ipynb)

# Cell 2: Mount Drive
from google.colab import drive
drive.mount('/content/drive')

# Cell 3: Extract (%%bash magic)
%%bash
cd /content
if [ ! -d "Alpaca_trading" ]; then
    unzip -q /content/drive/MyDrive/Colab_Projects/Alpaca_trading.zip
fi

# Cell 4: Path Setup
import os, sys
os.chdir('/content/Alpaca_trading')
sys.path.insert(0, '/content/Alpaca_trading')

# Cell 5: Dependencies
%pip install -q torch gymnasium alpaca-py pandas numpy scipy arch

# Cell 6: API Keys - MUST use broker's parser (handles "Key:"/"Secret:" labels)
ALPACA_KEYS_FILE = '/content/Alpaca_trading/config/API_key_500Paper.txt'
if os.path.exists(ALPACA_KEYS_FILE):
    from alpaca_trading.trading.broker import _read_keys_from_file
    parsed = _read_keys_from_file(ALPACA_KEYS_FILE)
    if parsed.get('key') and parsed.get('secret'):
        os.environ['APCA_API_KEY_ID'] = parsed['key']
        os.environ['APCA_API_SECRET_KEY'] = parsed['secret']
# NEVER use naive line reading (lines[0], lines[1]) - labels become values!

Data Loading Pattern

# Use the same pattern as training.ipynb
from alpaca_trading.data.caching_fetcher import prefetch_all_data

CACHE_DIR = '/content/drive/MyDrive/Colab_Projects/training_data'
prefetched_data = prefetch_all_data(
    SYMBOLS,
    force_refresh=False,
    cache_dir=CACHE_DIR,
    min_bars=2000,
    keys_file=API_KEYS_FILE,
)

Output Directory Pattern

# ALWAYS create directories before saving
from pathlib import Path

OUTPUT_DIR = '/content/drive/MyDrive/Colab_Projects/my_experiment'
Path(OUTPUT_DIR).mkdir(parents=True, exist_ok=True)
Path(f'{OUTPUT_DIR}/models').mkdir(parents=True, exist_ok=True)

# Then save
trainer.save(f'{OUTPUT_DIR}/models/{symbol}.pt')

GPU Cleanup Pattern

# After each training run (from training.ipynb)
import gc
trainer.cleanup()
del trainer, env, prices
gc.collect()
torch.cuda.empty_cache()

Failed Attempts

Checklist for New Notebooks

• Cells 1-6 copied exactly from training.ipynb
• Drive mount is separate cell (not combined)
• %%bash used for extraction (not Python)
• API keys loaded from config/ files
• Output directories created with mkdir
• GPU cleanup after each training run
• prefetchalldata() used for data loading

Files

notebooks/training.ipynb         # CANONICAL TEMPLATE
notebooks/agent_validation_analysis.ipynb  # Example following pattern

References

• Skill: colab-unzip-workflow - File paths and API keys
• Skill: persistent-cache-gap-filling - Data caching configuration
• Skill: gpu-memory-cleanup - GPU cleanup patterns