File size: 6,224 Bytes
2572f0f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
import os
import torch
import pandas as pd
import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.decomposition import PCA
from sklearn.metrics import silhouette_score
from torch.utils.data import DistributedSampler
from torchvision.datasets import ImageFolder

PLOTS_DIR = 'plots'
os.makedirs(PLOTS_DIR, exist_ok=True)

# 1. Reconstruct dataset file names matching the DistributedSampler order
print("Reconstructing dataset file list and sampler indices...")
dataset = ImageFolder('data/images/val')
sampler = DistributedSampler(dataset, num_replicas=1, rank=0)
indices = list(sampler)

# Get the list of file names in the exact sampler order
sampler_filenames = [dataset.imgs[idx][0] for idx in indices]
# Convert absolute/relative paths to match the parquet file_name column (e.g. val/class/file.jpg)
# dataset.imgs paths look like: 'data/images/val/03368_.../file.jpg'
cleaned_filenames = []
for p in sampler_filenames:
    parts = p.split(os.sep)
    # Find where 'val' starts
    if 'val' in parts:
        val_idx = parts.index('val')
        cleaned_filenames.append('/'.join(parts[val_idx:]))
    else:
        cleaned_filenames.append(p)

# 2. Load metadata and align
print("Loading and aligning metadata...")
df = pd.read_parquet('metadata/inat_world_model_master.parquet')
df_val = df[df['split'] == 'val'].copy()
df_val.set_index('file_name', inplace=True)

# Reorder metadata to match the sampler order
aligned_meta = []
for fname in cleaned_filenames:
    if fname in df_val.index:
        # If there are duplicates, take the first one
        row = df_val.loc[fname]
        if isinstance(row, pd.DataFrame):
            row = row.iloc[0]
        aligned_meta.append(row.to_dict())
    else:
        # Fallback empty dict
        aligned_meta.append({})

df_aligned = pd.DataFrame(aligned_meta)

# Drop rows where we have missing lifestyle/trophic level metadata
valid_mask = df_aligned['Primary.Lifestyle'].notna() & df_aligned['Trophic.Level'].notna()
print(f"Aligned metadata rows: {len(df_aligned)}, Valid rows for analysis: {valid_mask.sum()}")

# 3. Load embeddings
models_config = {
    'DINOv3': {
        'path': 'outputs/vit_large_patch16_dinov3/macro_val_30percent_rank_0.pt',
        'safe_name': 'vit_large_patch16_dinov3'
    },
    'SigLIP2': {
        'path': 'outputs/google_siglip2_so400m_patch14_384/macro_val_30percent_rank_0.pt',
        'safe_name': 'google_siglip2_so400m_patch14_384'
    },
    'BioCLIP2': {
        'path': 'outputs/hf_hub:imageomics_bioclip_2.5_vith14/macro_val_30percent_rank_0.pt',
        'safe_name': 'hf_hub:imageomics_bioclip_2.5_vith14'
    }
}

embeddings_data = {}
for name, cfg in models_config.items():
    if os.path.exists(cfg['path']):
        print(f"Loading {name} embeddings from {cfg['path']}...")
        data = torch.load(cfg['path'], map_location='cpu')
        embeddings = data['embeddings'].numpy()
        embeddings_data[name] = embeddings
    else:
        print(f"Warning: {cfg['path']} does not exist!")

# Subset metadata to match the length of the 30% macro files
# The 30% macro file contains exactly the first 30% of the validation set batches
n_samples = None
for name, embs in embeddings_data.items():
    n_samples = len(embs)
    break

if n_samples is None:
    print("Error: No embeddings loaded. Exiting.")
    exit(1)

print(f"Analyzing first {n_samples} samples...")
df_subset = df_aligned.iloc[:n_samples].copy()
valid_mask_sub = valid_mask.iloc[:n_samples].values

# Filter out rows with invalid metadata from both embeddings and metadata
df_filtered = df_subset[valid_mask_sub].copy()

# Apply PCA and calculate Silhouette Scores
results = {}
for name, embs in embeddings_data.items():
    embs_filtered = embs[valid_mask_sub]
    
    # 2D PCA Projection
    print(f"Running PCA for {name}...")
    pca = PCA(n_components=2, random_state=42)
    coords = pca.fit_transform(embs_filtered)
    
    # Calculate Silhouette Scores on the full-dimensional embeddings
    print(f"Calculating Silhouette Scores for {name}...")
    sil_lifestyle = silhouette_score(embs_filtered, df_filtered['Primary.Lifestyle'].astype(str))
    sil_trophic = silhouette_score(embs_filtered, df_filtered['Trophic.Level'].astype(str))
    
    results[name] = {
        'coords': coords,
        'sil_lifestyle': sil_lifestyle,
        'sil_trophic': sil_trophic
    }

# 4. Generate the 2x3 Plot Grid
sns.set_theme(style="whitegrid")
fig, axes = plt.subplots(2, 3, figsize=(18, 11))

for col_idx, model_name in enumerate(results.keys()):
    res = results[model_name]
    coords = res['coords']
    
    # Row 1: Primary Lifestyle
    ax1 = axes[0, col_idx]
    sns.scatterplot(
        x=coords[:, 0], y=coords[:, 1],
        hue=df_filtered['Primary.Lifestyle'],
        palette='Set1', alpha=0.5, s=8, ax=ax1, legend=(col_idx == 2)
    )
    ax1.set_title(f"{model_name}\nLifestyle Silhouette: {res['sil_lifestyle']:.3f}")
    if col_idx == 2:
        ax1.legend(title='Primary Lifestyle', bbox_to_anchor=(1.05, 1), loc='upper left')
    ax1.set_xlabel("PCA 1")
    ax1.set_ylabel("PCA 2")
    
    # Row 2: Trophic Level
    ax2 = axes[1, col_idx]
    sns.scatterplot(
        x=coords[:, 0], y=coords[:, 1],
        hue=df_filtered['Trophic.Level'],
        palette='Dark2', alpha=0.5, s=8, ax=ax2, legend=(col_idx == 2)
    )
    ax2.set_title(f"Trophic Level Silhouette: {res['sil_trophic']:.3f}")
    if col_idx == 2:
        ax2.legend(title='Trophic Level', bbox_to_anchor=(1.05, 1), loc='upper left')
    ax2.set_xlabel("PCA 1")
    ax2.set_ylabel("PCA 2")

plt.suptitle("Model Feature Space Representation vs. Ecological Metadata", y=0.98, fontsize=16, fontweight='bold')
plt.tight_layout()
fig.savefig(os.path.join(PLOTS_DIR, 'feature_representation_comparison.png'), bbox_inches='tight', dpi=200)
plt.close(fig)

print("\n=== SILHOUETTE SCORE SUMMARY ===")
for name, res in results.items():
    print(f"{name}:")
    print(f"  Primary Lifestyle Silhouette Score: {res['sil_lifestyle']:.4f}")
    print(f"  Trophic Level Silhouette Score:      {res['sil_trophic']:.4f}")

print("\nPlots generated successfully at plots/feature_representation_comparison.png")