-
Notifications
You must be signed in to change notification settings - Fork 21
/
model.py
163 lines (121 loc) · 5.94 KB
/
model.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
import datetime
import math
import numpy as np
import torch
from torch import nn
from torch.nn import Module, Parameter
import torch.nn.functional as F
from sklearn import metrics
from layers import *
from tqdm import tqdm
def trans_to_cuda(variable):
if torch.cuda.is_available():
return variable.cuda()
else:
return variable
def trans_to_cpu(variable):
if torch.cuda.is_available():
return variable.cpu()
else:
return variable
class HGNN_ATT(nn.Module):
def __init__(self, input_size, n_hid, output_size, dropout=0.3):
super(HGNN_ATT, self).__init__()
self.dropout = dropout
self.gat1 = HyperGraphAttentionLayerSparse(input_size, n_hid, dropout=self.dropout, alpha=0.2, transfer = False, concat=True)
self.gat2 = HyperGraphAttentionLayerSparse(n_hid, output_size, dropout=self.dropout, alpha=0.2, transfer = True, concat=False)
def forward(self, x, H):
x = self.gat1(x, H)
x = F.dropout(x, self.dropout, training=self.training)
x = self.gat2(x, H)
return x
class DocumentGraph(Module):
def __init__(self, opt, pre_trained_weight, class_weights, n_node, n_categories):
super(DocumentGraph, self).__init__()
self.hidden_size = opt.hiddenSize
self.n_node = n_node
self.n_categories = n_categories
self.batch_size = opt.batchSize
self.dropout = opt.dropout
self.initial_feature = opt.initialFeatureSize
self.normalization = opt.normalization
self.dataset = opt.dataset
self.embedding = nn.Embedding(self.n_node+1, self.initial_feature, padding_idx=0)
self.layer_normH = nn.LayerNorm(self.hidden_size, eps=1e-6)
if self.normalization:
self.layer_normC = nn.LayerNorm(self.n_categories, eps=1e-6)
self.prediction_transform = nn.Linear(self.hidden_size, self.n_categories, bias=True)
self.reset_parameters()
if opt.dataset == 'mr':
pre_trained_weight = torch.FloatTensor(pre_trained_weight)
self.embedding = nn.Embedding.from_pretrained(pre_trained_weight, freeze = False, padding_idx = 0)
self.hgnn = HGNN_ATT(self.initial_feature, self.initial_feature, self.hidden_size, dropout = self.dropout)
self.class_weights = class_weights
self.loss_function = nn.CrossEntropyLoss(weight = trans_to_cuda(torch.Tensor(self.class_weights).float()))
self.optimizer = torch.optim.Adam(self.parameters(), lr=opt.lr, weight_decay=opt.l2)
self.scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer, step_size=opt.lr_dc_step, gamma=opt.lr_dc)
def reset_parameters(self):
stdv = 1.0 / math.sqrt(self.hidden_size)
for weight in self.parameters():
weight.data.uniform_(-stdv, stdv)
def compute_scores(self, inputs, node_masks):
hidden = inputs * node_masks.view(node_masks.shape[0], -1, 1).float()
b = torch.sum(hidden * node_masks.view(node_masks.shape[0], -1, 1).float(),-2)/torch.sum(node_masks,-1).repeat(hidden.shape[2],1).transpose(0,1)
b = self.layer_normH(b)
b = self.prediction_transform(b)
pred = b
if self.normalization:
pred = self.layer_normC(b)
return pred
def forward(self, inputs, HT):
hidden = self.embedding(inputs)
nodes = self.hgnn(hidden, HT)
return nodes
def forward(model, alias_inputs, HT, items, targets, node_masks):
alias_inputs = trans_to_cuda(torch.Tensor(alias_inputs).long())
items = trans_to_cuda(torch.Tensor(items).long())
HT = trans_to_cuda(torch.Tensor(HT).float())
node_masks = trans_to_cuda(torch.Tensor(node_masks).float())
node = model(items, HT)
get = lambda i: node[i][alias_inputs[i]]
seq_hidden = torch.stack([get(i) for i in torch.arange(len(alias_inputs)).long()])
return targets, model.compute_scores(seq_hidden, node_masks)
def train_model(model, train_data, opt):
model.scheduler.step()
print('start training: ', datetime.datetime.now())
model.train()
total_loss = 0.0
slices = train_data.generate_batch(opt.batchSize, True)
for step in tqdm(range(len(slices)), total=len(slices), ncols=70, leave=False, unit='b'):
i = slices[step]
alias_inputs, HT, items, targets, node_masks = train_data.get_slice(i)
model.optimizer.zero_grad()
targets, scores = forward(model, alias_inputs, HT, items, targets, node_masks)
targets = trans_to_cuda(torch.Tensor(targets).long())
loss = model.loss_function(scores, targets)
loss.backward()
model.optimizer.step()
total_loss += loss
print('\tLoss:\t%.4f' % (total_loss))
def test_model(model, test_data, opt, verbose=True):
model.eval()
test_pred = []
test_labels = []
slices = test_data.generate_batch(10, False)
for step in tqdm(range(len(slices)), total=len(slices), ncols=70, leave=False, unit='b'):
i = slices[step]
alias_inputs, HT, items, targets, node_masks = test_data.get_slice(i)
targets, scores = forward(model, alias_inputs, HT, items, targets, node_masks)
pre_indices = scores.topk(1)[1]
test_labels += list(targets)
test_pred += list(trans_to_cpu(pre_indices).detach().numpy())
details = metrics.classification_report(test_labels, test_pred, digits=4)
acc = metrics.accuracy_score(test_labels, test_pred)
if verbose:
print("Test Precision, Recall and F1-Score...")
print(metrics.classification_report(test_labels, test_pred, digits=4))
print("Macro average Test Precision, Recall and F1-Score...")
print(metrics.precision_recall_fscore_support(test_labels, test_pred, average='macro'))
print("Micro average Test Precision, Recall and F1-Score...")
print(metrics.precision_recall_fscore_support(test_labels, test_pred, average='micro'))
return details,acc