-
-
Notifications
You must be signed in to change notification settings - Fork 57
/
train.py
223 lines (189 loc) · 8.31 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
import argparse
import time
from sys import platform
from models import *
from utils.datasets import *
from utils.utils import *
# batch_size 8: 32*17 = 544
# batch_size 4: 32*25 = 800 (1.47 vs 544) or 32*23 = 736
# batch_size 2: 32*35 = 1120 (1.40 vs 800, 2.06 cumulative)
# batch_size 1: 32*49 = 1568 (1.40 vs 1120, 2.88 cumulative)
targets_path = "utils/targets_c60.mat"
parser = argparse.ArgumentParser()
parser.add_argument("-epochs", type=int, default=999, help="number of epochs")
parser.add_argument("-batch_size", type=int, default=8, help="size of each image batch")
parser.add_argument("-cfg", type=str, default="cfg/c60_a30symmetric.cfg", help="cfg file path")
parser.add_argument("-img_size", type=int, default=32 * 25, help="size of each image dimension")
parser.add_argument("-resume", default=False, help="resume training flag")
opt = parser.parse_args()
print(opt)
def main(opt):
"""Initializes and trains a Darknet model for object detection with configurable parameters and data paths."""
os.makedirs("weights", exist_ok=True)
cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if cuda else "cpu")
random.seed(0)
np.random.seed(0)
torch.manual_seed(0)
if cuda:
torch.cuda.manual_seed(0)
torch.cuda.manual_seed_all(0)
torch.backends.cudnn.benchmark = True
# Configure run
if platform == "darwin": # MacOS (local)
train_path = "/Users/glennjocher/Downloads/DATA/xview/train_images"
else: # linux (GCP cloud)
train_path = "../train_images"
# Initialize model
model = Darknet(opt.cfg, opt.img_size)
# Get dataloader
dataloader = ListDataset(train_path, batch_size=opt.batch_size, img_size=opt.img_size, targets_path=targets_path)
# reload saved optimizer state
start_epoch = 0
best_loss = float("inf")
if opt.resume:
checkpoint = torch.load("weights/latest.pt", map_location="cpu")
model.load_state_dict(checkpoint["model"])
if torch.cuda.device_count() > 1:
print("Using ", torch.cuda.device_count(), " GPUs")
model = nn.DataParallel(model)
model.to(device).train()
# # Transfer learning
# for i, (name, p) in enumerate(model.named_parameters()):
# #name = name.replace('module_list.', '')
# #print('%4g %70s %9s %12g %20s %12g %12g' % (
# # i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))
# if p.shape[0] != 650: # not YOLO layer
# p.requires_grad = False
# Set optimizer
# optimizer = torch.optim.SGD(model.parameters(), lr=.001, momentum=.9, weight_decay=0.0005 * 0, nesterov=True)
# optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, model.parameters()))
optimizer = torch.optim.Adam(model.parameters())
optimizer.load_state_dict(checkpoint["optimizer"])
start_epoch = checkpoint["epoch"] + 1
best_loss = checkpoint["best_loss"]
del checkpoint # current, saved
else:
if torch.cuda.device_count() > 1:
print("Using ", torch.cuda.device_count(), " GPUs")
model = nn.DataParallel(model)
model.to(device).train()
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=1e-4, weight_decay=5e-4)
# Set scheduler
# scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, 24, eta_min=0.00001, last_epoch=-1)
# y = 0.001 * exp(-0.00921 * x) # 1e-4 @ 250, 1e-5 @ 500
# scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.99082, last_epoch=start_epoch - 1)
modelinfo(model)
t0, t1 = time.time(), time.time()
print(
"%10s"
* 16
% ("Epoch", "Batch", "x", "y", "w", "h", "conf", "cls", "total", "P", "R", "nGT", "TP", "FP", "FN", "time")
)
class_weights = xview_class_weights_hard_mining(range(60)).to(device)
n = 4 # number of pictures at a time
for epoch in range(opt.epochs):
epoch += start_epoch
# img_size = random.choice([19, 20, 21, 22, 23, 24, 25]) * 32
# dataloader = ListDataset(train_path, batch_size=opt.batch_size, img_size=img_size, targets_path=targets_path)
# print('Running image size %g' % img_size)
# Update scheduler
# if epoch % 25 == 0:
# scheduler.last_epoch = -1 # for cosine annealing, restart every 25 epochs
# scheduler.step()
# if epoch <= 100:
# for g in optimizer.param_groups:
# g['lr'] = 0.0005 * (0.992 ** epoch) # 1/10 th every 250 epochs
# g['lr'] = 0.001 * (0.9773 ** epoch) # 1/10 th every 100 epochs
# g['lr'] = 0.0005 * (0.955 ** epoch) # 1/10 th every 50 epochs
# g['lr'] = 0.0005 * (0.926 ** epoch) # 1/10 th every 30 epochs
ui = -1
rloss = defaultdict(float) # running loss
metrics = torch.zeros(4, 60)
for i, (imgs, targets) in enumerate(dataloader):
for j in range(len(imgs) // n):
targets_j = targets[j * n : j * n + n]
nGT = sum(len(x) for x in targets_j)
if nGT < 1:
continue
loss = model(
imgs[j * n : j * n + n].to(device),
targets_j,
requestPrecision=True,
weight=class_weights,
epoch=epoch,
)
optimizer.zero_grad()
loss.backward()
optimizer.step()
ui += 1
metrics += model.losses["metrics"]
for key, val in model.losses.items():
rloss[key] = (rloss[key] * ui + val) / (ui + 1)
# Precision
precision = metrics[0] / (metrics[0] + metrics[1] + 1e-16)
k = (metrics[0] + metrics[1]) > 0
mean_precision = precision[k].mean() if k.sum() > 0 else 0
# Recall
recall = metrics[0] / (metrics[0] + metrics[2] + 1e-16)
k = (metrics[0] + metrics[2]) > 0
mean_recall = recall[k].mean() if k.sum() > 0 else 0
s = ("%10s%10s" + "%10.3g" * 14) % (
f"{epoch:g}/{opt.epochs - 1:g}",
f"{i:g}/{len(dataloader) - 1:g}",
rloss["x"],
rloss["y"],
rloss["w"],
rloss["h"],
rloss["conf"],
rloss["cls"],
rloss["loss"],
mean_precision,
mean_recall,
model.losses["nGT"],
model.losses["TP"],
model.losses["FP"],
model.losses["FN"],
time.time() - t1,
)
t1 = time.time()
print(s)
# if i == 1:
# return
# # Update dynamic class weights
# new_weights = metrics[3]
# print(metrics[3])
# new_weights[new_weights == 0] = new_weights[new_weights > 0].min()
# new_weights = 1 / new_weights
# new_weights /= new_weights.sum()
# class_weights = class_weights * 0.9 + new_weights * 0.1
# class_weights /= class_weights.sum()
# print(1 / class_weights)
# Write epoch results
with open("results.txt", "a") as file:
file.write(s + "\n")
# Update best loss
loss_per_target = rloss["loss"] / rloss["nGT"]
if loss_per_target < best_loss:
best_loss = loss_per_target
# Save latest checkpoint
checkpoint = {
"epoch": epoch,
"best_loss": best_loss,
"model": model.state_dict(),
"optimizer": optimizer.state_dict(),
}
torch.save(checkpoint, "weights/latest.pt")
# Save best checkpoint
if best_loss == loss_per_target:
os.system("cp weights/latest.pt weights/best.pt")
# Save backup checkpoint
if (epoch > 0) & (epoch % 100 == 0):
os.system(f"cp weights/latest.pt weights/backup{epoch}.pt")
# Save final model
dt = time.time() - t0
print(f"Finished {epoch:g} epochs in {dt:.2f}s ({dt / (epoch + 1):.2f}s/epoch)")
if __name__ == "__main__":
torch.cuda.empty_cache()
main(opt)
torch.cuda.empty_cache()