class CenterNetCfg:
center_net: CenterNet = field(default_factory=CenterNet)
train: TrainCfg = field(default_factory=TrainCfg)
valid: ValidCfg = field(default_factory=ValidCfg)
gpus: Union[Tuple,
list] = (0, 1, 2, 3
) # gpu individual ids, not necessarily consecutive
class FasterRCNNCfg:
faster_rcnn: FasterRCNN = field(default_factory=FasterRCNN)
train: TrainCfg = field(default_factory=TrainCfg)
valid: ValidCfg = field(default_factory=ValidCfg)
# Dataset name. eg. 'coco', 'voc', 'voc_tiny'
dataset: str = 'voc_tiny'
# Path of the directory where the dataset is located.
dataset_root: str = '~/.mxnet/datasets/'
# Training with GPUs, you can specify (1,3) for example.
gpus: Union[Tuple, list] = (0, 1, 2, 3)
# Resume from previously saved parameters if not None.
# For example, you can resume from ./faster_rcnn_xxx_0123.params.
resume: str = ''
# Saving parameter prefix
save_prefix: str = ''
# Saving parameters epoch interval, best model will always be saved.
save_interval: int = 1
# Use MXNet Horovod for distributed training. Must be run with OpenMPI.
horovod: bool = False
# Number of data workers, you can use larger number to accelerate data loading,
# if your CPU and GPUs are powerful.
num_workers: int = 16
# KV store options. local, device, nccl, dist_sync, dist_device_sync,
# dist_async are available.
kv_store: str = 'nccl'
# Whether to disable hybridize the model. Memory usage and speed will decrese.
disable_hybridization: bool = False
class MaskRCNNCfg:
mask_rcnn: MaskRCNN = field(default_factory=MaskRCNN)
train: TrainCfg = field(default_factory=TrainCfg)
valid: ValidCfg = field(
default_factory=ValidCfg) # Dataset name. eg. 'coco', 'voc'
dataset = 'coco'
# Training with GPUs, you can specify (1,3) for example.
gpus = (0, )
# Resume from previously saved parameters if not None.
# For example, you can resume from ./faster_rcnn_xxx_0123.params.
resume = ''
# Saving parameter prefix
save_prefix = ''
# Saving parameters epoch interval, best model will always be saved.
save_interval = 1
# Use MXNet Horovod for distributed training. Must be run with OpenMPI.
horovod = False
# Number of data workers, you can use larger number to accelerate data loading,
# if your CPU and GPUs are powerful.
num_workers = 16
# KV store options. local, device, nccl, dist_sync, dist_device_sync,
# dist_async are available.
kv_store = 'nccl'
# Whether to disable hybridize the model. Memory usage and speed will decrese.
disable_hybridization = False
# Use NVIDIA MSCOCO API. Make sure you install first.
use_ext = False
class ImageClassificationCfg:
img_cls: ImageClassification = field(default_factory=ImageClassification)
train: TrainCfg = field(default_factory=TrainCfg)
valid: ValidCfg = field(default_factory=ValidCfg)
gpus: Union[Tuple,
list] = (0,
) # gpu individual ids, not necessarily consecutive
class TorchImageClassificationCfg:
img_cls : ImageClassification = field(default_factory=ImageClassification)
data: DataCfg = field(default_factory=DataCfg)
optimizer: OptimizerCfg = field(default_factory=OptimizerCfg)
train: TrainCfg = field(default_factory=TrainCfg)
augmentation: AugmentationCfg = field(default_factory=AugmentationCfg)
model_ema: ModelEMACfg = field(default_factory=ModelEMACfg)
misc: MiscCfg = field(default_factory=MiscCfg)
gpus : Union[Tuple, list] = (0, ) # gpu individual ids, not necessarily consecutive
class SSDCfg:
ssd: SSD = field(default_factory=SSD)
train: TrainCfg = field(default_factory=TrainCfg)
valid: ValidCfg = field(default_factory=ValidCfg)
# Dataset name. eg. 'coco', 'voc', 'voc_tiny'
dataset: str = 'voc_tiny'
# Path of the directory where the dataset is located.
dataset_root: str = '~/.mxnet/datasets/'
# Training with GPUs, you can specify (1,3) for example.
gpus: Union[Tuple, list] = (0, 1, 2, 3)
# Resume from previously saved parameters if not None.
# For example, you can resume from ./faster_rcnn_xxx_0123.params.
resume: str = ''
# Saving parameters epoch interval, best model will always be saved.
save_interval: int = 1
# Use MXNet Horovod for distributed training. Must be run with OpenMPI.
horovod: bool = False
# Number of data workers, you can use larger number to accelerate data loading,
# if your CPU and GPUs are powerful.
num_workers: int = 4
class CenterNet:
base_network: str = 'dla34_deconv' # base feature network
heads: CenterNetHead = field(default_factory=CenterNetHead)
scale: float = 4.0 # output vs input scaling ratio, e.g., input_h // feature_h
topk: int = 100 # topk detection results will be kept after inference
root: str = os.path.expanduser(os.path.join(
'~', '.mxnet', 'models')) # model zoo root dir
wh_weight: float = 0.1 # Loss weight for width/height
center_reg_weight: float = 1.0 # Center regression loss weight
data_shape: Tuple[int, int] = (512, 512)
# use the pre-trained detector for transfer learning(use preset, ignore other network settings)
transfer: str = 'center_net_resnet50_v1b_coco'