-
Notifications
You must be signed in to change notification settings - Fork 0
/
peak_backprop.py
43 lines (33 loc) · 1.26 KB
/
peak_backprop.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
import torch
import torch.nn.functional as F
from torch.autograd import Function
class PreHook(Function):
@staticmethod
def forward(ctx, input, offset):
ctx.save_for_backward(input, offset)
return input.clone()
@staticmethod
def backward(ctx, grad_output):
input, offset = ctx.saved_variables
return (input - offset) * grad_output, None
class PostHook(Function):
@staticmethod
def forward(ctx, input, norm_factor):
ctx.save_for_backward(norm_factor)
return input.clone()
@staticmethod
def backward(ctx, grad_output):
norm_factor, = ctx.saved_variables
eps = 1e-10
zero_mask = norm_factor < eps
grad_input = grad_output / (torch.abs(norm_factor) + eps)
grad_input[zero_mask.detach()] = 0
return None, grad_input
def pr_conv2d(self, input):
offset = input.min().detach()
input = PreHook.apply(input, offset)
resp = F.conv2d(input, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups).detach()
pos_weight = F.relu(self.weight).detach()
norm_factor = F.conv2d(input - offset, pos_weight, None, self.stride, self.padding, self.dilation, self.groups)
output = PostHook.apply(resp, norm_factor)
return output