CaliBench / SMART /toy_example /softece_Experiment.py
zhurong2333's picture
Add files using upload-large-folder tool
0293aec verified
Raw
History Blame Contribute Delete
43.1 kB
import sys
import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt
from tqdm import tqdm
import json
import time
# Add the parent directory to the path to import from PureLogits
current_dir = os.path.dirname(os.path.abspath(__file__))
parent_dir = os.path.dirname(current_dir)
grandparent_dir = os.path.dirname(parent_dir)
sys.path.append(grandparent_dir)
sys.path.append(parent_dir)
# Import necessary components from the calibrator
from calibrator.Component.metrics import (
BrierLoss, CrossEntropyLoss, MSELoss, SoftECE, ECE
)
# Set random seed for reproducibility
def set_seed(seed):
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Load logits and labels from cache directory
def load_data(cache_dir):
val_logits = np.load(os.path.join(cache_dir, "val_logits.npy"))
val_labels = np.load(os.path.join(cache_dir, "val_labels.npy"))
test_logits = np.load(os.path.join(cache_dir, "test_logits.npy"))
test_labels = np.load(os.path.join(cache_dir, "test_labels.npy"))
print(f"Loaded data: val_logits shape: {val_logits.shape}, val_labels shape: {val_labels.shape}")
print(f"Loaded data: test_logits shape: {test_logits.shape}, test_labels shape: {test_labels.shape}")
return val_logits, val_labels, test_logits, test_labels
# Compute hardness as the gap between top logit and runner-up
def compute_hardness(logits):
logits_tensor = torch.tensor(logits, dtype=torch.float32)
sorted_logits, _ = torch.sort(logits_tensor, descending=True)
logit_gap = (sorted_logits[0] - sorted_logits[1]).item()
return logit_gap
# Custom ECE wrapper to ensure it returns a tensor that can be backpropagated
class ECEWrapper(nn.Module):
def __init__(self, n_bins=15):
super(ECEWrapper, self).__init__()
self.ece = ECE(n_bins=n_bins)
def forward(self, logits, labels):
# ECE returns a scalar value, we need to wrap it in a tensor with requires_grad=True
ece_value = self.ece(logits, labels)
# Check if the result is already a tensor with grad
if isinstance(ece_value, torch.Tensor) and ece_value.requires_grad:
return ece_value
# Otherwise, create a tensor with requires_grad=True
if isinstance(ece_value, torch.Tensor):
return ece_value.clone().detach().requires_grad_(True)
else:
return torch.tensor(ece_value, requires_grad=True, device=logits.device)
# Custom Brier Loss implementation that ensures it's different from MSE
class CustomBrierLoss(nn.Module):
def __init__(self):
super(CustomBrierLoss, self).__init__()
def forward(self, logits, labels):
# Get predicted probabilities
outputs = F.softmax(logits, dim=1)
# Convert labels to one-hot
one_hot = torch.zeros(labels.size(0), outputs.size(1), device=labels.device)
one_hot.scatter_(1, labels.unsqueeze(1), 1)
# Compute Brier loss
brier_score = torch.mean(torch.sum((outputs - one_hot) ** 2, dim=1))
# Explicitly scale by 0.5 to differentiate from MSE
return brier_score * 0.5
# Experiment 1: Compare gradient stability of different calibration objectives
def experiment_1_gradient_stability(test_logits, test_labels, cache_dir, output_dir="results"):
print("\nExperiment 1: Comparing Gradient Stability of Different Calibration Objectives")
# Create output directory
os.makedirs(output_dir, exist_ok=True)
# Convert data to PyTorch tensors
logits_tensor = torch.tensor(test_logits, dtype=torch.float32)
labels_tensor = torch.tensor(test_labels, dtype=torch.long)
# Initialize different loss functions
soft_ece = SoftECE(n_bins=15)
traditional_ece = ECEWrapper(n_bins=15) # Use wrapper for ECE
brier_score = BrierLoss()
cross_entropy = CrossEntropyLoss()
mse_loss = MSELoss()
# Temperature parameter with gradient tracking
temperature = nn.Parameter(torch.ones(1))
# Sample a subset of data for visualization (to avoid clutter)
num_samples = 1000
indices = np.random.choice(len(test_logits), num_samples, replace=False)
# Confidence bins for analysis
confidence_bins = np.linspace(0.5, 1.0, 10)
bin_width = confidence_bins[1] - confidence_bins[0]
# Store gradients for each loss function across confidence levels
results = {
"SoftECE": {"grads": [], "conf_bins": []},
"ECE": {"grads": [], "conf_bins": []},
"Brier": {"grads": [], "conf_bins": []},
"CrossEntropy": {"grads": [], "conf_bins": []},
"MSE": {"grads": [], "conf_bins": []}
}
# Calculate gradients for different temperature values
temp_values = [0.5, 0.8, 1.0, 1.2, 1.5, 2.0]
for temp_val in temp_values:
print(f"\nAnalyzing gradients at temperature = {temp_val}")
# Set temperature to current value
with torch.no_grad():
temperature.fill_(temp_val)
# Process samples in batches to avoid memory issues
batch_size = 100
num_batches = (num_samples + batch_size - 1) // batch_size
for batch_idx in tqdm(range(num_batches)):
start_idx = batch_idx * batch_size
end_idx = min((batch_idx + 1) * batch_size, num_samples)
batch_indices = indices[start_idx:end_idx]
batch_logits = logits_tensor[batch_indices]
batch_labels = labels_tensor[batch_indices]
# Apply temperature scaling
scaled_logits = batch_logits / temperature
probs = F.softmax(scaled_logits, dim=1)
# Calculate confidence (max probability)
confidences = torch.max(probs, dim=1)[0].detach().numpy()
# Calculate gradients for each loss function
for loss_name, loss_fn in [
("SoftECE", soft_ece),
("ECE", traditional_ece),
("Brier", brier_score),
("CrossEntropy", cross_entropy),
("MSE", mse_loss)
]:
temperature.grad = None
if loss_name == "SoftECE":
loss = loss_fn(scaled_logits, batch_labels)
elif loss_name == "ECE":
loss = loss_fn(scaled_logits, batch_labels)
elif loss_name == "Brier":
loss = loss_fn(scaled_logits, batch_labels)
elif loss_name == "CrossEntropy":
loss = loss_fn(scaled_logits, batch_labels)
elif loss_name == "MSE":
loss = loss_fn(scaled_logits, batch_labels)
loss.backward(retain_graph=True)
# Store gradient magnitude for each sample
if temperature.grad is not None:
grad_magnitude = temperature.grad.item()
# Store gradients by confidence bin
for i, conf in enumerate(confidences):
bin_idx = np.digitize(conf, confidence_bins) - 1
if 0 <= bin_idx < len(confidence_bins):
results[loss_name]["grads"].append(grad_magnitude)
results[loss_name]["conf_bins"].append(confidence_bins[bin_idx])
# Plot gradient magnitudes across confidence levels
plt.figure(figsize=(10, 6))
for loss_name in ["SoftECE", "ECE", "Brier", "CrossEntropy", "MSE"]:
# Group gradients by confidence bin
binned_grads = {}
for grad, bin_val in zip(results[loss_name]["grads"], results[loss_name]["conf_bins"]):
if bin_val not in binned_grads:
binned_grads[bin_val] = []
binned_grads[bin_val].append(grad)
# Calculate mean and std of gradients for each bin
bin_centers = []
mean_grads = []
std_grads = []
for bin_val in sorted(binned_grads.keys()):
if binned_grads[bin_val]: # Ensure there are gradients for this bin
bin_centers.append(bin_val)
mean_grads.append(np.mean(np.abs(binned_grads[bin_val])))
std_grads.append(np.std(np.abs(binned_grads[bin_val])))
# Plot mean gradient magnitude with error bars
plt.errorbar(bin_centers, mean_grads, yerr=std_grads, label=loss_name, marker='o', capsize=4)
plt.xlabel('Confidence Level')
plt.ylabel('Gradient Magnitude')
plt.title('Gradient Stability Across Confidence Levels')
plt.legend()
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'gradient_stability.png'), dpi=300)
plt.close()
# Save results
with open(os.path.join(output_dir, 'gradient_stability_results.json'), 'w') as f:
# Convert numpy arrays to lists for JSON serialization
serializable_results = {}
for loss_name, data in results.items():
serializable_results[loss_name] = {
"grads": [float(g) for g in data["grads"]],
"conf_bins": [float(c) for c in data["conf_bins"]]
}
json.dump(serializable_results, f, indent=2)
print(f"Gradient stability analysis complete. Results saved to {output_dir}")
# Experiment 2: Gradient information content analysis
def experiment_2_gradient_information(test_logits, test_labels, cache_dir, output_dir="results"):
print("\nExperiment 2: Analyzing Gradient Information Content of Different Calibration Objectives")
# Create output directory
os.makedirs(output_dir, exist_ok=True)
# Convert data to PyTorch tensors
logits_tensor = torch.tensor(test_logits, dtype=torch.float32)
labels_tensor = torch.tensor(test_labels, dtype=torch.long)
# Sample a subset of data
num_samples = 2000
indices = np.random.choice(len(test_logits), num_samples, replace=False)
# Initialize different loss functions
soft_ece = SoftECE(n_bins=15, sigma=0.05)
traditional_ece = ECEWrapper(n_bins=15)
brier_score = BrierLoss()
cross_entropy = CrossEntropyLoss()
mse_loss = MSELoss()
# Create synthetic miscalibration by distorting temperatures
# We'll create 5 different scenarios ranging from underconfident to overconfident
temp_scenarios = [0.5, 0.8, 1.0, 1.2, 2.0]
# Store gradient direction consistency and correlation with calibration error
results = {
"temp_scenario": [],
"ece_values": [],
"gradient_consistency": {},
"gradient_correlation": {},
"gradient_entropy": {}
}
for loss_name in ["SoftECE", "ECE", "Brier", "CrossEntropy", "MSE"]:
results["gradient_consistency"][loss_name] = []
results["gradient_correlation"][loss_name] = []
results["gradient_entropy"][loss_name] = []
# Create confidence bins for analysis
confidence_bins = np.linspace(0.1, 1.0, 10)
for temperature_value in temp_scenarios:
print(f"\nAnalyzing scenario with temperature = {temperature_value}")
# Set up temperature parameter for this scenario
temperature = nn.Parameter(torch.ones(1) * temperature_value)
# Generate scaled logits for the whole dataset to calculate overall ECE
with torch.no_grad():
overall_scaled_logits = logits_tensor / temperature
overall_probs = F.softmax(overall_scaled_logits, dim=1)
ece_metric = ECE(n_bins=15)
ece_value = ece_metric(overall_scaled_logits, labels_tensor).item()
results["temp_scenario"].append(temperature_value)
results["ece_values"].append(ece_value)
print(f"Overall ECE for temperature {temperature_value}: {ece_value:.4f}")
# For each loss function, analyze gradient properties
for loss_name, loss_fn in [
("SoftECE", soft_ece),
("ECE", traditional_ece),
("Brier", brier_score),
("CrossEntropy", cross_entropy),
("MSE", mse_loss)
]:
print(f"Analyzing gradients for {loss_name}")
# Calculate gradients for each sample in batches
batch_size = 100
num_batches = (num_samples + batch_size - 1) // batch_size
# Store all gradients and corresponding confidences
all_grads = []
all_confidences = []
all_is_correct = []
for batch_idx in tqdm(range(num_batches)):
start_idx = batch_idx * batch_size
end_idx = min((batch_idx + 1) * batch_size, num_samples)
batch_indices = indices[start_idx:end_idx]
batch_logits = logits_tensor[batch_indices]
batch_labels = labels_tensor[batch_indices]
# Calculate gradients per sample
for i in range(len(batch_logits)):
sample_logits = batch_logits[i:i+1]
sample_label = batch_labels[i:i+1]
# Reset parameter to temperature_value for each sample
temperature = nn.Parameter(torch.ones(1) * temperature_value)
# Apply temperature scaling
scaled_logits = sample_logits / temperature
probs = F.softmax(scaled_logits, dim=1)
# Check if prediction is correct
pred = torch.argmax(probs, dim=1)
is_correct = (pred == sample_label).float().item()
# Calculate confidence
confidence = torch.max(probs).item()
# Calculate loss and gradient
temperature.grad = None
try:
loss = loss_fn(scaled_logits, sample_label)
loss.backward()
# Only store if gradient was successfully calculated
if temperature.grad is not None:
all_grads.append(temperature.grad.item())
all_confidences.append(confidence)
all_is_correct.append(is_correct)
except Exception as e:
# Skip samples that cause errors in gradient calculation
print(f" Skipping sample due to error: {e}")
continue
# Skip further analysis if we don't have enough gradient data
if len(all_grads) < 10:
print(f" Not enough gradient data for {loss_name}, skipping analysis")
continue
# Calculate gradient consistency per confidence bin
# (correlation between gradient direction and correctness)
bin_grad_consistency = []
bin_grad_entropy = []
for i in range(len(confidence_bins) - 1):
bin_start = confidence_bins[i]
bin_end = confidence_bins[i+1]
# Find samples in this bin
bin_indices = []
for j in range(len(all_confidences)):
if bin_start <= all_confidences[j] < bin_end:
bin_indices.append(j)
if len(bin_indices) > 5: # Only consider bins with enough samples
bin_grads = [all_grads[j] for j in bin_indices]
bin_correct = [all_is_correct[j] for j in bin_indices]
# Calculate consistency: avg sign of gradient * (conf - accuracy)
bin_conf = np.mean([all_confidences[j] for j in bin_indices])
bin_acc = np.mean(bin_correct)
calibration_error = bin_conf - bin_acc
# Compute correlation between gradient and calibration error
# A well-behaved gradient should be negative when confidence > accuracy
# and positive when confidence < accuracy
grad_signs = np.sign(bin_grads)
ce_sign = np.sign(calibration_error)
correct_direction = -1 * ce_sign # Gradient should point in opposite direction of error
consistency = np.mean(grad_signs == correct_direction)
bin_grad_consistency.append(consistency)
# Calculate entropy of gradient distribution in this bin
# Normalize gradients for comparison
if len(bin_grads) > 0 and np.std(bin_grads) > 0:
norm_grads = (bin_grads - np.mean(bin_grads)) / np.std(bin_grads)
# Use histogram to estimate entropy
hist, _ = np.histogram(norm_grads, bins=10, density=True)
hist = hist[hist > 0] # Avoid log(0)
entropy = -np.sum(hist * np.log(hist))
bin_grad_entropy.append(entropy)
# Store results
if bin_grad_consistency:
avg_consistency = np.mean(bin_grad_consistency)
results["gradient_consistency"][loss_name].append(avg_consistency)
print(f" {loss_name} gradient consistency: {avg_consistency:.4f}")
if bin_grad_entropy:
avg_entropy = np.mean(bin_grad_entropy)
results["gradient_entropy"][loss_name].append(avg_entropy)
print(f" {loss_name} gradient entropy: {avg_entropy:.4f}")
# Calculate correlation between gradient and calibration error
# For overall dataset
if len(all_grads) > 0:
accuracies = np.array(all_is_correct)
confidences = np.array(all_confidences)
grads = np.array(all_grads)
# Ensure we have valid data
if len(grads) > 0 and not np.all(np.isnan(grads)):
cal_errors = confidences - accuracies
# Correlation should be negative (gradient points opposite to error)
correlation = np.corrcoef(cal_errors, -grads)[0, 1]
if not np.isnan(correlation):
results["gradient_correlation"][loss_name].append(correlation)
print(f" {loss_name} gradient-error correlation: {correlation:.4f}")
# Plot results
# 1. Plot gradient consistency across different calibration scenarios
plt.figure(figsize=(10, 6))
for loss_name in ["SoftECE", "ECE", "Brier", "CrossEntropy", "MSE"]:
if loss_name in results["gradient_consistency"] and len(results["gradient_consistency"][loss_name]) > 0:
plt.plot(results["temp_scenario"], results["gradient_consistency"][loss_name],
marker='o', label=loss_name)
plt.xlabel('Temperature (Lower = More Overconfident)')
plt.ylabel('Gradient Direction Consistency')
plt.title('Gradient Consistency Across Calibration Scenarios')
plt.legend()
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'gradient_consistency.png'), dpi=300)
plt.close()
# 2. Plot gradient-error correlation
plt.figure(figsize=(10, 6))
for loss_name in ["SoftECE", "ECE", "Brier", "CrossEntropy", "MSE"]:
if loss_name in results["gradient_correlation"] and len(results["gradient_correlation"][loss_name]) > 0:
plt.plot(results["temp_scenario"], results["gradient_correlation"][loss_name],
marker='o', label=loss_name)
plt.xlabel('Temperature (Lower = More Overconfident)')
plt.ylabel('Correlation between Gradient and Calibration Error')
plt.title('Gradient-Error Correlation Across Calibration Scenarios')
plt.legend()
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'gradient_correlation.png'), dpi=300)
plt.close()
# 3. Plot gradient entropy (information content)
plt.figure(figsize=(10, 6))
for loss_name in ["SoftECE", "ECE", "Brier", "CrossEntropy", "MSE"]:
if loss_name in results["gradient_entropy"] and len(results["gradient_entropy"][loss_name]) > 0:
plt.plot(results["temp_scenario"], results["gradient_entropy"][loss_name],
marker='o', label=loss_name)
plt.xlabel('Temperature (Lower = More Overconfident)')
plt.ylabel('Gradient Distribution Entropy')
plt.title('Gradient Information Content Across Calibration Scenarios')
plt.legend()
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'gradient_entropy.png'), dpi=300)
plt.close()
# Save results
with open(os.path.join(output_dir, 'gradient_information_results.json'), 'w') as f:
# Ensure all data is serializable
serializable_results = {
"temp_scenario": list(results["temp_scenario"]),
"ece_values": list(results["ece_values"]),
"gradient_consistency": {k: list(v) for k, v in results["gradient_consistency"].items()},
"gradient_correlation": {k: list(v) for k, v in results["gradient_correlation"].items()},
"gradient_entropy": {k: list(v) for k, v in results["gradient_entropy"].items()}
}
json.dump(serializable_results, f, indent=2)
print(f"Gradient information analysis complete. Results saved to {output_dir}")
# Experiment 3: Sample efficiency and overfitting resistance analysis
def experiment_3_sample_efficiency(val_logits, val_labels, test_logits, test_labels, cache_dir, output_dir="results"):
print("\nExperiment 3: Analyzing Sample Efficiency and Overfitting Resistance")
# Create output directory
os.makedirs(output_dir, exist_ok=True)
# Convert data to PyTorch tensors
val_logits_tensor = torch.tensor(val_logits, dtype=torch.float32)
val_labels_tensor = torch.tensor(val_labels, dtype=torch.long)
test_logits_tensor = torch.tensor(test_logits, dtype=torch.float32)
test_labels_tensor = torch.tensor(test_labels, dtype=torch.long)
# Define different validation set sizes to test
val_set_sizes = [10, 20, 30, 40,50, 100, 150]
val_set_sizes = [min(size, len(val_logits)) for size in val_set_sizes]
# Define loss functions to compare
loss_functions = [
("SoftECE", lambda: SoftECE(n_bins=15, sigma=0.05)),
("Brier", lambda: CustomBrierLoss()),
("CrossEntropy", lambda: CrossEntropyLoss()),
("MSE", lambda: MSELoss())
]
# Number of training steps to monitor
max_steps = 100
step_interval = 5 # Record metrics every n steps
# Store results
results = {
"val_set_sizes": val_set_sizes,
"training_curves": {size: {} for size in val_set_sizes},
"final_metrics": {size: {} for size in val_set_sizes},
}
# Train on different validation set sizes
for val_size in val_set_sizes:
print(f"\nTraining with validation set size: {val_size}")
# Sample a subset of validation data
indices = np.random.choice(len(val_logits), val_size, replace=False)
subset_logits = val_logits_tensor[indices]
subset_labels = val_labels_tensor[indices]
# For each loss function
for loss_name, loss_fn_creator in loss_functions:
print(f" Training with {loss_name}")
# Initialize results storage for this scenario
if loss_name not in results["training_curves"][val_size]:
results["training_curves"][val_size][loss_name] = {
"steps": [],
"train_loss": [],
"train_ece": [],
"test_ece": [],
"test_acc": []
}
if loss_name not in results["final_metrics"][val_size]:
results["final_metrics"][val_size][loss_name] = {}
# Initialize temperature and optimizer
temperature = nn.Parameter(torch.ones(1))
optimizer = torch.optim.Adam([temperature], lr=0.01)
# Create loss function
loss_fn = loss_fn_creator()
# Training loop with tracking
for step in range(max_steps):
optimizer.zero_grad()
# Apply temperature scaling
scaled_logits = subset_logits / temperature
# Calculate loss
loss = loss_fn(scaled_logits, subset_labels)
# Backpropagation
loss.backward()
optimizer.step()
# Record metrics at specified intervals
if step % step_interval == 0 or step == max_steps - 1:
with torch.no_grad():
# Calculate training metrics
train_ece = ECE(n_bins=15)(scaled_logits, subset_labels).item()
# Calculate test metrics
test_scaled_logits = test_logits_tensor / temperature
test_probs = F.softmax(test_scaled_logits, dim=1)
test_ece = ECE(n_bins=15)(test_scaled_logits, test_labels_tensor).item()
test_acc = (torch.argmax(test_probs, dim=1) == test_labels_tensor).float().mean().item()
# Store metrics
results["training_curves"][val_size][loss_name]["steps"].append(step)
results["training_curves"][val_size][loss_name]["train_loss"].append(loss.item())
results["training_curves"][val_size][loss_name]["train_ece"].append(train_ece)
results["training_curves"][val_size][loss_name]["test_ece"].append(test_ece)
results["training_curves"][val_size][loss_name]["test_acc"].append(test_acc)
# Store final metrics
results["final_metrics"][val_size][loss_name] = {
"temperature": temperature.item(),
"train_ece": train_ece,
"test_ece": test_ece,
"test_acc": test_acc,
"train_test_ece_gap": abs(train_ece - test_ece) # Measure of overfitting
}
print(f" Final temperature: {temperature.item():.4f}, Test ECE: {test_ece:.4f}")
# Plot results
# 1. Plot final test ECE vs validation set size
plt.figure(figsize=(10, 6))
print("\nFinal Test ECE values for plotting:")
for loss_name, _ in loss_functions:
test_eces = [results["final_metrics"][size][loss_name]["test_ece"] for size in val_set_sizes]
print(f" {loss_name}: {test_eces}")
line, = plt.plot(val_set_sizes, test_eces, marker='o', linewidth=2, markersize=8)
plt.annotate(loss_name, (val_set_sizes[-1], test_eces[-1]),
xytext=(5, 0), textcoords='offset points', va='center')
plt.xlabel('Validation Set Size')
plt.ylabel('Test ECE')
plt.title('Sample Efficiency: Test ECE vs. Validation Set Size')
plt.legend([loss_name for loss_name, _ in loss_functions])
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'sample_efficiency_ece.png'), dpi=300)
plt.close()
# 2. Plot train-test ECE gap vs validation set size (measure of overfitting)
plt.figure(figsize=(10, 6))
print("\nTrain-Test ECE Gap values for plotting:")
for loss_name, _ in loss_functions:
gaps = [results["final_metrics"][size][loss_name]["train_test_ece_gap"] for size in val_set_sizes]
print(f" {loss_name}: {gaps}")
line, = plt.plot(val_set_sizes, gaps, marker='o', linewidth=2, markersize=8)
plt.annotate(loss_name, (val_set_sizes[-1], gaps[-1]),
xytext=(5, 0), textcoords='offset points', va='center')
plt.xlabel('Validation Set Size')
plt.ylabel('|Train ECE - Test ECE|')
plt.title('Overfitting Resistance: Train-Test Gap vs. Validation Set Size')
plt.legend([loss_name for loss_name, _ in loss_functions])
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'overfitting_resistance.png'), dpi=300)
plt.close()
# 3. Plot training curves for different loss functions on smallest dataset
smallest_size = val_set_sizes[0]
plt.figure(figsize=(12, 8))
for loss_name, _ in loss_functions:
steps = results["training_curves"][smallest_size][loss_name]["steps"]
test_eces = results["training_curves"][smallest_size][loss_name]["test_ece"]
plt.plot(steps, test_eces, marker='.', label=f"{loss_name}")
plt.xlabel('Training Step')
plt.ylabel('Test ECE')
plt.title(f'Training Stability with Small Dataset (n={smallest_size})')
plt.legend()
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, f'training_stability_{smallest_size}.png'), dpi=300)
plt.close()
# 4. Plot training curves for SoftECE across different dataset sizes
plt.figure(figsize=(12, 8))
for val_size in val_set_sizes:
steps = results["training_curves"][val_size]["SoftECE"]["steps"]
test_eces = results["training_curves"][val_size]["SoftECE"]["test_ece"]
plt.plot(steps, test_eces, marker='.', label=f"n={val_size}")
plt.xlabel('Training Step')
plt.ylabel('Test ECE')
plt.title('SoftECE Training Curves Across Dataset Sizes')
plt.legend()
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'softece_across_sizes.png'), dpi=300)
plt.close()
# Save results
with open(os.path.join(output_dir, 'sample_efficiency_results.json'), 'w') as f:
# Ensure all data is serializable
serializable_results = {
"val_set_sizes": results["val_set_sizes"],
"training_curves": {},
"final_metrics": {}
}
# Convert training curves
for size in results["training_curves"]:
serializable_results["training_curves"][str(size)] = {}
for loss_name in results["training_curves"][size]:
serializable_results["training_curves"][str(size)][loss_name] = {
k: [float(v_i) for v_i in v]
for k, v in results["training_curves"][size][loss_name].items()
}
# Convert final metrics
for size in results["final_metrics"]:
serializable_results["final_metrics"][str(size)] = {}
for loss_name in results["final_metrics"][size]:
serializable_results["final_metrics"][str(size)][loss_name] = {
k: float(v) for k, v in results["final_metrics"][size][loss_name].items()
}
json.dump(serializable_results, f, indent=2)
print(f"Sample efficiency analysis complete. Results saved to {output_dir}")
# Experiment 4: Bias-variance tradeoff visualization
def experiment_4_bias_variance_tradeoff(val_logits, val_labels, test_logits, test_labels, cache_dir, output_dir="results"):
print("\nExperiment 4: Visualizing Bias-Variance Tradeoff")
# Create output directory
os.makedirs(output_dir, exist_ok=True)
# Convert data to PyTorch tensors
val_logits_tensor = torch.tensor(val_logits, dtype=torch.float32)
val_labels_tensor = torch.tensor(val_labels, dtype=torch.long)
test_logits_tensor = torch.tensor(test_logits, dtype=torch.float32)
test_labels_tensor = torch.tensor(test_labels, dtype=torch.long)
# Define loss functions to compare
loss_functions = [
("SoftECE", lambda: SoftECE(n_bins=15, sigma=0.05)),
("ECE", lambda: ECEWrapper(n_bins=15)),
("Brier", lambda: CustomBrierLoss()),
("CrossEntropy", lambda: CrossEntropyLoss()),
("MSE", lambda: MSELoss())
]
# Bootstrap parameters
n_bootstraps = 20 # Number of bootstrap samples
bootstrap_size = min(500, len(val_logits)) # Size of each bootstrap sample
# Store results
results = {
"bootstrap_metrics": {loss_name: {
"temperatures": [],
"test_eces": [],
"calibrated_confidences": []
} for loss_name, _ in loss_functions},
"aggregated_metrics": {loss_name: {} for loss_name, _ in loss_functions}
}
# Create bootstrap samples
print(f"Creating {n_bootstraps} bootstrap samples of size {bootstrap_size}")
bootstrap_indices = []
for i in range(n_bootstraps):
# Sample with replacement
indices = np.random.choice(len(val_logits), bootstrap_size, replace=True)
bootstrap_indices.append(indices)
# Train on each bootstrap sample
for bootstrap_idx, indices in enumerate(bootstrap_indices):
print(f"\nTraining on bootstrap sample {bootstrap_idx+1}/{n_bootstraps}")
# Get bootstrap sample
bootstrap_logits = val_logits_tensor[indices]
bootstrap_labels = val_labels_tensor[indices]
# Train each loss function on this bootstrap sample
for loss_name, loss_fn_creator in loss_functions:
# Initialize temperature and optimizer
temperature = nn.Parameter(torch.ones(1))
optimizer = torch.optim.Adam([temperature], lr=0.01)
# Create loss function
loss_fn = loss_fn_creator()
# Train for a fixed number of epochs
epochs = 100
for epoch in range(epochs):
optimizer.zero_grad()
# Apply temperature scaling
scaled_logits = bootstrap_logits / temperature
# Calculate loss
loss = loss_fn(scaled_logits, bootstrap_labels)
# Backpropagation
loss.backward()
optimizer.step()
# Evaluate on test set
with torch.no_grad():
# Apply learned temperature
scaled_test_logits = test_logits_tensor / temperature
test_probs = F.softmax(scaled_test_logits, dim=1)
# Calculate metrics
test_ece = ECE(n_bins=15)(scaled_test_logits, test_labels_tensor).item()
# Store max confidences from a random subset for visualization
subset_size = min(1000, len(test_logits))
subset_indices = np.random.choice(len(test_logits), subset_size, replace=False)
subset_confs = torch.max(test_probs[subset_indices], dim=1)[0].cpu().numpy()
# Store results for this bootstrap run
results["bootstrap_metrics"][loss_name]["temperatures"].append(temperature.item())
results["bootstrap_metrics"][loss_name]["test_eces"].append(test_ece)
results["bootstrap_metrics"][loss_name]["calibrated_confidences"].append(subset_confs)
print(f" {loss_name}: Temperature = {temperature.item():.4f}, Test ECE = {test_ece:.4f}")
# Calculate aggregated metrics
for loss_name, _ in loss_functions:
temps = results["bootstrap_metrics"][loss_name]["temperatures"]
eces = results["bootstrap_metrics"][loss_name]["test_eces"]
# Calculate statistics
results["aggregated_metrics"][loss_name] = {
"temperature_mean": np.mean(temps),
"temperature_std": np.std(temps),
"temperature_cv": np.std(temps) / np.mean(temps) if np.mean(temps) > 0 else 0, # Coefficient of variation
"ece_mean": np.mean(eces),
"ece_std": np.std(eces),
"ece_cv": np.std(eces) / np.mean(eces) if np.mean(eces) > 0 else 0
}
# Prepare plots
# 1. Box plots of learned temperatures
plt.figure(figsize=(10, 6))
temp_data = [results["bootstrap_metrics"][loss_name]["temperatures"] for loss_name, _ in loss_functions]
plt.boxplot(temp_data, labels=[loss_name for loss_name, _ in loss_functions])
plt.ylabel('Temperature Value')
plt.title('Distribution of Learned Temperatures Across Bootstrap Samples')
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'temperature_distribution.png'), dpi=300)
plt.close()
# 2. Box plots of test ECEs
plt.figure(figsize=(10, 6))
ece_data = [results["bootstrap_metrics"][loss_name]["test_eces"] for loss_name, _ in loss_functions]
plt.boxplot(ece_data, labels=[loss_name for loss_name, _ in loss_functions])
plt.ylabel('Test ECE')
plt.title('Distribution of Test ECE Across Bootstrap Samples')
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'ece_distribution.png'), dpi=300)
plt.close()
# 3. Scatter plot of temperature variance vs ECE mean
plt.figure(figsize=(10, 6))
temp_vars = [results["aggregated_metrics"][loss_name]["temperature_cv"] for loss_name, _ in loss_functions]
ece_means = [results["aggregated_metrics"][loss_name]["ece_mean"] for loss_name, _ in loss_functions]
plt.scatter(temp_vars, ece_means, s=100)
# Add labels to points
for i, (loss_name, _) in enumerate(loss_functions):
plt.annotate(loss_name, (temp_vars[i], ece_means[i]),
textcoords="offset points", xytext=(0,10), ha='center')
plt.xlabel('Temperature Coefficient of Variation')
plt.ylabel('Mean Test ECE')
plt.title('Bias-Variance Tradeoff: Parameter Stability vs. Calibration Performance')
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'bias_variance_tradeoff.png'), dpi=300)
plt.close()
# 4. Histograms of calibrated confidences for each method
# We'll use the last bootstrap sample for visualization
plt.figure(figsize=(15, 10))
for i, (loss_name, _) in enumerate(loss_functions):
plt.subplot(2, 3, i+1)
confidences = results["bootstrap_metrics"][loss_name]["calibrated_confidences"][-1]
plt.hist(confidences, bins=20, alpha=0.7)
plt.title(f'{loss_name} Calibrated Confidences')
plt.xlabel('Confidence')
plt.ylabel('Count')
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'calibrated_confidence_distribution.png'), dpi=300)
plt.close()
# 5. Violin plots of ECE distribution
plt.figure(figsize=(12, 6))
# Create violin plot
plt.violinplot(ece_data, showmeans=True, showmedians=True)
# Add labels
plt.xticks(range(1, len(loss_functions) + 1), [loss_name for loss_name, _ in loss_functions])
plt.ylabel('Test ECE')
plt.title('Density Distribution of Test ECE Across Bootstrap Samples')
plt.grid(True, alpha=0.3)
plt.tight_layout()
plt.savefig(os.path.join(output_dir, 'ece_density_distribution.png'), dpi=300)
plt.close()
# Save results
with open(os.path.join(output_dir, 'bias_variance_results.json'), 'w') as f:
# Ensure all data is serializable
serializable_results = {
"bootstrap_metrics": {},
"aggregated_metrics": {}
}
# Convert bootstrap metrics
for loss_name in results["bootstrap_metrics"]:
serializable_results["bootstrap_metrics"][loss_name] = {
"temperatures": [float(t) for t in results["bootstrap_metrics"][loss_name]["temperatures"]],
"test_eces": [float(e) for e in results["bootstrap_metrics"][loss_name]["test_eces"]],
# Don't save all confidences as it's too much data
"temperature_mean": float(np.mean(results["bootstrap_metrics"][loss_name]["temperatures"])),
"temperature_std": float(np.std(results["bootstrap_metrics"][loss_name]["temperatures"])),
"ece_mean": float(np.mean(results["bootstrap_metrics"][loss_name]["test_eces"])),
"ece_std": float(np.std(results["bootstrap_metrics"][loss_name]["test_eces"]))
}
# Convert aggregated metrics
for loss_name in results["aggregated_metrics"]:
serializable_results["aggregated_metrics"][loss_name] = {
k: float(v) for k, v in results["aggregated_metrics"][loss_name].items()
}
json.dump(serializable_results, f, indent=2)
print(f"Bias-variance tradeoff analysis complete. Results saved to {output_dir}")
def main():
# Set random seed
set_seed(42)
# Define cache directory containing logits and labels
cache_dir = "/hdd/haolan/shats/PureLogits/cache/imagenet_vit_b_16_seed1_vs0.2"
# cache_dir = "/hdd/haolan/shats/PureLogits/cache/cifar100_resnet50_cross_entropy_seed1"
# cache_dir = "/hdd/haolan/shats/PureLogits/cache/imagenet_resnet50_seed1_vs0.2"
# cache_dir = "/hdd/haolan/shats/PureLogits/cache/cifar10_resnet50_cross_entropy_seed1"
# Load data
val_logits, val_labels, test_logits, test_labels = load_data(cache_dir)
# Create output directory
output_dir = "experiment_results/softece_validation"
os.makedirs(output_dir, exist_ok=True)
# Run experiments
# experiment_1_gradient_stability(test_logits, test_labels, cache_dir, output_dir)
# experiment_2_gradient_information(test_logits, test_labels, cache_dir, output_dir)
experiment_3_sample_efficiency(val_logits, val_labels, test_logits, test_labels, cache_dir, output_dir)
# experiment_4_bias_variance_tradeoff(val_logits, val_labels, test_logits, test_labels, cache_dir, output_dir)
if __name__ == "__main__":
start_time = time.time()
main()
elapsed_time = time.time() - start_time
print(f"Experiments completed in {elapsed_time/60:.2f} minutes")