File size: 10,568 Bytes
58d9bc3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Activation Energy Analysis for HPS Model
Generates a visualization showing:
1. Pre-GAP spatial activation energy heatmap (channels x spatial positions)
2. Per-channel energy before vs after GAP
3. Overall energy loss quantification
"""
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import matplotlib.gridspec as gridspec
from matplotlib.colors import LogNorm
import warnings
warnings.filterwarnings('ignore')

# Load the HPS model
print("Loading HPS model...")
model = tf.keras.models.load_model('/home/ubuntu/models/hps_model.h5', compile=False)

# Find the GAP layer and the layer before it
gap_layer = None
pre_gap_layer = None
for i, layer in enumerate(model.layers):
    if 'global_average_pooling' in layer.name.lower() or isinstance(layer, tf.keras.layers.GlobalAveragePooling1D):
        gap_layer = layer
        pre_gap_layer = model.layers[i - 1]
        print(f"Found GAP layer: {gap_layer.name} (index {i})")
        print(f"Pre-GAP layer: {pre_gap_layer.name} (index {i-1})")
        break

if gap_layer is None:
    # Try to find by output shape
    for i, layer in enumerate(model.layers):
        name = layer.name.lower()
        if 'pool' in name or 'gap' in name:
            print(f"  Candidate pooling layer: {layer.name}, output: {layer.output_shape}")
    
    # Search more broadly
    for i, layer in enumerate(model.layers):
        try:
            out_shape = layer.output_shape
            if isinstance(out_shape, tuple) and len(out_shape) == 2:
                prev = model.layers[i-1]
                prev_shape = prev.output_shape
                if isinstance(prev_shape, tuple) and len(prev_shape) == 3:
                    print(f"Found dimension reduction: {prev.name} {prev_shape} -> {layer.name} {out_shape}")
                    gap_layer = layer
                    pre_gap_layer = prev
                    break
        except:
            continue

print(f"\nPre-GAP output shape: {pre_gap_layer.output.shape}")
print(f"GAP output shape: {gap_layer.output.shape}")

# Create a sub-model that outputs pre-GAP activations
input_layer = model.input
pre_gap_output = pre_gap_layer.output
gap_output = gap_layer.output

sub_model = tf.keras.Model(inputs=input_layer, outputs=[pre_gap_output, gap_output])

# Generate synthetic input data (random traces, same shape as ASCAD)
# We use random data because we're analyzing the model's learned spatial structure,
# not the data itself. The weights determine how energy is distributed.
input_shape = model.input_shape[1:]
print(f"Input shape: {input_shape}")

# Use 100 random traces as stated in the report
np.random.seed(42)
batch_size = 100
X_batch = np.random.randn(batch_size, *input_shape).astype(np.float32)

print("Running forward pass...")
pre_gap_acts, gap_acts = sub_model.predict(X_batch, verbose=0)
print(f"Pre-GAP activations shape: {pre_gap_acts.shape}")  # (100, spatial, channels)
print(f"Post-GAP activations shape: {gap_acts.shape}")  # (100, channels)

# Compute activation energy metrics
# Energy = sum of squared activations (L2 energy)
spatial_dim = pre_gap_acts.shape[1]
channel_dim = pre_gap_acts.shape[2]

# Per-position, per-channel energy (averaged over batch)
# Shape: (spatial, channels)
energy_map = np.mean(pre_gap_acts ** 2, axis=0)  # (spatial, channels)
print(f"Energy map shape: {energy_map.shape}")

# Per-channel spatial variance (what GAP destroys)
per_channel_variance = np.mean(np.var(pre_gap_acts, axis=1), axis=0)  # (channels,)
per_channel_mean_energy = np.mean(np.mean(pre_gap_acts ** 2, axis=1), axis=0)  # (channels,)

# Total energy before and after GAP
total_energy_pre = np.mean(np.sum(pre_gap_acts ** 2, axis=(1, 2)))
total_energy_post = np.mean(np.sum(gap_acts ** 2, axis=1)) * spatial_dim  # scale back
energy_retained = total_energy_post / total_energy_pre * 100
energy_lost = 100 - energy_retained

# Variance-based analysis
total_variance_pre = np.sum(per_channel_variance)
mean_variance = np.mean(per_channel_variance)

print(f"\n=== ENERGY ANALYSIS ===")
print(f"Total pre-GAP energy: {total_energy_pre:.2f}")
print(f"Total post-GAP energy (scaled): {total_energy_post:.2f}")
print(f"Energy retained: {energy_retained:.1f}%")
print(f"Energy lost: {energy_lost:.1f}%")
print(f"Mean per-channel spatial variance: {mean_variance:.2f}")
print(f"Total spatial variance: {total_variance_pre:.2f}")

# Sort channels by variance for visualization
channel_variance_order = np.argsort(per_channel_variance)[::-1]

# ============================================================
# Generate the figure
# ============================================================
fig = plt.figure(figsize=(16, 14))
gs = gridspec.GridSpec(3, 2, height_ratios=[1.2, 1, 1], hspace=0.35, wspace=0.3)

# --- Panel (a): Pre-GAP Activation Energy Heatmap ---
ax1 = fig.add_subplot(gs[0, :])
# Show top 64 channels by variance for readability
top_channels = channel_variance_order[:64]
energy_subset = energy_map[:, top_channels].T  # (64, spatial)

# Downsample spatial dimension for readability
spatial_bins = 100
bin_size = spatial_dim // spatial_bins
energy_binned = np.zeros((len(top_channels), spatial_bins))
for b in range(spatial_bins):
    start = b * bin_size
    end = min((b + 1) * bin_size, spatial_dim)
    energy_binned[:, b] = np.mean(energy_subset[:, start:end], axis=1)

im = ax1.imshow(energy_binned, aspect='auto', cmap='hot', interpolation='nearest')
ax1.set_xlabel('Spatial Position (binned into 100 segments)', fontsize=11)
ax1.set_ylabel('Channel (sorted by variance)', fontsize=11)
ax1.set_title('(a) Pre-GAP Activation Energy Map (Top 64 Channels by Spatial Variance)', fontsize=13, fontweight='bold')
cbar = plt.colorbar(im, ax=ax1, shrink=0.8)
cbar.set_label('Mean Squared Activation', fontsize=10)

# Add annotation showing spatial structure exists
ax1.text(0.02, 0.95, f'Spatial dim: {spatial_dim} positions × {channel_dim} channels',
         transform=ax1.transAxes, fontsize=9, color='white', va='top',
         bbox=dict(boxstyle='round', facecolor='black', alpha=0.7))

# --- Panel (b): Per-Channel Energy Before vs After GAP ---
ax2 = fig.add_subplot(gs[1, 0])
# Show per-channel: total energy vs GAP-retained energy
channels_sorted = channel_variance_order[:64]
pre_energy_sorted = per_channel_mean_energy[channels_sorted]
post_energy_sorted = np.mean(gap_acts ** 2, axis=0)[channels_sorted]

x_pos = np.arange(len(channels_sorted))
width = 0.35
bars1 = ax2.bar(x_pos - width/2, pre_energy_sorted, width, label='Pre-GAP (per position)', color='#e74c3c', alpha=0.8)
bars2 = ax2.bar(x_pos + width/2, post_energy_sorted, width, label='Post-GAP (averaged)', color='#3498db', alpha=0.8)
ax2.set_xlabel('Channel Index (sorted by variance)', fontsize=10)
ax2.set_ylabel('Mean Squared Activation', fontsize=10)
ax2.set_title('(b) Per-Channel Energy: Before vs After GAP', fontsize=12, fontweight='bold')
ax2.legend(fontsize=9)
ax2.set_xticks(x_pos[::8])
ax2.set_xticklabels([str(c) for c in channels_sorted[::8]], fontsize=8)

# --- Panel (c): Spatial Variance Distribution ---
ax3 = fig.add_subplot(gs[1, 1])
variance_sorted = np.sort(per_channel_variance)[::-1]
ax3.bar(range(len(variance_sorted)), variance_sorted, color='#e67e22', alpha=0.8, width=1.0)
ax3.axhline(y=mean_variance, color='red', linestyle='--', linewidth=2, label=f'Mean variance: {mean_variance:.2f}')
ax3.set_xlabel('Channel (sorted by variance)', fontsize=10)
ax3.set_ylabel('Spatial Variance', fontsize=10)
ax3.set_title('(c) Per-Channel Spatial Variance (Destroyed by GAP)', fontsize=12, fontweight='bold')
ax3.legend(fontsize=10)

# --- Panel (d): Before/After GAP Summary ---
ax4 = fig.add_subplot(gs[2, 0])
# Show the averaging effect: pick one high-variance channel
example_channel = channel_variance_order[0]
example_activations = np.mean(pre_gap_acts[:, :, example_channel], axis=0)  # (spatial,)
gap_value = np.mean(gap_acts[:, example_channel])

ax4.plot(range(spatial_dim), example_activations, color='#e74c3c', alpha=0.6, linewidth=0.5, label='Pre-GAP spatial activations')
ax4.axhline(y=gap_value, color='#3498db', linewidth=3, linestyle='-', label=f'Post-GAP value: {gap_value:.4f}')
ax4.fill_between(range(spatial_dim), example_activations, gap_value, alpha=0.15, color='red')
ax4.set_xlabel('Spatial Position', fontsize=10)
ax4.set_ylabel('Activation Value', fontsize=10)
ax4.set_title(f'(d) GAP Collapse: Channel {example_channel} (Highest Variance)', fontsize=12, fontweight='bold')
ax4.legend(fontsize=9, loc='upper right')
ax4.text(0.02, 0.05, 'Red shaded area = information destroyed by averaging',
         transform=ax4.transAxes, fontsize=9, style='italic', color='#c0392b')

# --- Panel (e): Overall Energy Budget ---
ax5 = fig.add_subplot(gs[2, 1])
labels = ['Energy\nRetained', 'Energy\nDestroyed']
sizes = [energy_retained, energy_lost]
colors = ['#3498db', '#e74c3c']
explode = (0, 0.05)

wedges, texts, autotexts = ax5.pie(sizes, explode=explode, labels=labels, colors=colors,
                                     autopct='%1.1f%%', shadow=True, startangle=90,
                                     textprops={'fontsize': 12})
for autotext in autotexts:
    autotext.set_fontsize(14)
    autotext.set_fontweight('bold')
ax5.set_title('(e) Activation Energy Budget After GAP', fontsize=12, fontweight='bold')

# Add text box with exact numbers
textstr = (f'Pre-GAP: {spatial_dim} × {channel_dim} = {spatial_dim * channel_dim:,} values\n'
           f'Post-GAP: 1 × {channel_dim} = {channel_dim} values\n'
           f'Compression ratio: {spatial_dim}:1\n'
           f'Mean spatial variance: {mean_variance:.2f}')
props = dict(boxstyle='round', facecolor='wheat', alpha=0.8)
ax5.text(-0.1, -0.15, textstr, transform=ax5.transAxes, fontsize=9,
         verticalalignment='top', bbox=props)

plt.suptitle('HPS Model: Activation Energy Analysis Before and After Global Average Pooling',
             fontsize=15, fontweight='bold', y=0.98)

plt.savefig('/home/ubuntu/figures/fig7_activation_energy.png', dpi=200, bbox_inches='tight',
            facecolor='white', edgecolor='none')
plt.close()

print(f"\nFigure saved to /home/ubuntu/figures/fig7_activation_energy.png")
print(f"\nKey numbers for report:")
print(f"  Spatial dimension: {spatial_dim}")
print(f"  Channel dimension: {channel_dim}")
print(f"  Energy retained: {energy_retained:.1f}%")
print(f"  Energy destroyed: {energy_lost:.1f}%")
print(f"  Mean per-channel variance: {mean_variance:.2f}")
print(f"  Compression ratio: {spatial_dim}:1")