-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlocal_conv_ELL.py
164 lines (144 loc) · 6.49 KB
/
local_conv_ELL.py
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
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from linearFA import LinearFA
from surrogate_gradient import ExponentialSurroGrad
class LocalLossBlockConv(nn.Module):
def __init__(self,
ch_in,
ch_out,
kernel_size,
stride,
padding,
dim_out,
dim_decoder,
num_classes,
threshold=None,
decay=0.2,
print_stats=False,
first_layer=False,
bias=False,
random_matrix=False,
fa=False,
last_layer=False):
super(LocalLossBlockConv, self).__init__()
self.ch_in = ch_in
self.ch_out = ch_out
self.num_classes = num_classes
self.first_layer = first_layer
self.last_layer = last_layer
self.bias = bias
self.thresh = threshold
self.decay = decay
self.dim_in_decoder = dim_decoder
self.random_matrix = random_matrix
self.fa = fa
self.is_print_stats = print_stats
self.encoder = nn.Conv2d(ch_in, ch_out, kernel_size, stride=stride, padding=padding, bias=self.bias)
if True:
# Resolve average-pooling kernel size in order for flattened dim to match self.dim_in_decoder
ks_h, ks_w = 1, 1
dim_out_h, dim_out_w = dim_out, dim_out
dim_in_decoder = ch_out * dim_out_h * dim_out_w
while dim_in_decoder > self.dim_in_decoder and ks_h < dim_out:
ks_h *= 2
dim_out_h = math.ceil(dim_out / ks_h)
dim_in_decoder = ch_out * dim_out_h * dim_out_w
if dim_in_decoder > self.dim_in_decoder:
ks_w *= 2
dim_out_w = math.ceil(dim_out / ks_w)
dim_in_decoder = ch_out * dim_out_h * dim_out_w
if ks_h > 1 or ks_w > 1:
pad_h = (ks_h * (dim_out_h - dim_out // ks_h)) // 2
pad_w = (ks_w * (dim_out_w - dim_out // ks_w)) // 2
self.avg_pool = nn.AvgPool2d((ks_h, ks_w), padding=(pad_h, pad_w))
print(self.avg_pool)
else:
self.avg_pool = None
# Auxiliary classifier
if self.fa:
self.decoder_y = LinearFA(dim_in_decoder, num_classes, bias=self.bias)
else:
self.decoder_y = nn.Linear(dim_in_decoder, num_classes, bias=self.bias)
if self.random_matrix:
self.decoder_y.weight.requires_grad = False
if self.decoder_y.bias is not None:
self.decoder_y.bias.requires_grad = False
self.optimizer = torch.optim.Adam(self.parameters(), lr=0, weight_decay=0.0)
self.clear_stats()
def clear_stats(self):
self.loss_pred = 0.0
self.correct = 0
self.examples = 0
def print_stats(self):
err = 100.0 * float(self.examples - self.correct) / self.examples
stats = '{},loss_pred={:.4f}, error={:.3f}%, num_examples={}\n'.format(
self.encoder,
self.loss_pred / self.examples,
err,
self.examples)
return stats, err
def set_learning_rate(self, lr):
self.lr = lr
for param_group in self.optimizer.param_groups:
param_group['lr'] = self.lr
def optim_zero_grad(self):
self.optimizer.zero_grad()
def optim_step(self):
self.optimizer.step()
def forward(self, x, y, y_onehot, time_window=10):
# Loop over time window T
for step in range(time_window):
# Compute the input current
if self.first_layer:
h = self.encoder(x)
else:
h = self.encoder(x[step])
# Membrane integration and spike firing
if step == 0:
mem = h
spike = ExponentialSurroGrad.apply(mem, self.thresh)
spike_return = spike.clone() # Output spikes
if self.training or self.is_print_stats:
if self.avg_pool is not None:
mem_pool = self.avg_pool(
spike) # Average pooling operation is applied first in order to reduce the number of auxilinary weights
else:
mem_pool = spike
y_hat_mem = self.decoder_y(mem_pool.view(mem_pool.size(0), -1))
y_hat_spike = ExponentialSurroGrad.apply(y_hat_mem, self.thresh)
spike_sum = y_hat_spike # Accumulate spikes for decision
else:
# Detach the membrane and spike from the computation graph because ELL doesn't consider temporal dependent gradients.
mem = mem.detach() * self.decay + h - spike.detach() * self.thresh * self.decay # Membrane integration
spike = ExponentialSurroGrad.apply(mem, self.thresh)
spike_return = torch.cat((spike_return, spike), dim=0)
if self.training or self.is_print_stats:
if self.avg_pool is not None:
mem_pool = self.avg_pool(spike)
else:
mem_pool = spike
y_hat_mem = y_hat_mem.detach() * self.decay + self.decoder_y(
mem_pool.view(mem_pool.size(0), -1)) - y_hat_spike.detach() * self.thresh * self.decay
y_hat_spike = ExponentialSurroGrad.apply(y_hat_mem, self.thresh)
spike_sum = spike_sum + y_hat_spike
# Compute loss and update parameters during training
if self.training or self.is_print_stats:
loss_sup = F.mse_loss(y_hat_spike, y_onehot.detach())
if step == 0:
loss = loss_sup
else:
loss = loss + loss_sup
if self.training:
self.optimizer.zero_grad()
loss_sup.backward(retain_graph=False)
self.optimizer.step() # Update all parameters
spike_return = spike_return.view(-1, *spike.size())
if self.is_print_stats:
self.loss_pred += loss.item() * y_onehot.size(0)
self.correct += spike_sum.max(1)[1].eq(y).cpu().sum()
self.examples += y_onehot.size(0)
if self.last_layer:
return (100.0 * float(self.examples - self.correct) / self.examples), loss.item()
return spike_return.detach(), loss.item()