Verify Training Pipeline

Verify Training Pipeline

Systematically verify that a CVlization training example is complete, properly structured, and functional.

When to Use

• Validating a new or modified training example
• Debugging training pipeline issues
• Ensuring example completeness before commits
• Verifying example works after CVlization updates

Important Context

Shared GPU Environment: This machine may be used by multiple users simultaneously. Before running GPU-intensive training:

1. Check GPU memory availability with nvidia-smi
2. Wait for sufficient VRAM and low GPU utilization if needed
3. Consider stopping other processes if you have permission
4. If CUDA OOM errors occur, wait and retry when GPU is less busy

Verification Checklist

1. Structure Verification

Check that the example directory contains all required files:

# Navigate to example directory
cd examples/<capability>/<task>/<framework>/

# Expected structure:
# .
# ├── example.yaml        # Required: CVL metadata
# ├── Dockerfile          # Required: Container definition
# ├── build.sh            # Required: Build script
# ├── train.sh            # Required: Training script
# ├── train.py            # Required: Training code
# ├── README.md           # Recommended: Documentation
# ├── requirements.txt    # Optional: Python dependencies
# ├── data/               # Optional: Data directory
# └── outputs/            # Created at runtime

Key files to check:

• example.yaml - Must have: name, capability, stability, presets (build, train)
• Dockerfile - Should copy necessary files and install dependencies
• build.sh - Must set SCRIPT_DIR and call docker build
• train.sh - Must mount volumes correctly, pass environment variables, and forward CUDA_VISIBLE_DEVICES to the container (use ${CUDA_VISIBLE_DEVICES:+--env "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES"} so it's only set when the host has it set)
• .gitignore - Must cover all runtime artifacts (weights, checkpoints, results). Note: CVL_OUTPUTS maps to the workspace root (the example dir itself), NOT outputs/, so files like *.pt, result.txt etc. will land in the example directory directly. Verify with git status after a test run.

2. Build Verification

# Option 1: Build using script directly
./build.sh

# Option 2: Build using CVL CLI (recommended)
cvl run <example-name> build

# Verify image was created
docker images | grep <example-name>

# Expected: Image appears with recent timestamp

What to check:

• Build completes without errors (both methods)
• All dependencies install successfully
• Image size is reasonable (check for unnecessary files)
• cvl info <example-name> shows correct metadata

3. Training Verification

Start training and monitor for proper initialization:

# Option 1: Run training using script directly
./train.sh

# Option 2: Run training using CVL CLI (recommended)
cvl run <example-name> train

# With custom parameters (if supported)
BATCH_SIZE=2 NUM_EPOCHS=1 ./train.sh

Immediate checks (first 30-60 seconds):

• Container starts without errors
• Dataset loads successfully
• Model initializes (check GPU memory with nvidia-smi)
• Training loop begins (first batch processes)
• Logs are being written

4. Metrics Verification

Monitor metrics appropriate to the task type:

Generative Tasks (LLM, Text Generation, Image Generation)

• Primary metric: train/loss (should decrease over time)
• Target: Loss consistently decreasing, not NaN/Inf
• Typical range: Depends on task (LLM: 2-5 initial, <1 after convergence)
• Check for: Gradient explosions, NaN losses

# For LLM/generative models
tail -f logs/train.log | grep -i "loss\|iter\|step"

Classification Tasks (Image, Text, Document)

• Primary metrics: train/loss, train/accuracy, val/accuracy
• Target: Accuracy increasing, loss decreasing
• Typical range: Accuracy 0-100%, converges based on task difficulty
• Check for: Overfitting (train acc >> val acc)

# Watch accuracy metrics
tail -f lightning_logs/version_0/metrics.csv
# or for WandB
tail -f logs/train.log | grep -i "accuracy\|acc"

Object Detection Tasks

• Primary metrics: train/loss, val/map (mean Average Precision), val/map_50
• Target: mAP increasing, loss decreasing
• Typical range: mAP 0-100, good models achieve 30-90% depending on dataset
• Components: loss_classifier, loss_box_reg, loss_objectness, loss_rpn_box_reg

# Monitor detection metrics
tail -f logs/train.log | grep -i "map\|loss_classifier\|loss_box"

Segmentation Tasks (Semantic, Instance, Panoptic)

• Primary metrics: train/loss, val/iou (Intersection over Union), val/pixel_accuracy
• Target: IoU increasing (>0.5 is decent, >0.7 is good), loss decreasing
• Typical range: IoU 0-1, pixel accuracy 0-100%
• Variants: mIoU (mean IoU across classes)
• Panoptic-specific: PQ (Panoptic Quality), SQ (Segmentation Quality), RQ (Recognition Quality). PQ = SQ * RQ. Reported separately for "Things" (instances) and "Stuff" (amorphous regions). Pretrained COCO models may start with PQ ~28-30 on tiny subsets.

# Monitor segmentation metrics
tail -f lightning_logs/version_0/metrics.csv | grep -i "iou\|pixel"

Fine-tuning / Transfer Learning

• Primary metrics: train/loss, eval/loss, task-specific metrics
• Target: Both losses decreasing, eval loss not diverging from train loss
• Check for: Catastrophic forgetting, adapter convergence
• Special: For LoRA/DoRA, verify adapters are saved

# Check if adapters are being saved
ls -la outputs/*/lora_adapters/
# Should contain: adapter_config.json, adapter_model.safetensors

5. Runtime Checks

GPU VRAM Usage Monitoring (REQUIRED):

Before, during, and after training, actively monitor GPU VRAM usage:

# In another terminal, watch GPU memory in real-time
watch -n 1 nvidia-smi

# Or get detailed memory breakdown
nvidia-smi --query-gpu=index,name,memory.used,memory.total,memory.free,utilization.gpu --format=csv,noheader,nounits

# Record peak VRAM usage during training
nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | awk '{print $1 " MB"}'

Expected metrics:

• GPU memory usage: 60-95% of available VRAM (adjust batch size if 100% or <30%)
• GPU utilization: 70-100% during training steps
• Temperature: Stable (<85°C)
• Memory behavior: Should stabilize after model loading, spike during forward/backward passes

What to record for verification metadata:

• Peak VRAM usage in GB (e.g., "7.4GB VRAM" or "3.2GB VRAM")
• Percentage of total VRAM (e.g., "32%" for 7.4GB on 24GB GPU)
• GPU utilization percentage (e.g., "100% GPU utilization")

Troubleshooting:

• CUDA OOM: Reduce BATCH_SIZE, MAX_SEQ_LEN, or model size
• Low GPU utilization (<50%): Check data loading bottlenecks, increase batch size
• Memory keeps growing: Possible memory leak, check gradient accumulation

Docker Container Health:

# List running containers
docker ps

# Check logs for errors
docker logs <container-name-or-id>

# Verify mounts
docker inspect <container-id> | grep -A 10 Mounts
# Should see: workspace, cvlization_repo, huggingface cache

Output Directory:

# Check outputs are being written
ls -la outputs/ logs/ lightning_logs/
# Expected: Checkpoints, logs, or saved models appearing

# For WandB integration
ls -la wandb/
# Expected: run-<timestamp>-<id> directories

6. Lazy Downloading & Caching Verification

Verify that datasets and pretrained weights are cached properly:

# Check CVlization dataset cache
ls -la ~/.cache/cvlization/data/
# Expected: Dataset archives and extracted folders
# Examples: coco_panoptic_tiny/, stanford_background/, etc.

# Check framework-specific caches
ls -la ~/.cache/torch/hub/checkpoints/       # PyTorch pretrained weights
ls -la ~/.cache/huggingface/                 # HuggingFace models

# Verify no repeated downloads on second run
# First run: Should see "Downloading..." messages
./train.sh 2>&1 | tee first_run.log

# Clean workspace data (but keep cache)
rm -rf ./data/

# Second run: Should NOT download again, uses cache
./train.sh 2>&1 | tee second_run.log

# Verify no download messages in second run
grep -i "download" second_run.log
# Expected: Minimal or no download activity (weights already cached)

What to verify:

• Training data downloads to ~/.cache/cvlization/data/ (not ./data/)
• Pretrained weights cached by framework (PyTorch: ~/.cache/torch/, HuggingFace: ~/.cache/huggingface/)
• Second run reuses cached files without re-downloading
• Check train.py for data_dir parameter passed to dataset builders

7. Quick Validation Test

For fast verification (useful during development):

# Run 1 epoch with limited data
MAX_TRAIN_SAMPLES=10 NUM_EPOCHS=1 ./train.sh

# Expected runtime: 1-5 minutes
# Verify: Completes without errors, metrics logged

8. Update Verification Metadata

After successful verification, update the example.yaml with verification metadata:

First, check GPU info:

# Get GPU model and VRAM
nvidia-smi --query-gpu=name,memory.total --format=csv,noheader

Format:

verification:
  last_verified: 2025-10-25
  last_verification_note: "Verified build, training initialization, lazy downloading, and metrics logging on [GPU_MODEL] ([VRAM]GB VRAM)"

What to include in the note:

• What was verified: build, training, metrics
• Key aspects: lazy downloading, caching, GPU utilization
• GPU info: Dynamically determine GPU model and VRAM using nvidia-smi (e.g., "A10 GPU (24GB VRAM)", "RTX 4090 (24GB)")
  • If no GPU: Use "CPU-only"
• VRAM usage: Peak VRAM used during training (e.g., "GPU usage: 7.4GB VRAM (32%), 100% GPU utilization")
  • Get with: nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits
  • Convert to GB and calculate percentage of total VRAM
• Training extent: e.g., "1 epoch quick test" or "Full 10 epoch training"
• Any limitations: e.g., "CUDA OOM on full batch size"

Example complete entry:

name: pose-estimation-mmpose
docker: mmpose
capability: perception/pose_estimation
# ... other fields ...

verification:
  last_verified: 2025-10-25
  last_verification_note: "Verified build, CVL CLI integration, and lazy downloading to ~/.cache/cvlization/data/. Training not fully verified due to GPU memory constraints (CUDA OOM on shared GPU)."

When to update:

• After completing full verification checklist (steps 1-7)
• Partial verification is acceptable - note what was verified
• When re-verifying after CVlization updates or fixes

Common Issues and Fixes

Build Failures

# Issue: Dockerfile can't find files
# Fix: Check COPY paths are relative to Dockerfile location

# Issue: Dependency conflicts
# Fix: Check requirements.txt versions, update base image

# Issue: Large build context
# Fix: Add .dockerignore file

Training Failures

# Issue: CUDA out of memory
# Fix: Reduce BATCH_SIZE, MAX_SEQ_LEN, or image size

# Issue: Dataset not found
# Fix: Check data/ directory exists, run data preparation script

# Issue: Permission denied on outputs
# Fix: Ensure output directories are created before docker run

Metric Issues

# Issue: Loss is NaN
# Fix: Reduce learning rate, check data normalization, verify labels

# Issue: No metrics logged
# Fix: Check training script has logging configured (wandb/tensorboard)

# Issue: Loss not decreasing
# Fix: Verify learning rate, check data quality, increase epochs

Example Commands

Perception - Object Detection

cd examples/perception/object_detection/torchvision
./build.sh
./train.sh
# Monitor: train/loss, val/map, val/map_50
# Success: mAP > 0.3 after a few epochs

Perception - Semantic Segmentation

cd examples/perception/segmentation/semantic_torchvision
./build.sh
./train.sh
# Monitor: train/loss, val/iou, val/pixel_accuracy
# Success: IoU > 0.5, pixel_accuracy > 80%

Generative - LLM Training

cd examples/generative/llm/nanogpt
./build.sh
./train.sh
# Monitor: train/loss, val/loss, iter time
# Success: Loss decreasing from ~4.0 to <2.0

Document AI - Fine-tuning

cd examples/perception/doc_ai/granite_docling_finetune
./build.sh
MAX_TRAIN_SAMPLES=20 NUM_EPOCHS=1 ./train.sh
# Monitor: train/loss, eval/loss
# Success: Both losses decrease, adapters saved to outputs/

CVL Integration

These examples integrate with the CVL command system:

# List all available examples
cvl list

# Get example info
cvl info granite_docling_finetune

# Run example directly (uses example.yaml presets)
cvl run granite_docling_finetune build
cvl run granite_docling_finetune train

Success Criteria

A training pipeline passes verification when:

1. ✅ Structure: All required files present, example.yaml valid
2. ✅ Build: Docker image builds without errors (both ./build.sh and cvl run <name> build)
3. ✅ Start: Training starts, dataset loads, model initializes (both ./train.sh and cvl run <name> train)
4. ✅ Metrics Improve: Training loss decreases OR model accuracy/mAP/IoU improves over epochs
5. ✅ Central Caching: Training data cached to ~/.cache/cvlization/data/ (NOT to local ./data/), pretrained weights cached to framework-specific locations (~/.cache/torch/, ~/.cache/huggingface/)
6. ✅ Lazy Downloading: Datasets and pretrained weights download only when needed, avoiding repeated downloads on subsequent runs
7. ✅ Outputs: Checkpoints/adapters/logs saved correctly; .gitignore covers all runtime artifacts (weights land in example dir root via CVL_OUTPUTS, not outputs/)
8. ✅ CVL CLI: cvl info <name> shows correct metadata, build and train presets work
9. ✅ Documentation: README explains how to use the example
10. ✅ Verification Metadata: example.yaml updated with verification field containing last_verified date and last_verification_note

Related Files

Check these files for debugging:

• train.py - Core training logic
• Dockerfile - Environment setup
• requirements.txt - Python dependencies
• example.yaml - CVL metadata and presets
• README.md - Usage instructions

Tips

• Use MAX_TRAIN_SAMPLES=<small_number> for fast validation
• Monitor GPU memory with nvidia-smi in separate terminal
• Check docker logs <container> if training hangs
• For WandB integration, set WANDB_API_KEY environment variable
• Most examples support environment variable overrides (check train.sh)
• Test wandb without API key: Use WANDB_MODE=offline ./train.sh --track ... to verify wandb hooks fire correctly without needing a real API key. Check for wandb: Synced N W&B file(s), M media file(s) in the output to confirm images and scalars were logged.
• Root-owned files from Docker: Files created inside Docker containers (e.g., cached datasets, generated masks, output checkpoints) are owned by root and can't be deleted from the host without sudo. To clean them up: docker run --rm --mount "type=bind,src=${HOME}/.cache/cvlization,dst=/data" <image> rm -rf /data/path/to/stale/files

Source: kungfuai/cvlization — distributed by TomeVault.