-
Notifications
You must be signed in to change notification settings - Fork 0
/
setup_yolov5.py
201 lines (173 loc) · 8.2 KB
/
setup_yolov5.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
import os
import sys
from pathlib import Path
from shutil import rmtree, copytree, copy2
from glob import glob
import pkg_resources
import importlib
from multiprocessing import cpu_count
import yaml
import numpy as np
from config import Config, parser
from utils.general import quietly_run, listify, sizify, autotype
from utils import detection
from utils.detection import write_yolov5_labels, write_dataset_yaml
from metadata import get_metadata
DEBUG = False
cd_yolov5 = False
parser.add_argument("--repo")
parser.add_argument("--branch")
parser.add_argument("--wheels", help="local path to wheels (pycocotools, required packages)")
parser.add_argument("--run", "--launch", action="store_true", help="run YOLOv5 after setup")
# Read config file and parser_args
parser_args, _ = parser.parse_known_args(sys.argv)
print("[ √ ] Config file:", parser_args.config_file)
cfg = Config('configs/yolov5_setup')
if parser_args.config_file: cfg.update(parser_args.config_file)
cfg.repo = parser_args.repo or cfg.repo
cfg.branch = parser_args.branch or cfg.branch
cfg.wheels = parser_args.wheels or cfg.wheels
cfg.use_folds = parser_args.use_folds or cfg.use_folds
cfg.epochs = parser_args.epochs or cfg.epochs
cfg.size = sizify(parser_args.size or cfg.size)
cfg.keep_unlabelled = cfg.keep_unlabelled or False
for key, value in listify(parser_args.set):
autotype(cfg, key, value)
if DEBUG: print(cfg)
# Installs
if os.path.exists('yolov5'):
rmtree('yolov5')
if cfg.branch:
quietly_run(f'git clone -b {cfg.branch} {cfg.repo}')
else:
quietly_run(f'git clone {cfg.repo}')
quietly_run('pip install -r yolov5/requirements.txt')
print(f"[ √ ] YOLOv5 branch {cfg.branch} from {cfg.repo}")
if cfg.wheels:
whls = glob(f'{cfg.wheels}/pycocotools-*.whl')
if whls:
if len(whls) > 1:
print('WARNING: found several wheels for pycocotools, using', whls[0])
quietly_run(f'pip install {whls[0]}')
else:
quietly_run('pip install pycocotools')
print("[ √ ] pycocotools:", pkg_resources.require("pycocotools")[0].version)
if cd_yolov5: os.chdir('yolov5')
data_path = Path(os.getcwd()) / Path('data' if cd_yolov5 else 'yolov5/data')
assert data_path.exists()
# Import project (code, constant settings)
project = importlib.import_module(f'projects.{cfg.project}') if cfg.project else None
if project:
print("[ √ ] Project:", cfg.project)
project.init(cfg)
# Metadata, labels
metadata = get_metadata(cfg, project)
if DEBUG:
print(f"Row 1 / {len(metadata)} in metadata after get_metadata():")
print(metadata.iloc[0])
if hasattr(project, 'get_yolov5_labels'):
metadata = project.get_yolov5_labels(metadata, cfg)
else:
metadata = detection.get_yolov5_labels(metadata, cfg)
if DEBUG:
print(f"Row 1 / {len(metadata)} in metadata after get_yolov5_labels():")
print(metadata.iloc[0])
# Dump labels for external validation
cfg.out_dir = Path(cfg.out_dir)
os.makedirs(cfg.out_dir, exist_ok=True)
metadata.to_json(cfg.out_dir / 'labels.json')
# Create file structure required by YOLOv5 datasets:
# - Separate folders for train/valid images
# - Symlinks to actual image file location (superfast)
# - Labels in separate .txt files, one per image (slow)
#
cfg.use_folds = sorted(set(cfg.use_folds)) # duplicate folds would overwrite results
for i, fold in enumerate(cfg.use_folds):
path = data_path / f'fold{fold}'
print(f"Preparing data for fold {fold} ({i+1}/{len(cfg.use_folds)} to go)")
metadata['is_valid'] = metadata.fold == fold
os.makedirs(path / 'labels' / 'train')
os.makedirs(path / 'images' / 'train')
os.makedirs(path / 'labels' / 'valid')
os.makedirs(path / 'images' / 'valid')
write_dataset_yaml(cfg, path)
for image_path in metadata.loc[~ metadata.is_valid, 'image_path'].unique():
os.symlink(cfg.image_root / image_path, path / 'images' / 'train' / Path(image_path).name)
for image_path in metadata.loc[metadata.is_valid, 'image_path'].unique():
os.symlink(cfg.image_root / image_path, path / 'images' / 'valid' / Path(image_path).name)
if cfg.train_on_all:
image_path = metadata.image_path.iloc[0] # YOLOv5 needs at least one valid image
os.symlink(cfg.image_root / image_path, path / 'images' / 'valid' / Path(image_path).name)
if cfg.n_bg_images:
print(f"Adding {cfg.n_bg_images} unlabelled background images to train.")
for image_path in np.random.choice(cfg.bg_images, size=cfg.n_bg_images, replace=False):
os.symlink(cfg.image_root / image_path, path / 'images' / 'train' / Path(image_path).name)
for split in ('train', 'valid'):
print(f"Writing label txt files for {split}...")
write_yolov5_labels(metadata, split, path)
# Write hyperparameter yaml file
hyp_yaml = data_path / 'hyps' / 'hyp.yaml'
with open(data_path / 'hyps' / 'hyp.scratch.yaml', 'r') as fp:
hyp = yaml.safe_load(fp)
hyp.update(dict(lr0=cfg.lr, lrf=0.1,
hsv_h=cfg.hue or 0.0, hsv_s=cfg.saturation or 0.0, hsv_v=cfg.value or 0.0,
scale=0.6, flipud=0.0, fliplr=0.5 if cfg.hflip else 0,
perspective=cfg.perspective or 0.0,
anchor_t=cfg.anchor_t or 4.0,
cls=cfg.cls or 0.5,
box=cfg.box or 0.05))
with open(hyp_yaml, 'w', encoding='utf8') as fp:
yaml.dump(hyp, fp, default_flow_style=False, allow_unicode=True)
print("[ √ ] YOLOv5 🚀 is ready for launch!")
if not parser_args.run:
quit()
assert len(cfg.use_folds) == 1
fold = f'fold{cfg.use_folds[0]}'
os.chdir(data_path.parent)
quietly_run('wandb disabled')
# Prepare restart files
if cfg.rst_path and cfg.rst_name:
checkpoint_file = cfg.out_dir / fold / 'weights' / f'{cfg.rst_name}.pt'
if Path(cfg.rst_path).name == 'weights':
rst_fold = Path(cfg.rst_path).parent.name
if rst_fold != fold:
print(f"WARNING: restart from potentially different fold: {rst_fold}")
copytree(Path(cfg.rst_path).parent, cfg.out_dir / fold, dirs_exist_ok=True)
assert os.path.exists(checkpoint_file), f'{cfg.rst_name}.pt not found in {cfg.rst_pth}'
assert os.path.isfile(checkpoint_file), f'{checkpoint_file} is no file'
assert os.path.exists(cfg.out_dir / fold / 'opt.yaml'), f'opt.yaml missing for resume'
else:
os.makedirs(cfg.out_dir / fold / 'weights', exist_ok=True)
copy2(Path(cfg.rst_path) / f'{cfg.rst_name}.pt', checkpoint_file)
if (Path(cfg.rst_path) / f'{cfg.rst_name}_opt.yaml').exists():
copy2(Path(cfg.rst_path) / f'{cfg.rst_name}_opt.yaml', cfg.out_dir / fold / 'opt.yaml')
elif (Path(cfg.rst_path).parent / 'opt.yaml').exists():
copy2(Path(cfg.rst_path).parent / 'opt.yaml', cfg.out_dir / fold / 'opt.yaml')
elif (Path(cfg.rst_path) / 'opt.yaml').exists():
copy2(Path(cfg.rst_path) / 'opt.yaml', cfg.out_dir / fold / 'opt.yaml')
# Compose YOLOv5 command-line options
weights_or_resume = (f'--resume {checkpoint_file}' if cfg.rst_path and cfg.rst_name else
f'--weights {cfg.pretrained}' if cfg.pretrained else
f'--weights "" --cfg {cfg.arch_name}.yaml')
size = f'--img {max(cfg.size)}'
rect = '--rect' if cfg.rectangular else ''
batch = f'--batch {cfg.bs}'
epochs = f'--epochs {cfg.epochs}'
hyp = f'--hyp {hyp_yaml.relative_to(data_path.parent)}'
adam = f'--adam' if cfg.use_adam else ''
data = f'--data {fold}/dataset.yaml'
image_weights = f'--image-weights' if cfg.class_weights and (cfg.n_classes > 1) else ''
multi_scale = '--multi-scale' if cfg.multiscale else ''
noautoanchor = '--noautoanchor' if cfg.noautoanchor else ''
project = f'--project {cfg.out_dir}'
name = f'--name {fold}'
workers = f'--workers {cpu_count()}'
exist_ok = '--exist-ok'
val = ('--noval' if cfg.train_on_all else
f'--val_iou_thres {cfg.val_iou_thres} --val_conf_thres {cfg.val_conf_thres} --val_max_det {cfg.val_max_det}')
cache = '--cache' if cfg.cache_images else ''
aux_loss = f'--aux_loss {cfg.aux_loss} --noautoanchor' if cfg.aux_loss else ''
print(f"\nRunning YOLOv5 on {fold}...")
cml_options = ' '.join([weights_or_resume, size, rect, batch, epochs, hyp, adam, data, image_weights,
multi_scale, noautoanchor, project, name, workers, exist_ok, val, cache, aux_loss])
quietly_run(f'python train.py {cml_options}', debug=True)