AmberLJC's picture
Upload experiment.py with huggingface_hub
86f312f verified
Raw
History Blame Contribute Delete
16.4 kB
"""
Gradient Clipping Experiment
This script demonstrates how gradient clipping stabilizes training by preventing
sudden large weight updates caused by rare, high-loss data points.
Experiment Setup:
- Simple model: Embedding(4, 16) -> Linear(16, 4)
- Vocabulary: ['A', 'B', 'C', 'D']
- Dataset: 1000 samples with imbalanced targets (990 'A', 10 'B')
- Compare training with and without gradient clipping
"""
import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np
import matplotlib.pyplot as plt
import random
# Set seeds for reproducibility
SEED = 42
def set_seeds(seed=SEED):
"""Set all random seeds for reproducibility."""
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
# =============================================================================
# 1. MODEL DEFINITION
# =============================================================================
class SimpleNextTokenModel(nn.Module):
"""
Simple model that takes a token index and predicts the next token.
Architecture: Embedding -> Linear
"""
def __init__(self, vocab_size=4, embedding_dim=16):
super().__init__()
self.embedding = nn.Embedding(vocab_size, embedding_dim)
self.linear = nn.Linear(embedding_dim, vocab_size)
def forward(self, x):
"""
Args:
x: Token indices of shape (batch_size,)
Returns:
Logits of shape (batch_size, vocab_size)
"""
embedded = self.embedding(x) # (batch_size, embedding_dim)
logits = self.linear(embedded) # (batch_size, vocab_size)
return logits
# =============================================================================
# 2. DATASET CREATION
# =============================================================================
def create_imbalanced_dataset(n_samples=1000, n_rare=10, seed=SEED):
"""
Create a synthetic dataset with imbalanced targets.
Args:
n_samples: Total number of samples
n_rare: Number of rare 'B' samples
seed: Random seed for reproducibility
Returns:
inputs: Random token indices (0-3)
targets: 990 'A' (0) and 10 'B' (1)
rare_indices: Indices where target is 'B'
"""
# Set seed for reproducibility
set_seeds(seed)
vocab = {'A': 0, 'B': 1, 'C': 2, 'D': 3}
# Random input tokens
inputs = torch.randint(0, 4, (n_samples,))
# Create imbalanced targets: mostly 'A' (0), few 'B' (1)
targets = torch.zeros(n_samples, dtype=torch.long) # All 'A' initially
# Randomly select indices for rare 'B' samples
rare_indices = random.sample(range(n_samples), n_rare)
targets[rare_indices] = 1 # Set to 'B'
return inputs, targets, sorted(rare_indices)
# =============================================================================
# 3. UTILITY FUNCTIONS
# =============================================================================
def compute_weight_norm(model):
"""Compute L2 norm of all model weights."""
total_norm = 0.0
for param in model.parameters():
total_norm += param.data.norm(2).item() ** 2
return total_norm ** 0.5
def get_initial_weights(seed=SEED):
"""Get initial weights for reproducible model initialization."""
set_seeds(seed)
model = SimpleNextTokenModel(vocab_size=4, embedding_dim=16)
return {name: param.clone() for name, param in model.state_dict().items()}
def train_epoch(model, optimizer, criterion, inputs, targets, clip_grad=False, max_norm=1.0):
"""
Train for one epoch, recording metrics at each step.
Args:
model: The neural network
optimizer: SGD optimizer
criterion: CrossEntropyLoss
inputs: Input token indices
targets: Target token indices
clip_grad: Whether to apply gradient clipping
max_norm: Maximum gradient norm (if clipping)
Returns:
losses: List of losses per step
grad_norms: List of gradient norms per step (before clipping)
weight_norms: List of weight norms per step
"""
model.train()
losses = []
grad_norms = []
weight_norms = []
# Train on each sample individually to see the effect of rare samples
for i in range(len(inputs)):
x = inputs[i:i+1] # Single sample
y = targets[i:i+1]
optimizer.zero_grad()
# Forward pass
logits = model(x)
loss = criterion(logits, y)
# Backward pass
loss.backward()
# Compute gradient norm BEFORE clipping
# Use a large value to just compute the norm without clipping
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), float('inf'))
# Apply gradient clipping if requested
if clip_grad:
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
# Update weights
optimizer.step()
# Record metrics
losses.append(loss.item())
grad_norms.append(grad_norm.item())
weight_norms.append(compute_weight_norm(model))
return losses, grad_norms, weight_norms
# =============================================================================
# 4. TRAINING FUNCTIONS
# =============================================================================
def run_training(inputs, targets, rare_indices, clip_grad=False, max_norm=1.0, n_epochs=3, lr=0.1, init_weights=None):
"""
Run complete training loop.
Args:
inputs: Input token indices
targets: Target token indices
rare_indices: Indices of rare 'B' samples
clip_grad: Whether to apply gradient clipping
max_norm: Maximum gradient norm threshold
n_epochs: Number of training epochs
lr: Learning rate
init_weights: Initial model weights for reproducibility
Returns:
all_losses, all_grad_norms, all_weight_norms: Metrics across all steps
"""
# Create fresh model with same initial weights
set_seeds(SEED)
model = SimpleNextTokenModel(vocab_size=4, embedding_dim=16)
if init_weights:
model.load_state_dict(init_weights)
optimizer = optim.SGD(model.parameters(), lr=lr)
criterion = nn.CrossEntropyLoss()
all_losses = []
all_grad_norms = []
all_weight_norms = []
mode = "WITH" if clip_grad else "WITHOUT"
print(f"\n{'='*60}")
print(f"Training {mode} gradient clipping (max_norm={max_norm})")
print(f"{'='*60}")
for epoch in range(n_epochs):
losses, grad_norms, weight_norms = train_epoch(
model, optimizer, criterion, inputs, targets,
clip_grad=clip_grad, max_norm=max_norm
)
all_losses.extend(losses)
all_grad_norms.extend(grad_norms)
all_weight_norms.extend(weight_norms)
avg_loss = np.mean(losses)
max_grad = np.max(grad_norms)
print(f"Epoch {epoch+1}/{n_epochs}: Avg Loss={avg_loss:.4f}, Max Grad Norm={max_grad:.4f}")
return all_losses, all_grad_norms, all_weight_norms
# =============================================================================
# 5. PLOTTING FUNCTIONS
# =============================================================================
def plot_metrics(losses, grad_norms, weight_norms, title, filename, rare_indices=None, n_samples=1000):
"""
Plot training metrics: loss, gradient norm, and weight norm.
Args:
losses: List of losses per step
grad_norms: List of gradient norms per step
weight_norms: List of weight norms per step
title: Plot title
filename: Output filename
rare_indices: Indices of rare 'B' samples (for highlighting)
n_samples: Number of samples per epoch
"""
fig, axes = plt.subplots(3, 1, figsize=(12, 10), sharex=True)
steps = range(len(losses))
n_epochs = len(losses) // n_samples
# Plot 1: Training Loss
axes[0].plot(steps, losses, 'b-', alpha=0.7, linewidth=0.5)
axes[0].set_ylabel('Training Loss', fontsize=12)
axes[0].set_title(title, fontsize=14, fontweight='bold')
axes[0].grid(True, alpha=0.3)
# Highlight rare sample positions
if rare_indices:
for epoch in range(n_epochs):
for idx in rare_indices:
step = epoch * n_samples + idx
if step < len(losses):
axes[0].axvline(x=step, color='red', alpha=0.3, linewidth=0.5)
# Plot 2: Gradient Norm
axes[1].plot(steps, grad_norms, 'g-', alpha=0.7, linewidth=0.5)
axes[1].set_ylabel('Gradient L2 Norm', fontsize=12)
axes[1].grid(True, alpha=0.3)
# Add horizontal line at clipping threshold
if "With" in title or "WITH" in title:
axes[1].axhline(y=1.0, color='red', linestyle='--', label='Clip threshold (1.0)')
axes[1].legend()
if rare_indices:
for epoch in range(n_epochs):
for idx in rare_indices:
step = epoch * n_samples + idx
if step < len(grad_norms):
axes[1].axvline(x=step, color='red', alpha=0.3, linewidth=0.5)
# Plot 3: Weight Norm
axes[2].plot(steps, weight_norms, 'm-', alpha=0.7, linewidth=0.5)
axes[2].set_ylabel('Weight L2 Norm', fontsize=12)
axes[2].set_xlabel('Training Step', fontsize=12)
axes[2].grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(filename, dpi=150, bbox_inches='tight')
plt.close()
print(f"Plot saved to: {filename}")
def plot_comparison(metrics_no_clip, metrics_with_clip, rare_indices, filename, n_samples=1000):
"""
Create side-by-side comparison plot.
Args:
metrics_no_clip: (losses, grad_norms, weight_norms) without clipping
metrics_with_clip: (losses, grad_norms, weight_norms) with clipping
rare_indices: Indices of rare 'B' samples
filename: Output filename
n_samples: Number of samples per epoch
"""
fig, axes = plt.subplots(3, 2, figsize=(16, 12))
losses_no, grads_no, weights_no = metrics_no_clip
losses_with, grads_with, weights_with = metrics_with_clip
steps = range(len(losses_no))
n_epochs = len(losses_no) // n_samples
# Column 1: Without Clipping
axes[0, 0].plot(steps, losses_no, 'b-', alpha=0.7, linewidth=0.5)
axes[0, 0].set_ylabel('Training Loss', fontsize=11)
axes[0, 0].set_title('WITHOUT Gradient Clipping', fontsize=13, fontweight='bold', color='red')
axes[0, 0].grid(True, alpha=0.3)
axes[1, 0].plot(steps, grads_no, 'g-', alpha=0.7, linewidth=0.5)
axes[1, 0].set_ylabel('Gradient L2 Norm', fontsize=11)
axes[1, 0].grid(True, alpha=0.3)
axes[2, 0].plot(steps, weights_no, 'm-', alpha=0.7, linewidth=0.5)
axes[2, 0].set_ylabel('Weight L2 Norm', fontsize=11)
axes[2, 0].set_xlabel('Training Step', fontsize=11)
axes[2, 0].grid(True, alpha=0.3)
# Column 2: With Clipping
axes[0, 1].plot(steps, losses_with, 'b-', alpha=0.7, linewidth=0.5)
axes[0, 1].set_title('WITH Gradient Clipping (max_norm=1.0)', fontsize=13, fontweight='bold', color='green')
axes[0, 1].grid(True, alpha=0.3)
axes[1, 1].plot(steps, grads_with, 'g-', alpha=0.7, linewidth=0.5)
axes[1, 1].axhline(y=1.0, color='red', linestyle='--', linewidth=2, label='Clip threshold')
axes[1, 1].legend(loc='upper right')
axes[1, 1].grid(True, alpha=0.3)
axes[2, 1].plot(steps, weights_with, 'm-', alpha=0.7, linewidth=0.5)
axes[2, 1].set_xlabel('Training Step', fontsize=11)
axes[2, 1].grid(True, alpha=0.3)
# Highlight rare sample positions in all plots
for col in range(2):
for row in range(3):
for epoch in range(n_epochs):
for idx in rare_indices:
step = epoch * n_samples + idx
if step < len(losses_no):
axes[row, col].axvline(x=step, color='red', alpha=0.2, linewidth=0.5)
# Add legend for rare samples
axes[0, 0].axvline(x=-100, color='red', alpha=0.5, linewidth=2, label="Rare 'B' samples")
axes[0, 0].legend(loc='upper right')
# Add overall title
fig.suptitle('Effect of Gradient Clipping on Training Stability\n(Red lines indicate rare "B" samples)',
fontsize=14, fontweight='bold', y=1.02)
plt.tight_layout()
plt.savefig(filename, dpi=150, bbox_inches='tight')
plt.close()
print(f"Comparison plot saved to: {filename}")
# =============================================================================
# 6. MAIN EXECUTION
# =============================================================================
def main():
print("="*60)
print("GRADIENT CLIPPING EXPERIMENT")
print("="*60)
print("\nThis experiment demonstrates how gradient clipping stabilizes")
print("training by preventing sudden large weight updates caused by")
print("rare, high-loss data points.\n")
# Create dataset ONCE (used for both runs)
inputs, targets, rare_indices = create_imbalanced_dataset(n_samples=1000, n_rare=10, seed=SEED)
print(f"Dataset created:")
print(f" Total samples: {len(inputs)}")
print(f" Target 'A' (0): {(targets == 0).sum().item()}")
print(f" Target 'B' (1): {(targets == 1).sum().item()}")
print(f" Rare 'B' indices: {rare_indices}")
# Get initial weights (same for both runs)
init_weights = get_initial_weights(seed=SEED)
# Run training WITHOUT gradient clipping
losses_no_clip, grads_no_clip, weights_no_clip = run_training(
inputs, targets, rare_indices,
clip_grad=False, n_epochs=3, lr=0.1, init_weights=init_weights
)
# Run training WITH gradient clipping
losses_with_clip, grads_with_clip, weights_with_clip = run_training(
inputs, targets, rare_indices,
clip_grad=True, max_norm=1.0, n_epochs=3, lr=0.1, init_weights=init_weights
)
# Generate individual plots
print("\n" + "="*60)
print("GENERATING PLOTS")
print("="*60)
plot_metrics(
losses_no_clip, grads_no_clip, weights_no_clip,
"Training WITHOUT Gradient Clipping",
"no_clipping.png",
rare_indices
)
plot_metrics(
losses_with_clip, grads_with_clip, weights_with_clip,
"Training WITH Gradient Clipping (max_norm=1.0)",
"with_clipping.png",
rare_indices
)
# Generate comparison plot
plot_comparison(
(losses_no_clip, grads_no_clip, weights_no_clip),
(losses_with_clip, grads_with_clip, weights_with_clip),
rare_indices,
"comparison.png"
)
# Print summary statistics
print("\n" + "="*60)
print("SUMMARY STATISTICS")
print("="*60)
print("\nWithout Gradient Clipping:")
print(f" Max Gradient Norm: {max(grads_no_clip):.4f}")
print(f" Mean Gradient Norm: {np.mean(grads_no_clip):.4f}")
print(f" Std Gradient Norm: {np.std(grads_no_clip):.4f}")
print(f" Final Weight Norm: {weights_no_clip[-1]:.4f}")
print(f" Final Loss: {losses_no_clip[-1]:.4f}")
print("\nWith Gradient Clipping (max_norm=1.0):")
print(f" Max Gradient Norm: {max(grads_with_clip):.4f}")
print(f" Mean Gradient Norm: {np.mean(grads_with_clip):.4f}")
print(f" Std Gradient Norm: {np.std(grads_with_clip):.4f}")
print(f" Final Weight Norm: {weights_with_clip[-1]:.4f}")
print(f" Final Loss: {losses_with_clip[-1]:.4f}")
# Return statistics for report
return {
'no_clip': {
'max_grad': max(grads_no_clip),
'mean_grad': np.mean(grads_no_clip),
'std_grad': np.std(grads_no_clip),
'final_weight': weights_no_clip[-1],
'final_loss': losses_no_clip[-1]
},
'with_clip': {
'max_grad': max(grads_with_clip),
'mean_grad': np.mean(grads_with_clip),
'std_grad': np.std(grads_with_clip),
'final_weight': weights_with_clip[-1],
'final_loss': losses_with_clip[-1]
},
'rare_indices': rare_indices
}
if __name__ == "__main__":
stats = main()
print("\n" + "="*60)
print("EXPERIMENT COMPLETE!")
print("="*60)