-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
192 lines (157 loc) · 6.63 KB
/
train.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
#
import argparse
import os
import pickle
import pprint
import numpy as np
import torch
import tqdm
from torch.nn.modules.loss import CrossEntropyLoss
from torch.utils.data.dataloader import DataLoader
import torch.nn.functional as F
from model.drpt import DRPT
from parameters import parser, YML_PATH
from loss import loss_calu
import time
from datetime import datetime
# from test import *
import test as test
from dataset import CompositionDataset
from utils import *
def train_model(model, optimizer, config, train_dataset, val_dataset, test_dataset):
train_dataloader = DataLoader(
train_dataset,
num_workers = 16,
batch_size=config.train_batch_size,
shuffle=True
)
model.train()
best_loss = 1e5
best_metric = 0
# scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=2, gamma=0.5)
attr2idx = train_dataset.attr2idx
obj2idx = train_dataset.obj2idx
train_pairs = torch.tensor([(attr2idx[attr], obj2idx[obj])
for attr, obj in train_dataset.train_pairs]).cuda()
train_losses = []
# print("learning_rate now is: {}".format(optimizer.state_dict()['param_groups']))
for i in range(config.epoch_start, config.epochs):
progress_bar = tqdm.tqdm(
total=len(train_dataloader), desc="epoch % 3d" % (i + 1)
)
epoch_train_losses = []
# model.store_params_update_scope()
for bid, batch in enumerate(train_dataloader):
# if bid > 1:
# break
predict, loss = model(batch, train_pairs)
# normalize loss to account for batch accumulation
loss = loss / config.gradient_accumulation_steps
# backward pass
loss.backward()
# weights update
if ((bid + 1) % config.gradient_accumulation_steps == 0) or (bid + 1 == len(train_dataloader)):
optimizer.step()
optimizer.zero_grad()
epoch_train_losses.append(loss.item())
progress_bar.set_postfix({"train loss": np.mean(epoch_train_losses[-50:])})
progress_bar.update()
# scheduler.step()
progress_bar.close()
progress_bar.write(f"epoch {i +1} train loss {np.mean(epoch_train_losses)}")
train_losses.append(np.mean(epoch_train_losses))
if (i + 1) % config.save_every_n == 0:
torch.save(model.state_dict(), os.path.join(config.save_path, f"epoch_{i}.pt"))
if config.update == True:
model.update_status(i)
if i < config.jump_epoch:
continue
print("Evaluating val dataset:")
logging.info("Evaluating val dataset:")
loss_avg, val_result = evaluate(model, val_dataset)
print("Now status is {}".format(model.train_status))
logging.info("Now status is {}".format(model.train_status))
print("Loss average on val dataset: {}".format(loss_avg))
print("Evaluating test dataset:")
logging.info("Evaluating test dataset:")
evaluate(model, test_dataset)
if config.best_model_metric == "best_loss":
if loss_avg.cpu().float() < best_loss:
best_loss = loss_avg.cpu().float()
print("Evaluating test dataset:")
evaluate(model, test_dataset)
torch.save(model.state_dict(), os.path.join(
config.save_path, f"best.pt"
))
else:
if val_result[config.best_model_metric] > best_metric:
best_metric = val_result[config.best_model_metric]
print("Evaluating test dataset:")
evaluate(model, test_dataset)
torch.save(model.state_dict(), os.path.join(
config.save_path, f"best.pt"
))
if i + 1 == config.epochs:
print("Evaluating test dataset on Closed World")
model.load_state_dict(torch.load(os.path.join(config.save_path, f"best.pt")))
evaluate(model, test_dataset)
if config.save_model:
torch.save(model.state_dict(), os.path.join(config.save_path, f'final_model.pt'))
def evaluate(model, dataset):
model.eval()
evaluator = test.Evaluator(dataset, model=None)
all_logits, all_attr_gt, all_obj_gt, all_pair_gt, loss_avg = test.predict_logits(
model, dataset, config)
test_stats = test.test(
dataset,
evaluator,
all_logits,
all_attr_gt,
all_obj_gt,
all_pair_gt,
config
)
result = ""
key_set = ["best_seen", "best_unseen", "AUC", "best_hm", "attr_acc", "obj_acc"]
for key in test_stats:
if key in key_set:
result = result + key + " " + str(round(test_stats[key], 4)) + "| "
print(result)
logging.info(result)
model.train()
return loss_avg, test_stats
if __name__ == "__main__":
config = parser.parse_args()
log_time = datetime.now()
os.makedirs(os.path.join("logs", config.dataset), exist_ok=True)
set_log("logs/" + config.dataset + '/' + str(log_time) + ".log")
load_args(YML_PATH[config.dataset], config)
logging.warning(config.log_id)
logging.info(config)
print(config)
# set the seed value
set_seed(config.seed)
dataset_path = config.dataset_path
train_dataset = CompositionDataset(dataset_path,
phase='train',
split='compositional-split-natural')
val_dataset = CompositionDataset(dataset_path,
phase='val',
split='compositional-split-natural')
test_dataset = CompositionDataset(dataset_path,
phase='test',
split='compositional-split-natural')
allattrs = train_dataset.attrs
allobj = train_dataset.objs
classes = [cla.replace(".", " ").lower() for cla in allobj]
attributes = [attr.replace(".", " ").lower() for attr in allattrs]
offset = len(attributes)
ent_attr, ent_obj = train_dataset.ent_attr, train_dataset.ent_obj
model = DRPT(config, attributes=attributes, classes=classes, offset=offset, ent_attr=ent_attr, ent_obj=ent_obj).cuda()
optimizer = torch.optim.Adam(model.parameters(), lr=config.lr, weight_decay=config.weight_decay)
# model.load_state_dict(torch.load(config.load_model))
os.makedirs(config.save_path, exist_ok=True)
train_model(model, optimizer, config, train_dataset, val_dataset, test_dataset)
with open(os.path.join(config.save_path, "config.pkl"), "wb") as fp:
pickle.dump(config, fp)
print("done!")