-
-
Notifications
You must be signed in to change notification settings - Fork 3.4k
/
train.py
380 lines (321 loc) · 15.2 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
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
import argparse
import time
import torch.distributed as dist
import torch.optim as optim
from torch.utils.data import DataLoader
import test # Import test.py to get mAP after each epoch
from models import *
from utils.datasets import *
from utils.utils import *
# Hyperparameters
# 0.861 0.956 0.936 0.897 1.51 10.39 0.1367 0.01057 0.01181 0.8409 0.1287 0.001028 -3.441 0.9127 0.0004841
hyp = {'k': 10.39, # loss multiple
'xy': 0.1367, # xy loss fraction
'wh': 0.01057, # wh loss fraction
'cls': 0.01181, # cls loss fraction
'conf': 0.8409, # conf loss fraction
'iou_t': 0.1287, # iou target-anchor training threshold
'lr0': 0.001028, # initial learning rate
'lrf': -3.441, # final learning rate = lr0 * (10 ** lrf)
'momentum': 0.9127, # SGD momentum
'weight_decay': 0.0004841, # optimizer weight decay
}
# 0.856 0.95 0.935 0.887 1.3 8.488 0.1081 0.01351 0.01351 0.8649 0.1 0.001 -3 0.9 0.0005
# hyp = {'k': 8.4875, # loss multiple
# 'xy': 0.108108, # xy loss fraction
# 'wh': 0.013514, # wh loss fraction
# 'cls': 0.013514, # cls loss fraction
# 'conf': 0.86486, # conf loss fraction
# 'iou_t': 0.1, # iou target-anchor training threshold
# 'lr0': 0.001, # initial learning rate
# 'lrf': -3., # final learning rate = lr0 * (10 ** lrf)
# 'momentum': 0.9, # SGD momentum
# 'weight_decay': 0.0005, # optimizer weight decay
# }
def train(
cfg,
data_cfg,
img_size=416,
resume=False,
epochs=273, # 500200 batches at bs 64, dataset length 117263
batch_size=16,
accumulate=1,
multi_scale=False,
freeze_backbone=False,
transfer=False # Transfer learning (train only YOLO layers)
):
init_seeds()
weights = 'weights' + os.sep
latest = weights + 'latest.pt'
best = weights + 'best.pt'
device = torch_utils.select_device()
if multi_scale:
img_size = 608 # initiate with maximum multi_scale size
opt.num_workers = 0 # bug https://github.com/ultralytics/yolov3/issues/174
else:
torch.backends.cudnn.benchmark = True # unsuitable for multiscale
# Configure run
data_dict = parse_data_cfg(data_cfg)
train_path = data_dict['train']
nc = int(data_dict['classes']) # number of classes
# Initialize model
model = Darknet(cfg, img_size).to(device)
# Optimizer
optimizer = optim.SGD(model.parameters(), lr=hyp['lr0'], momentum=hyp['momentum'], weight_decay=hyp['weight_decay'])
cutoff = -1 # backbone reaches to cutoff layer
start_epoch = 0
best_loss = float('inf')
nf = int(model.module_defs[model.yolo_layers[0] - 1]['filters']) # yolo layer size (i.e. 255)
if resume: # Load previously saved model
if transfer: # Transfer learning
chkpt = torch.load(weights + 'yolov3-spp.pt', map_location=device)
model.load_state_dict({k: v for k, v in chkpt['model'].items() if v.numel() > 1 and v.shape[0] != 255},
strict=False)
for p in model.parameters():
p.requires_grad = True if p.shape[0] == nf else False
else: # resume from latest.pt
chkpt = torch.load(latest, map_location=device) # load checkpoint
model.load_state_dict(chkpt['model'])
start_epoch = chkpt['epoch'] + 1
if chkpt['optimizer'] is not None:
optimizer.load_state_dict(chkpt['optimizer'])
best_loss = chkpt['best_loss']
del chkpt
else: # Initialize model with backbone (optional)
if '-tiny.cfg' in cfg:
cutoff = load_darknet_weights(model, weights + 'yolov3-tiny.conv.15')
else:
cutoff = load_darknet_weights(model, weights + 'darknet53.conv.74')
# Scheduler https://github.com/ultralytics/yolov3/issues/238
# lf = lambda x: 1 - x / epochs # linear ramp to zero
# lf = lambda x: 10 ** (hyp['lrf'] * x / epochs) # exp ramp
lf = lambda x: 1 - 10 ** (hyp['lrf'] * (1 - x / epochs)) # inverse exp ramp
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lf, last_epoch=start_epoch - 1)
# scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[218, 245], gamma=0.1, last_epoch=start_epoch-1)
# # Plot lr schedule
# y = []
# for _ in range(epochs):
# scheduler.step()
# y.append(optimizer.param_groups[0]['lr'])
# plt.plot(y, label='LambdaLR')
# plt.xlabel('epoch')
# plt.xlabel('LR')
# plt.tight_layout()
# plt.savefig('LR.png', dpi=300)
# Dataset
dataset = LoadImagesAndLabels(train_path, img_size, batch_size, augment=True)
# Initialize distributed training
if torch.cuda.device_count() > 1:
dist.init_process_group(backend=opt.backend, init_method=opt.dist_url, world_size=opt.world_size, rank=opt.rank)
model = torch.nn.parallel.DistributedDataParallel(model)
# sampler = torch.utils.data.distributed.DistributedSampler(dataset)
# Dataloader
dataloader = DataLoader(dataset,
batch_size=batch_size,
num_workers=opt.num_workers,
shuffle=False, # disable rectangular training if True
pin_memory=True,
collate_fn=dataset.collate_fn)
# Mixed precision training https://github.com/NVIDIA/apex
# install help: https://github.com/NVIDIA/apex/issues/259
mixed_precision = False
if mixed_precision:
from apex import amp
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
# Start training
t, t0 = time.time(), time.time()
model.hyp = hyp # attach hyperparameters to model
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights
model_info(model)
nb = len(dataloader)
results = (0, 0, 0, 0, 0) # P, R, mAP, F1, test_loss
n_burnin = min(round(nb / 5 + 1), 1000) # burn-in batches
os.remove('train_batch0.jpg') if os.path.exists('train_batch0.jpg') else None
os.remove('test_batch0.jpg') if os.path.exists('test_batch0.jpg') else None
for epoch in range(start_epoch, epochs):
model.train()
print(('\n%8s%12s' + '%10s' * 7) % ('Epoch', 'Batch', 'xy', 'wh', 'conf', 'cls', 'total', 'nTargets', 'time'))
# Update scheduler
scheduler.step()
# Freeze backbone at epoch 0, unfreeze at epoch 1
if freeze_backbone and epoch < 2:
for name, p in model.named_parameters():
if int(name.split('.')[1]) < cutoff: # if layer < 75
p.requires_grad = False if epoch == 0 else True
mloss = torch.zeros(5).to(device) # mean losses
for i, (imgs, targets, _, _) in enumerate(dataloader):
imgs = imgs.to(device)
targets = targets.to(device)
nt = len(targets)
# Plot images with bounding boxes
if epoch == 0 and i == 0:
plot_images(imgs=imgs, targets=targets, fname='train_batch0.jpg')
# SGD burn-in
if epoch == 0 and i <= n_burnin:
lr = hyp['lr0'] * (i / n_burnin) ** 4
for x in optimizer.param_groups:
x['lr'] = lr
# Run model
pred = model(imgs)
# Compute loss
loss, loss_items = compute_loss(pred, targets, model)
if torch.isnan(loss):
print('WARNING: nan loss detected, ending training')
return results
# Compute gradient
if mixed_precision:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# Accumulate gradient for x batches before optimizing
if (i + 1) % accumulate == 0 or (i + 1) == nb:
optimizer.step()
optimizer.zero_grad()
# Update running mean of tracked metrics
mloss = (mloss * i + loss_items) / (i + 1)
# Print batch results
s = ('%8s%12s' + '%10.3g' * 7) % (
'%g/%g' % (epoch, epochs - 1),
'%g/%g' % (i, nb - 1), *mloss, nt, time.time() - t)
t = time.time()
print(s)
# Multi-Scale training (320 - 608 pixels) every 10 batches
if multi_scale and (i + 1) % 10 == 0:
dataset.img_size = random.choice(range(10, 20)) * 32
print('multi_scale img_size = %g' % dataset.img_size)
# Calculate mAP (always test final epoch, skip first 5 if opt.nosave)
if not (opt.notest or (opt.nosave and epoch < 5)) or epoch == epochs - 1:
with torch.no_grad():
results = test.test(cfg, data_cfg, batch_size=batch_size, img_size=img_size, model=model,
conf_thres=0.1)
# Write epoch results
with open('results.txt', 'a') as file:
file.write(s + '%11.3g' * 5 % results + '\n') # P, R, mAP, F1, test_loss
# Update best loss
test_loss = results[4]
if test_loss < best_loss:
best_loss = test_loss
# Save training results
save = (not opt.nosave) or (epoch == epochs - 1)
if save:
# Create checkpoint
chkpt = {'epoch': epoch,
'best_loss': best_loss,
'model': model.module.state_dict() if type(
model) is nn.parallel.DistributedDataParallel else model.state_dict(),
'optimizer': optimizer.state_dict()}
# Save latest checkpoint
torch.save(chkpt, latest)
# Save best checkpoint
if best_loss == test_loss:
torch.save(chkpt, best)
# Save backup every 10 epochs (optional)
if epoch > 0 and epoch % 10 == 0:
torch.save(chkpt, weights + 'backup%g.pt' % epoch)
# Delete checkpoint
del chkpt
dt = (time.time() - t0) / 3600
print('%g epochs completed in %.3f hours.' % (epoch - start_epoch, dt))
return results
def print_mutation(hyp, results):
# Write mutation results
a = '%11s' * len(hyp) % tuple(hyp.keys()) # hyperparam keys
b = '%11.4g' * len(hyp) % tuple(hyp.values()) # hyperparam values
c = '%11.3g' * len(results) % results # results (P, R, mAP, F1, test_loss)
print('\n%s\n%s\nEvolved fitness: %s\n' % (a, b, c))
with open('evolve.txt', 'a') as f:
f.write(c + b + '\n')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--epochs', type=int, default=273, help='number of epochs')
parser.add_argument('--batch-size', type=int, default=16, help='size of each image batch')
parser.add_argument('--accumulate', type=int, default=1, help='accumulate gradient x batches before optimizing')
parser.add_argument('--cfg', type=str, default='cfg/yolov3-spp.cfg', help='cfg file path')
parser.add_argument('--data-cfg', type=str, default='data/coco.data', help='coco.data file path')
parser.add_argument('--multi-scale', action='store_true', help='random image sizes per batch 320 - 608')
parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)')
parser.add_argument('--resume', action='store_true', help='resume training flag')
parser.add_argument('--transfer', action='store_true', help='transfer learning flag')
parser.add_argument('--num-workers', type=int, default=2, help='number of Pytorch DataLoader workers')
parser.add_argument('--dist-url', default='tcp://127.0.0.1:9999', type=str, help='distributed training init method')
parser.add_argument('--rank', default=0, type=int, help='distributed training node rank')
parser.add_argument('--world-size', default=1, type=int, help='number of nodes for distributed training')
parser.add_argument('--backend', default='nccl', type=str, help='distributed backend')
parser.add_argument('--nosave', action='store_true', help='do not save training results')
parser.add_argument('--notest', action='store_true', help='only test final epoch')
parser.add_argument('--evolve', action='store_true', help='run hyperparameter evolution')
parser.add_argument('--var', default=0, type=int, help='debug variable')
opt = parser.parse_args()
print(opt, end='\n\n')
if opt.evolve:
opt.notest = True # save time by only testing final epoch
opt.nosave = True # do not save checkpoints
# Train
results = train(
opt.cfg,
opt.data_cfg,
img_size=opt.img_size,
resume=opt.resume or opt.transfer,
transfer=opt.transfer,
epochs=opt.epochs,
batch_size=opt.batch_size,
accumulate=opt.accumulate,
multi_scale=opt.multi_scale,
)
# Evolve hyperparameters (optional)
if opt.evolve:
best_fitness = results[2] # use mAP for fitness
# Write mutation results
print_mutation(hyp, results)
gen = 50 # generations to evolve
for _ in range(gen):
# Mutate hyperparameters
old_hyp = hyp.copy()
init_seeds(seed=int(time.time()))
s = [.2, .2, .2, .2, .2, .3, .2, .2, .02, .3]
for i, k in enumerate(hyp.keys()):
x = (np.random.randn(1) * s[i] + 1) ** 1.1 # plt.hist(x.ravel(), 100)
hyp[k] = hyp[k] * float(x) # vary by about 30% 1sigma
# Clip to limits
keys = ['iou_t', 'momentum', 'weight_decay']
limits = [(0, 0.90), (0.80, 0.95), (0, 0.01)]
for k, v in zip(keys, limits):
hyp[k] = np.clip(hyp[k], v[0], v[1])
# Normalize loss components (sum to 1)
keys = ['xy', 'wh', 'cls', 'conf']
s = sum([v for k, v in hyp.items() if k in keys])
for k in keys:
hyp[k] /= s
# Determine mutation fitness
results = train(
opt.cfg,
opt.data_cfg,
img_size=opt.img_size,
resume=opt.resume or opt.transfer,
transfer=opt.transfer,
epochs=opt.epochs,
batch_size=opt.batch_size,
accumulate=opt.accumulate,
multi_scale=opt.multi_scale,
)
mutation_fitness = results[2]
# Write mutation results
print_mutation(hyp, results)
# Update hyperparameters if fitness improved
if mutation_fitness > best_fitness:
# Fitness improved!
print('Fitness improved!')
best_fitness = mutation_fitness
else:
hyp = old_hyp.copy() # reset hyp to
# # Plot results
# import numpy as np
# import matplotlib.pyplot as plt
#
# a = np.loadtxt('evolve.txt')
# x = a[:, 3]
# fig = plt.figure(figsize=(14, 7))
# for i in range(1, 10):
# plt.subplot(2, 5, i)
# plt.plot(x, a[:, i + 5], '.')