VLM-Lens / src /concepts /pca_separation.py
marstin's picture
[martin-dev] add demo v1 test
d425e71
raw
history blame
13 kB
"""PCA scatter plot visualization for VLM concept analysis.
Creates 2D scatter plots of concepts and targets in PCA space for interpretability.
"""
from __future__ import annotations
import os
from typing import Optional
import matplotlib.pyplot as plt
import numpy as np
from pca import (apply_pca_to_layer, extract_concept_from_filename,
group_tensors_by_concept, load_tensors_by_layer)
def create_pca_scatter_plots(
target_db_path: str,
concept_db_path: str,
layer_names: Optional[list[str]] = None,
output_dir: str = 'output',
figsize: tuple[int, int] = (12, 8),
alpha: float = 0.7,
target_marker_size: int = 100,
concept_marker_size: int = 50
) -> None:
"""Create 2D PCA scatter plots for concepts and targets.
Args:
target_db_path: Path to target images database
concept_db_path: Path to concept images database
layer_names: List of layer names to visualize (None for all layers)
output_dir: Directory to save plots
figsize: Figure size (width, height)
alpha: Transparency for concept points
target_marker_size: Size of target markers
concept_marker_size: Size of concept markers
"""
print('Creating PCA scatter plots...')
# Load tensors from both databases
print(f'Loading tensors from {target_db_path}...')
target_tensors = load_tensors_by_layer(target_db_path, 'cpu')
print(f'Loading tensors from {concept_db_path}...')
concept_tensors = load_tensors_by_layer(concept_db_path, 'cpu')
# Find common layers
common_layers = set(target_tensors.keys()) & set(concept_tensors.keys())
print(f'Found {len(common_layers)} common layers: {sorted(common_layers)}')
if not common_layers:
print('No common layers found between databases!')
return
# Determine which layers to visualize
if layer_names is None:
layers_to_analyze = sorted(common_layers)
else:
if isinstance(layer_names, str):
layer_names = [layer_names]
layers_to_analyze = [layer for layer in layer_names if layer in common_layers]
os.makedirs(output_dir, exist_ok=True)
# Create plots for each layer
for layer in layers_to_analyze:
print(f'\nProcessing layer: {layer}')
target_layer_tensors = target_tensors[layer]
concept_layer_tensors = concept_tensors[layer]
if not target_layer_tensors or not concept_layer_tensors:
print(f'Skipping layer {layer} - insufficient data')
continue
# Apply PCA with 2 components
print(' Applying PCA with 2 components...')
transformed_targets, transformed_concepts, pca_model = apply_pca_to_layer(
target_layer_tensors, concept_layer_tensors, n_components=2
)
if pca_model is None:
print(f' Failed to apply PCA for layer {layer}')
continue
# Group concepts for coloring
concept_groups = group_tensors_by_concept(transformed_concepts)
# Create the plot
fig, ax = plt.subplots(figsize=figsize)
# Define colors for concepts (use a colormap)
concept_names = sorted(concept_groups.keys())
colors = plt.cm.Set3(np.linspace(0, 1, len(concept_names)))
color_map = dict(zip(concept_names, colors))
# Plot concept prototypes
for concept_name, concept_data in concept_groups.items():
concept_coords = np.array([data[0] for data in concept_data])
ax.scatter(
concept_coords[:, 0],
concept_coords[:, 1],
c=[color_map[concept_name]],
s=concept_marker_size,
alpha=alpha,
label=f'{concept_name} (prototypes)',
marker='o',
edgecolors='white',
linewidth=0.5
)
# Plot targets
target_coords = np.array([data[0] for data in transformed_targets])
target_concepts = []
# Extract target concepts for coloring
for data in transformed_targets:
target_concept = extract_concept_from_filename(data[3]) # data[3] is image_filename
target_concepts.append(target_concept)
# Plot targets with concept-based coloring
for i, (coord, target_concept) in enumerate(zip(target_coords, target_concepts)):
if target_concept in color_map:
color = color_map[target_concept]
label = f'{target_concept} (target)' if i == 0 or target_concept != target_concepts[i-1] else None
else:
color = 'black'
label = 'Unknown (target)' if i == 0 else None
ax.scatter(
coord[0],
coord[1],
c=[color],
s=target_marker_size,
alpha=0.9,
marker='^', # Triangle for targets
edgecolors='black',
linewidth=1.0,
label=label
)
# Customize the plot
ax.set_xlabel(f'PC1 ({pca_model.explained_variance_ratio_[0]:.3f} variance explained)')
ax.set_ylabel(f'PC2 ({pca_model.explained_variance_ratio_[1]:.3f} variance explained)')
ax.set_title(f'PCA Visualization: Concepts vs Targets\nLayer: {layer}')
ax.grid(True, alpha=0.3)
# Create legend with better organization
handles, labels = ax.get_legend_handles_labels()
# Separate prototype and target entries
prototype_handles, prototype_labels = [], []
target_handles, target_labels = [], []
for handle, label in zip(handles, labels):
if '(prototypes)' in label:
prototype_handles.append(handle)
prototype_labels.append(label.replace(' (prototypes)', ''))
elif '(target)' in label:
target_handles.append(handle)
target_labels.append(label.replace(' (target)', ''))
# Create two-column legend
if prototype_handles and target_handles:
legend1 = ax.legend(
prototype_handles,
[f'{label} (○)' for label in prototype_labels],
title='Concept Prototypes',
loc='upper left',
bbox_to_anchor=(1.02, 1.0),
fontsize=9
)
ax.add_artist(legend1)
ax.legend(
target_handles,
[f'{label} (△)' for label in target_labels],
title='Target Images',
loc='upper left',
bbox_to_anchor=(1.02, 0.6),
fontsize=9
)
else:
ax.legend(bbox_to_anchor=(1.02, 1.0), loc='upper left', fontsize=9)
# Add statistics text
stats_text = (
f'Total variance explained: {pca_model.explained_variance_ratio_.sum():.3f}\n'
f'Concepts: {len(concept_groups)}\n'
f'Prototypes: {len(transformed_concepts)}\n'
f'Targets: {len(transformed_targets)}'
)
ax.text(
0.02, 0.98,
stats_text,
transform=ax.transAxes,
verticalalignment='top',
bbox=dict(boxstyle='round', facecolor='white', alpha=0.8),
fontsize=9
)
plt.tight_layout()
# Save plot
plot_filename = f'{output_dir}/pca_scatter_layer_{layer.replace("/", "_")}.png'
plt.savefig(plot_filename, dpi=300, bbox_inches='tight')
plt.close()
print(f' Plot saved: {plot_filename}')
# Print summary statistics
print(f' Variance explained: PC1={pca_model.explained_variance_ratio_[0]:.3f}, '
f'PC2={pca_model.explained_variance_ratio_[1]:.3f}, '
f'Total={pca_model.explained_variance_ratio_.sum():.3f}')
print(f' Plotted {len(concept_groups)} concept groups with {len(transformed_concepts)} prototypes')
print(f' Plotted {len(transformed_targets)} target images')
print(f'\nPCA scatter plots complete. Plots saved in {output_dir}/')
def create_concept_separation_analysis(
target_db_path: str,
concept_db_path: str,
layer_names: Optional[list[str]] = None,
output_dir: str = 'output'
) -> None:
"""Analyze concept separation in PCA space.
Args:
target_db_path: Path to target images database
concept_db_path: Path to concept images database
layer_names: List of layer names to analyze (None for all layers)
output_dir: Directory to save analysis
"""
print('\nAnalyzing concept separation in PCA space...')
# Load tensors
target_tensors = load_tensors_by_layer(target_db_path, 'cpu')
concept_tensors = load_tensors_by_layer(concept_db_path, 'cpu')
common_layers = set(target_tensors.keys()) & set(concept_tensors.keys())
if layer_names is None:
layers_to_analyze = sorted(common_layers)
else:
if isinstance(layer_names, str):
layer_names = [layer_names]
layers_to_analyze = [layer for layer in layer_names if layer in common_layers]
os.makedirs(output_dir, exist_ok=True)
with open(f'{output_dir}/pca_separation_analysis.txt', 'w') as f:
f.write('PCA Concept Separation Analysis\n')
f.write('=' * 40 + '\n\n')
for layer in layers_to_analyze:
target_layer_tensors = target_tensors[layer]
concept_layer_tensors = concept_tensors[layer]
if not concept_layer_tensors:
continue
# Apply PCA
_, transformed_concepts, pca_model = apply_pca_to_layer(
target_layer_tensors, concept_layer_tensors, n_components=2
)
if pca_model is None:
continue
f.write(f'Layer: {layer}\n')
f.write('-' * 20 + '\n')
# Group concepts
concept_groups = group_tensors_by_concept(transformed_concepts)
# Calculate concept centroids in PCA space
concept_centroids = {}
for concept_name, concept_data in concept_groups.items():
coords = np.array([data[0] for data in concept_data])
concept_centroids[concept_name] = np.mean(coords, axis=0)
# Calculate pairwise distances between concept centroids
concept_names = list(concept_centroids.keys())
f.write('Concept centroid distances in PC1-PC2 space:\n')
for i, concept1 in enumerate(concept_names):
for j, concept2 in enumerate(concept_names[i+1:], i+1):
centroid1 = concept_centroids[concept1]
centroid2 = concept_centroids[concept2]
distance = np.linalg.norm(centroid1 - centroid2)
f.write(f' {concept1} - {concept2}: {distance:.3f}\n')
# Calculate within-concept scatter
f.write('\nWithin-concept scatter (std dev):\n')
for concept_name, concept_data in concept_groups.items():
coords = np.array([data[0] for data in concept_data])
if len(coords) > 1:
std_pc1 = np.std(coords[:, 0])
std_pc2 = np.std(coords[:, 1])
f.write(f' {concept_name}: PC1={std_pc1:.3f}, PC2={std_pc2:.3f}\n')
f.write('\nPCA Statistics:\n')
f.write(f' PC1 variance explained: {pca_model.explained_variance_ratio_[0]:.3f}\n')
f.write(f' PC2 variance explained: {pca_model.explained_variance_ratio_[1]:.3f}\n')
f.write(f' Total variance explained: {pca_model.explained_variance_ratio_.sum():.3f}\n')
f.write('\n\n')
print(f'Separation analysis saved to {output_dir}/pca_separation_analysis.txt')
if __name__ == '__main__':
# Configuration
target_db_path = 'output/llava.db'
concept_db_path = 'output/llava-concepts-colors.db'
# Visualization parameters
layer_names = None # None for all layers, or specify: ['layer_name1', 'layer_name2']
print('=' * 60)
print('VLM PCA VISUALIZATION')
print('=' * 60)
try:
# Create scatter plots
create_pca_scatter_plots(
target_db_path=target_db_path,
concept_db_path=concept_db_path,
layer_names=layer_names,
output_dir='output',
figsize=(12, 8),
alpha=0.7,
target_marker_size=100,
concept_marker_size=50
)
# Analyze concept separation
create_concept_separation_analysis(
target_db_path=target_db_path,
concept_db_path=concept_db_path,
layer_names=layer_names,
output_dir='output'
)
print('\nVisualization complete!')
except Exception as e:
print(f'Error during visualization: {e}')
import traceback
traceback.print_exc()