-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
250 lines (215 loc) · 8.89 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
import sys
import os
os.environ["CURL_CA_BUNDLE"] = "/etc/ssl/certs/ca-certificates.crt" # A workaround in case this happens: https://github.com/mapbox/rasterio/issues/1289
import time
import datetime
import argparse
import copy
import random
import math
import logging
import numpy as np
import pandas as pd
from pathlib import Path
import torch
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = True
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import models
import config
import utils
from dataloaders.StreamingDatasets import StreamingGeospatialDataset, StreamingValidationDataset
from dataloaders.data_agu import *
from loss import *
NUM_WORKERS = 4
NUM_CHIPS_PER_TILE = config.NUM_CHIPS_PER_TILE
CHIP_SIZE = config.TRAIN_CHIP_SIZE
INIT_LR = 0.001
parser = argparse.ArgumentParser(description='DFC2021 baseline training script')
parser.add_argument('-t', '--train_fn', type=str, required=True)
parser.add_argument('-v', '--valid_fn', type=str, required=True)
parser.add_argument('-o', '--output_dir', type=str, required=True)
parser.add_argument('-bb', '--backbone', default='efficientnet-b0',
choices=(
'efficientnet-b0',
'efficientnet-b1',
'efficientnet-b2',
'efficientnet-b3',
'efficientnet-b4',
'efficientnet-b5',
'efficientnet-b6',
'efficientnet-b7',
'efficientnet-b0'
),
help='Backbone to use'
)
## Training arguments
# parser.add_argument('--gpu', type=int, default=0, help='The ID of the GPU to use')
parser.add_argument('-g', '--gpu', type=str, default=None, help='The indices of GPUs to enable (default: all)')
parser.add_argument('-bs', '--batch_size', type=int, default=32, help='Batch size to use for training (default: 32)')
parser.add_argument('-e', '--num_epochs', type=int, default=50, help='Number of epochs to train for (default: 50)')
parser.add_argument('-s', '--seed', type=int, default=0, help='Random seed to pass to numpy and torch (default: 0)')
args = parser.parse_args()
def weights_init(model, seed=7):
np.random.seed(seed)
torch.manual_seed(seed)
random.seed(seed)
torch.cuda.manual_seed_all(seed)
for m in model.modules():
if isinstance(m, nn.Conv2d) or isinstance(m, nn.ConvTranspose2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2.0 / n))
if m.bias is not None:
m.bias.data.zero_()
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.Linear):
m.weight.data.normal_(0, 0.01)
if m.bias is not None:
m.bias.data.zero_()
def main():
# print("Starting DFC2021 baseline training script at %s" % (str(datetime.datetime.now())))
#-------------------
# Setup
#-------------------
assert os.path.exists(args.train_fn)
assert os.path.exists(args.valid_fn)
now_time = datetime.datetime.now()
time_str = datetime.datetime.strftime(now_time, '%m-%d_%H-%M-%S')
# output path
# output_dir = Path(args.output_dir).parent / time_str / Path(args.output_dir).stem
output_dir = Path(args.output_dir)
output_dir.mkdir(exist_ok=True, parents=True)
logger = utils.init_logger(output_dir / 'info.log')
# if os.path.isfile(args.output_dir):
# print("A file was passed as `--output_dir`, please pass a directory!")
# return
#
# if os.path.exists(args.output_dir) and len(os.listdir(args.output_dir)):
# if args.overwrite:
# print("WARNING! The output directory, %s, already exists, we might overwrite data in it!" % (args.output_dir))
# else:
# print("The output directory, %s, already exists and isn't empty. We don't want to overwrite and existing results, exiting..." % (args.output_dir))
# return
# else:
# print("The output directory doesn't exist or is empty.")
# os.makedirs(args.output_dir, exist_ok=True)
if args.gpu is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
n_gpu = torch.cuda.device_count()
device = torch.device('cuda:0' if n_gpu > 0 else 'cpu')
device_ids = list(range(n_gpu))
np.random.seed(args.seed)
torch.manual_seed(args.seed)
#-------------------
# Load input data
#-------------------
train_dataframe = pd.read_csv(args.train_fn)
train_image_fns = train_dataframe["image_fn"].values
train_label_fns = train_dataframe["label_fn"].values
train_groups = train_dataframe["group"].values
train_dataset = StreamingGeospatialDataset(
imagery_fns=train_image_fns, label_fns=train_label_fns, groups=train_groups, chip_size=CHIP_SIZE,
num_chips_per_tile=NUM_CHIPS_PER_TILE, transform=transform, nodata_check=nodata_check
)
valid_dataframe = pd.read_csv(args.valid_fn)
valid_image_fns = valid_dataframe["image_fn"].values
valid_label_fns = valid_dataframe["label_fn"].values
valid_groups = valid_dataframe["group"].values
valid_dataset = StreamingValidationDataset(
imagery_fns=valid_image_fns, label_fns=valid_label_fns, groups=valid_groups, chip_size=CHIP_SIZE,
stride=CHIP_SIZE, transform=transform, nodata_check=nodata_check
)
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
num_workers=NUM_WORKERS,
pin_memory=True,
)
valid_dataloader = torch.utils.data.DataLoader(
valid_dataset,
batch_size=args.batch_size,
num_workers=NUM_WORKERS,
pin_memory=True,
)
num_training_images_per_epoch = int(len(train_image_fns) * NUM_CHIPS_PER_TILE)
# print("We will be training with %d batches per epoch" % (num_training_batches_per_epoch))
#-------------------
# Setup training
#-------------------
# if args.model == "unet":
# model = models.get_unet()
# elif args.model == "fcn":
# model = models.get_fcn()
# else:
# raise ValueError("Invalid model")
model = models.isCNN(args.backbone)
weights_init(model, seed=args.seed)
model = model.to(device)
if len(device_ids) > 1:
model = torch.nn.DataParallel(model, device_ids=device_ids)
trainable_params = filter(lambda p: p.requires_grad, model.parameters())
optimizer = optim.AdamW(trainable_params, lr=INIT_LR, amsgrad=True, weight_decay=5e-4)
lr_criterion = nn.CrossEntropyLoss(ignore_index=0) # todo
hr_criterion = hr_loss
# criterion = balanced_ce_loss
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, "min", factor=0.5, patience=3, min_lr=0.0000001)
# factor=0.5, patience=3, min_lr=0.0000001
logger.info("Trainable parameters: {}".format(utils.count_parameters(model)))
#-------------------
# Model training
#-------------------
train_loss_total_epochs, valid_loss_total_epochs, epoch_lr = [], [], []
best_loss = 1e50
num_times_lr_dropped = 0
# model_checkpoints = []
# temp_model_fn = os.path.join(output_dir, "most_recent_model.pt")
for epoch in range(args.num_epochs):
lr = utils.get_lr(optimizer)
train_loss_epoch, valid_loss_epoch = utils.fit(
model,
device,
train_dataloader,
valid_dataloader,
num_training_images_per_epoch,
optimizer,
lr_criterion,
hr_criterion,
epoch,
logger)
scheduler.step(valid_loss_epoch)
if epoch % config.SAVE_PERIOD == 0 and epoch != 0:
temp_model_fn = output_dir / 'checkpoint-epoch{}.pth'.format(epoch+1)
torch.save(model.state_dict(), temp_model_fn)
if valid_loss_epoch < best_loss:
logger.info("Saving model_best.pth...")
temp_model_fn = output_dir / 'model_best.pth'
torch.save(model.state_dict(), temp_model_fn)
best_loss = valid_loss_epoch
if utils.get_lr(optimizer) < lr:
num_times_lr_dropped += 1
print("")
print("Learning rate dropped")
print("")
train_loss_total_epochs.append(train_loss_epoch)
valid_loss_total_epochs.append(valid_loss_epoch)
epoch_lr.append(lr)
# if num_times_lr_dropped == 4:
# break
#-------------------
# Save everything
#-------------------
# save_obj = {
# 'args': args,
# 'training_task_losses': training_task_losses,
# "checkpoints": model_checkpoints
# }
#
# save_obj_fn = "results.pt"
# with open(os.path.join(output_dir, save_obj_fn), 'wb') as f:
# torch.save(save_obj, f)
if __name__ == "__main__":
main()