Model Post-processing Parameters

The following tables list the configuration parameters required by each model.

**Table 1** Parameters for post-processing of the YOLOv3 model (**yolov3_tf_bs1_fp16.cfg**)
Parameter	Description	Default Value	Value Range
CLASS_NUM	Number of classes.	80	None
BIASES_NUM	Number of anchor widths and heights (18 indicates nine anchors, and each anchor corresponds to a pair of widths and heights).	18	[0, 100]
BIASE	Every two numbers form the width and height of an anchor. For example, 10, 13 indicates the width and height of the first anchor.	10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326	None
SCORE_THRESH	Threshold for determining whether an object is of a certain class. If the value is greater than the threshold, the object is considered as the class.	0.3	[0.0, 1.0]
OBJECTNESS_THRESH	Threshold for determining whether it is an object. If the value is greater than the threshold, it is an object.	0.3	[0.0, 1.0]
IOU_THRESH	IOU threshold of two enclosures. If the threshold is exceeded, the two frames are considered as one frame.	0.45	[0.0, 1.0]
YOLO_TYPE	Number of output tensors. 3 indicates that three feature maps are output.	3	[0, 16]
ANCHOR_DIM	Number of anchor frames corresponding to each feature map.	3	[0, 16]
MODEL_TYPE	Data layout format. 0 indicates NHWC, 1 indicates NCHW, and 2 indicates NCHWC.	0	None
FRAMEWORK	The value is of the string type. It can be MindSpore, PyTorch, TensorFlow, or Caffe.	TensorFlow	None
SEPARATE_SCORE_THRESH	Threshold corresponding to each class.	The number of SCORE_THRESH (threshold) is determined by CLASS_NUM (quantity). The threshold is separated by commas (,). (The number of thresholds is equal to the value of CLASS_NUM.)	None

**Table 2** Parameters for post-processing of the ResNet-50 model (**resnet50_aipp_tf.cfg**)
Parameter	Description	Default Value	Value Range
CLASS_NUM	Number of classes.	1001	[0, 2000]
SOFTMAX	Indicates whether to perform softmax calculation in postprocessing. It is of the Boolean type.	false	None
TOP_K	Top K classes with the highest possibility.	1	[0, 16]

**Table 3** Parameters for post-processing of the Faster R-CNN model (**faster_rcnn_uncut.cfg**)
Parameter	Description	Default Value	Value Range
CLASS_NUM	Number of classes.	91	[0, 1000]
SCORE_THRESH	Threshold for determining whether an object is of a certain class. If the value is greater than the threshold, the object is considered as the class.	0.5	[0.0, 1.0]
IOU_THRESH	IOU threshold of two enclosures. If the threshold is exceeded, the two frames are considered as one frame.	0.45	[0.0, 1.0]
SEPARATE_SCORE_THRESH	Threshold corresponding to each class.	The number of SCORE_THRESH (threshold) is determined by CLASS_NUM (quantity). The threshold is separated by commas (,). (The number of thresholds is equal to the value of CLASS_NUM.)	None
MODEL_TYPE	The options are as follows: 0: original 1: nms_cut (No non-maximum value suppression is performed on the model.) 2: FPN	0	None
FRAMEWORK	The value can be: TensorFlow MindSpore PyTorch	TensorFlow	None
NMS_FINISHED	Attribute value specific to the old framework, of the Boolean type. false: The model does not contain the NMS operator and NMS operations need to be performed in postprocessing. true: The model contains the NMS operator and NMS operations do not need to be performed in postprocessing.	true	None
Note: The MODEL_TYPE and FRAMEWORK parameters are described as follows: original adapts to the TensorFlow framework. nms_cut adapts to the TensorFlow or MindSpore framework. FPN adapts to the PyTorch framework.

**Table 4** Parameters of the SSD_VGG model post-processing (**ssd_vgg16_caffe_release.cfg**)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	5
SCORE_THRESH	Target threshold.	0.4
SEPARATE_SCORE_THRESH	Threshold corresponding to each class.	The number of SCORE_THRESH (threshold) is determined by CLASS_NUM (quantity). The threshold is separated by commas (,). (The number of thresholds is equal to the value of CLASS_NUM.)

**Table 5** Parameters for postprocessing of the SSD-MobileNet-v1-FPN model (**ssd_mobilenetv1_fpn.cfg**)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	3
SCORE_THRESH	Threshold for determining whether an object is of a certain class. If the value is greater than the threshold, the object is considered as the class.	0.5
SEPARATE_SCORE_THRESH	Threshold corresponding to each class.	The number of SCORE_THRESH (threshold) is determined by CLASS_NUM (quantity). The threshold is separated by commas (,). (The number of thresholds is equal to the value of CLASS_NUM.)

**Table 6** Parameters for postprocessing of the CRNN model (**crnn_ssh_2.cfg**)
Parameter	Description	Default Value	Value Range
CLASS_NUM	Number of classes.	0	[0, 10000]
OBJECT_NUM	Maximum number of characters that can be detected.	0	[0, 1000]
BLANK_INDEX	Index value of a blank character.	0	[0, 10000]
WITH_ARGMAX	Whether argmax has been performed on the model backbone.	false	None

**Table 7** **(Old framework)** Parameters of the ResNet feature model post-processing (**resnet_feature_caffe_release.cfg**)
Parameter	Description	Default Value
ACTIVATION_FUNCTION	Activation function used to activate the model output data.	None

**Table 8** **(Old framework)** Parameters of post-processing of the ResNet model with multiple classification properties (**resnet_attribute_caffe_release.cfg**)
Parameter	Description	Default Value
ATTRIBUTE_NUM	Number of model output properties.	5
ACTIVATION_FUNCTION	Type of the activation function. Currently, only the sigmoid function is supported.	None
ATTRIBUTE_INDEX	Index of a model output property. Ensure that the number of indexes is the same as the value of ATTRIBUTE_NUM.	None

**Table 9** **(Old framework)** Parameters of post-processing of the ResNet model with binary property classification (**resnet_attribute_caffe_release.cfg**)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	5

**Table 10** **(Old framework)** Parameters of YOLOv4 model post-processing (**yolov4_pt_bs1_fp16.cfg**)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	80
BIASES_NUM	Number of anchor widths and heights (18 indicates nine anchors, and each anchor corresponds to a pair of widths and heights).	18
BIASES	Every two numbers form the width and height of an anchor. For example, 10, 13 indicates the width and height of the first anchor.	10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326
SCORE_THRESH	Threshold for determining whether an object is of a certain class. If the value is greater than the threshold, the object is considered as the class.	0.3
OBJECTNESS_THRESH	Threshold for determining whether it is an object. If the value is greater than the threshold, it is an object.	0.3
IOU_THRESH	IOU threshold of two enclosures. If the threshold is exceeded, the two frames are considered as one frame.	0.45
YOLO_TYPE	Number of output tensors. 3 indicates that three feature maps are output.	3
ANCHOR_DIM	Number of anchor frames corresponding to each feature map.	3
MODEL_TYPE	Data layout format. 0 indicates NHWC, and 1 indicates NCHW.	0
FRAMEWORK_TYPE	Model framework. The value 0 indicates PyTorch, and the value 1 indicates MindSpore.	0
SEPARATE_SCORE_THRESH	Threshold corresponding to each class.	The number of SCORE_THRESH (threshold) is determined by CLASS_NUM (quantity). The threshold is separated by commas (,). (The number of thresholds is equal to the value of CLASS_NUM.)

**Table 11** Parameters of the YOLOv4 model post-processing (**yolov4_pt_bs1_fp16.cfg**)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	80
BIASES_NUM	Number of anchor widths and heights (18 indicates nine anchors, and each anchor corresponds to a pair of widths and heights).	18
BIASES	Every two numbers form the width and height of an anchor. For example, 10, 13 indicates the width and height of the first anchor.	10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326
SCORE_THRESH	Threshold for determining whether an object is of a certain class. If the value is greater than the threshold, the object is considered as the class.	0.3
OBJECTNESS_THRESH	Threshold for determining whether it is an object. If the value is greater than the threshold, it is an object.	0.3
IOU_THRESH	IOU threshold of two enclosures. If the threshold is exceeded, the two frames are considered as one frame.	0.45
YOLO_TYPE	Number of output tensors. 3 indicates that three feature maps are output.	3
ANCHOR_DIM	Number of anchor frames corresponding to each feature map.	3
MODEL_TYPE	Data layout format. 0 indicates NHWC, 1 indicates NCHW, and 2 indicates NCHWC.	0
FRAMEWORK	The value is of the string type. It can be MindSpore, PyTorch, TensorFlow, or Caffe.	MindSpore
SEPARATE_SCORE_THRESH	Threshold corresponding to each class.	The number of SCORE_THRESH (threshold) is determined by CLASS_NUM (quantity). The threshold is separated by commas (,). (The number of thresholds is equal to the value of CLASS_NUM.)
YOLO_VERSION	YOLO model version in use.	(Mandatory) YOLO_VERSION = 4

**Table 12** Parameters of the YOLOv5 model post-processing (**yolov5_pt_bs1_fp32.cfg**)
Parameter	Description	Default Value	Value Range
CLASS_NUM	Number of classes.	80	[0, 1000]
BIASES_NUM	Number of anchor widths and heights (18 indicates nine anchors, and each anchor corresponds to a pair of widths and heights).	18	[0, 1000]
BIASE	Every two numbers form the width and height of an anchor. For example, 10, 13 indicates the width and height of the first anchor.	10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326	None
SCORE_THRESH	Threshold for determining whether an object is of a certain class. If the value is greater than the threshold, the object is considered as the class.	0.3	[0.0, 1.0]
OBJECTNESS_THRESH	Threshold for determining whether it is an object. If the value is greater than the threshold, it is an object.	0.3	[0.0, 1.0]
IOU_THRESH	IOU threshold of two enclosures. If the threshold is exceeded, the two frames are considered as one frame.	0.45	[0.0, 1.0]
YOLO_TYPE	Number of output tensors. 3 indicates that three feature maps are output.	3	[0, 1000]
ANCHOR_DIM	Number of anchor frames corresponding to each feature map.	3	[0, 1000]
MODEL_TYPE	Data layout format. 0 indicates NHWC, 1 indicates NCHW, and 2 indicates NCHWC (Currently, only models in PyTorch framework are supported.).	2	[0, 1000]
FRAMEWORK	The value is of the string type. It can be MindSpore, PyTorch, TensorFlow, or Caffe.	PyTorch	None
SEPARATE_SCORE_THRESH	Threshold corresponding to each class.	The number of SCORE_THRESH (threshold) is determined by CLASS_NUM (quantity). The threshold is separated by commas (,). (The number of thresholds is equal to the value of CLASS_NUM.)	None
YOLO_VERSION	YOLO model version in use.	(Mandatory) YOLO_VERSION=5	None

**Table 13** **(Old framework)** Parameters of the FasterRCNN-Fpn or CascadeRCNN-Fpn model post-processing (**fasterrcnn.cfg** or **cascadercnn.cfg**)
Parameter	Description	Default Value
SCORE_THRESH	Threshold for determining whether an object is of a certain class. If the value is greater than the threshold, the object is considered as the class. Value 0.3 is used in the test.	0.5
FPN_SWITCH	FPN switch. This parameter must be set to true for both models.	false

**Table 14** Parameters for postprocessing of the DeepLabV3+ (TensorFlow) model (**deeplabv3.cfg**)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	21
FRAMEWORK_TYPE	Deep learning framework type.	0: TensorFlow framework

**Table 15** Parameters for postprocessing of the CTPN model (**ctpn_tf.cfg**)
Parameter	Description	Default Value	Value Range
IS_ORIENTED	Indicates whether to detect box tilt.	false	None
BOX_IOU_THRESH	IOU threshold of the small box.	0.7	[0.0, 1.0]
TEXT_IOU_THRESH	IOU threshold of the final text box.	0.2	[0.0, 1.0]
TEXT_PROPOSALS_MIN_SCORE	Minimum score for filtering the small box.	0.7	[0.0, 1.0]
LINE_MIN_SCORE	Minimum score for filtering the final text box.	0.9	[0.0, 1.0]
IS_MINDSPORE	Whether the MindSpore framework is used.	false	None

**Table 16** Parameters for postprocessing of the CTPN model (**ctpn_mindspore.cfg**)
Parameter	Description	Default Value	Value Range
IS_ORIENTED	Indicates whether to detect box tilt.	false	None
BOX_IOU_THRESH	IOU threshold of the small box.	0.7	[0.0, 1.0]
TEXT_IOU_THRESH	IOU threshold of the final text box.	0.2	[0.0, 1.0]
TEXT_PROPOSALS_MIN_SCORE	Minimum score for filtering the small box.	0.7	[0.0, 1.0]
LINE_MIN_SCORE	Minimum score for filtering the final text box.	0.9	[0.0, 1.0]
IS_MINDSPORE	Whether the MindSpore framework is used.	true	None

**Table 17** Parameters for postprocessing of the ResNet-18 model (**resnet18_aipp_tf.cfg**)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	2

**Table 18** Parameters for postprocessing of the DeepLabv3 (MindSpore) model (**deeplabv3.cfg**)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	21
MODEL_TYPE	Layout of the model inference output data. The value 0 indicates NHWC, and the value 1 indicates NCHW.	1
FRAMEWORK_TYPE	Deep learning framework type.	2: MindSpore framework

**Table 19** Parameters for postprocessing of the BERT-Base (Uncased) model (bert.cfg)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	2
CHECK_MODEL	Checks model compatibility.	false

**Table 20** Parameters for postprocessing of the DeepLabV3+ (PyTorch) model (**deeplabv3.cfg**)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	21
CHECK_MODEL	Checks model compatibility.	true
MODEL_TYPE	Layout of the model inference output data. The value 0 indicates NHWC, and the value 1 indicates NCHW.	1
FRAMEWORK_TYPE	Deep learning framework type.	Select 1 for the PyTorch framework.

**Table 21** Parameters for postprocessing of the U-Net model (**unet_simple.cfg**)
Parameter	Description	Default Value
CLASS_NUM	Number of classes.	2
POST_TYPE	Model post-processing mode. The value 0 indicates that argmax is performed on the model logits output (NHWC type). The value 1 indicates that the model argmax output (NHW type) is transparently transmitted.	1
RESIZE_TYPE	Interpolation restoration mode of the pixel image. Currently, only the following two modes are supported: 0: Interpolation restoration is not performed. 1: The nearest neighbor interpolation restoration is performed.	1

**Table 22** Parameters for postprocessing of the Mask R-CNN model (**mask_rcnn_2017.cfg**)
Parameter	Description	Default Value	Value Range
CLASS_NUM	Total number of inference categories, excluding the background.	80	[0, 100]
SCORE_THRESH	Confidence score threshold, which can be adjusted based on service scenarios.	0.7	[0.0, 1.0]
IOU_THRESH	IOU threshold, which can be adjusted based on service scenarios.	0.5	[0.0, 1.0]
RPN_MAX_NUM	Maximum number of region proposal networks.	1000	[0, 1000]
MAX_PER_IMG	Maximum value of the prediction box of each image sorted by confidence.	128	[0, 150]
MASK_THREAD_BINARY	Mask threshold of the input RCNN.	0.5	[0.0, 1.0]
MASK_SHAPE_SIZE	Mask shape in mask_rcnn. Only a single parameter is supported to indicate a square.	28	[0, 100]
MODEL_TYPE	The options are as follows: 0: MindSpore 1: PyTorch	0	None
SEPARATE_SCORE_THRESH	Threshold corresponding to each class.	The number of SCORE_THRESH (threshold) is determined by CLASS_NUM (quantity). The threshold is separated by commas (,). (The number of thresholds is equal to the value of CLASS_NUM.)	None

**Table 23** Parameters for postprocessing of the Ssd_Mobilenet_v1_Fpn_for_MindSpore model (**ssd_mobilenetv1_fpn.cfg**)
Parameter	Description	Default Value	Value Range
CLASS_NUM	Number of classes.	81	[0, 100]
SCORE_THRESH	Threshold for determining whether an object is of a certain class. If the value is greater than the threshold, the object is considered as the class.	0.5	[0.0, 1.0]
IOU_THRESH	Threshold of the object overlap degree. If the value is greater than the threshold, two object frames correspond to the same object.	0.6	[0.0, 1.0]
SEPARATE_SCORE_THRESH	Threshold corresponding to each class.	The number of SCORE_THRESH (threshold) is determined by CLASS_NUM (quantity). The threshold is separated by commas (,). (The number of thresholds is equal to the value of CLASS_NUM.)	None

**Table 24** Parameters for postprocessing of the OpenPose model (**openpose.cfg**)
Parameter	Description	Default Value	Value Range
KEYPOINT_NUM	Number of key points, including the background (a background is counted as one).	19	[0, 100]
FILTER_SIZE	Length (or width) of the Gaussian filter kernel.	25	[0, 100]
SIGMA	Variance of the Gaussian filter kernel.	3	[0, 10]

**Table 25** Parameters for postprocessing the HigherHRnet model (**higherhrnet.cfg**)
Parameter	Description	Default Value	Value Range
KEYPOINT_NUM	Number of key points.	17	[0, 20]
SCORE_THRESH	Key point threshold.	0.1	[0.0, 1.0]

**Table 26** Parameters for postprocessing of the Unet++ model (**unet_nested.cfg**)
Parameter	Description	Default Value	Value Range
CLASS_NUM	Number of classes.	3	None
POST_TYPE	Model post-processing mode. The value 0 indicates that argmax is performed on the model logits output (NHWC type). The value 1 indicates that the model argmax output (NHW type) is transparently transmitted.	1	[0, 16]
RESIZE_TYPE	Interpolation restoration mode of the pixel image. Currently, only the following two modes are supported: 0: Interpolation restoration is not performed. 1: The nearest neighbor interpolation restoration is performed.	1	[0, 16]

**Table 27** Parameters for postprocessing of RetinaNet model (**retinanet_tf.cfg**)
Parameter	Description	Default Value	Value Range
CLASS_NUM	Number of dataset types. (The default number of COCO datasets is 80.)	80	[0, 100]
MODEL_TYPE	Model type. Currently, only the TensorFlow model is supported.	0	[0, 100]
SCORE_THRESH	Score threshold.	0.5	[0.0, 1.0]

Parent topic: Existing Supported Models