class config:
dataset_name = 'COCO'
network = 'resnet50_retinanet'
pretrained = False
num_classes = 80
input_image_size = 800
model = retinanet.__dict__[network](**{
'pretrained': pretrained,
'num_classes': num_classes,
})
decoder = RetinaDecoder()
val_dataset = CocoDetection(COCO2017_path,
set_name='val2017',
transform=transforms.Compose([
Normalize(),
RetinaStyleResize(
resize=input_image_size,
multi_scale=False,
multi_scale_range=[0.8, 1.0]),
]))
seed = 0
batch_size = 16
num_workers = 16
trained_model_path = 'detection_training/coco/retinanet_res50_resize800_multi_ciou/checkpoints/resnet50_retinanet-epoch12-mAP0.355.pth'
class config:
dataset_name = 'COCO'
network = 'resnet50_fcos'
pretrained = False
num_classes = 80
input_image_size = 800
model = fcos.__dict__[network](**{
'pretrained': pretrained,
'num_classes': num_classes,
})
decoder = FCOSDecoder()
val_dataset = CocoDetection(COCO2017_path,
set_name='val2017',
transform=transforms.Compose([
Normalize(),
RetinaStyleResize(
resize=input_image_size,
multi_scale=False,
multi_scale_range=[0.8, 1.0]),
]))
seed = 0
# when testing,using nn.DataParallel mode,batch_size is total size
batch_size = 16
num_workers = 16
trained_model_path = 'detection_training/coco/fcos_res50_resize800_multi_ciou/checkpoints/resnet50_fcos-epoch12-mAP0.37735844188817935.pth'
class config:
dataset_name = 'COCO'
network = 'resnet18_centernet'
pretrained = False
num_classes = 80
input_image_size = 512
model = centernet.__dict__[network](**{
'pretrained': pretrained,
'num_classes': num_classes,
})
decoder = CenterNetDecoder()
val_dataset = CocoDetection(COCO2017_path,
set_name='val2017',
transform=transforms.Compose([
Normalize(),
YoloStyleResize(
resize=input_image_size,
multi_scale=False,
multi_scale_range=[0.6, 1.4]),
]))
seed = 0
batch_size = 64
num_workers = 16
trained_model_path = ''
class config:
dataset_name = 'COCO'
network = 'yolov3'
pretrained = False
num_classes = 80
input_image_size = 416
model = yolov3.__dict__[network](**{
'pretrained':
pretrained,
'anchor_sizes': [[10, 13], [16, 30], [33, 23], [30, 61], [62, 45],
[59, 119], [116, 90], [156, 198], [373, 326]],
'strides': [8, 16, 32],
'num_classes':
num_classes,
})
decoder = Yolov4Decoder()
val_dataset = CocoDetection(COCO2017_path,
set_name='val2017',
transform=transforms.Compose([
Normalize(),
YoloStyleResize(
resize=input_image_size,
multi_scale=False,
multi_scale_range=[0.8, 1.0]),
]))
seed = 0
batch_size = 64
num_workers = 16
trained_model_path = ''
class config:
dataset_name = 'VOC'
network = 'resnet50_retinanet'
pretrained = False
num_classes = 20
input_image_size = 400
model = retinanet.__dict__[network](**{
'pretrained': pretrained,
'num_classes': num_classes,
})
decoder = RetinaDecoder()
val_dataset = VocDetection(root_dir=VOCdataset_path,
image_sets=[('2007', 'test')],
transform=transforms.Compose([
Normalize(),
RetinaStyleResize(
resize=input_image_size,
multi_scale=False,
multi_scale_range=[0.8, 1.0]),
]),
keep_difficult=False)
seed = 0
batch_size = 64
num_workers = 16
eval_iou_threshold = 0.5
trained_model_path = 'detection_training/voc/retinanet_res50_resize400_multi_ciou/checkpoints/resnet50_retinanet-epoch12-mAP0.7711355083699912.pth'
class config:
dataset_name = 'COCO'
network = 'resnet50_fcos'
pretrained = False
num_classes = 80
input_image_size = 400
model = fcos.__dict__[network](**{
'pretrained': pretrained,
'num_classes': num_classes,
})
criterion = FCOSLoss(box_loss_iou_type='CIoU', box_loss_weight=1.)
decoder = FCOSDecoder()
train_dataset = CocoDetection(COCO2017_path,
set_name='train2017',
transform=transforms.Compose([
RandomHorizontalFlip(flip_prob=0.5),
Normalize(),
RetinaStyleResize(
resize=input_image_size,
multi_scale=True,
multi_scale_range=[0.8, 1.0]),
]))
val_dataset = CocoDetection(COCO2017_path,
set_name='val2017',
transform=transforms.Compose([
Normalize(),
RetinaStyleResize(
resize=input_image_size,
multi_scale=False,
multi_scale_range=[0.8, 1.0]),
]))
seed = 0
# batch_size is total size in DataParallel mode
# batch_size is per gpu node size in DistributedDataParallel mode
batch_size = 16
num_workers = 16
# choose 'SGD' or 'AdamW'
optimizer = 'AdamW'
# 'AdamW' doesn't need gamma and momentum variable
gamma = 0.1
momentum = 0.9
# choose 'MultiStepLR' or 'CosineLR'
# milestones only use in 'MultiStepLR'
scheduler = 'MultiStepLR'
lr = 1e-4
weight_decay = 1e-3
milestones = [8, 11]
warm_up_epochs = 0
epochs = 12
eval_epoch = [1, 3, 5, 8, 11, 12]
print_interval = 100
# only in DistributedDataParallel mode can use sync_bn
distributed = True
sync_bn = False
apex = True
class config:
dataset_name = 'COCO'
network = 'resnet18_centernet'
pretrained = False
num_classes = 80
input_image_size = 512
model = centernet.__dict__[network](**{
'pretrained': pretrained,
'num_classes': num_classes,
})
criterion = CenterNetLoss()
decoder = CenterNetDecoder()
train_dataset = CocoDetection(COCO2017_path,
set_name='train2017',
transform=transforms.Compose([
RandomHorizontalFlip(flip_prob=0.5),
Normalize(),
YoloStyleResize(
resize=input_image_size,
multi_scale=True,
multi_scale_range=[0.6, 1.4]),
]))
val_dataset = CocoDetection(COCO2017_path,
set_name='val2017',
transform=transforms.Compose([
Normalize(),
YoloStyleResize(
resize=input_image_size,
multi_scale=False,
multi_scale_range=[0.6, 1.4]),
]))
seed = 0
# batch_size is total size in DataParallel mode
# batch_size is per gpu node size in DistributedDataParallel mode
batch_size = 64
num_workers = 16
# choose 'SGD' or 'AdamW'
optimizer = 'AdamW'
# 'AdamW' doesn't need gamma and momentum variable
gamma = 0.1
momentum = 0.9
# choose 'MultiStepLR' or 'CosineLR'
# milestones only use in 'MultiStepLR'
scheduler = 'MultiStepLR'
lr = 1e-4
weight_decay = 1e-3
milestones = [90, 120]
warm_up_epochs = 0
epochs = 140
eval_epoch = [i * 5 for i in range(epochs // 5)]
print_interval = 100
# only in DistributedDataParallel mode can use sync_bn
distributed = True
sync_bn = False
apex = True
class config:
dataset_name = 'VOC'
network = 'resnet50_retinanet'
pretrained = False
num_classes = 20
input_image_size = 400
model = retinanet.__dict__[network](**{
'pretrained': pretrained,
'num_classes': num_classes,
})
criterion = RetinaLoss()
decoder = RetinaDecoder()
train_dataset = VocDetection(root_dir=VOCdataset_path,
image_sets=[('2007', 'trainval'),
('2012', 'trainval')],
transform=transforms.Compose([
RandomHorizontalFlip(flip_prob=0.5),
Normalize(),
RetinaStyleResize(
resize=input_image_size,
multi_scale=True,
multi_scale_range=[0.8, 1.0]),
]),
keep_difficult=False)
val_dataset = VocDetection(root_dir=VOCdataset_path,
image_sets=[('2007', 'test')],
transform=transforms.Compose([
Normalize(),
RetinaStyleResize(
resize=input_image_size,
multi_scale=False,
multi_scale_range=[0.8, 1.0]),
]),
keep_difficult=False)
seed = 0
# batch_size is total size in DataParallel mode
# batch_size is per gpu node size in DistributedDataParallel mode
batch_size = 16
num_workers = 16
# choose 'SGD' or 'AdamW'
optimizer = 'AdamW'
# 'AdamW' doesn't need gamma and momentum variable
gamma = 0.1
momentum = 0.9
# choose 'MultiStepLR' or 'CosineLR'
# milestones only use in 'MultiStepLR'
scheduler = 'MultiStepLR'
lr = 1e-4
weight_decay = 1e-3
milestones = [8, 11]
warm_up_epochs = 0
epochs = 12
eval_epoch = [1, 3, 5, 8, 11, 12]
eval_iou_threshold = 0.5
print_interval = 100
# only in DistributedDataParallel mode can use sync_bn
distributed = True
sync_bn = False
apex = True
np.random.seed(seed)
# for cpu gpu
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
from simpleAICV.detection.common import RandomHorizontalFlip, RandomCrop, RandomTranslate, Normalize, YoloStyleResize, RetinaStyleResize, DetectionCollater
cocodataset = CocoDetection(
COCO2017_path,
set_name='train2017',
transform=transforms.Compose([
RandomHorizontalFlip(),
RandomCrop(),
RandomTranslate(),
Normalize(),
#YoloStyleResize(resize=640, multi_scale=True),
RetinaStyleResize(resize=800, multi_scale=True),
]))
count = 0
for per_sample in tqdm(cocodataset):
print(per_sample['image'].shape, per_sample['annots'].shape,
per_sample['scale'], per_sample['origin_hw'])
print(per_sample['image'].dtype, per_sample['annots'].dtype,
per_sample['scale'].dtype, per_sample['origin_hw'].dtype)
if count < 10:
count += 1
else:
break
class config:
dataset_name = 'COCO'
network = 'yolov4'
pretrained = False
num_classes = 80
input_image_size = 416
model = yolov4.__dict__[network](**{
'pretrained':
pretrained,
'anchor_sizes': [[10, 13], [16, 30], [33, 23], [30, 61], [62, 45],
[59, 119], [116, 90], [156, 198], [373, 326]],
'strides': [8, 16, 32],
'num_classes':
num_classes,
})
criterion = Yolov4Loss(
anchor_sizes=[[10, 13], [16, 30], [33, 23], [30, 61], [62, 45],
[59, 119], [116, 90], [156, 198], [373, 326]],
strides=[8, 16, 32],
)
decoder = Yolov4Decoder()
train_dataset = CocoDetection(COCO2017_path,
set_name='train2017',
transform=transforms.Compose([
RandomHorizontalFlip(flip_prob=0.5),
Normalize(),
YoloStyleResize(
resize=input_image_size,
multi_scale=True,
multi_scale_range=[0.8, 1.0]),
]))
val_dataset = CocoDetection(COCO2017_path,
set_name='val2017',
transform=transforms.Compose([
Normalize(),
YoloStyleResize(
resize=input_image_size,
multi_scale=False,
multi_scale_range=[0.8, 1.0]),
]))
seed = 0
# batch_size is total size in DataParallel mode
# batch_size is per gpu node size in DistributedDataParallel mode
batch_size = 32
num_workers = 16
# choose 'SGD' or 'AdamW'
optimizer = 'AdamW'
# 'AdamW' doesn't need gamma and momentum variable
gamma = 0.1
momentum = 0.9
# choose 'MultiStepLR' or 'CosineLR'
# milestones only use in 'MultiStepLR'
scheduler = 'MultiStepLR'
lr = 1e-4
weight_decay = 1e-3
milestones = [60, 90]
warm_up_epochs = 0
epochs = 100
eval_epoch = [1, 2, 3, 4, 5]
print_interval = 10
# only in DistributedDataParallel mode can use sync_bn
distributed = True
sync_bn = False
apex = True