-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
261 lines (218 loc) · 12.2 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
import os
import sys
import random
import pprint
import platform
import argparse
from pathlib import Path
from copy import deepcopy
from datetime import datetime
from collections import defaultdict
import torch
import numpy as np
from torch.cuda import amp
from tqdm import tqdm, trange
from thop import profile
from torch import nn
from torch import optim
from torch.backends import cudnn
from torch.utils.data import DataLoader, distributed
import torch.distributed as dist
import torch.multiprocessing as mp
from torch.nn.parallel import DistributedDataParallel as DDP
ROOT = Path(__file__).resolve().parents[0]
OS_SYSTEM = platform.system()
TIMESTAMP = datetime.today().strftime("%Y-%m-%d_%H-%M")
cudnn.benchmark = True
SEED = 2023
random.seed(SEED)
torch.manual_seed(SEED)
from dataloader import Dataset, BasicTransform, AugmentTransform
from model import YoloModel
from utils import (YoloLoss, Evaluator, ModelEMA,
resume_state, generate_random_color, set_lr,
build_basic_logger, setup_worker_logging, setup_primary_logging, de_parallel)
from val import validate, result_analyis
def setup(rank, world_size):
if OS_SYSTEM == "Linux":
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "12345"
dist.init_process_group("nccl", rank=rank, world_size=world_size)
def cleanup():
if OS_SYSTEM == "Linux":
dist.destroy_process_group()
def train(args, dataloader, model, ema, criterion, optimizer, scaler):
loss_type = ["multipart", "obj", "noobj", "txty", "twth", "cls"]
losses = defaultdict(float)
model.train()
optimizer.zero_grad()
for i, minibatch in enumerate(dataloader):
ni = i + len(dataloader) * (epoch - 1)
if ni <= args.nw:
args.grad_accumulate = max(1, np.interp(ni, [0, args.nw], [1, args.nominal_batch_size / args.batch_size]).round())
set_lr(optimizer, args.base_lr * pow(ni / (args.nw), 4))
images, labels = minibatch[1], minibatch[2]
if args.multiscale:
if ni % 10 == 0 and ni > 0:
args.train_size = random.randint(10, 19) * 32
model.module.set_grid_xy(input_size=args.train_size) if hasattr(model, "module") else model.set_grid_xy(input_size=args.train_size)
criterion.set_grid_xy(input_size=args.train_size)
images = nn.functional.interpolate(images, size=args.train_size, mode="bilinear")
with amp.autocast(enabled=not args.no_amp):
predictions = model(images.cuda(args.rank, non_blocking=True))
loss = criterion(predictions=predictions, labels=labels)
scaler.scale((loss[0] / args.grad_accumulate) * args.world_size).backward()
if ni - args.last_opt_step >= args.grad_accumulate:
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
if ema is not None:
ema.update(model)
args.last_opt_step = ni
for loss_name, loss_value in zip(loss_type, loss):
if not torch.isfinite(loss_value) and loss_name != "multipart":
print(f"############## {loss_name} Loss is Nan/Inf ! {loss_value} ##############")
sys.exit(0)
else:
losses[loss_name] += loss_value.item()
del images, predictions
torch.cuda.empty_cache()
loss_str = f"[Train-Epoch:{epoch:03d}] "
for loss_name in loss_type:
losses[loss_name] /= len(dataloader)
loss_str += f"{loss_name}: {losses[loss_name]:.4f} "
return loss_str
def parse_args(make_dirs=True):
parser = argparse.ArgumentParser()
parser.add_argument("--exp", type=str, required=True, help="Name to log training")
parser.add_argument("--data", type=str, default="toy.yaml", help="Path to data.yaml")
parser.add_argument("--img-size", type=int, default=416, help="Model input size")
parser.add_argument("--batch-size", type=int, default=32, help="Batch size")
parser.add_argument("--num-epochs", type=int, default=200, help="Number of training epochs")
parser.add_argument("--warmup", type=int, default=1, help="Epochs for warming up training")
parser.add_argument("--base-lr", type=float, default=0.001, help="Base learning rate")
parser.add_argument("--lr-decay", nargs="+", default=[100, 150], type=int, help="Epoch to learning rate decay")
parser.add_argument("--momentum", type=float, default=0.9, help="Momentum")
parser.add_argument("--weight-decay", type=float, default=0.0005, help="Weight decay")
parser.add_argument("--label-smoothing", type=float, default=0.1, help="Label smoothing")
parser.add_argument("--conf-thres", type=float, default=0.001, help="Threshold to filter confidence score")
parser.add_argument("--nms-thres", type=float, default=0.6, help="Threshold to filter Box IoU of NMS process")
parser.add_argument("--img-interval", type=int, default=10, help="Interval to log train/val image")
parser.add_argument("--workers", type=int, default=8, help="Number of workers used in dataloader")
parser.add_argument("--world-size", type=int, default=1, help="Number of available GPU devices")
parser.add_argument("--rank", type=int, default=0, help="Process id for computation")
parser.add_argument("--no-amp", action="store_true", help="Use of FP32 training (default: AMP training)")
parser.add_argument("--multiscale", action="store_true", help="Multi-scale training")
parser.add_argument("--scratch", action="store_true", help="Scratch training without pretrained weights")
parser.add_argument("--resume", action="store_true", help="Name to resume path")
args = parser.parse_args()
args.data = ROOT / "data" / args.data
args.exp_path = ROOT / "experiment" / args.exp
args.weight_dir = args.exp_path / "weight"
args.img_log_dir = args.exp_path / "train-image"
args.load_path = args.weight_dir / "last.pt" if args.resume else None
assert args.world_size > 0, "Executable GPU machine does not exist, This training supports on CUDA available environment."
if make_dirs:
os.makedirs(args.weight_dir, exist_ok=True)
os.makedirs(args.img_log_dir, exist_ok=True)
return args
def main_work(rank, world_size, args, logger):
################################### Init Process ####################################
setup(rank, world_size)
torch.manual_seed(SEED)
torch.cuda.set_device(rank)
if OS_SYSTEM == "Linux":
import logging
setup_worker_logging(rank, logger)
else:
logging = logger
################################### Init Instance ###################################
global epoch
args.rank = rank
args.last_opt_step = -1
args.nominal_batch_size = 64
args.batch_size = args.batch_size // world_size
args.train_size = 608 if args.multiscale else args.img_size
args.grad_accumulate = max(round(args.nominal_batch_size / args.batch_size), 1)
args.workers = min([os.cpu_count() // max(world_size, 1), args.batch_size if args.batch_size > 1 else 0, args.workers])
train_dataset = Dataset(yaml_path=args.data, phase="train")
train_transformer = AugmentTransform(input_size=args.train_size)
train_dataset.load_transformer(transformer=train_transformer)
train_sampler = distributed.DistributedSampler(train_dataset, num_replicas=world_size, rank=args.rank, shuffle=True)
train_loader = DataLoader(dataset=train_dataset, collate_fn=Dataset.collate_fn, batch_size=args.batch_size,
shuffle=False, pin_memory=True, num_workers=args.workers, sampler=train_sampler)
val_dataset = Dataset(yaml_path=args.data, phase="val")
val_transformer = BasicTransform(input_size=args.img_size)
val_dataset.load_transformer(transformer=val_transformer)
val_loader = DataLoader(dataset=val_dataset, collate_fn=Dataset.collate_fn, batch_size=args.batch_size,
shuffle=False, pin_memory=True, num_workers=args.workers)
args.anchors = train_dataset.anchors
args.class_list = train_dataset.class_list
args.color_list = generate_random_color(len(args.class_list))
args.nw = max(round(args.warmup * len(train_loader)), 100)
args.mAP_filepath = val_dataset.mAP_filepath
model = YoloModel(input_size=args.img_size, num_classes=len(args.class_list), anchors=args.anchors, pretrained=not args.scratch).cuda(args.rank)
macs, params = profile(deepcopy(model), inputs=(torch.randn(1, 3, args.img_size, args.img_size).cuda(args.rank),), verbose=False)
model.set_grid_xy(input_size=args.train_size)
criterion = YoloLoss(input_size=args.train_size, anchors=model.anchors, label_smoothing=args.label_smoothing)
optimizer = optim.SGD(model.parameters(), lr=args.base_lr, momentum=args.momentum, weight_decay=args.weight_decay)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.lr_decay, gamma=0.1)
evaluator = Evaluator(annotation_file=args.mAP_filepath)
scaler = amp.GradScaler(enabled=not args.no_amp)
ema = ModelEMA(model=model) if args.rank == 0 else None
#################################### Load Model #####################################
if args.resume:
assert args.load_path.is_file(), "Not exist trained weights in the directory path !"
start_epoch = resume_state(args.load_path, args.rank, model, ema, optimizer, scheduler, scaler)
else:
start_epoch = 1
if args.rank == 0:
logging.warning(f"[Arguments]\n{pprint.pformat(vars(args))}\n")
logging.warning(f"Architecture Info - Params(M): {params/1e+6:.2f}, FLOPs(B): {2*macs/1E+9:.2f}")
#################################### Train Model ####################################
if OS_SYSTEM == "Linux":
model = DDP(model, device_ids=[args.rank])
dist.barrier()
if args.rank == 0:
progress_bar = trange(start_epoch, args.num_epochs+1, total=args.num_epochs, initial=start_epoch, ncols=115)
else:
progress_bar = range(start_epoch, args.num_epochs+1)
best_epoch, best_score, best_mAP_str, mAP_dict = 0, 0, "", None
for epoch in progress_bar:
if args.rank == 0:
train_loader = tqdm(train_loader, desc=f"[TRAIN:{epoch:03d}/{args.num_epochs:03d}]", ncols=115, leave=False)
train_sampler.set_epoch(epoch)
train_loss_str = train(args=args, dataloader=train_loader, model=model, ema=ema, criterion=criterion, optimizer=optimizer, scaler=scaler)
if args.rank == 0:
logging.warning(train_loss_str)
save_opt = {"running_epoch": epoch,
"class_list": args.class_list,
"model_state": deepcopy(de_parallel(model)).state_dict(),
"ema_state": deepcopy(ema.module).state_dict(),
"ema_update": ema.updates,
"optimizer_state": optimizer.state_dict(),
"scheduler_state": scheduler.state_dict(),
"scaler_state_dict": scaler.state_dict()}
torch.save(save_opt, args.weight_dir / "last.pt")
if epoch % 10 == 0:
val_loader = tqdm(val_loader, desc=f"[VAL:{epoch:03d}/{args.num_epochs:03d}]", ncols=115, leave=False)
mAP_dict, eval_text = validate(args=args, dataloader=val_loader, model=ema.module, evaluator=evaluator, epoch=epoch)
ap50 = mAP_dict["all"]["mAP_50"]
logging.warning(eval_text)
if ap50 > best_score:
result_analyis(args=args, mAP_dict=mAP_dict["all"])
best_epoch, best_score, best_mAP_str = epoch, ap50, eval_text
torch.save(save_opt, args.weight_dir / "best.pt")
scheduler.step()
if mAP_dict and args.rank == 0:
logging.warning(f"[Best mAP at {best_epoch}]{best_mAP_str}")
cleanup()
if __name__ == "__main__":
args = parse_args(make_dirs=True)
if OS_SYSTEM == "Linux":
torch.multiprocessing.set_start_method("spawn", force=True)
logger = setup_primary_logging(args.exp_path / "train.log")
mp.spawn(main_work, args=(args.world_size, args, logger), nprocs=args.world_size, join=True)
else:
logger = build_basic_logger(args.exp_path / "train.log")
main_work(rank=0, world_size=1, args=args, logger=logger)