-
Notifications
You must be signed in to change notification settings - Fork 13
/
qat.py
205 lines (189 loc) · 10 KB
/
qat.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
from super_gradients.training.models.detection_models.pp_yolo_e import PPYoloEPostPredictionCallback
from super_gradients.training.datasets.detection_datasets.coco_format_detection import COCOFormatDetectionDataset
from super_gradients.training.transforms.transforms import DetectionMosaic, DetectionRandomAffine, DetectionHSV, \
DetectionHorizontalFlip, DetectionPaddedRescale, DetectionStandardize, DetectionTargetsFormatTransform
from super_gradients.training.datasets.datasets_utils import worker_init_reset_seed
from super_gradients.training.utils.detection_utils import CrowdDetectionCollateFN
from super_gradients.training.pre_launch_callbacks import modify_params_for_qat
from super_gradients.training.metrics import DetectionMetrics_050
from super_gradients.training.losses import PPYoloELoss
from super_gradients.training import dataloaders
from super_gradients.training import Trainer
from super_gradients.training import models
import argparse
import torch
import time
import yaml
import json
import os
if __name__ == '__main__':
ap = argparse.ArgumentParser()
ap.add_argument("-i", "--data", type=str, required=True,
help="path to data.yaml")
ap.add_argument("-b", "--batch", type=int, default=6,
help="Training batch size")
ap.add_argument("-e", "--epoch", type=int, default=100,
help="Training number of epochs")
ap.add_argument("-j", "--worker", type=int, default=2,
help="Training number of workers")
ap.add_argument("-m", "--model", type=str, default='yolo_nas_s',
choices=['yolo_nas_s', 'yolo_nas_m', 'yolo_nas_l'],
help="Model type (eg: yolo_nas_s)")
ap.add_argument("-w", "--weight", type=str, required=True,
help="path to pre-trained model weight [ckpt_best.pth]")
ap.add_argument("-s", "--size", type=int, default=640,
help="input image size")
ap.add_argument("--gpus", action='store_true',
help="Run on all gpus")
ap.add_argument("--cpu", action='store_true',
help="Run on CPU")
# train_params
ap.add_argument("--warmup_mode", type=str, default='linear_epoch_step',
help="Warmup Mode")
ap.add_argument("--warmup_initial_lr", type=float, default=1e-6,
help="Warmup Initial LR")
ap.add_argument("--lr_warmup_epochs", type=int, default=3,
help="LR Warmup Epochs")
ap.add_argument("--initial_lr", type=float, default=5e-4,
help="Inital LR")
ap.add_argument("--lr_mode", type=str, default='cosine',
help="LR Mode")
ap.add_argument("--cosine_final_lr_ratio", type=float, default=0.1,
help="Cosine Final LR Ratio")
ap.add_argument("--optimizer", type=str, default='AdamW',
help="Optimizer")
ap.add_argument("--weight_decay", type=float, default=0.0001,
help="Weight Decay")
args = vars(ap.parse_args())
# Quantization Aware Training INFO
print("\x1b[6;37;41m [INFO] Quantization Aware Training \x1b[0m")
print("\x1b[1;37;41m [WARNING]: Quantization Aware Training Requires a Large Amount of System RAM \x1b[0m")
# Start Time
s_time = time.time()
# Load Path Params
yaml_params = yaml.safe_load(open(args['data'], 'r'))
with open(os.path.join(yaml_params['Dir'], yaml_params['labels']['train'])) as f:
no_class = len(json.load(f)['categories'])
f.close()
print(f"\033[1m[INFO] Number of Classes: {no_class}\033[0m")
# Training on GPU or CPU
_, name = args['weight'].split('/')[-3:-1]
if args['cpu']:
print('[INFO] Training on \033[1mCPU\033[0m')
trainer = Trainer(experiment_name=name, ckpt_root_dir='qat', device='cpu')
elif args['gpus']:
print(f'[INFO] Training on GPU: \033[1m{torch.cuda.get_device_name()}\033[0m')
trainer = Trainer(experiment_name=name, ckpt_root_dir='qat', multi_gpu=args['gpus'])
else:
print(f'[INFO] Training on GPU: \033[1m{torch.cuda.get_device_name()}\033[0m')
trainer = Trainer(experiment_name=name, ckpt_root_dir='qat')
# Load best model
best_model = models.get(args['model'],
num_classes=no_class,
checkpoint_path=args['weight'])
# Reain Dataset
trainset = COCOFormatDetectionDataset(data_dir=yaml_params['Dir'],
images_dir=yaml_params['images']['train'],
json_annotation_file=yaml_params['labels']['train'],
input_dim=(args['size'], args['size']),
ignore_empty_annotations=False,
transforms=[
DetectionMosaic(prob=1., input_dim=(args['size'], args['size'])),
DetectionRandomAffine(degrees=0., scales=(0.5, 1.5), shear=0.,
target_size=(args['size'], args['size']),
filter_box_candidates=False, border_value=128),
DetectionHSV(prob=1., hgain=5, vgain=30, sgain=30),
DetectionHorizontalFlip(prob=0.5),
DetectionPaddedRescale(input_dim=(args['size'], args['size']), max_targets=300),
DetectionStandardize(max_value=255),
DetectionTargetsFormatTransform(max_targets=300, input_dim=(args['size'], args['size']),
output_format="LABEL_CXCYWH")
])
train_dataloader_params = {
"shuffle": True,
"batch_size": args['batch'],
"drop_last": False,
"pin_memory": True,
"collate_fn": CrowdDetectionCollateFN(),
"worker_init_fn": worker_init_reset_seed,
"min_samples": 512
}
# Valid Data
valset = COCOFormatDetectionDataset(data_dir=yaml_params['Dir'],
images_dir=yaml_params['images']['val'],
json_annotation_file=yaml_params['labels']['val'],
input_dim=(args['size'], args['size']),
ignore_empty_annotations=False,
transforms=[
DetectionPaddedRescale(input_dim=(args['size'], args['size']), max_targets=300),
DetectionStandardize(max_value=255),
DetectionTargetsFormatTransform(max_targets=300, input_dim=(args['size'], args['size']),
output_format="LABEL_CXCYWH")
])
val_dataloader_params = {
"shuffle": False,
"batch_size": int(args['batch']*2),
"num_workers": args['worker'],
"drop_last": False,
"pin_memory": True,
"collate_fn": CrowdDetectionCollateFN(),
"worker_init_fn": worker_init_reset_seed
}
train_params = {
'silent_mode': False,
"average_best_models":True,
"warmup_mode": args['warmup_mode'],
"warmup_initial_lr": args['warmup_initial_lr'],
"lr_warmup_epochs": args['lr_warmup_epochs'],
"initial_lr": args['initial_lr'],
"lr_mode": args['lr_mode'],
"cosine_final_lr_ratio": args['cosine_final_lr_ratio'],
"optimizer": args['optimizer'],
"optimizer_params": {"weight_decay": args['weight_decay']},
"zero_weight_decay_on_bias_and_bn": True,
"ema": True,
"ema_params": {"decay": 0.9, "decay_type": "threshold"},
"max_epochs": args['epoch'],
"mixed_precision": True,
"loss": PPYoloELoss(
use_static_assigner=False,
num_classes=no_class,
reg_max=16
),
"valid_metrics_list": [
DetectionMetrics_050(
score_thres=0.1,
top_k_predictions=300,
num_cls=no_class,
normalize_targets=True,
post_prediction_callback=PPYoloEPostPredictionCallback(
score_threshold=0.01,
nms_top_k=1000,
max_predictions=300,
nms_threshold=0.7
)
)
],
"metric_to_watch": 'mAP@0.50'
}
# Quantization Aware Training
print("\x1b[1;37;41m [INFO]: Launching Quantization Aware Training \x1b[0m")
train_params, trainset, valset, train_dataloader_params, val_dataloader_params = modify_params_for_qat(
train_params, trainset, valset, train_dataloader_params, val_dataloader_params
)
# Print Training Params
print('[INFO] Training Params:\n', train_params)
train_loader = dataloaders.get(dataset=trainset,
dataloader_params=train_dataloader_params)
valid_loader = dataloaders.get(dataset=valset,
dataloader_params=val_dataloader_params)
# Quantization Aware Training
trainer.qat(
model=best_model,
training_params=train_params,
train_loader=train_loader,
valid_loader=valid_loader,
calib_loader=train_loader
)
print(f'[INFO] Training Completed in \033[1m{(time.time()-s_time)/3600} Hours\033[0m')
print("\x1b[1;37;42m [SUCCESS]: Quantization Aware Training Completed \x1b[0m")