-
Notifications
You must be signed in to change notification settings - Fork 3
/
trainer.py
207 lines (174 loc) · 8.14 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
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
import sys
sys.path.append('core')
import os
from tqdm import tqdm
import torch
import torch.nn as nn
import torch.distributed as dist
from utils.utils import ensure_folder
def reduce_list(lists, nprocs):
new_lists = {}
for key, value in lists.items():
rt = value.clone()
dist.all_reduce(rt, op=dist.ReduceOp.SUM)
rt /= nprocs
new_lists[key] = rt.item()
return new_lists
def reduce_tensor(tensor, nprocs):
rt = tensor.clone()
dist.all_reduce(rt, op=dist.ReduceOp.SUM)
rt /= nprocs
return rt
class Trainer:
def __init__(self, args, model, loss=None, optimizer=None, logger=None, lr_scheduler=None, scaler=None):
self.args = args
self.model = model
self.loss = loss
self.optimizer = optimizer
self.logger = logger
self.lr_scheduler = lr_scheduler
self.scaler = scaler
if self.logger is None:
def print_line(line, subname=None):
if self.args.local_rank == 0:
print(line)
self.log_info = print_line
else:
self.log_info = self.logger.log_info
def weight_fix(self, way, refer_dict=None):
# fix weights
if way == 'checkpoint':
assert refer_dict is not None
for n, p in self.model.named_parameters():
if n in refer_dict.keys():
p.requires_grad = False
elif way == 'encoder':
for n, p in self.model.named_parameters():
if 'fnet' in n or 'cnet' in n or 'enet' in n or 'fusion' in n:
p.requires_grad = False
elif way == 'event':
for n, p in self.model.named_parameters():
if 'enet' in n or 'fusion' in n:
p.requires_grad = False
elif way == 'eventencoder':
for n, p in self.model.named_parameters():
if 'enet' in n:
p.requires_grad = False
elif way == 'eventfusion':
for n, p in self.model.named_parameters():
if 'fusion' in n:
p.requires_grad = False
elif way == 'imageencoder':
for n, p in self.model.named_parameters():
if 'fnet' in n or 'cnet' in n:
p.requires_grad = False
elif way == 'raft':
for n, p in self.model.named_parameters():
if 'fnet' in n or 'cnet' in n or 'update_block' in n:
p.requires_grad = False
elif way == 'allencoder':
for n, p in self.model.named_parameters():
if 'enet' in n or 'fusion' in n or 'fnet' in n or 'cnet' in n:
p.requires_grad = False
elif way == 'update':
for n, p in self.model.named_parameters():
if 'update_block' in n:
p.requires_grad = False
self.log_info("Weight fix way: {} complete.".format(way if way != "" else "None"), "trainer")
def partial_load(self, path, weight_fix=None, not_load=False):
# partial parameters loading
assert path != ''
load_dict = torch.load(path, map_location=torch.device("cpu"))
try:
if "model" not in load_dict.keys():
pretrained_dict = {k: v for k, v in load_dict.items() if k in self.model.state_dict().keys() \
and k != 'module.update_block.encoder.conv.weight' \
and k != 'module.update_block.encoder.conv.bias' \
and not k.startswith('module.update_block.flow_enc')}
else:
pretrained_dict = {k: v for k, v in load_dict.pop("model").items() if k in self.model.state_dict().keys() \
and k != 'module.update_block.encoder.conv.weight' \
and k != 'module.update_block.encoder.conv.bias' \
and not k.startswith('module.update_block.flow_enc')}
assert len(pretrained_dict.keys()) > 0
if not not_load:
self.model.load_state_dict(pretrained_dict, strict=False)
self.log_info("Partial load model from {} complete.".format(path), "trainer")
else:
self.log_info("Partial load dict from {} only for weight fix, but not load to model.".format(path), "trainer")
except:
raise KeyError("'model' not in or mismatch state_dict.keys(), please check partial checkpoint path {}".format(path))
self.weight_fix(weight_fix, pretrained_dict)
def load(self, path, only_model=True):
assert path != ''
state_dict = torch.load(path, map_location=torch.device("cpu"))
try:
if "model" not in state_dict.keys():
self.model.load_state_dict(state_dict)
else:
self.model.load_state_dict(state_dict.pop("model"))
except:
raise KeyError("'model' not in or mismatch state_dict.keys(), please check checkpoint path {}".format(path))
index = 0
if not only_model:
try:
self.optimizer.load_state_dict(state_dict.pop("optimizer"))
except:
self.log_info("'optimizer' not in state_dict.keys(), skip it.", "trainer")
try:
self.lr_scheduler.load_state_dict(state_dict.pop("lr_scheduler"))
except:
self.log_info("'lr_scheduler' not in state_dict.keys(), skip it.", "trainer")
try:
index = state_dict.pop("index")
except:
self.log_info("'index' not in state_dict.keys(), set to 0.", "trainer")
self.log_info("Load model/optimizer/index from {} complete, index {}".format(path, index), "trainer")
else:
self.log_info("Load model from {} complete, index {}".format(path, index), "trainer")
return index
def store(self, path, name, index=None):
if path != "" and name != "":
checkpoint = {}
checkpoint["model"] = self.model.state_dict()
checkpoint["optimizer"] = self.optimizer.state_dict()
checkpoint["lr_scheduler"] = self.lr_scheduler.state_dict()
checkpoint["index"] = index
ensure_folder(path)
save_path = os.path.join(path, "{}_{}.pth".format(name, checkpoint["index"]))
torch.save(checkpoint, save_path)
self.log_info("<<< Save model to {} complete".format(save_path), "trainer")
def run_epoch(self, dataloader):
self.model.train()
if self.args.local_rank == 0:
self.bar = tqdm(total=len(dataloader), position=0, leave=True)
for index, batch in enumerate(dataloader):
for key in batch.keys():
if torch.is_tensor(batch[key]):
batch[key] = batch[key].cuda(self.args.gpus[self.args.local_rank] \
if self.args.local_rank != -1 else 0, non_blocking=True)
# output = self.model(batch['img1'], batch['img2'], self.args.iters)
output = self.model(batch, self.args.iters)
loss = self.loss(output, batch)
self.optimizer.zero_grad()
torch.distributed.barrier()
reduced_loss = reduce_list(loss, self.args.nprocs)
self.scaler.scale(loss['loss']).backward()
self.scaler.unscale_(self.optimizer)
nn.utils.clip_grad_norm_(self.model.parameters(), self.args.clip)
self.scaler.step(self.optimizer)
self.lr_scheduler.step()
self.scaler.update()
if self.args.local_rank == 0:
self.bar_update(reduced_loss)
if self.logger is not None:
self.logger.push(reduced_loss, 'loss', last=False)
self.logger.push({'lr': self.optimizer.state_dict()['param_groups'][0]['lr']})
if self.args.local_rank == 0:
self.bar.close()
def bar_update(self, loss):
loss_description = ""
for data, key in zip(loss.values(), loss.keys()):
loss_description += "{}:{:5.4f}, ".format(key, data) if 'px' not in key else "{}:{:4.3f}, ".format(key, data)
self.bar.set_description(loss_description)
self.bar.update(1)