Training on Google Colab Skill

This skill enables efficient model training using Google Colab's GPU resources while maintaining code and data synchronization with the local project via Google Drive.

When to Use

• •Need GPU acceleration for model training
• •Training experiments that exceed local hardware capabilities
• •Long-running training jobs that benefit from cloud execution
• •Testing hyperparameter variations at scale

Prerequisites

1. •Google account with Colab Pro (recommended) or Colab Free
2. •Google Drive with project folder structure
3. •Training data uploaded to Google Drive
4. •Local project configured for Drive sync (see docs/reference/colab-pro-setup.md)

Workflow

Step 1: Prepare Training Configuration

Create or update a training config in training/configs/:

# training/configs/experiment_name.yaml
model:
  name: resnet50
  num_classes: 2
  pretrained: true

training:
  epochs: 100
  batch_size: 32
  learning_rate: 0.001
  optimizer: adam
  
data:
  train_path: /content/drive/MyDrive/traina/data/train
  val_path: /content/drive/MyDrive/traina/data/val
  
augmentation:
  enabled: true
  normalize: false  # Set based on production requirements

Step 2: Sync Code to Google Drive

Before syncing, perform a Pre-flight Check:

1. •Run a local smoke test (e.g., python smoke_test.py or set DRY_RUN=True in your notebook).
2. •Verify that 1 epoch runs on CPU with a few batches.
3. •Only sync after local verification passes.

# From project root
./training/scripts/sync_to_drive.sh

Step 3: Launch Colab Notebook

1. •Open training/notebooks/colab_training.ipynb in Google Colab
2. •Connect to GPU runtime (Runtime → Change runtime type → GPU)
3. •Mount Google Drive:
   python

   from google.colab import drive
   drive.mount('/content/drive')
   

4. •Execute training cells

Step 4: Download Results

After training completes:

# Sync trained models back from Drive
./training/scripts/sync_from_drive.sh \
  --source "drive/MyDrive/traina/experiments/" \
  --dest "training/experiments/"

Configuration Options

Environment Modes

GPU Optimization

# Enable mixed precision for faster training
from torch.cuda.amp import autocast, GradScaler

scaler = GradScaler()

for batch in dataloader:
    optimizer.zero_grad()
    
    with autocast():
        outputs = model(batch)
        loss = criterion(outputs, targets)
    
    scaler.scale(loss).backward()
    scaler.step(optimizer)
    scaler.update()

Best Practices

1. •

   Version Control

   • •Git tracks code and configs locally
   • •Drive stores datasets and model weights
   • •Never commit large .pt files to git
2. •

   Experiment Tracking

   • •Use descriptive experiment names
   • •Log all hyperparameters
   • •Save training curves and metrics
3. •

   Data Management

   • •Keep training data in Drive under versioned folders
   • •Use symbolic links for large datasets
   • •Validate data integrity before training
4. •

   Cost Optimization

   • •Colab Free: Limited GPU hours per day
   • •Colab Pro: Faster GPUs, longer sessions
   • •Monitor runtime to avoid losing progress
5. •

   "Run All" Configuration Pattern

   • •Use config cells at the top of notebooks with skip flags
   • •Allows selective experiment execution when using "Run All"
   • •Example pattern:
   python

   # === EXPERIMENT CONFIGURATION ===
   DRY_RUN = False           # Smoke test mode
   SKIP_EXPERIMENT_1 = True  # Skip if inconclusive/slow
   THRESHOLD_MODEL = 'production'  # Model selection
   

   • •Each experiment cell checks its skip flag before executing
   • •Ensures reproducible "Run All" behavior

Troubleshooting

Related Resources

• •PyTorch Training Tutorial
• •Colab Pro Setup Guide
• •Benchmark Latency Skill
• •Evaluate Model Skill