File size: 3,042 Bytes
6d11ed7 | 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 | """
Gate visualization — renders each hidden node's receptive field
(W1 column) as a 28x28 image.
Shows which pixel combinations open each gate.
"""
import os, pickle
import numpy as np
try:
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
HAS_PLT = True
except ImportError:
HAS_PLT = False
print('matplotlib not installed. Install with: pip install matplotlib')
def load_ckpt(path):
with open(path, 'rb') as f:
state = pickle.load(f)
return state['W1'], state['W2']
def visualize_gates(W1, W2, out_dir, top_k=32):
"""Visualize receptive fields of top_k gates ranked by W2 influence."""
n_h = W1.shape[1]
# gate importance: L1 norm of W2 row (influence on output)
importance = np.abs(W2).sum(axis=1)
top_idx = np.argsort(importance)[::-1][:top_k]
# each gate's preferred output class
preferred_class = W2.argmax(axis=1)
# grid plot
cols = 8
rows = (top_k + cols - 1) // cols
fig, axes = plt.subplots(rows, cols, figsize=(cols * 1.5, rows * 1.8))
axes = axes.flatten()
for i, idx in enumerate(top_idx):
rf = W1[:, idx].reshape(28, 28)
axes[i].imshow(rf, cmap='RdBu_r', vmin=-rf.max(), vmax=rf.max())
axes[i].set_title(f'g{idx}→{preferred_class[idx]}', fontsize=8)
axes[i].axis('off')
for i in range(len(top_idx), len(axes)):
axes[i].axis('off')
plt.suptitle(f'Top {top_k} Gate Receptive Fields (W1 columns, 28x28)\n'
f'Title: gate_id → preferred_digit', fontsize=10)
plt.tight_layout()
path = os.path.join(out_dir, 'gate_receptive_fields.png')
plt.savefig(path, dpi=150, bbox_inches='tight')
plt.close()
print(f'Saved: {path}')
def visualize_votes(W2, out_dir):
"""Heatmap of each gate's output vote (W2 row)."""
# top 32 gates
importance = np.abs(W2).sum(axis=1)
top_idx = np.argsort(importance)[::-1][:32]
fig, ax = plt.subplots(figsize=(6, 8))
im = ax.imshow(W2[top_idx], cmap='RdBu_r', aspect='auto')
ax.set_xlabel('Output digit')
ax.set_ylabel('Gate (sorted by importance)')
ax.set_xticks(range(10))
ax.set_yticks(range(len(top_idx)))
ax.set_yticklabels([f'g{i}' for i in top_idx], fontsize=7)
plt.colorbar(im, ax=ax, label='Vote strength')
plt.title('Gate Output Votes (W2 rows)')
plt.tight_layout()
path = os.path.join(out_dir, 'gate_output_votes.png')
plt.savefig(path, dpi=150, bbox_inches='tight')
plt.close()
print(f'Saved: {path}')
if __name__ == '__main__':
if not HAS_PLT:
raise SystemExit(1)
ckpt = os.path.join(os.path.dirname(__file__), '..', 'data', 'mnist_mat_ckpt.pkl')
if not os.path.exists(ckpt):
print(f'Checkpoint not found: {ckpt}')
raise SystemExit(1)
W1, W2 = load_ckpt(ckpt)
out_dir = os.path.join(os.path.dirname(__file__), '..', 'figures')
os.makedirs(out_dir, exist_ok=True)
visualize_gates(W1, W2, out_dir)
visualize_votes(W2, out_dir)
print('Done.')
|