File size: 9,895 Bytes
bb049db | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 | #!/usr/bin/env python3
"""Visualize causal vs bidirectional ROI attention masks to illustrate the method."""
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
import numpy as np
def make_attention_matrix(seq_labels, bidir_range=None, bidir_mode='full'):
"""
Build attention matrix.
- Default: causal (lower triangular)
- bidir_range: (start, end) indices for bidirectional tokens
- bidir_mode:
'full' = bidir tokens attend to ALL tokens (past + future)
'mutual' = bidir tokens attend to each other + causal to rest
"""
n = len(seq_labels)
# Start with causal mask (lower triangular)
mask = np.tril(np.ones((n, n)))
if bidir_range is not None:
rs, re = bidir_range
if bidir_mode == 'full':
# ROI queries attend to full sequence
for i in range(rs, re):
mask[i, :] = 1.0
elif bidir_mode == 'mutual':
# ROI tokens only see each other bidirectionally
# Keep causal for non-ROI keys
for i in range(rs, re):
for j in range(rs, re):
mask[i, j] = 1.0 # ROI↔ROI bidir
return mask
def plot_mask(ax, mask, seq_labels, title, token_colors, bidir_cells=None):
"""Plot a single attention mask matrix with bidir cells highlighted."""
n = len(seq_labels)
# Create colored matrix
colored = np.ones((n, n, 3)) # white = masked
for i in range(n):
for j in range(n):
if mask[i, j] > 0:
colored[i, j] = token_colors[i]
ax.imshow(colored, aspect='equal', origin='upper')
# Mark bidir cells with a distinct pattern (small red dot)
if bidir_cells is not None:
for (i, j) in bidir_cells:
if mask[i, j] > 0:
ax.plot(j, i, 's', color='red', markersize=3, alpha=0.5)
# Grid lines
for i in range(n + 1):
ax.axhline(i - 0.5, color='gray', linewidth=0.5, alpha=0.4)
ax.axvline(i - 0.5, color='gray', linewidth=0.5, alpha=0.4)
ax.set_xticks(range(n))
ax.set_xticklabels(seq_labels, rotation=45, ha='right', fontsize=7)
ax.set_yticks(range(n))
ax.set_yticklabels(seq_labels, fontsize=7)
ax.set_xlabel('Key (attends to)', fontsize=9)
ax.set_ylabel('Query (token)', fontsize=9)
ax.set_title(title, fontsize=10, fontweight='bold', pad=10)
def main():
# === Inference sequence: [sys, visual(source), ROI crops, text query] ===
seq_labels = [
'sys₁', 'sys₂',
'v₁', 'v₂', 'v₃', 'v₄', 'v₅', 'v₆',
'roi₁', 'roi₂', 'roi₃', 'roi₄',
'q₁', 'q₂', 'q₃',
]
n = len(seq_labels)
color_map = {
'sys': np.array([0.7, 0.7, 0.7]),
'v': np.array([0.35, 0.55, 0.85]),
'roi': np.array([0.95, 0.55, 0.20]),
'q': np.array([0.40, 0.75, 0.40]),
'a': np.array([0.85, 0.35, 0.50]),
}
token_colors = []
for label in seq_labels:
if label.startswith('sys'): token_colors.append(color_map['sys'])
elif label.startswith('v'): token_colors.append(color_map['v'])
elif label.startswith('roi'): token_colors.append(color_map['roi'])
elif label.startswith('q'): token_colors.append(color_map['q'])
roi_start = seq_labels.index('roi₁')
roi_end = seq_labels.index('roi₄') + 1
# ===== Figure 1: 3-way comparison (inference) =====
fig, axes = plt.subplots(1, 3, figsize=(18, 6))
# (a) Standard causal
mask_causal = make_attention_matrix(seq_labels)
plot_mask(axes[0], mask_causal, seq_labels,
'(a) Standard Causal\n(baseline)', token_colors)
# (b) ROI↔ROI mutual bidir (ROI tokens see each other, causal to rest)
mask_mutual = make_attention_matrix(seq_labels,
bidir_range=(roi_start, roi_end),
bidir_mode='mutual')
# Collect bidir-only cells (above diagonal within ROI block)
bidir_cells_mutual = []
for i in range(roi_start, roi_end):
for j in range(roi_start, roi_end):
if j > i: # above diagonal = non-causal = the bidir addition
bidir_cells_mutual.append((i, j))
plot_mask(axes[1], mask_mutual, seq_labels,
'(b) ROI↔ROI Bidir\n(layers K→35)', token_colors,
bidir_cells=bidir_cells_mutual)
# Highlight ROI↔ROI block
rect1 = mpatches.FancyBboxPatch(
(roi_start - 0.5, roi_start - 0.5),
roi_end - roi_start, roi_end - roi_start,
linewidth=2.5, edgecolor='red', facecolor='none',
boxstyle='round,pad=0', linestyle='--'
)
axes[1].add_patch(rect1)
axes[1].annotate('mutual bidir\n(ROI↔ROI only)',
xy=(roi_end + 0.5, roi_start + 1), fontsize=8,
color='red', fontweight='bold')
# (c) ROI→ALL bidir (ROI sees everything including future)
mask_full = make_attention_matrix(seq_labels,
bidir_range=(roi_start, roi_end),
bidir_mode='full')
bidir_cells_full = []
for i in range(roi_start, roi_end):
for j in range(n):
if j > i:
bidir_cells_full.append((i, j))
plot_mask(axes[2], mask_full, seq_labels,
'(c) ROI→All Bidir\n(layers K→35)', token_colors,
bidir_cells=bidir_cells_full)
rect2 = mpatches.FancyBboxPatch(
(-0.5, roi_start - 0.5), n, roi_end - roi_start,
linewidth=2.5, edgecolor='red', facecolor='none',
boxstyle='round,pad=0', linestyle='--'
)
axes[2].add_patch(rect2)
axes[2].annotate('full bidir\n(ROI sees all)',
xy=(n - 1, roi_start + 1), fontsize=8,
color='red', fontweight='bold', ha='right')
# Legend
legend_patches = [
mpatches.Patch(color=color_map['sys'], label='System tokens'),
mpatches.Patch(color=color_map['v'], label='Visual tokens (source)'),
mpatches.Patch(color=color_map['roi'], label='ROI crop tokens'),
mpatches.Patch(color=color_map['q'], label='Text query tokens'),
mpatches.Patch(facecolor='white', edgecolor='gray', label='Masked (cannot attend)'),
plt.Line2D([0], [0], marker='s', color='red', markersize=6,
linestyle='', alpha=0.5, label='Bidir addition (non-causal)'),
]
fig.legend(handles=legend_patches, loc='lower center', ncol=6,
fontsize=8, bbox_to_anchor=(0.5, -0.02))
plt.suptitle('Selective Bidirectional Visual Attention for SD-RPN (Inference, Layers K→35)',
fontsize=13, fontweight='bold', y=1.02)
plt.tight_layout()
plt.savefig('/opt/tiger/thothvl_pretrain/visualize/attention_mask_comparison.png',
dpi=200, bbox_inches='tight', pad_inches=0.3)
print("Saved: visualize/attention_mask_comparison.png")
# ===== Figure 2: Training masks (layers 0→K-1 vs K→35) =====
train_labels = [
'sys₁', 'sys₂',
'v₁', 'v₂', 'v₃', 'v₄', 'v₅', 'v₆',
'q₁', 'q₂',
'a₁', 'a₂', 'a₃',
]
train_colors = []
for label in train_labels:
if label.startswith('sys'): train_colors.append(color_map['sys'])
elif label.startswith('v'): train_colors.append(color_map['v'])
elif label.startswith('q'): train_colors.append(color_map['q'])
elif label.startswith('a'): train_colors.append(color_map['a'])
vis_s = train_labels.index('v₁')
vis_e = train_labels.index('v₆') + 1
fig2, axes2 = plt.subplots(1, 2, figsize=(14, 6))
# Layers 0→K-1: standard causal
mask_train_early = make_attention_matrix(train_labels)
plot_mask(axes2[0], mask_train_early, train_labels,
'Layers 0→K-1: Standard Causal', train_colors)
# Layers K→35: visual↔visual mutual bidir
mask_train_late = make_attention_matrix(train_labels,
bidir_range=(vis_s, vis_e),
bidir_mode='mutual')
bidir_cells_train = []
for i in range(vis_s, vis_e):
for j in range(vis_s, vis_e):
if j > i:
bidir_cells_train.append((i, j))
plot_mask(axes2[1], mask_train_late, train_labels,
'Layers K→35: Visual↔Visual Bidir', train_colors,
bidir_cells=bidir_cells_train)
rect3 = mpatches.FancyBboxPatch(
(vis_s - 0.5, vis_s - 0.5),
vis_e - vis_s, vis_e - vis_s,
linewidth=2.5, edgecolor='red', facecolor='none',
boxstyle='round,pad=0', linestyle='--'
)
axes2[1].add_patch(rect3)
axes2[1].annotate('mutual bidir\n(vis↔vis)',
xy=(vis_e + 0.3, vis_s + 1), fontsize=8,
color='red', fontweight='bold')
legend_patches2 = [
mpatches.Patch(color=color_map['sys'], label='System'),
mpatches.Patch(color=color_map['v'], label='Visual'),
mpatches.Patch(color=color_map['q'], label='Query'),
mpatches.Patch(color=color_map['a'], label='Answer (target)'),
plt.Line2D([0], [0], marker='s', color='red', markersize=6,
linestyle='', alpha=0.5, label='Bidir addition'),
]
fig2.legend(handles=legend_patches2, loc='lower center', ncol=5,
fontsize=9, bbox_to_anchor=(0.5, -0.02))
plt.suptitle('Training: Bidirectional Visual Attention (K=24)',
fontsize=14, fontweight='bold', y=1.02)
plt.tight_layout()
plt.savefig('/opt/tiger/thothvl_pretrain/visualize/attention_mask_training.png',
dpi=200, bbox_inches='tight', pad_inches=0.3)
print("Saved: visualize/attention_mask_training.png")
if __name__ == '__main__':
main()
|