File size: 3,313 Bytes
79e28b7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 | import argparse
import torch
import torch.nn as nn
def run_proof():
print("======================================================================")
print("ZYMATICA | Radical Coordinate Resonance Alignment (RCRA) Loss Proof")
print("======================================================================\n")
vocab_size = 128
batch_size = 4
K_TOP = 16 # K-Top parameter (simplified for demonstration)
coord_alpha = 0.8
print(f"[1] Instantiating Vocab Coordinate Radicals Map (size {vocab_size}x3)...")
# Setup coordinates: domain, subdomain, polarity
# Normalized between 0 and 1
torch.manual_seed(42)
coords_tensor = torch.rand((vocab_size, 3), dtype=torch.float32)
# 2. Setup synthetic forward pass outputs (logits and targets)
print(f"\n[2] Simulating Forward Pass Output Logits (requires_grad=True)...")
logits = torch.randn((batch_size, vocab_size), dtype=torch.float32, requires_grad=True)
targets = torch.randint(0, vocab_size, (batch_size,), dtype=torch.long)
print(f" - Logits shape: {logits.shape}")
print(f" - Targets: {targets.tolist()}")
# 3. Calculate Cross-Entropy Loss
print("\n[3] Computing Standard Cross-Entropy Loss...")
loss_ce_fct = nn.CrossEntropyLoss()
loss_ce = loss_ce_fct(logits, targets)
print(f" - Cross-Entropy Loss: {loss_ce.item():.4f}")
# 4. Calculate Radical Coordinate Resonance Loss (RCRA)
print("\n[4] Computing Cuneiform-U Radical Coordinate Resonance Loss...")
# Get top-K predicted logits and indices
topk_logits, topk_indices = torch.topk(logits, k=K_TOP, dim=-1)
probs = torch.softmax(topk_logits, dim=-1)
# Lookup coordinates of top-K predicted indices
# Shape: (batch_size, K, 3)
topk_coords = coords_tensor[topk_indices]
# Calculate predicted coordinates (weighted average)
# Shape: (batch_size, 1, 3) -> squeeze to (batch_size, 3)
pred_coords = torch.bmm(probs.unsqueeze(1), topk_coords).squeeze(1)
# Lookup target coordinates
# Shape: (batch_size, 3)
target_coords = coords_tensor[targets]
# Compute MSE loss over coordinates
loss_coord = torch.mean((pred_coords - target_coords) ** 2)
print(f" - Expected coordinate vectors (first batch): {pred_coords[0].tolist()}")
print(f" - Target coordinate vectors (first batch): {target_coords[0].tolist()}")
print(f" - Coordinate Resonance Loss: {loss_coord.item():.6f}")
# 5. Combine losses and backpropagate
print("\n[5] Combining Losses and Running Backpropagation...")
total_loss = loss_ce + coord_alpha * loss_coord
print(f" - Total Combined Loss: {total_loss.item():.4f}")
# Run backpropagation
total_loss.backward()
# Check if gradients flow back to logits successfully
grad_norm = logits.grad.norm().item()
print(f" - Logits gradient norm after backward: {grad_norm:.6f}")
assert grad_norm > 0, "Gradient flow failed! Logits received zero gradients."
print("\n[VERIFICATION] RCRA loss function and gradient flow verified.")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Zymatica RCRA Loss Proof")
parser.add_argument("--test", action="store_true", help="Run test mode")
args = parser.parse_args()
run_proof()
|