/
models.py
217 lines (184 loc) · 8.65 KB
/
models.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
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
import os
import numpy as np
import math
import torch
from torch.autograd import Variable
import torch.nn as nn
import torch.nn.functional as F
path = 'train'
if not os.path.isdir(path):
os.makedirs(path)
def init_truncated_normal(model, aux_str=''):
if model is None: return None
init_path = '{path}/{in_dim:d}_{out_dim:d}{aux_str}.pth' \
.format(path=path, in_dim=model.in_features, out_dim=model.out_features, aux_str=aux_str)
if os.path.isfile(init_path):
model.load_state_dict(torch.load(init_path))
print('load init weight: {init_path}'.format(init_path=init_path))
else:
if isinstance(model, nn.ModuleList):
[truncated_normal(sub) for sub in model]
else:
truncated_normal(model)
print('generate init weight: {init_path}'.format(init_path=init_path))
torch.save(model.state_dict(), init_path)
print('save init weight: {init_path}'.format(init_path=init_path))
return model
def truncated_normal(model):
std = math.sqrt(2./(model.in_features + model.out_features))
if model.bias is not None:
model.bias.data.zero_()
model.weight.data.normal_(std=std)
truncate_me = (model.weight.data > 2.*std) | (model.weight.data < -2.*std)
while truncate_me.sum() > 0:
model.weight.data[truncate_me] = torch.normal(std=std*torch.ones(truncate_me.sum()))
truncate_me = (model.weight.data > 2.*std) | (model.weight.data < -2.*std)
return model
class DeepLinearReLU(nn.Sequential):
def __init__(self, dims, no_last_relu=False, init=True):
super(DeepLinearReLU, self).__init__()
for d in range(len(dims)):
if d == 0: continue
sub = nn.Linear(dims[d-1], dims[d])
if init: sub = init_truncated_normal(sub, '_d{d:d}'.format(d=d))
self.add_module('fc{d}'.format(d=d), sub)
if no_last_relu and d == len(dims)-1: continue
self.add_module('relu{d}'.format(d=d), nn.ReLU())
class FLModel(nn.Linear):
def reset_parameters(self):
init_truncated_normal(self)
def forward(self, input, input_norm=False):
if input_norm: input = F.normalize(input, p=2, dim=1)
return super(FLModel, self).forward(input)
class TDModel(nn.ModuleList):
def __init__(self, in_dim, out_dims, ns=False, input_norm=False, relu=False, softmax='n'):
super(TDModel, self).__init__()
self.input_norm = input_norm
self.relu = relu
self.softmax = softmax
self += [nn.Linear(in_dim, out_dim+int(ns)) for out_dim in out_dims]
self.in_features = in_dim
self.out_features = sum(out_dims)
if ns: self.out_features += len(out_dims)
self.bias = None
init_truncated_normal(self, '_td')
def forward(self, input, m=-1):
if self.input_norm: input = F.normalize(input, p=2, dim=1)
if m < 0:
return torch.cat([self.postprocess(sub(input)) for sub in self], dim=1)
else:
return self.postprocess(self[m](input))
def postprocess(self, input):
if self.relu:
input = F.relu(input)
if self.softmax == 'l':
input = F.log_softmax(input, dim=1)
elif self.softmax == 's':
input = F.softmax(input, dim=1)
return input
class LOOLoss(nn.Module):
def __init__(self, T, opts):
super(LOOLoss, self).__init__()
self.gpu = opts.gpu
self.loo = opts.loo if 'LOO' in opts.method else 0.
self.label_smooth = opts.label_smooth
self.kld_u_const = math.log(len(T['wnids']))
self.relevant = [torch.from_numpy(rel) for rel in T['relevant']]
self.labels_relevant = torch.from_numpy(T['labels_relevant'].astype(np.uint8))
ch_slice = T['ch_slice']
if opts.class_wise:
num_children = T['num_children']
num_supers = len(num_children)
self.class_weight = torch.zeros(ch_slice[-1])
for m, num_ch in enumerate(num_children):
self.class_weight[ch_slice[m]:ch_slice[m+1]] = 1. / (num_ch * num_supers)
else:
self.class_weight = torch.ones(ch_slice[-1]) / ch_slice[-1]
def forward(self, input, target): # input = Variable(logits), target = labels
loss = Variable(torch.zeros(1).cuda()) if self.gpu else Variable(torch.zeros(1))
# novel loss
if self.loo > 0.:
target_novel = self.labels_relevant[target]
for i, rel in enumerate(self.relevant):
if target_novel[:,i].any():
relevant_loc = target_novel[:,i].nonzero().view(-1)
loss += -F.log_softmax(input[relevant_loc][:, rel], dim=1)[:,0].mean() * self.class_weight[i]
loss *= self.loo
# known loss
log_probs = F.log_softmax(input, dim=1)
loss += F.nll_loss(log_probs, Variable(target))
# regularization
if self.label_smooth > 0.:
loss -= (log_probs.mean() + self.kld_u_const) * self.label_smooth
return loss
def cuda(self, device=None):
super(LOOLoss, self).cuda(device)
self.relevant = [rel.cuda(device) for rel in self.relevant]
self.labels_relevant = self.labels_relevant.cuda(device)
return self
class TDLoss(nn.Module):
def __init__(self, T, opts):
super(TDLoss, self).__init__()
self.gpu = opts.gpu
self.label_smooth = opts.label_smooth
self.ex_smooth = opts.ex_smooth if opts.method == 'TD' else 0.
self.class_wise = opts.class_wise
self.novel_score = opts.novel_score
self.labels_ch = torch.from_numpy(T['labels_ch'])
self.labels_in = torch.from_numpy(T['labels_in'].astype(np.uint8))
self.labels_out = torch.from_numpy(T['labels_out'].astype(np.uint8))
self.root = T['root'] - len(T['wnids_leaf'])
self.num_children = T['num_children']
self.ch_slice = T['ch_slice']
self.kld_u_const = [math.log(num_ch) for num_ch in self.num_children]
def forward(self, input, target, m): # input = Variable(logits), target = labels
loss = Variable(torch.zeros(1).cuda()) if self.gpu else Variable(torch.zeros(1))
# known loss
log_probs = F.log_softmax(input, dim=1)
num_inputs = 0
for i_ch in range(self.num_children[m]):
target_ch = self.labels_ch[:, self.ch_slice[m]+i_ch][target]
num_inputs_ch = (target_ch >= 0).sum()
if num_inputs_ch > 0:
num_inputs += num_inputs_ch
loss += F.nll_loss(log_probs, Variable(target_ch), size_average=self.class_wise, ignore_index=-1)
# training model parameters of novel classes
if self.novel_score:
# novel loss
if m != self.root and self.ex_smooth > 0.: # root does not have exclusive leaves
target_out = self.labels_out[:,m][target]
num_inputs_ch = target_out.sum()
if num_inputs_ch > 0:
num_inputs += num_inputs_ch
target_ch = self.num_children[m]*target_out.long()
loss += F.nll_loss(log_probs, Variable(target_ch),
size_average=self.class_wise, ignore_index=0) * self.ex_smooth
# known & novel loss normalization
if self.class_wise:
loss /= self.num_children[m] + (m != self.root)
elif num_inputs > 0:
loss /= num_inputs
# smoothing output of novel classes
else:
# known loss normalization
if self.class_wise:
loss /= self.num_children[m]
elif num_inputs > 0:
loss /= num_inputs
# novel loss
if m != self.root and self.ex_smooth > 0.: # root does not have exclusive leaves
target_out = self.labels_out[:,m][target]
if target_out.any():
loss -= (log_probs[target_out.nonzero().view(-1)].mean() + self.kld_u_const[m]) * self.ex_smooth
# regularization
if self.label_smooth > 0.:
target_in = self.labels_in[:,m][target]
if target_in.any():
loss -= (log_probs[target_in.nonzero().view(-1)].mean() + self.kld_u_const[m]) * self.label_smooth
return loss
def cuda(self, device=None):
super(TDLoss, self).cuda(device)
self.labels_ch = self.labels_ch.cuda(device)
self.labels_in = self.labels_in.cuda(device)
self.labels_out = self.labels_out.cuda(device)
return self