-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrainer.py
161 lines (147 loc) · 9.02 KB
/
trainer.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
from dataset import Dataset
from models import *
from utils import *
import torch
import torch.nn as nn
import torch.nn.functional as F
import os
class Trainer:
def __init__(self, dataset, model_name, args):
self.device = args.device
self.model_name = model_name
if self.model_name == 'TransE':
self.model = TransE(dataset.num_ent(), dataset.num_rel(), args.dim, self.device)
if self.model_name == 'TransH':
self.model = TransH(dataset.num_ent(), dataset.num_rel(), args.dim, self.device)
if self.model_name == 'TransR':
self.model = TransR(dataset.num_ent(), dataset.num_rel(), args.dim, self.device)
if self.model_name == 'TransD':
self.model = TransD(dataset.num_ent(), dataset.num_rel(), args.dim, self.device)
if self.model_name == 'DistMult':
self.model = DistMult(dataset.num_ent(), dataset.num_rel(), args.dim, self.device)
if self.model_name == 'ComplEx':
self.model = ComplEx(dataset.num_ent(), dataset.num_rel(), args.dim, self.device)
if self.model_name == 'SimplE':
self.model = SimplE(dataset.num_ent(), dataset.num_rel(), args.dim, self.device, dataset, args)
if self.model_name == 'SimplE2':
self.model = SimplE2(dataset.num_ent(), dataset.num_rel(), args.dim, self.device)
if self.model_name == 'ConvE':
self.model = ConvE(dataset.num_ent(), dataset.num_rel(), args, self.device)
if self.model_name == 'ConvKB2D':
self.model = ConvKB2D(dataset.num_ent(), dataset.num_rel(), args, self.device)
if self.model_name == 'RGCN':
self.model = RGCN(dataset.num_ent(), dataset.num_rel(), args, self.device)
if self.model_name == 'CompGCN':
self.model = CompGCN_DistMult(dataset.num_ent(), dataset.num_rel(), args, self.device)
self.dataset = dataset
self.args = args
def train(self):
self.model.train()
optimizer = torch.optim.Adam(
self.model.parameters(),
lr=self.args.lr
)
for epoch in range(1, self.args.ne + 1):
last_batch = False
total_loss = 0.0
while not last_batch:
if self.model_name == 'ConvE':
batch = torch.tensor(self.dataset.next_pos_batch(self.args.batch_size))
else:
batch = self.dataset.next_batch(self.args.batch_size, neg_ratio=self.args.neg_ratio, neg_sampler=self.args.neg_sampler, device = self.device)
last_batch = self.dataset.was_last_batch()
optimizer.zero_grad()
hs = (batch[:,0]).clone().detach().long().to(self.device)
rs = (batch[:,1]).clone().detach().long().to(self.device)
ts = (batch[:,2]).clone().detach().long().to(self.device)
if self.model_name == 'RGCN':
train_data = generate_sampled_graph_and_labels(self.dataset.data["train"], self.args.batch_size, self.args.graph_split_size, \
self.dataset.num_ent(), self.dataset.num_rel(), self.args.neg_ratio)
entity_embedding = self.model(train_data.entity, train_data.edge_index, train_data.edge_type, train_data.edge_norm)
loss = self.model.score_loss(entity_embedding, train_data.samples, train_data.labels) + self.args.reg * self.model.reg_loss(entity_embedding)
elif self.model_name != 'ConvE':
scores = self.model.forward(hs, rs, ts)
if last_batch:
nb_pos = self.dataset.data["train"].shape[0] % self.args.batch_size
pos_scores, neg_scores = scores[:nb_pos], scores[nb_pos:]
else:
pos_scores, neg_scores = scores[:self.args.batch_size], scores[self.args.batch_size:]
loss = self.model._loss(pos_scores, neg_scores, self.args.neg_ratio)
if self.args.reg != 0.0 :
if self.model_name != 'SimplE':
loss += self.args.reg*self.model._regularization(batch[:, 0].to(self.device), batch[:, 1].to(self.device), batch[:, 2].to(self.device))
else:
loss = self.model.calc_loss(hs, rs, ts)
loss.backward()
optimizer.step()
total_loss += loss.cpu().item()
if epoch % self.args.save_each == 0:
print("Loss in iteration " + str(epoch) + ": " + str(total_loss))
self.save_model(self.model_name, epoch)
def save_model(self, model, chkpnt):
print("Saving the model")
directory = "models/" + self.dataset.name + "/" + model + "/"
if not os.path.exists(directory):
os.makedirs(directory)
torch.save(self.model.state_dict(), directory + "dim="+str(self.args.dim) + \
"_lr="+str(self.args.lr) + "_neg="+str(self.args.neg_ratio) + "_bs="+str(self.args.batch_size) + "_reg="+str(self.args.reg) + "__epoch="+str(chkpnt) + ".pt")
def resume_training(self):
directory = "models/" + self.dataset.name + "/" + self.model_name + "/"
resume_epoch = self.args.resume_epoch
if resume_epoch == 0:
resume_epoch = max([int(f[-11:].split('=')[-1].split('.')[0]) for f in os.listdir("models/" + self.dataset.name + "/" + self.model_name + "/")])
model_path = directory + "dim="+str(self.args.dim) + \
"_lr="+str(self.args.lr) + "_neg="+str(self.args.neg_ratio) + "_bs="+str(self.args.batch_size) + "_reg="+str(self.args.reg) + "__epoch="+str(resume_epoch) + ".pt"
else:
model_path = directory + "dim="+str(self.args.dim) + \
"_lr="+str(self.args.lr) + "_neg="+str(self.args.neg_ratio) + "_bs="+str(self.args.batch_size) + "_reg="+str(self.args.reg) + "__epoch="+str(resume_epoch) + ".pt"
print('Resuming from ' + str(model_path))
self.model.load_state_dict(torch.load(model_path))
self.model.train()
optimizer = torch.optim.Adam(
self.model.parameters(),
lr=self.args.lr
)
for epoch in range(resume_epoch + 1, resume_epoch + self.args.ne + 1):
if self.model == 'CompGCN':
loss = run_epoch(epoch, val_mrr = 0)
else:
last_batch = False
total_loss = 0.0
while not last_batch:
if self.model_name == 'ConvE':
batch = torch.tensor(self.dataset.next_pos_batch(self.args.batch_size))
else:
batch = self.dataset.next_batch(self.args.batch_size, neg_ratio=self.args.neg_ratio, neg_sampler=self.args.neg_sampler, device = self.device)
last_batch = self.dataset.was_last_batch()
optimizer.zero_grad()
hs = (batch[:,0]).clone().detach().long().to(self.device)
rs = (batch[:,1]).clone().detach().long().to(self.device)
ts = (batch[:,2]).clone().detach().long().to(self.device)
chunks = self.args.neg_ratio + 1
if self.model == 'CompGCN':
pred = self.model.forward(hs, rs)
loss = self.model.loss(pred, label)
if self.model_name == 'RGCN':
train_data = generate_sampled_graph_and_labels(self.dataset.data["train"], self.args.batch_size, self.args.graph_split_size, \
self.dataset.num_ent(), self.dataset.num_rel(), self.args.neg_ratio)
entity_embedding = self.model(train_data.entity, train_data.edge_index, train_data.edge_type, train_data.edge_norm)
loss = self.model.score_loss(entity_embedding, train_data.samples, train_data.labels) + self.args.reg * self.model.reg_loss(entity_embedding)
elif self.model_name != 'ConvE':
scores = self.model.forward(hs, rs, ts)
if last_batch:
nb_pos = self.dataset.data["train"].shape[0] % self.args.batch_size
pos_scores, neg_scores = scores[:nb_pos], scores[nb_pos:]
else:
pos_scores, neg_scores = scores[:self.args.batch_size], scores[self.args.batch_size:]
loss = self.model._loss(pos_scores, neg_scores, self.args.neg_ratio)
if self.args.reg != 0.0 :
loss += self.args.reg*self.model._regularization(batch[:, 0], batch[:, 1], batch[:, 2])
else:
loss = self.model.calc_loss(hs, rs, ts)
loss.backward()
optimizer.step()
total_loss += loss.cpu().item()
print("Loss in iteration " + str(epoch) + ": " + str(total_loss))
if epoch % self.args.save_each == 0:
self.save_model(self.model_name, epoch)