forked from facebookresearch/neuralvolumes
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
173 lines (153 loc) · 6.57 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
# Copyright (c) Facebook, Inc. and its affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
#
"""Train an autoencoder."""
import math
import sys
import time
import gc
import wandb
import numpy as np
import os
from eval.NeuralVolumePlotter.NeuralVolumeBuilder import NeuralVolumeBuilder
from utils.EnvUtils import EnvUtils
from utils.TrainUtils import TrainUtils
import torch.backends.cudnn
sys.dont_write_bytecode = True
torch.backends.cudnn.benchmark = True # gotta go fast!
if __name__ == "__main__":
train_utils = TrainUtils()
args = train_utils.parse_cmd_arguments()
config_path = train_utils.prepare_and_get_configpath(args)
outpath, log = train_utils.get_outpath_and_print_infos(config_path, args)
trainprofile, progressprof = train_utils.load_profiles(config_path)
train_dataloader, test_dataloader, dataset = train_utils.build_datasets(trainprofile, progressprof, args)
writer, ae, aeoptim, lossweights = train_utils.get_writer_autencoder_optimizer_lossweights(
trainprofile,
progressprof,
dataset,
args,
outpath
)
start_training_time = time.time()
# train
starttime = time.time()
# evalpoints = np.geomspace(1., trainprofile.get_maxiter(), 100).astype(np.int32)
evalpoints = np.linspace(100, trainprofile.get_maxiter(), (trainprofile.get_maxiter() - 100) // 100)
iternum = log.iternum
prevloss = np.inf
env_utils = EnvUtils()
env = env_utils.get_env()
epochs_to_learn = 10000
run = wandb.init(
project=env["wandb"]["project"],
entity=env["wandb"]["entity"],
name=os.path.basename(outpath),
config={
"experiment_path": outpath,
"learning_rate": trainprofile.get_lr(),
"epochs": epochs_to_learn,
"batch_size": trainprofile.get_batchsize(),
"mode": "offline"
}
)
print("wandb sync " + os.path.dirname(run.dir))
train_with_ground_truth_loss = trainprofile.get_should_train_with_ground_truth()
for epoch in range(epochs_to_learn):
for data in train_dataloader:
# forward
output = ae(iternum, lossweights.keys(), **{k: x.to("cuda") for k, x in data.items()})
ground_truth_loss_train = None
if 'ground_truth_loss' in output.keys():
ground_truth_loss_train = output['ground_truth_loss']
train_loss = train_utils.calculate_final_loss_from_output(
output=output,
lossweights=lossweights,
ground_turth_loss=ground_truth_loss_train,
iternum=iternum,
train_with_ground_truth=train_with_ground_truth_loss
)
starttime = train_utils.print_iteration_infos(
iternum=iternum,
loss=train_loss,
output=output,
starttime=starttime)
# update parameters
aeoptim.zero_grad()
train_loss.backward()
aeoptim.step()
# check for loss explosion
if train_loss.item() > 20 * prevloss or not np.isfinite(train_loss.item()):
print("Unstable loss function; resetting")
ae.module.load_state_dict(torch.load("{}/aeparams.pt".format(outpath)), strict=False)
aeoptim = trainprofile.get_optimizer(ae.module)
prevloss = train_loss.item()
test_loss = None
testoutput = None
np_img = None
test_batch = None
ground_truth_loss_test = None
# save intermediate results
if (iternum < 100 and iternum % 20 == 0) or (100 <= iternum <= 1000 and iternum % 100 == 0) or iternum % 1000 == 0 or iternum in [0, 1, 2, 3, 4, 5]:
test_batch, testoutput = train_utils.get_testbatch_testoutput(
iternum=iternum,
progressprof=progressprof,
test_dataloader=test_dataloader,
ae=ae,
lossweights=lossweights
)
ground_truth_loss_test = None
if 'ground_truth_loss' in testoutput.keys():
ground_truth_loss_test = testoutput['ground_truth_loss']
test_loss = train_utils.calculate_final_loss_from_output(
output=testoutput,
lossweights=lossweights,
ground_turth_loss=ground_truth_loss_test,
iternum=iternum,
train_with_ground_truth=train_with_ground_truth_loss
)
np_img = train_utils.save_model_and_validation_pictures(
iternum=iternum,
outpath=outpath,
ae=ae,
test_batch=test_batch,
testoutput=testoutput,
trainprofile=trainprofile,
data=data,
writer=writer)
train_utils.save_wandb_info(
iternum=iternum,
train_loss=train_loss,
train_output=output,
test_loss=test_loss,
test_output=testoutput,
validation_img=np_img,
ground_truth_loss_train=ground_truth_loss_train,
ground_truth_loss_test=ground_truth_loss_test,
wandb=wandb)
iternum += 1
torch.cuda.empty_cache()
del train_loss
del output
if test_batch is not None:
del test_batch
if test_loss is not None:
del test_loss
if testoutput is not None:
del testoutput
gc.collect()
if iternum >= trainprofile.get_maxiter():
break
torch.save(ae.module.state_dict(), "{}/last_aeparams.pt".format(outpath))
def format_time_of_sec(time_needed_in_sec: float) -> str:
time_needed_in_min = float(time_needed_in_sec) / 60.0
time_needed_in_hours = time_needed_in_min / 60.0
mins_formated = time_needed_in_min - int(math.floor(time_needed_in_hours)) * 60
time_needed_in_days = time_needed_in_hours / 24.0
days_formated = int(math.floor(time_needed_in_days))
hours_formated = time_needed_in_hours - days_formated * 24
return "{}days {}h {}min".format(days_formated, hours_formated, mins_formated)
print("max iterations reached. finish training! -> Needed: for {} iterations".format(format_time_of_sec(time.time() - start_training_time), iternum))