class Config(object):
log = './log' # Path to save log
checkpoint_path = './checkpoints' # Path to store checkpoint model
resume = './checkpoints/latest.pth' # load checkpoint model
evaluate = None # evaluate model path
train_dataset_path = os.path.join(COCO2017_path, 'images/train2017')
val_dataset_path = os.path.join(COCO2017_path, 'images/val2017')
dataset_annotations_path = os.path.join(COCO2017_path, 'annotations')
network = "resnet101_fcos"
pretrained = False
num_classes = 80
seed = 0
input_image_size = 667
train_dataset = CocoDetection(image_root_dir=train_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="train2017",
transform=transforms.Compose([
RandomFlip(flip_prob=0.5),
RandomCrop(crop_prob=0.5),
RandomTranslate(translate_prob=0.5),
Resize(resize=input_image_size),
]))
val_dataset = CocoDetection(image_root_dir=val_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="val2017",
transform=transforms.Compose([
Resize(resize=input_image_size),
]))
epochs = 12
per_node_batch_size = 8
lr = 1e-4
num_workers = 4
print_interval = 100
apex = True
sync_bn = False
class Config(object):
log = './log' # Path to save log
checkpoint_path = './checkpoints' # Path to store checkpoint model
resume = './checkpoints/latest.pth' # load checkpoint model
evaluate = None # evaluate model path
train_dataset_path = os.path.join(COCO2017_path, 'images/train2017')
val_dataset_path = os.path.join(COCO2017_path, 'images/val2017')
dataset_annotations_path = os.path.join(COCO2017_path, 'annotations')
network = "resnet18_centernet"
pretrained = False
num_classes = 80
seed = 0
input_image_size = 512
use_multi_scale = True
multi_scale_range = [0.6, 1.4]
stride = 4
train_dataset = CocoDetection(image_root_dir=train_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="train2017",
transform=transforms.Compose([
RandomFlip(flip_prob=0.5),
RandomCrop(crop_prob=0.5),
RandomTranslate(translate_prob=0.5),
Normalize(),
]))
val_dataset = CocoDetection(image_root_dir=val_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="val2017",
transform=transforms.Compose([
Normalize(),
Resize(resize=input_image_size),
]))
epochs = 140
milestones = [90, 120]
per_node_batch_size = 24
lr = 5e-4
num_workers = 4
print_interval = 100
apex = True
sync_bn = False
def test_model(args):
print(args)
if args.use_gpu:
# use one Graphics card to test
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
if not torch.cuda.is_available():
raise Exception("need gpu to test network!")
torch.cuda.empty_cache()
if args.seed is not None:
random.seed(args.seed)
if args.use_gpu:
torch.cuda.manual_seed_all(args.seed)
cudnn.deterministic = True
if args.use_gpu:
cudnn.benchmark = True
cudnn.enabled = True
coco_val_dataset = CocoDetection(
image_root_dir=os.path.join(COCO2017_path, 'images/val2017'),
annotation_root_dir=os.path.join(COCO2017_path, 'annotations'),
set="val2017",
transform=transforms.Compose([
Normalize(),
Resize(resize=args.input_image_size),
]))
if args.detector == "retinanet":
model = _retinanet(args.backbone, args.use_pretrained_model,
args.pretrained_model_path, args.num_classes)
decoder = RetinaDecoder(image_w=args.input_image_size,
image_h=args.input_image_size,
min_score_threshold=args.min_score_threshold)
elif args.detector == "fcos":
model = _fcos(args.backbone, args.use_pretrained_model,
args.pretrained_model_path, args.num_classes)
decoder = FCOSDecoder(image_w=args.input_image_size,
image_h=args.input_image_size,
min_score_threshold=args.min_score_threshold)
elif args.detector == "centernet":
model = _centernet(args.backbone, args.use_pretrained_model,
args.pretrained_model_path, args.num_classes)
decoder = CenterNetDecoder(
image_w=args.input_image_size,
image_h=args.input_image_size,
min_score_threshold=args.min_score_threshold)
elif args.detector == "yolov3":
model = _yolov3(args.backbone, args.use_pretrained_model,
args.pretrained_model_path, args.num_classes)
decoder = YOLOV3Decoder(image_w=args.input_image_size,
image_h=args.input_image_size,
min_score_threshold=args.min_score_threshold)
else:
print("unsupport detection model!")
return
flops_input = torch.randn(1, 3, args.input_image_size,
args.input_image_size)
flops, params = profile(model, inputs=(flops_input, ))
flops, params = clever_format([flops, params], "%.3f")
print(
f"backbone:{args.backbone},detector: '{args.detector}', flops: {flops}, params: {params}"
)
if args.use_gpu:
model = model.cuda()
decoder = decoder.cuda()
model = nn.DataParallel(model)
print(f"start eval.")
all_eval_result = validate(coco_val_dataset, model, decoder, args)
print(f"eval done.")
if all_eval_result is not None:
print(
f"val: backbone: {args.backbone}, detector: {args.detector}, IoU=0.5:0.95,area=all,maxDets=100,mAP:{all_eval_result[0]:.3f}, IoU=0.5,area=all,maxDets=100,mAP:{all_eval_result[1]:.3f}, IoU=0.75,area=all,maxDets=100,mAP:{all_eval_result[2]:.3f}, IoU=0.5:0.95,area=small,maxDets=100,mAP:{all_eval_result[3]:.3f}, IoU=0.5:0.95,area=medium,maxDets=100,mAP:{all_eval_result[4]:.3f}, IoU=0.5:0.95,area=large,maxDets=100,mAP:{all_eval_result[5]:.3f}, IoU=0.5:0.95,area=all,maxDets=1,mAR:{all_eval_result[6]:.3f}, IoU=0.5:0.95,area=all,maxDets=10,mAR:{all_eval_result[7]:.3f}, IoU=0.5:0.95,area=all,maxDets=100,mAR:{all_eval_result[8]:.3f}, IoU=0.5:0.95,area=small,maxDets=100,mAR:{all_eval_result[9]:.3f}, IoU=0.5:0.95,area=medium,maxDets=100,mAR:{all_eval_result[10]:.3f}, IoU=0.5:0.95,area=large,maxDets=100,mAR:{all_eval_result[11]:.3f}"
)
return
class Config(object):
log = './log_obj365' # Path to save log
checkpoint_path = './checkpoints_obj365' # Path to store checkpoint model
resume = './checkpoints_obj365/latest.pth' # load checkpoint model
evaluate = None # evaluate model path
train_dataset_path = os.path.join('/home/jovyan/data-vol-polefs-1/small_sample/obj365', 'images/train2017')
val_dataset_path = os.path.join('/home/jovyan/data-vol-polefs-1/small_sample/obj365', 'images/val2017')
dataset_annotations_path = os.path.join('/home/jovyan/data-vol-polefs-1/small_sample/obj365', 'annotations')
network = "resnet50_centernet"
pretrained = False
num_classes = [365]
seed = 0
input_image_size = 512
multi_head = False
#must use at the same time
#use mlp layer after head
cls_mlp = False
#load the params to head2
load_head = False
#use selayer before head
selayer = False
#load the params to head2
load_head = False
#use ttf head in centernet head
use_ttf = False
pre_model_dir = None
use_multi_scale = False
multi_scale_range = [0.6, 1.4]
stride = 4
train_dataset = CocoDetection(image_root_dir=train_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="train2017",
transform=transforms.Compose([
RandomFlip(flip_prob=0.5),
RandomCrop(crop_prob=0.5),
RandomTranslate(translate_prob=0.5),
Normalize(),
]))
val_dataset = CocoDetection(image_root_dir=val_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="val2017",
transform=transforms.Compose([
Normalize(),
Resize(resize=input_image_size),
]))
epochs = 140
milestones = [90, 120]
per_node_batch_size = 8
lr = 5e-4
num_workers = 16
print_interval = 100
apex = True
sync_bn = False
class Config(object):
version = 10
log = './log_ctn_multi_v' + str(version) # Path to save log
checkpoint_path = './checkpoints_multi_v' + str(version) # Path to store checkpoint model
resume = checkpoint_path + '/latest.pth' # load checkpoint model
evaluate = None # evaluate model path
train_dataset_path = os.path.join('/home/jovyan/data-vol-polefs-1/dataset/taibiao', 'images/train2017')
val_dataset_path = os.path.join('/home/jovyan/data-vol-polefs-1/dataset/taibiao', 'images/val2017')
dataset_annotations_path = os.path.join('/home/jovyan/data-vol-polefs-1/dataset/taibiao', 'annotations')
# train_dataset_path = os.path.join('/home/jovyan/data-vol-polefs-1/dataset/lanmu', 'images/train2017')
# val_dataset_path = os.path.join('/home/jovyan/data-vol-polefs-1/dataset/lanmu', 'images/val2017')
# dataset_annotations_path = os.path.join('/home/jovyan/data-vol-polefs-1/dataset/lanmu', 'annotations')
# train_dataset_path = os.path.join('/home/jovyan/data-vol-polefs-2/datasets_shenhe/scripts/03-08', 'crop_images')
# val_dataset_path = os.path.join('/home/jovyan/data-vol-polefs-2/datasets_shenhe/scripts/03-08', 'train2017')
# dataset_annotations_path = os.path.join('/home/jovyan/data-vol-polefs-2/datasets_shenhe/scripts/03-08', 'annotations')
network = "resnet50_centernet"
pretrained = False
#*********************************************************************************
multi_head = True
#must use at the same time
#use mlp layer after head
cls_mlp = False
#load the params to head2
load_head = False
#use selayer before head
selayer = True
#use ttf head in centernet head
use_ttf = True
# pre_model_dir = '/home/jovyan/data-vol-1/zhangze/code/pytorch-ImageNet-CIFAR-COCO-VOC-training/detection_experiments/resnet18_centernet_coco_distributed_apex_resize512/checkpoints/best.pth'
pre_model_dir = "/home/jovyan/data-vol-1/zhangze/code/multi_task/train/checkpoints_multi_v2/best.pth"
#head1 head2 classes
num_classes = [41, 1]
#*********************************************************************************
seed = 0
input_image_size = 512
use_multi_scale = False
multi_scale_range = [0.6, 1.4]
stride = 4
train_dataset = CocoDetection(image_root_dir=train_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="train2017",
transform=transforms.Compose([
RandomFlip(flip_prob=0.5),
RandomCrop(crop_prob=0.5),
RandomTranslate(translate_prob=0.5),
Normalize(),
]))
val_dataset = CocoDetection(image_root_dir=val_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="val2017",
transform=transforms.Compose([
Normalize(),
Resize(resize=input_image_size),
]))
epochs = 140
milestones = [90, 120]
per_node_batch_size = 16
lr = 5e-4
# lr = 1e-5
num_workers = 4
print_interval = 100
apex = True
sync_bn = False
class Config(object):
version = 1
log = './multi_task/log_' + str(version) # Path to save log
checkpoint_path = '/home/jovyan/data-vol-polefs-1/small_sample/multi_task/checkpoints/v{}'.format(version) # Path to store checkpoint model
resume = '/home/jovyan/data-vol-polefs-1/small_sample/multi_task/checkpoints/v{}/latest.pth'.format(version) # load checkpoint model
pre_model_dir = '/home/jovyan/data-vol-polefs-1/small_sample/multi_task/checkpoints/v1/best.pth'.format(version-1)
evaluate = None # evaluate model path
base_path = '/home/jovyan/data-vol-polefs-1/small_sample/dataset'
train_dataset_path = os.path.join(base_path, 'images/images')
val_dataset_path = os.path.join(base_path, 'images/images')
dataset_annotations_path = os.path.join("/home/jovyan/data-vol-polefs-1/small_sample/multi_task", 'annotations/v{}'.format(version))
network = "resnet50_yolof"
seed = 0
#resize
input_image_size = 667
num_classes = 1
train_dataset = CocoDetection(image_root_dir=train_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="train",
transform=transforms.Compose([
RandomFlip(flip_prob=0.5),
RandomCrop(crop_prob=0.5),
RandomTranslate(translate_prob=0.5),
Normalize(),
Resize(resize=input_image_size)
]))
val_dataset = CocoDetection(image_root_dir=val_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="val",
transform=transforms.Compose([
Normalize(),
Resize(resize=input_image_size),
]))
#***********************************************************#
#use the pretrained backbone
pretrained=True
#freeze the backbone and neck
# freeze = False
#load the previous model to train,if v1:use the coco_pretrained,else use the latest version model
# previous = False
#train multi task head
multi_task = False
#***********************************************************#
#fpn
#use yolof neck
use_yolof = True
#fpn encode channels
fpn_out_channels=512
use_p5=True
#***********************************************************#
#head
class_num=1
#use gn in head
use_GN_head=False
prior=0.01
cnt_on_reg=True
#yolof
yolof_encoder_channels = 512
#ttf up
use_ttf = True
ttf_out_channels = [256, 128]
#
dubble_run = True
#***********************************************************#
#training
epochs = 24
per_node_batch_size = 4
lr = 1e-4
num_workers = 4
print_interval = 100
eval_interval = 4
apex = True
sync_bn = False
#down sample strides
strides=[16]
#limit in decoder
limit_range=[[-1,99999]]
# limit_range=[[-1,667]]
#scales parameter in head
scales = [1.0]
#***********************************************************#
#inference
score_threshold=0.05
nms_iou_threshold=0.6
max_detection_num=1000
class Config(object):
log = './log_obj365' # Path to save log
checkpoint_path = './checkpoints_obj365' # Path to store checkpoint model
resume = './checkpoints_obj365/latest.pth' # load checkpoint model
evaluate = None # evaluate model path
train_dataset_path = os.path.join(
'/home/jovyan/data-vol-polefs-1/small_sample/obj365',
'images/train2017')
# val_dataset_path = os.path.join('/home/jovyan/data-vol-polefs-1/small_sample/obj365', 'images/val2017')
val_dataset_path = os.path.join(
'/home/jovyan/data-vol-polefs-1/dataset/coco', 'val2017')
dataset_annotations_path = os.path.join(
'/home/jovyan/data-vol-polefs-1/small_sample/obj365', 'annotations')
network = "resnet50_centernet"
pretrained = False
num_classes = [365]
seed = 0
input_image_size = 512
max_object_num = 700
multi_head = False
#must use at the same time
#use mlp layer after head
cls_mlp = False
#load the params to head2
load_head = False
#use selayer before head
selayer = False
#load the params to head2
load_head = False
#use ttf head in centernet head
use_ttf = False
pre_model_dir = None
use_multi_scale = False
multi_scale_range = [0.6, 1.4]
stride = 4
train_dataset = CocoDetection2(
data_dir=
'/home/jovyan/data-vol-1/zhangze/code/CenterNet-better/datasets',
split='train',
split_ratio=1.0,
img_size=input_image_size)
val_dataset = CocoDetection(image_root_dir=val_dataset_path,
annotation_root_dir=dataset_annotations_path,
set="val2017",
transform=transforms.Compose([
Normalize(),
Resize(resize=input_image_size),
]))
epochs = 140
milestones = [90, 120]
per_node_batch_size = 8
lr = 5e-3
num_workers = 8
print_interval = 100
apex = False
sync_bn = False