cs249r_book/tinytorch/tests/20_capstone/README.md

Capstone Integration Tests - Module 20

Comprehensive integration tests that validate the ENTIRE TinyTorch learning journey.

Overview

The capstone tests verify that all 19 previous modules work together to build production-ready ML systems.

Test Coverage

Priority 1: Complete ML Pipeline

• test_complete_ml_pipeline_end_to_end: Full data → model → training → evaluation
• Validates: Modules 01-08 integration

Priority 2: Model Architecture

• test_mlp_architecture_integration: Multi-layer perceptron
• test_cnn_architecture_integration: CNN with Conv2d, pooling, flatten
• test_transformer_architecture_integration: Attention, embeddings, positional encoding

Priority 3: Training Convergence

• test_xor_convergence: Classic XOR problem
• test_binary_classification_convergence: Real binary classification

Priority 4: Optimization & Deployment

• test_quantization_pipeline: INT8 quantization
• test_pruning_pipeline: Weight pruning
• test_combined_optimization_deployment: Quantization + pruning together

Priority 5: Gradient Flow & Performance

• test_deep_network_gradient_flow: Gradients through all layer types
• test_memory_efficiency: Reasonable memory usage
• test_training_performance: Training speed meets expectations

Running Tests

# Run all capstone tests
pytest tests/20_capstone/ -v

# Run specific test class
pytest tests/20_capstone/test_capstone_core.py::TestCompleteMLPipeline -v

Test Philosophy

Tests follow production ML workflow patterns:

1. Data Creation → Representative datasets
2. Model Building → Real architectures (MLP, CNN, Transformer)
3. Training → Actual convergence (loss decreases, accuracy improves)
4. Evaluation → Real metrics
5. Optimization → Production techniques (quantization, pruning)

Success Criteria

For capstone tests to pass, students must have:

1. Built all 19 modules correctly
2. Integrated modules properly
3. Implemented autograd correctly (gradients flow everywhere)
4. Created working optimizers
5. Validated on real tasks (models actually learn)

What This Tests That Unit Tests Don't

Aspect	Unit Tests	Capstone Tests
Scope	Single module	All 19 modules together
Integration	Module isolation	Cross-module integration
Real workflows	Synthetic checks	Production ML pipelines
Learning	Correctness only	Models must converge
Deployment	Not tested	Quantization, pruning tested