/
resnet.py
55 lines (45 loc) 路 1.77 KB
/
resnet.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
# coding: utf8
from __future__ import unicode_literals
from .model import Model
class Residual(Model):
def __init__(self, layer):
Model.__init__(self)
self._layers.append(layer)
self.on_data_hooks.append(on_data)
@property
def nO(self):
return self._layers[-1].nO
def predict(self, X):
Y = self._layers[0](X)
if isinstance(X, list) or isinstance(X, tuple):
return [X[i] + Y[i] for i in range(len(X))]
elif isinstance(X, tuple) and isinstance(Y, tuple) and len(X) == 2:
assert X[1].sum() == Y[1].sum()
assert Y[1].sum() == Y[0].shape[0], (Y[1].sum(), Y[0].shape[0])
return (X[0] + Y[0], Y[1])
else:
return X + Y
def begin_update(self, X, drop=0.0):
y, bp_y = self._layers[0].begin_update(X, drop=drop)
if isinstance(X, list):
output = [X[i] + y[i] for i in range(len(X))]
elif isinstance(X, tuple) and isinstance(y, tuple) and len(X) == 2:
# Handle case where we have (data, lengths) tuple
assert X[1].sum() == y[1].sum()
assert y[1].sum() == y[0].shape[0], (y[1].sum(), y[0].shape[0])
output = (X[0] + y[0], y[1])
else:
output = X + y
def residual_bwd(d_output, sgd=None):
dX = bp_y(d_output, sgd)
if isinstance(d_output, list) or isinstance(d_output, tuple):
return [d_output[i] + dX[i] for i in range(len(d_output))]
else:
return d_output + dX
return output, residual_bwd
def on_data(self, X, y=None):
for layer in self._layers:
for hook in layer.on_data_hooks:
hook(layer, X, y)
if hasattr(layer, "W"):
layer.W.fill(0)