-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
345 lines (262 loc) · 8.35 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
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
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
from collections import OrderedDict
import os
import json
import time
from argparse import ArgumentParser
# argument parsing
parser = ArgumentParser()
parser.add_argument(
"config_path",
help="path to the experiment config file for this training script"
)
parser.add_argument(
"--seed",
default=0,
type=int,
help="random seed, can be specified as an arg or in the config."
)
parser.add_argument(
"--gpu",
type=int,
default=None,
help="gpu override for debugging to set the gpu to use."
)
args = parser.parse_args()
# load config
config = open(args.config_path)
config = json.load(config)
# set gpu
# MUST BE DONE BEFORE TORCH IS IMPORTED
if args.gpu is not None:
config["gpu_id"] = args.gpu
elif "gpu_id" in config and (
type(config["gpu_id"]) == int
or
type(config["gpu_id"]) == list
):
pass
else:
config["gpu_id"] = 0
print(f"gpus allowed to be used: ", config["gpu_id"])
# will have brackets if read from json
os.environ["CUDA_VISIBLE_DEVICES"] = str(
config["gpu_id"]
).replace("[", "").replace("]", "")
import torch
from utils.data_utils import (
Data, get_preprocessing_transforms, TRAIN_DATASETS
)
from utils.eval_utils import TopKError
from utils.train_utils import (
OPTIMIZER_MAPPING,
SCHEDULER_MAPPING,
AverageMeter,
ProgressMeter,
save_checkpoint,
save_state_dict
)
from models.model_utils import model_generator, load_weights_from_file
assert config["id_dataset"]["name"] in TRAIN_DATASETS, "not valid train set"
# set random seed
# CL arg overrides value in config file
if args.seed is not None:
torch.manual_seed(args.seed)
# add seed into config dictionary
config["seed"] = args.seed
elif "seed" in config and type(config["seed"]) == int:
torch.manual_seed(config['seed'])
# no seed in config or as CL arg
else:
torch.manual_seed(0)
config["seed"] = 0
# set training device, defaults to cuda
dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# dev = torch.device("cpu")
print(f"using {dev} for training")
# load training dataset
# more training arguments are passed directly from the configuration json
training_data = Data(
**config["id_dataset"],
transforms=get_preprocessing_transforms(config["id_dataset"]["name"])
)
# load the model
model = model_generator(
config["model"]["model_type"],
**config["model"]["model_params"]
)
# training loss
# only standard one hot CE
criterion = torch.nn.CrossEntropyLoss()
# pretrained weights if path supplied
if (
"pretrained_path" in config["model"]
and
config["model"]["pretrained_path"] is not None
):
if "keep_last_layer" in config["model"]:
# where pretrained weights are
load_weights_from_file(
model,
config["model"]["pretrained_path"],
keep_last_layer=config["model"]["keep_last_layer"]
)
else:
load_weights_from_file(
model,
config["model"]["pretrained_path"],
keep_last_layer=True
)
# directory to save weights from training
if (
"weights_path" in config["model"]
and
config["model"]["weights_path"] is not None
):
# make a directory if it doesn't already exist
if not os.path.exists(config["model"]["weights_path"]):
os.mkdir(config["model"]["weights_path"])
# multigpu
if (
config["data_parallel"]
and torch.cuda.device_count() > 1
and dev.type == "cuda"
):
model = torch.nn.DataParallel(model)
multi_gpu = True
else:
multi_gpu = False
model.to(dev)
# optimizer and scheduler
optimizer = OPTIMIZER_MAPPING[config["train_params"]["optimizer"]](
model.parameters(), **config["train_params"]["optimizer_params"]
)
scheduler = SCHEDULER_MAPPING[config["train_params"]["lr_scheduler"]](
optimizer, **config["train_params"]["lr_scheduler_params"]
)
def train_epoch(train_loader, model, criterion, optimizer, epoch:int):
"""Train the model for one epoch of the dataset."""
batch_time = AverageMeter('Time', ':6.3f')
data_time = AverageMeter('Data', ':6.3f')
losses = AverageMeter('Loss', ':.2e')
top1 = AverageMeter('Err@1', ':6.2f')
top5 = AverageMeter('Err@5', ':6.2f')
ece = AverageMeter("ECE", ":6.2f")
top1_calc = TopKError(k=1)
top5_calc = TopKError(k=5)
progress = ProgressMeter(
len(train_loader),
[batch_time, data_time, losses, top1, top5, ece],
prefix=f"Epoch: [{epoch}]")
# switch to train
model.train()
start = time.time()
for i, (inputs, targets) in enumerate(train_loader):
# measure data loading time
data_time.update(time.time() - start)
# move data to correct device
inputs, targets = inputs.to(dev), targets.to(dev)
# compute output
outputs = model(inputs)
loss = criterion(outputs, targets)
# measure accuracy and record loss
# note that outputs should be logits
# targets should be labels (no distillation)
err1 = top1_calc(targets, outputs)
err5 = top5_calc(targets, outputs)
batch_size = inputs.size(0) # may be smaller for last batch of epoch
losses.update(loss.item(), batch_size)
top1.update(err1, batch_size)
top5.update(err5, batch_size)
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - start)
start = time.time()
# print every 5 batches
if i % 5 == 0:
progress.display(i)
def evaluate_epoch(val_loader, model, criterion, epoch: int) -> dict:
"""Evaluate the model for one epoch of the validation dataset."""
batch_time = AverageMeter('Time', ':6.3f')
meters = {
"losses": AverageMeter('Loss', ':.4e'),
"top1": AverageMeter('Err@1', ':6.2f'),
"top5": AverageMeter('Err@5', ':6.2f'),
"ece": AverageMeter("ECE", ":6.2f")
}
top1_calc = TopKError(k=1)
top5_calc = TopKError(k=5)
progress = ProgressMeter(
len(val_loader),
[
batch_time,
meters["losses"],
meters["top1"],
meters["top5"]
],
prefix=f"Epoch: [{epoch}]")
# switch to evaluation mode (e.g. freezes bn stats)
model.eval()
start = time.time()
with torch.no_grad():
for i, (inputs, targets) in enumerate(val_loader):
# move data to correct device
inputs, targets = inputs.to(dev), targets.to(dev)
outputs = model(inputs) # just logits no features
loss = criterion(outputs, targets)
# measure accuracy and record loss
# note that outputs should be logits
# targets should be labels (no distillation)
err1 = top1_calc(targets, outputs)
err5 = top5_calc(targets, outputs)
batch_size = inputs.size(0) # may be smaller for last batch
meters["losses"].update(loss.item(), batch_size)
meters["top1"].update(err1, batch_size)
meters["top5"].update(err5, batch_size)
# measure elapsed time
batch_time.update(time.time() - start)
start = time.time()
if i % 20 == 0:
progress.display(i)
eval_res = {}
eval_res = {
"err1": meters["top1"].avg,
"err5": meters["top5"].avg,
"ece": meters["ece"].avg,
"loss": meters["losses"].avg
}
return eval_res
# training loop
for epoch in range(config["train_params"]["num_epochs"]):
train_epoch(
training_data.train_loader,
model,
criterion,
optimizer,
epoch
)
# reduce learning rate if at correct epoch
scheduler.step()
if training_data.val_size > 0:
res = evaluate_epoch(
training_data.val_loader,
model,
criterion,
epoch
)
# TODO add loading from checkpoint
save_checkpoint(
{
'epoch': epoch + 1,
'config': args.config_path,
'state_dict': model.state_dict(),
'optimizer' : optimizer.state_dict(),
"scheduler": scheduler.state_dict()
},
False,
config
)
# save at the end for future use
save_state_dict(model, config=config, is_best=False)