jasonfan's picture
2026-03-19: ICL code
64bce2a verified
import matplotlib.pyplot as plt
import torch
import numpy as np
from mpl_toolkits.mplot3d import Axes3D
import os
def plot_losses(losses, title="Training Loss", xlabel="Iteration", ylabel="Loss", save_path=None):
"""
Plot training losses.
Args:
losses (list): List of loss values
title (str): Plot title
xlabel (str): X-axis label
ylabel (str): Y-axis label
save_path (str): Optional path to save the figure
"""
plt.figure(figsize=(10, 6))
plt.plot(losses, 'o-')
plt.xlabel(xlabel)
plt.ylabel(ylabel)
plt.title(title)
plt.grid(True)
if save_path:
plt.savefig(save_path)
print(f"Loss plot saved to {save_path}")
plt.show()
def plot_multiple_losses(loss_dict, title="Training Losses", xlabel="Iteration", ylabel="Loss", save_path=None):
"""
Plot multiple loss curves for comparison.
Args:
loss_dict (dict): Dictionary mapping loss names to lists of loss values
title (str): Plot title
xlabel (str): X-axis label
ylabel (str): Y-axis label
save_path (str): Optional path to save the figure
"""
plt.figure(figsize=(12, 7))
for name, losses in loss_dict.items():
plt.plot(losses, '-', label=name)
plt.xlabel(xlabel)
plt.ylabel(ylabel)
plt.title(title)
plt.grid(True)
plt.legend()
if save_path:
plt.savefig(save_path)
print(f"Multiple losses plot saved to {save_path}")
plt.show()
def plot_multiple_accuracies(acc_dict, title="Classification Accuracy", xlabel="Percent Labeled", ylabel="Accuracy", save_path=None):
"""
Plot multiple accuracy curves for comparison.
Args:
acc_dict (dict): Dictionary mapping method names to [x_values, accuracy_values]
title (str): Plot title
xlabel (str): X-axis label
ylabel (str): Y-axis label
save_path (str): Optional path to save the figure
"""
plt.figure(figsize=(14, 8))
markers = ['o', 's', '^', 'v', 'd', '*', 'x', '+', 'h', 'p']
colors = plt.cm.tab10.colors
for i, (name, (x_values, acc_values)) in enumerate(acc_dict.items()):
marker = markers[i % len(markers)]
color = colors[i % len(colors)]
plt.plot(x_values, acc_values, marker=marker, color=color, linewidth=2, markersize=8, label=name)
plt.xlabel(xlabel, fontsize=14)
plt.ylabel(ylabel, fontsize=14)
plt.title(title, fontsize=16)
plt.grid(True)
plt.legend(fontsize=12)
plt.ylim(0, 1)
if save_path:
plt.savefig(save_path, dpi=300, bbox_inches='tight')
print(f"Accuracy comparison plot saved to {save_path}")
plt.show()
def plot_swiss_roll_3d(xyz, labels, title="Swiss Roll in 3D", save_path=None, figsize=(10, 8)):
"""
Plot 3D Swiss roll with color-coded labels.
Args:
xyz (numpy.ndarray): 3D coordinates of points
labels (numpy.ndarray): Labels for coloring points
title (str): Plot title
save_path (str): Optional path to save the figure
figsize (tuple): Figure size
"""
fig = plt.figure(figsize=figsize)
ax = fig.add_subplot(111, projection='3d')
scatter = ax.scatter(
xyz[:, 0],
xyz[:, 1],
xyz[:, 2],
c=labels,
cmap='bwr',
s=40
)
legend = ax.legend(*scatter.legend_elements(), title="Classes")
ax.add_artist(legend)
ax.set_xlabel("X")
ax.set_ylabel("Y")
ax.set_zlabel("Z")
ax.set_title(title)
if save_path:
plt.savefig(save_path)
print(f"3D plot saved to {save_path}")
plt.show()
def plot_laplacian_comparison(real_lap, pred_lap, title="Laplacian Comparison", save_path=None):
"""
Plot comparison between real and predicted Laplacian matrices.
Args:
real_lap (numpy.ndarray): Real Laplacian matrix
pred_lap (numpy.ndarray): Predicted Laplacian matrix
title (str): Plot title
save_path (str): Optional path to save the figure
"""
# Convert to numpy if tensors
if isinstance(real_lap, torch.Tensor):
real_lap = real_lap.cpu().numpy()
if isinstance(pred_lap, torch.Tensor):
pred_lap = pred_lap.cpu().numpy()
diff = pred_lap - real_lap
fig, axs = plt.subplots(1, 3, figsize=(18, 6))
# Plot real Laplacian
im0 = axs[0].imshow(real_lap, cmap='viridis')
axs[0].set_title("Real Laplacian")
plt.colorbar(im0, ax=axs[0])
# Plot predicted Laplacian
im1 = axs[1].imshow(pred_lap, cmap='viridis')
axs[1].set_title("Predicted Laplacian")
plt.colorbar(im1, ax=axs[1])
# Plot difference
im2 = axs[2].imshow(diff, cmap='bwr')
axs[2].set_title("Difference (Predicted - Real)")
plt.colorbar(im2, ax=axs[2])
plt.suptitle(title, fontsize=16)
plt.tight_layout()
if save_path:
plt.savefig(save_path)
print(f"Laplacian comparison plot saved to {save_path}")
plt.show()
# Return Frobenius norm of the difference
return np.sqrt(np.sum(diff**2))
def plot_eigenvector_comparison(real_evecs, pred_evecs, k=3, title="Eigenvector Comparison", save_path=None):
"""
Plot comparison between real and predicted eigenvectors.
Args:
real_evecs (numpy.ndarray): Real eigenvectors
pred_evecs (numpy.ndarray): Predicted eigenvectors
k (int): Number of eigenvectors to plot
title (str): Plot title
save_path (str): Optional path to save the figure
"""
# Convert to numpy if tensors
if isinstance(real_evecs, torch.Tensor):
real_evecs = real_evecs.cpu().numpy()
if isinstance(pred_evecs, torch.Tensor):
pred_evecs = pred_evecs.cpu().numpy()
# Normalize eigenvectors
real_evecs_norm = real_evecs / np.linalg.norm(real_evecs, axis=0, keepdims=True)
pred_evecs_norm = pred_evecs / np.linalg.norm(pred_evecs, axis=0, keepdims=True)
# Plot comparisons
fig, axs = plt.subplots(k, 1, figsize=(10, 4*k))
if k == 1:
axs = [axs]
for i in range(k):
# Handle sign ambiguity
corr_pos = np.corrcoef(real_evecs_norm[:, i], pred_evecs_norm[:, i])[0, 1]
corr_neg = np.corrcoef(real_evecs_norm[:, i], -pred_evecs_norm[:, i])[0, 1]
pred_ev_to_plot = -pred_evecs_norm[:, i] if abs(corr_neg) > abs(corr_pos) else pred_evecs_norm[:, i]
corr = max(abs(corr_pos), abs(corr_neg))
axs[i].plot(real_evecs_norm[:, i], 'b-', label='Real')
axs[i].plot(pred_ev_to_plot, 'r--', label='Predicted')
axs[i].grid(True)
axs[i].set_title(f"Eigenvector {i+1} (Correlation: {corr:.4f})")
axs[i].legend()
plt.suptitle(title, fontsize=16)
plt.tight_layout()
if save_path:
plt.savefig(save_path)
print(f"Eigenvector comparison plot saved to {save_path}")
plt.show()
def plot_embedding_3d(embeddings, labels, title="3D Embedding", save_path=None):
"""
Plot 3D embedding of points with color-coded labels.
Args:
embeddings (numpy.ndarray): Embedding coordinates (n_samples, 3)
labels (numpy.ndarray): Labels for coloring points
title (str): Plot title
save_path (str): Optional path to save the figure
"""
fig = plt.figure(figsize=(10, 8))
ax = fig.add_subplot(111, projection='3d')
scatter = ax.scatter(
embeddings[:, 0],
embeddings[:, 1],
embeddings[:, 2],
c=labels,
cmap='bwr',
s=40
)
legend = ax.legend(*scatter.legend_elements(), title="Classes")
ax.add_artist(legend)
ax.set_xlabel("Component 1")
ax.set_ylabel("Component 2")
ax.set_zlabel("Component 3")
ax.set_title(title)
if save_path:
plt.savefig(save_path)
print(f"3D embedding plot saved to {save_path}")
plt.show()
def create_multi_embedding_plot(embeddings_dict, labels, title="Embedding Comparison", save_path=None):
"""
Create a multi-panel plot comparing different embeddings.
Args:
embeddings_dict (dict): Dictionary mapping names to embedding arrays
labels (numpy.ndarray): Labels for coloring points
title (str): Plot title
save_path (str): Optional path to save the figure
"""
n_embeddings = len(embeddings_dict)
n_cols = min(3, n_embeddings)
n_rows = (n_embeddings + n_cols - 1) // n_cols
fig = plt.figure(figsize=(6*n_cols, 5*n_rows))
for i, (name, embedding) in enumerate(embeddings_dict.items()):
ax = fig.add_subplot(n_rows, n_cols, i+1, projection='3d')
scatter = ax.scatter(
embedding[:, 0],
embedding[:, 1],
embedding[:, 2],
c=labels,
cmap='bwr',
s=30
)
if i == 0:
legend = ax.legend(*scatter.legend_elements(), title="Classes")
ax.add_artist(legend)
ax.set_xlabel("Component 1")
ax.set_ylabel("Component 2")
ax.set_zlabel("Component 3")
ax.set_title(name)
plt.suptitle(title, fontsize=16)
plt.tight_layout(rect=[0, 0, 1, 0.95]) # Adjust for the super title
if save_path:
plt.savefig(save_path, dpi=300, bbox_inches='tight')
print(f"Multi-embedding plot saved to {save_path}")
plt.show()
def plot_accuracy_vs_context(acc_results, title="Accuracy vs Context Size", xlabel="Context Size", ylabel="Accuracy", save_path=None):
"""
Plot accuracy as a function of context size.
Args:
acc_results (dict): Dictionary mapping context size to accuracy
title (str): Plot title
xlabel (str): X-axis label
ylabel (str): Y-axis label
save_path (str): Optional path to save the figure
"""
import matplotlib.pyplot as plt
# Extract data
context_sizes = sorted(acc_results.keys())
accuracies = [acc_results[c] for c in context_sizes]
# Create plot
plt.figure(figsize=(10, 6))
plt.plot(context_sizes, accuracies, 'o-', linewidth=2, markersize=8)
plt.xlabel(xlabel, fontsize=14)
plt.ylabel(ylabel, fontsize=14)
plt.title(title, fontsize=16)
plt.grid(True)
plt.ylim(0, 1)
# Add specific annotations for min and max accuracies
min_acc = min(accuracies)
max_acc = max(accuracies)
min_idx = accuracies.index(min_acc)
max_idx = accuracies.index(max_acc)
plt.annotate(f'Min: {min_acc:.3f}',
xy=(context_sizes[min_idx], min_acc),
xytext=(10, 20),
textcoords='offset points',
arrowprops=dict(arrowstyle='->'))
plt.annotate(f'Max: {max_acc:.3f}',
xy=(context_sizes[max_idx], max_acc),
xytext=(10, -20),
textcoords='offset points',
arrowprops=dict(arrowstyle='->'))
if save_path:
plt.savefig(save_path, dpi=300, bbox_inches='tight')
print(f"Accuracy vs context plot saved to {save_path}")
plt.show()
def plot_embedding_comparison(raw_coords, embedding, labels, title="Embedding Comparison", save_path=None):
"""
Create side-by-side plots comparing raw coordinates with embedding.
Args:
raw_coords (numpy.ndarray): Raw coordinates [n_samples, 2]
embedding (numpy.ndarray): Embedding coordinates [n_samples, n_components]
labels (numpy.ndarray): Labels for coloring points
title (str): Plot title
save_path (str): Optional path to save the figure
"""
import matplotlib.pyplot as plt
import numpy as np
# Create figure
fig, axs = plt.subplots(1, 2, figsize=(14, 6))
# Plot raw coordinates
scatter1 = axs[0].scatter(
raw_coords[:, 0],
raw_coords[:, 1],
c=labels,
cmap='bwr',
s=40,
alpha=0.8
)
axs[0].set_xlabel("X")
axs[0].set_ylabel("Y")
axs[0].set_title("Raw Coordinates")
axs[0].grid(True)
# Plot embedding (first 2 components)
if embedding.shape[1] >= 2:
scatter2 = axs[1].scatter(
embedding[:, 0],
embedding[:, 1],
c=labels,
cmap='bwr',
s=40,
alpha=0.8
)
axs[1].set_xlabel("Component 1")
axs[1].set_ylabel("Component 2")
axs[1].set_title("Embedding")
axs[1].grid(True)
# Add legend
legend = axs[0].legend(*scatter1.legend_elements(), title="Classes")
axs[0].add_artist(legend)
plt.suptitle(title, fontsize=16)
plt.tight_layout()
if save_path:
plt.savefig(save_path, dpi=300, bbox_inches='tight')
print(f"Embedding comparison plot saved to {save_path}")
plt.show()
def plot_swiss_roll_colored_by_embedding(xyz, embedding, component_idx=0, title="Swiss Roll Colored by Embedding", save_path=None):
"""
Plot 3D Swiss roll colored by embedding component values.
Args:
xyz (numpy.ndarray): 3D coordinates of points
embedding (numpy.ndarray): Embedding values
component_idx (int): Index of embedding component to use for coloring
title (str): Plot title
save_path (str): Optional path to save the figure
"""
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
import numpy as np
# Extract the selected component
if embedding.ndim > 1:
color_values = embedding[:, component_idx]
else:
color_values = embedding
# Normalize values for coloring
normalized_values = (color_values - np.min(color_values)) / (np.max(color_values) - np.min(color_values))
fig = plt.figure(figsize=(10, 8))
ax = fig.add_subplot(111, projection='3d')
scatter = ax.scatter(
xyz[:, 0],
xyz[:, 1],
xyz[:, 2],
c=normalized_values,
cmap='viridis',
s=40
)
# Add a color bar
cbar = plt.colorbar(scatter)
cbar.set_label(f"Normalized Component {component_idx+1} Values")
ax.set_xlabel("X")
ax.set_ylabel("Y")
ax.set_zlabel("Z")
ax.set_title(title)
if save_path:
plt.savefig(save_path)
print(f"3D embedding plot saved to {save_path}")
plt.show()
import matplotlib.pyplot as plt
import numpy as np
import os
import datetime
def plot_test_results(args, results, n_samples=100, figure_dir='../figures'):
"""
Plot test results across different context sizes.
Args:
args: Command line arguments containing test_mode, data_type, etc.
results: Dictionary containing the test results for each context size
n_samples: Total number of samples per graph
figure_dir: Directory to save the figures
"""
# Create timestamp for unique filenames
timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
# Create directory structure based on test_mode and data_type
specific_dir = os.path.join(figure_dir, f"test_mode_{args.test_mode}", f"data_{args.data_type}")
os.makedirs(specific_dir, exist_ok=True)
# Extract context sizes and convert to percentages
context_sizes = list(results.keys())
context_percentages = [ctx_size / n_samples * 100 for ctx_size in context_sizes]
# Extract average accuracies and standard deviations for each method
avg_icl = [sum(results[ctx]['icl_accs']) / len(results[ctx]['icl_accs']) for ctx in context_sizes]
avg_lr = [sum(results[ctx]['lr_accs']) / len(results[ctx]['lr_accs']) for ctx in context_sizes]
avg_rkhs = [sum(results[ctx]['rkhs_accs']) / len(results[ctx]['rkhs_accs']) for ctx in context_sizes]
avg_raw_lr = [sum(results[ctx]['raw_lr_accs']) / len(results[ctx]['raw_lr_accs']) for ctx in context_sizes]
avg_raw_rkhs = [sum(results[ctx]['raw_rkhs_accs']) / len(results[ctx]['raw_rkhs_accs']) for ctx in context_sizes]
std_icl = [np.std(results[ctx]['icl_accs']) for ctx in context_sizes]
std_lr = [np.std(results[ctx]['lr_accs']) for ctx in context_sizes]
std_rkhs = [np.std(results[ctx]['rkhs_accs']) for ctx in context_sizes]
std_raw_lr = [np.std(results[ctx]['raw_lr_accs']) for ctx in context_sizes]
std_raw_rkhs = [np.std(results[ctx]['raw_rkhs_accs']) for ctx in context_sizes]
# Generate a base filename with test details
base_filename = f"mode_{args.test_mode}_data_{args.data_type}_samples_{n_samples}_{timestamp}"
# Create the main plot
plt.figure(figsize=(10, 6))
# Plot each method with error bars
plt.errorbar(context_percentages, avg_icl, yerr=std_icl, fmt='o-', label='ICL', capsize=4, linewidth=2)
plt.errorbar(context_percentages, avg_lr, yerr=std_lr, fmt='s-', label='LR (PE)', capsize=4, linewidth=2)
plt.errorbar(context_percentages, avg_rkhs, yerr=std_rkhs, fmt='^-', label='RKHS (PE)', capsize=4, linewidth=2)
plt.errorbar(context_percentages, avg_raw_lr, yerr=std_raw_lr, fmt='d--', label='LR (Raw)', capsize=4, linewidth=2)
plt.errorbar(context_percentages, avg_raw_rkhs, yerr=std_raw_rkhs, fmt='v--', label='RKHS (Raw)', capsize=4, linewidth=2)
# Add labels and title
plt.xlabel('Context Size (% of Total Samples)')
plt.ylabel('Accuracy')
title_str = f'Classification Performance vs. Context Size\n(Mode: {args.test_mode}, Data: {args.data_type})'
plt.title(title_str)
plt.grid(True, alpha=0.3)
plt.legend(loc='lower right')
# Adjust x-axis to show percentages
plt.xticks(context_percentages)
# Save the figure
plt.tight_layout()
fig_path = os.path.join(specific_dir, f"{base_filename}_all_methods.png")
plt.savefig(fig_path, dpi=300)
print(f"Figure saved to {fig_path}")
# Create a second plot: Just PE methods vs Raw methods
plt.figure(figsize=(10, 6))
# Group PE and Raw methods
plt.errorbar(context_percentages, avg_lr, yerr=std_lr, fmt='s-', label='LR (PE)', capsize=4, linewidth=2, color='blue')
plt.errorbar(context_percentages, avg_rkhs, yerr=std_rkhs, fmt='^-', label='RKHS (PE)', capsize=4, linewidth=2, color='green')
plt.errorbar(context_percentages, avg_raw_lr, yerr=std_raw_lr, fmt='s--', label='LR (Raw)', capsize=4, linewidth=2, color='blue', alpha=0.5)
plt.errorbar(context_percentages, avg_raw_rkhs, yerr=std_raw_rkhs, fmt='^--', label='RKHS (Raw)', capsize=4, linewidth=2, color='green', alpha=0.5)
# Add labels and title
plt.xlabel('Context Size (% of Total Samples)')
plt.ylabel('Accuracy')
title_str = f'PE Features vs. Raw Features Performance\n(Mode: {args.test_mode}, Data: {args.data_type})'
plt.title(title_str)
plt.grid(True, alpha=0.3)
plt.legend(loc='lower right')
# Adjust x-axis to show percentages
plt.xticks(context_percentages)
# Save the figure
plt.tight_layout()
fig_path = os.path.join(specific_dir, f"{base_filename}_pe_vs_raw.png")
plt.savefig(fig_path, dpi=300)
print(f"Figure saved to {fig_path}")
# Create a third plot: ICL vs best traditional method
plt.figure(figsize=(10, 6))
import matplotlib.pyplot as plt
import numpy as np
import os
import datetime
import json
def plot_individual_accuracies(args, results, n_samples=100, figure_dir='../figures'):
"""
Plot individual accuracy metrics across different context sizes.
Args:
args: Command line arguments containing test_mode, data_type, etc.
results: Dictionary containing the test results for each context size
n_samples: Total number of samples per graph
figure_dir: Directory to save the figures
"""
# Create timestamp for unique filenames
timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
# Create directory structure based on test_mode and data_type
specific_dir = os.path.join(figure_dir, f"test_mode_{args.test_mode}", f"data_{args.data_type}", "individual_plots")
os.makedirs(specific_dir, exist_ok=True)
# Extract context sizes and convert to percentages
context_sizes = list(results.keys())
context_percentages = [ctx_size / n_samples * 100 for ctx_size in context_sizes]
# Extract average accuracies and standard deviations for each method
avg_icl = [sum(results[ctx]['icl_accs']) / len(results[ctx]['icl_accs']) for ctx in context_sizes]
avg_lr = [sum(results[ctx]['lr_accs']) / len(results[ctx]['lr_accs']) for ctx in context_sizes]
avg_rkhs = [sum(results[ctx]['rkhs_accs']) / len(results[ctx]['rkhs_accs']) for ctx in context_sizes]
avg_raw_lr = [sum(results[ctx]['raw_lr_accs']) / len(results[ctx]['raw_lr_accs']) for ctx in context_sizes]
avg_raw_rkhs = [sum(results[ctx]['raw_rkhs_accs']) / len(results[ctx]['raw_rkhs_accs']) for ctx in context_sizes]
std_icl = [np.std(results[ctx]['icl_accs']) for ctx in context_sizes]
std_lr = [np.std(results[ctx]['lr_accs']) for ctx in context_sizes]
std_rkhs = [np.std(results[ctx]['rkhs_accs']) for ctx in context_sizes]
std_raw_lr = [np.std(results[ctx]['raw_lr_accs']) for ctx in context_sizes]
std_raw_rkhs = [np.std(results[ctx]['raw_rkhs_accs']) for ctx in context_sizes]
# Define metrics to plot individually
metrics = [
{'name': 'ICL', 'avg': avg_icl, 'std': std_icl, 'color': 'red', 'marker': 'o'},
{'name': 'LR (PE)', 'avg': avg_lr, 'std': std_lr, 'color': 'blue', 'marker': 's'},
{'name': 'RKHS (PE)', 'avg': avg_rkhs, 'std': std_rkhs, 'color': 'green', 'marker': '^'},
{'name': 'LR (Raw)', 'avg': avg_raw_lr, 'std': std_raw_lr, 'color': 'purple', 'marker': 'd'},
{'name': 'RKHS (Raw)', 'avg': avg_raw_rkhs, 'std': std_raw_rkhs, 'color': 'orange', 'marker': 'v'}
]
# Plot each metric individually
for metric in metrics:
plt.figure(figsize=(8, 5))
# Plot with error bars
plt.errorbar(
context_percentages,
metric['avg'],
yerr=metric['std'],
fmt=f"{metric['marker']}-",
label=metric['name'],
capsize=4,
linewidth=2,
color=metric['color']
)
# Add labels and title
plt.xlabel('Context Size (% of Total Samples)')
plt.ylabel('Accuracy')
title_str = f"{metric['name']} Accuracy vs. Context Size\n(Mode: {args.test_mode}, Data: {args.data_type})"
plt.title(title_str)
plt.grid(True, alpha=0.3)
# Generate filename
safe_name = metric['name'].replace(' ', '_').replace('(', '').replace(')', '')
base_filename = f"mode_{args.test_mode}_data_{args.data_type}_samples_{n_samples}_{timestamp}"
filename = f"{base_filename}_{safe_name}.png"
# Adjust x-axis to show percentages
plt.xticks(context_percentages)
# Set y-axis range from 0.5 to 1.0 for better comparison
plt.ylim([0.5, 1.0])
# Add exact values as text annotations
for i, (x, y, std) in enumerate(zip(context_percentages, metric['avg'], metric['std'])):
plt.annotate(
f"{y:.3f}±{std:.3f}",
xy=(x, y),
xytext=(0, 10),
textcoords='offset points',
ha='center',
fontsize=8,
bbox=dict(boxstyle='round,pad=0.3', fc='white', alpha=0.7)
)
# Save the figure
plt.tight_layout()
fig_path = os.path.join(specific_dir, filename)
plt.savefig(fig_path, dpi=300)
print(f"Individual {metric['name']} plot saved to {fig_path}")
plt.close()
# Create a tabular summary and save as CSV
import pandas as pd
# Create DataFrame
summary_data = {
'Context_Size': context_sizes,
'Context_Percentage': context_percentages,
'ICL_Mean': avg_icl,
'ICL_Std': std_icl,
'LR_PE_Mean': avg_lr,
'LR_PE_Std': std_lr,
'RKHS_PE_Mean': avg_rkhs,
'RKHS_PE_Std': std_rkhs,
'LR_Raw_Mean': avg_raw_lr,
'LR_Raw_Std': std_raw_lr,
'RKHS_Raw_Mean': avg_raw_rkhs,
'RKHS_Raw_Std': std_raw_rkhs
}
df = pd.DataFrame(summary_data)
# Save as CSV
csv_path = os.path.join(specific_dir, f"{base_filename}_summary.csv")
df.to_csv(csv_path, index=False)
print(f"Summary data saved to {csv_path}")
# Also generate a formatted table as text
text_table = "Context Size | ICL | LR (PE) | RKHS (PE) | LR (Raw) | RKHS (Raw)\n"
text_table += "------------|-----|---------|-----------|----------|------------\n"
for i, ctx in enumerate(context_sizes):
text_table += f"{ctx:^12} | {avg_icl[i]:.3f}±{std_icl[i]:.3f} | {avg_lr[i]:.3f}±{std_lr[i]:.3f} | "
text_table += f"{avg_rkhs[i]:.3f}±{std_rkhs[i]:.3f} | {avg_raw_lr[i]:.3f}±{std_raw_lr[i]:.3f} | "
text_table += f"{avg_raw_rkhs[i]:.3f}±{std_raw_rkhs[i]:.3f}\n"
# Save table as txt
txt_path = os.path.join(specific_dir, f"{base_filename}_table.txt")
with open(txt_path, 'w') as f:
f.write(text_table)
print(f"Formatted table saved to {txt_path}")
return specific_dir