mirror of
https://github.com/MLSysBook/TinyTorch.git
synced 2026-04-28 23:18:22 -05:00
c7dbf68dcf
Major fixes for complete training pipeline functionality: Core Components Fixed: - Parameter class: Now wraps Variables with requires_grad=True for proper gradient tracking - Variable.sum(): Essential for scalar loss computation from multi-element tensors - Gradient handling: Fixed memoryview issues in autograd and activations - Tensor indexing: Added __getitem__ support for weight inspection Training Results: - XOR learning: 100% accuracy (4/4) - network successfully learns XOR function - Linear regression: Weight=1.991 (target=2.0), Bias=0.980 (target=1.0) - Integration tests: 21/22 passing (95.5% success rate) - Module tests: All individual modules passing - General functionality: 4/5 tests passing with core training working Technical Details: - Fixed gradient data access patterns throughout activations.py - Added safe memoryview handling in Variable.backward() - Implemented proper Parameter-Variable delegation - Added Tensor subscripting for debugging access(https://claude.ai/code)
87 lines
2.7 KiB
Python
87 lines
2.7 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
Test Loss Connection to Computational Graph
|
|
"""
|
|
|
|
import numpy as np
|
|
import sys
|
|
import os
|
|
|
|
# Add TinyTorch to path
|
|
sys.path.insert(0, os.path.join(os.path.dirname(__file__), 'tinytorch'))
|
|
|
|
from tinytorch.core.autograd import Variable
|
|
from tinytorch.core.layers import matmul
|
|
|
|
print("🔍 TESTING LOSS CONNECTION TO COMPUTATIONAL GRAPH")
|
|
print("=" * 50)
|
|
|
|
def test_loss_connection():
|
|
"""Test how to properly create scalar loss that maintains graph."""
|
|
print("\n1. Testing Different Loss Creation Methods...")
|
|
|
|
x = Variable([[2.0, 3.0]], requires_grad=True)
|
|
w = Variable([[0.1], [0.2]], requires_grad=True)
|
|
|
|
# Matrix multiplication
|
|
result = matmul(x, w) # [[0.8]]
|
|
print(f"result: {result}")
|
|
print(f"result.grad_fn: {result.grad_fn}")
|
|
|
|
# Method 1: Extract scalar and create new Variable (BROKEN)
|
|
print("\n--- Method 1: Extract scalar and create new Variable ---")
|
|
x.grad = None
|
|
w.grad = None
|
|
|
|
scalar_value = np.sum(result.data.data if hasattr(result.data, 'data') else result.data)
|
|
loss1 = Variable(scalar_value)
|
|
print(f"loss1: {loss1}")
|
|
print(f"loss1.grad_fn: {loss1.grad_fn}")
|
|
|
|
loss1.backward()
|
|
print(f"Method 1 - x.grad: {x.grad}, w.grad: {w.grad}")
|
|
|
|
# Method 2: Use Variable operations to create scalar (SHOULD WORK)
|
|
print("\n--- Method 2: Use Variable operations to create scalar ---")
|
|
x.grad = None
|
|
w.grad = None
|
|
|
|
# Sum using Variable operations (maintains graph)
|
|
if hasattr(Variable, 'sum'):
|
|
loss2 = Variable.sum(result)
|
|
else:
|
|
# Create our own sum operation
|
|
loss2 = result # If result is already scalar, use it directly
|
|
if result.data.data.size > 1:
|
|
# If result has multiple elements, sum manually while preserving graph
|
|
# For now, just use the first element if it's scalar
|
|
pass
|
|
|
|
print(f"loss2: {loss2}")
|
|
print(f"loss2.grad_fn: {loss2.grad_fn}")
|
|
|
|
loss2.backward()
|
|
print(f"Method 2 - x.grad: {x.grad}, w.grad: {w.grad}")
|
|
|
|
# Method 3: Direct backward on result (KNOWN TO WORK)
|
|
print("\n--- Method 3: Direct backward on result ---")
|
|
x.grad = None
|
|
w.grad = None
|
|
|
|
result.backward()
|
|
print(f"Method 3 - x.grad: {x.grad}, w.grad: {w.grad}")
|
|
|
|
def test_sum_operation():
|
|
"""Test if Variable has sum operation."""
|
|
print("\n2. Testing Variable Sum Operation...")
|
|
|
|
# Check if Variable class has sum method
|
|
print(f"Variable has sum method: {hasattr(Variable, 'sum')}")
|
|
|
|
# Try to create a Variable and see what methods it has
|
|
v = Variable([[1.0, 2.0, 3.0]])
|
|
print(f"Variable methods: {[method for method in dir(v) if not method.startswith('_')]}")
|
|
|
|
if __name__ == "__main__":
|
|
test_loss_connection()
|
|
test_sum_operation() |