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
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
|
"""
Convolutional SNN with Lyapunov regularization for image classification.
Properly handles spatial structure:
- Input: (B, C, H, W) static image OR (B, T, C, H, W) spike tensor
- Uses Conv-LIF layers to preserve spatial hierarchy
- Rate encoding converts images to spike trains
Based on standard SNN vision practices:
- Rate/Poisson encoding for input
- Conv → BatchNorm → LIF → Pool architecture
- Time comes from encoding + LIF dynamics, not flattening
"""
from typing import Dict, List, Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
import snntorch as snn
from snntorch import surrogate
class RateEncoder(nn.Module):
"""
Rate (Poisson/Bernoulli) encoder for static images.
Converts intensity x ∈ [0,1] to spike probability per timestep.
Each pixel independently fires with P(spike) = x * gain.
Args:
T: Number of timesteps
gain: Scaling factor for firing probability (default 1.0)
Input: (B, C, H, W) normalized image in [0, 1]
Output: (B, T, C, H, W) binary spike tensor
"""
def __init__(self, T: int = 25, gain: float = 1.0):
super().__init__()
self.T = T
self.gain = gain
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Args:
x: (B, C, H, W) image tensor, values in [0, 1]
Returns:
spikes: (B, T, C, H, W) binary spike tensor
"""
# Clamp to valid probability range
prob = (x * self.gain).clamp(0, 1)
# Expand for T timesteps: (B, C, H, W) -> (B, T, C, H, W)
prob = prob.unsqueeze(1).expand(-1, self.T, -1, -1, -1)
# Sample spikes
spikes = torch.bernoulli(prob)
return spikes
class DirectEncoder(nn.Module):
"""
Direct encoding - feed static image as constant current.
Common in surrogate gradient papers: no spike encoding at input,
let spiking emerge from the network dynamics.
Input: (B, C, H, W) image
Output: (B, T, C, H, W) repeated image (as analog current)
"""
def __init__(self, T: int = 25):
super().__init__()
self.T = T
def forward(self, x: torch.Tensor) -> torch.Tensor:
# Simply repeat across time
return x.unsqueeze(1).expand(-1, self.T, -1, -1, -1)
class ConvLIFBlock(nn.Module):
"""
Conv → BatchNorm → LIF block.
Maintains spatial structure while adding spiking dynamics.
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
stride: int = 1,
padding: int = 1,
beta: float = 0.9,
threshold: float = 1.0,
spike_grad=None,
):
super().__init__()
if spike_grad is None:
spike_grad = surrogate.fast_sigmoid(slope=25)
self.conv = nn.Conv2d(
in_channels, out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
bias=False,
)
self.bn = nn.BatchNorm2d(out_channels)
self.lif = snn.Leaky(
beta=beta,
threshold=threshold,
spike_grad=spike_grad,
init_hidden=False,
)
def forward(
self,
x: torch.Tensor,
mem: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Args:
x: (B, C_in, H, W) input (spikes or current)
mem: (B, C_out, H', W') membrane potential
Returns:
spk: (B, C_out, H', W') output spikes
mem: (B, C_out, H', W') updated membrane
"""
cur = self.bn(self.conv(x))
spk, mem = self.lif(cur, mem)
return spk, mem
class ConvLyapunovSNN(nn.Module):
"""
Convolutional SNN with Lyapunov exponent regularization.
Architecture for CIFAR-10 (32x32x3):
Input → Encoder → [Conv-LIF-Pool] × N → FC → Output
Properly preserves spatial structure for hierarchical feature learning.
Args:
in_channels: Input channels (3 for RGB)
num_classes: Output classes
channels: List of channel sizes for conv layers
T: Number of timesteps
beta: LIF membrane decay
threshold: LIF firing threshold
encoding: 'rate', 'direct', or 'none' (pre-encoded input)
encoding_gain: Gain for rate encoding
"""
def __init__(
self,
in_channels: int = 3,
num_classes: int = 10,
channels: List[int] = [64, 128, 256],
T: int = 25,
beta: float = 0.9,
threshold: float = 1.0,
encoding: str = 'rate',
encoding_gain: float = 1.0,
dropout: float = 0.2,
):
super().__init__()
self.T = T
self.encoding_type = encoding
self.channels = channels
self.num_layers = len(channels)
# Input encoder
if encoding == 'rate':
self.encoder = RateEncoder(T=T, gain=encoding_gain)
elif encoding == 'direct':
self.encoder = DirectEncoder(T=T)
else:
self.encoder = None # Expect pre-encoded (B, T, C, H, W) input
# Build conv-LIF layers
self.blocks = nn.ModuleList()
self.pools = nn.ModuleList()
ch_in = in_channels
for ch_out in channels:
self.blocks.append(
ConvLIFBlock(ch_in, ch_out, beta=beta, threshold=threshold)
)
self.pools.append(nn.AvgPool2d(2))
ch_in = ch_out
# Calculate output spatial size after pooling
# CIFAR: 32 -> 16 -> 8 -> 4 (for 3 layers)
spatial_size = 32 // (2 ** len(channels))
fc_input = channels[-1] * spatial_size * spatial_size
# Fully connected readout
self.dropout = nn.Dropout(dropout)
self.fc = nn.Linear(fc_input, num_classes)
self._init_weights()
def _init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight)
if m.bias is not None:
nn.init.zeros_(m.bias)
def _init_mem(self, batch_size: int, device, dtype) -> List[torch.Tensor]:
"""Initialize membrane potentials for all layers."""
mems = []
H, W = 32, 32
for i, ch in enumerate(self.channels):
H, W = H // 2, W // 2 # After pooling
# Actually we need size BEFORE pooling for LIF
H_pre, W_pre = H * 2, W * 2
mems.append(torch.zeros(batch_size, ch, H_pre, W_pre, device=device, dtype=dtype))
return mems
def forward(
self,
x: torch.Tensor,
compute_lyapunov: bool = False,
lyap_eps: float = 1e-4,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Dict]]:
"""
Forward pass with optional Lyapunov computation.
Args:
x: Input tensor
- If encoder: (B, C, H, W) static image
- If no encoder: (B, T, C, H, W) pre-encoded spikes
compute_lyapunov: Whether to compute Lyapunov exponent
lyap_eps: Perturbation magnitude
Returns:
logits: (B, num_classes)
lyap_est: Scalar Lyapunov estimate or None
recordings: Optional dict with spike recordings
"""
# Encode input if needed
if self.encoder is not None:
x = self.encoder(x) # (B, C, H, W) -> (B, T, C, H, W)
B, T, C, H, W = x.shape
device, dtype = x.device, x.dtype
# Initialize membrane potentials
mems = self._init_mem(B, device, dtype)
# For accumulating output spikes
spike_sum = None
# Lyapunov setup
if compute_lyapunov:
mems_p = [m + lyap_eps * torch.randn_like(m) for m in mems]
lyap_accum = torch.zeros(B, device=device, dtype=dtype)
# Time loop
for t in range(T):
x_t = x[:, t] # (B, C, H, W)
# Forward through conv-LIF blocks
h = x_t
new_mems = []
for i, (block, pool) in enumerate(zip(self.blocks, self.pools)):
h, mem = block(h, mems[i])
new_mems.append(mem)
h = pool(h) # Spatial downsampling
mems = new_mems
# Accumulate final layer spikes
if spike_sum is None:
spike_sum = h.view(B, -1)
else:
spike_sum = spike_sum + h.view(B, -1)
# Lyapunov computation
if compute_lyapunov:
h_p = x_t
new_mems_p = []
for i, (block, pool) in enumerate(zip(self.blocks, self.pools)):
h_p, mem_p = block(h_p, mems_p[i])
new_mems_p.append(mem_p)
h_p = pool(h_p)
# Compute divergence
delta_sq = torch.zeros(B, device=device, dtype=dtype)
for i in range(self.num_layers):
diff = new_mems_p[i] - new_mems[i]
delta_sq += (diff ** 2).sum(dim=(1, 2, 3))
delta = torch.sqrt(delta_sq + 1e-12)
lyap_accum = lyap_accum + torch.log(delta / lyap_eps + 1e-12)
# Renormalize perturbation
for i in range(self.num_layers):
diff = new_mems_p[i] - new_mems[i]
norm = torch.sqrt((diff ** 2).sum(dim=(1, 2, 3), keepdim=True) + 1e-12)
# Broadcast norm to spatial dimensions
norm = norm.view(B, 1, 1, 1)
new_mems_p[i] = new_mems[i] + lyap_eps * diff / norm
mems_p = new_mems_p
# Readout
out = self.dropout(spike_sum)
logits = self.fc(out)
if compute_lyapunov:
lyap_est = (lyap_accum / T).mean()
else:
lyap_est = None
return logits, lyap_est, None
class VGGLyapunovSNN(nn.Module):
"""
VGG-style deep Conv-SNN with Lyapunov regularization.
Deeper architecture for more challenging benchmarks.
Uses multiple conv layers between pooling to increase depth.
Architecture (VGG-9 style):
[Conv-LIF × 2, Pool] → [Conv-LIF × 2, Pool] → [Conv-LIF × 2, Pool] → FC
"""
def __init__(
self,
in_channels: int = 3,
num_classes: int = 10,
T: int = 25,
beta: float = 0.9,
threshold: float = 1.0,
encoding: str = 'rate',
dropout: float = 0.3,
):
super().__init__()
self.T = T
self.encoding_type = encoding
spike_grad = surrogate.fast_sigmoid(slope=25)
if encoding == 'rate':
self.encoder = RateEncoder(T=T)
elif encoding == 'direct':
self.encoder = DirectEncoder(T=T)
else:
self.encoder = None
# VGG-style blocks: (in_ch, out_ch, num_convs)
block_configs = [
(in_channels, 64, 2), # 32x32 -> 16x16
(64, 128, 2), # 16x16 -> 8x8
(128, 256, 2), # 8x8 -> 4x4
]
self.blocks = nn.ModuleList()
for in_ch, out_ch, n_convs in block_configs:
layers = []
for i in range(n_convs):
ch_in = in_ch if i == 0 else out_ch
layers.append(nn.Conv2d(ch_in, out_ch, 3, padding=1, bias=False))
layers.append(nn.BatchNorm2d(out_ch))
self.blocks.append(nn.ModuleList(layers))
# LIF neurons for each conv layer
self.lifs = nn.ModuleList([
snn.Leaky(beta=beta, threshold=threshold, spike_grad=spike_grad, init_hidden=False)
for _ in range(6) # 2 convs × 3 blocks
])
self.pools = nn.ModuleList([nn.AvgPool2d(2) for _ in range(3)])
# FC layers
self.fc1 = nn.Linear(256 * 4 * 4, 512)
self.lif_fc = snn.Leaky(beta=beta, threshold=threshold, spike_grad=spike_grad, init_hidden=False)
self.dropout = nn.Dropout(dropout)
self.fc2 = nn.Linear(512, num_classes)
self._init_weights()
def _init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out')
elif isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight)
if m.bias is not None:
nn.init.zeros_(m.bias)
def forward(
self,
x: torch.Tensor,
compute_lyapunov: bool = False,
lyap_eps: float = 1e-4,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Dict]]:
if self.encoder is not None:
x = self.encoder(x)
B, T, C, H, W = x.shape
device, dtype = x.device, x.dtype
# Initialize all membrane potentials
# For each conv layer output
mem_shapes = [
(B, 64, 32, 32), (B, 64, 32, 32), # Block 1
(B, 128, 16, 16), (B, 128, 16, 16), # Block 2
(B, 256, 8, 8), (B, 256, 8, 8), # Block 3
(B, 512), # FC
]
mems = [torch.zeros(s, device=device, dtype=dtype) for s in mem_shapes]
spike_sum = torch.zeros(B, 512, device=device, dtype=dtype)
if compute_lyapunov:
mems_p = [m + lyap_eps * torch.randn_like(m) for m in mems]
lyap_accum = torch.zeros(B, device=device, dtype=dtype)
for t in range(T):
h = x[:, t]
lif_idx = 0
for block_idx, (block_layers, pool) in enumerate(zip(self.blocks, self.pools)):
for i in range(0, len(block_layers), 2): # Conv, BN pairs
conv, bn = block_layers[i], block_layers[i + 1]
h = bn(conv(h))
h, mems[lif_idx] = self.lifs[lif_idx](h, mems[lif_idx])
lif_idx += 1
h = pool(h)
# FC layers
h = h.view(B, -1)
h = self.fc1(h)
h, mems[6] = self.lif_fc(h, mems[6])
spike_sum = spike_sum + h
# Lyapunov (simplified - just on last layer)
if compute_lyapunov:
diff = mems[6] - mems_p[6] if t > 0 else torch.zeros_like(mems[6])
delta = torch.norm(diff.view(B, -1), dim=1) + 1e-12
if t > 0:
lyap_accum = lyap_accum + torch.log(delta / lyap_eps + 1e-12)
mems_p[6] = mems[6] + lyap_eps * torch.randn_like(mems[6])
out = self.dropout(spike_sum)
logits = self.fc2(out)
lyap_est = (lyap_accum / T).mean() if compute_lyapunov else None
return logits, lyap_est, None
def create_conv_snn(
model_type: str = 'simple',
**kwargs,
) -> nn.Module:
"""
Factory function for Conv-SNN models.
Args:
model_type: 'simple' (3-layer) or 'vgg' (6-layer VGG-style)
**kwargs: Arguments passed to model constructor
"""
if model_type == 'simple':
return ConvLyapunovSNN(**kwargs)
elif model_type == 'vgg':
return VGGLyapunovSNN(**kwargs)
else:
raise ValueError(f"Unknown model_type: {model_type}")
|