-
Notifications
You must be signed in to change notification settings - Fork 0
/
forward_wrapper.py
147 lines (125 loc) · 6.19 KB
/
forward_wrapper.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
from utils import *
import autograd_wl
"""
Wrapper for variance reduction opts
"""
class ForwardWrapper(nn.Module):
def __init__(self, n_nodes, n_hid, n_layers, n_classes, concat=False):
super(ForwardWrapper, self).__init__()
self.n_layers = n_layers
if concat:
self.hiddens = torch.zeros(n_layers, n_nodes, 2*n_hid)
else:
self.hiddens = torch.zeros(n_layers, n_nodes, n_hid)
def forward_full(self, net, x, adjs, sampled_nodes):
for ell in range(len(net.gcs)):
x = net.gcs[ell](x, adjs[ell])
self.hiddens[ell,sampled_nodes[ell]] = x.cpu().detach()
x = net.relu(x)
x = net.dropout(x)
x = net.gc_out(x)
return x
def forward_mini(self, net, x, adjs, sampled_nodes, x_exact, adjs_exact, input_exact_nodes):
cached_outputs = []
for ell in range(len(net.gcs)):
x_bar = x if ell == 0 else net.dropout(net.relu(self.hiddens[ell-1,sampled_nodes[ell-1]].to(x)))
x_bar_exact = x_exact[input_exact_nodes[ell]] if ell == 0 else net.dropout(net.relu(self.hiddens[ell-1,input_exact_nodes[ell]].to(x)))
x = net.gcs[ell](x, adjs[ell]) - net.gcs[ell](x_bar, adjs[ell]) + net.gcs[ell](x_bar_exact, adjs_exact[ell])
cached_outputs += [x.detach().cpu()]
x = net.relu(x)
x = net.dropout(x)
x = net.gc_out(x)
for ell in range(len(net.gcs)):
self.hiddens[ell, sampled_nodes[ell]] = cached_outputs[ell]
return x
def calculate_sample_grad(self, net, x, adjs, sampled_nodes, targets, batch_nodes):
outputs = self.forward_full(net, x, adjs, sampled_nodes)
loss = net.loss_f(outputs, targets[batch_nodes])
loss.backward()
grad_per_sample = autograd_wl.calculate_sample_grad(net.gc_out)
return grad_per_sample.cpu().numpy()
def partial_grad(self, net, x, adjs, sampled_nodes, x_exact, adjs_exact, input_exact_nodes, targets, weight=None):
outputs = self.forward_mini(net, x, adjs, sampled_nodes, x_exact, adjs_exact, input_exact_nodes)
if weight is None:
loss = net.loss_f(outputs, targets)
else:
if net.multi_class:
loss = net.loss_f_vec(outputs, targets)
loss = loss.mean(1) * weight
else:
loss = net.loss_f_vec(outputs, targets) * weight
loss = loss.sum()
loss.backward()
return loss.detach()
def partial_grad_with_norm(self, net, x, adjs, sampled_nodes, x_exact, adjs_exact, input_exact_nodes, targets, weight):
num_samples = targets.size(0)
outputs = self.forward_mini(net, x, adjs, sampled_nodes, x_exact, adjs_exact, input_exact_nodes)
if net.multi_class:
loss = net.loss_f_vec(outputs, targets)
loss = loss.mean(1) * weight
else:
loss = net.loss_f_vec(outputs, targets) * weight
loss = loss.sum()
loss.backward()
grad_per_sample = autograd_wl.calculate_sample_grad(net.gc_out)
grad_per_sample = grad_per_sample*(1/weight/num_samples)
return loss.detach(), grad_per_sample.cpu().numpy()
class ForwardWrapper_v2(nn.Module):
def __init__(self, n_nodes, n_hid, n_layers, n_classes, concat=False):
super(ForwardWrapper_v2, self).__init__()
self.n_layers = n_layers
if concat:
self.hiddens = torch.zeros(n_layers, n_nodes, 2*n_hid)
else:
self.hiddens = torch.zeros(n_layers, n_nodes, n_hid)
def forward_full(self, net, x, adjs, sampled_nodes):
for ell in range(len(net.gcs)):
x = net.gcs[ell](x, adjs[ell])
self.hiddens[ell,sampled_nodes[ell]] = x.cpu().detach()
x = net.dropout(net.relu(x))
x = net.gc_out(x)
return x
def forward_mini(self, net, x, adjs, sampled_nodes, x_exact, adjs_exact, input_exact_nodes):
cached_outputs = []
for ell in range(len(net.gcs)):
x_bar_exact = x_exact[input_exact_nodes[ell]] if ell == 0 else net.dropout(net.relu(self.hiddens[ell-1,input_exact_nodes[ell]].to(x)))
x = torch.cat([x, x_bar_exact], dim=0)
x = net.gcs[ell](x, adjs_exact[ell])
cached_outputs += [x.detach().cpu()]
x = net.dropout(net.relu(x))
x = net.gc_out(x)
for ell in range(len(net.gcs)):
self.hiddens[ell, sampled_nodes[ell]] = cached_outputs[ell]
return x
def calculate_sample_grad(self, net, x, adjs, sampled_nodes, targets, batch_nodes):
outputs = self.forward_full(net, x, adjs, sampled_nodes)
loss = net.loss_f(outputs, targets[batch_nodes])
loss.backward()
grad_per_sample = autograd_wl.calculate_sample_grad(net.gc_out)
return grad_per_sample.cpu().numpy()
def partial_grad(self, net, x, adjs, sampled_nodes, x_exact, adjs_exact, input_exact_nodes, targets, weight=None):
outputs = self.forward_mini(net, x, adjs, sampled_nodes, x_exact, adjs_exact, input_exact_nodes)
if weight is None:
loss = net.loss_f(outputs, targets)
else:
if net.multi_class:
loss = net.loss_f_vec(outputs, targets)
loss = loss.mean(1) * weight
else:
loss = net.loss_f_vec(outputs, targets) * weight
loss = loss.sum()
loss.backward()
return loss.detach()
def partial_grad_with_norm(self, net, x, adjs, sampled_nodes, x_exact, adjs_exact, input_exact_nodes, targets, weight):
num_samples = targets.size(0)
outputs = self.forward_mini(net, x, adjs, sampled_nodes, x_exact, adjs_exact, input_exact_nodes)
if net.multi_class:
loss = net.loss_f_vec(outputs, targets)
loss = loss.mean(1) * weight
else:
loss = net.loss_f_vec(outputs, targets) * weight
loss = loss.sum()
loss.backward()
grad_per_sample = autograd_wl.calculate_sample_grad(net.gc_out)
grad_per_sample = grad_per_sample*(1/weight/num_samples)
return loss.detach(), grad_per_sample.cpu().numpy()