aclmdlDesc

aclmdlDesc* aclmdlCreateDesc()

desc = acl.mdl.create_desc()

Creates data of the aclmdlDesc type.

None

desc: int, address of an object of the aclmdlDesc type.

None

None

aclError aclmdlDestroyDesc(aclmdlDesc *modelDesc)

ret = acl.mdl.destroy_desc(model_desc)

Destroys data of the aclmdlDesc type.

model_desc: pointer to the aclmdlDesc to be destroyed.

ret: int, error code.

• 0 indicates success.
• Other values indicate failure.

Destruction API: The released or destroyed resources are no longer available after this API is called. You are advised to set the released or destroyed resources to invalid values (for example, 0 or None) after calling this API.

None

aclError aclmdlGetDesc(aclmdlDesc *modelDesc, uint32_t modelId)

ret = acl.mdl.get_desc(model_desc, model_id)

Obtains model data of the aclmdlDesc type using the model ID.

model_desc: int, pointer to the data of the aclmdlDesc type.

model_id: int, model ID.

ret: int, error code.

• 0 indicates success.
• Other values indicate failure.

None

None

size_t aclmdlGetNumInputs(aclmdlDesc *modelDesc)

size = acl.mdl.get_num_inputs(model_desc)

Obtains the number of input models based on data of the aclmdlDesc type.

model_desc: int, pointer to the data of the aclmdlDesc type.

size: int, number of input models.

None

None

size_t aclmdlGetNumOutputs(aclmdlDesc *modelDesc)

size = acl.mdl.get_num_outputs(model_desc)

Obtains the number of output models based on data of the aclmdlDesc type.

model_desc: int, pointer to the data of the aclmdlDesc type.

size: int, number of input models.

None

None

size_t aclmdlGetInputSizeByIndex(aclmdlDesc *modelDesc, size_t index)

size = acl.mdl.get_input_size_by_index(model_desc, index)

Obtains the input size in bytes based on data of the aclmdlDesc type.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input to be obtained, starting from 0.

size: int, size of the input, in bytes.

None

None

size_t aclmdlGetOutputSizeByIndex(aclmdlDesc *modelDesc, size_t index)

size = acl.mdl.get_output_size_by_index(model_desc, index)

Obtains the output size in bytes based on data of the aclmdlDesc type.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input to be obtained, starting from 0.

size: int, size of the input, in bytes.

None

None

aclError aclmdlGetInputDims(const aclmdlDesc *modelDesc, size_t index, aclmdlIODims *dims)

dims, ret = acl.mdl.get_input_dims(model_desc, index)

Obtains dimension information of the input tensor based on the model description.

If the model contains static AIPP configuration information, call acl.mdl.get_input_dims or acl.mdl.get.input_dims_v2 to obtain the dimension information as required. The differences between the two interfaces are as follows:

• acl.mdl.get_input_dims: The obtained dimension values are consistent with those of the input image. For details, see Table 9-1.

• acl.mdl.get_input_dims_v2: The obtained value of the H dimension is that of the input image multiplied by a coefficient, and the value of the C dimension varies according to the image format. For details, see Table 9-2 in acl.mdl.get_input_dims_v2. The multiplied dimension values are consistent with those obtained by calling acl.mdl.get_input_size_by_index.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input whose Dims is to be obtained, starting from 0.

dims: dict, maximum supported combinations for dynamic-batch or dynamic-size scenarios based on the dictionary array.

• dims = {

  "name": xxx, #tensor name

  "dimCount": xxx, #Number of dimensions in shape

  "dims": [xx, xx, xx] # dimension information

  }

  In dynamic batch and dynamic resolution scenarios, if the number of batches, or the width and height of the input tensor are –1, the tensor is dynamic. For example, if the format of the input tensor is NCHW, then the tensor dimension [–1, 3, 224, 224] indicates the batch is dynamic, and the tensor dimension [1, 3, –1, –1] indicates the resolution is dynamic. The italic numbers are examples only.

ret: int, error code.

• 0 indicates success.
• Other values indicate failure.

None

None

Image Format

dims Format

dims Shape

YUV420SP_U8

NHWC

n,h,w,c

XRGB8888_U8

NHWC

n,h,w,c

RGB888_U8

NHWC

n,h,w,c

YUV400_U8

NHWC

n,h,w,c

ARGB8888_U8

NHWC

n,h,w,c

YUYV_U8

NHWC

n,h,w,c

YUV422SP_U8

NHWC

n,h,w,c

AYUV444_U8

NHWC

n,h,w,c

aclError aclmdlGetInputDimsV2(const aclmdlDesc *modelDesc, size_t index, aclmdlIODims *dims)

dims, ret = acl.mdl.get_input_dims(model_desc, index)

Obtains dimension information of the input tensor based on the model description.

If the model contains static AIPP configuration information, call acl.mdl.get_input_dims or acl.mdl.get.input_dims_v2 to obtain the dimension information as required. The differences between the two interfaces are as follows:

• acl.mdl.get_input_dims: The obtained dimension values are consistent with those of the input image. For details, see Table 9-1.

• acl.mdl.get_input_dims_v2: The obtained value of the H dimension is that of the input image multiplied by a coefficient, and the value of the C dimension varies according to the image format. For details, see Table 9-2 in acl.mdl.get_input_dims_v2. The multiplied dimension values are consistent with those obtained by calling aclmdlGetInputSizeByIndex.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input to be obtained, starting from 0.

dims: dict, maximum supported combinations for dynamic-batch or dynamic-size scenarios based on the dictionary array.

• dims = {

  "name": xxx, #tensor name

  "dimCount": xxx, #Number of dimensions in shape

  "dims": [xx, xx, xx] # dimension information

  }

  In dynamic batch and dynamic resolution scenarios, if the number of batches, or the width and height of the input tensor are –1, the tensor is dynamic. For example, if the format of the input tensor is NCHW, then the tensor dimension [–1,3,224,224] indicates the batch is dynamic, and the tensor dimension [1,3,–1,–1] indicates the resolution is dynamic. The italic numbers are examples only.

  In the static AIPP scenario, this API applies only to the NHWC format. Table 9-2 describes the mapping between the supported input image formats and dims shapes.

ret: int, error code.

• 0 indicates success.
• Other values indicate failure.

None

None

YUV420SP_U8

NHWC

"n,h*1.5,w,1"

XRGB8888_U8

NHWC

"n,h,w,4"

RGB888_U8

NHWC

"n,h,w,3"

YUV400_U8

NHWC

"n,h,w,1"

ARGB8888_U8

NHWC

"n,h,w,4"

YUYV_U8

NHWC

"n,h,w,2"

YUV422SP_U8

NHWC

"n,h*2,w,1"

AYUV444_U8

NHWC

"n,h,w,4"

aclError aclmdlGetOutputDims(const aclmdlDesc *modelDesc, size_t index, aclmdlIODims *dims)

acl.mdl.get_output_dims(model_desc, index)

Obtains tensor dimension information of an output based on the model description.

In the fixed shape scenario, this API is used to obtain the dimension information of the output tensor of a model.

This API is used to obtain the dimension information of the maximum batch/image size choice in the dynamic shape (dynamic batch or dynamic resolution) scenario.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input to be obtained, starting from 0.

dims: dict, maximum supported combinations for dynamic-batch or dynamic-size scenarios based on the dictionary array.

• dims = {

  "name": xxx, #tensor name

  "dimCount": xxx, #Number of dimensions in shape

  "dims": [xx, xx, xx] # dimension information

  }

ret: int, error code.

• 0 indicates success.
• Other values indicate failure.

None

None

const char *aclmdlGetInputNameByIndex(const aclmdlDesc *modelDesc, size_t index)

input_name = acl.mdl.get_input_name_by_index(model_desc, index)

Obtains the name of an input in a model.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input to be obtained, starting from 0.

input_name: str, input name.

None

None

const char *aclmdlGetOutputNameByIndex(const aclmdlDesc *modelDesc, size_t index)

output_name = acl.mdl.get_output_name_by_index(model_desc, index)

Obtains the name of an output in a model.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input to be obtained, starting from 0.

output_name: str, output name.

None

None

aclFormat aclmdlGetInputFormat(const aclmdlDesc *modelDesc, size_t index)

format_info = acl.mdl.get_input_format(model_desc, index)

Obtains the format of an input in a model.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input to be obtained, starting from 0.

format_info: int, format of the specified input.

• –1: ACL_FORMAT_UNDEFINED (unknown format, default value)

• 0: ACL_FORMAT_NCHW, in NCHW format
• 1: ACL_FORMAT_NHWC, in NHWC format
• 2: ACL_FORMAT_ND, which indicates that all formats are supported. Exercise cautions when using operators except for single-input operators that are processed inherently such as square and tanh.
• 3: ACL_FORMAT_NC1HWC0, Huawei proprietary 5D data format
• 4: ACL_FORMAT_FRACTAL_Z, convolution weight format

None

None

aclFormat aclmdlGetOutputFormat(const aclmdlDesc *modelDesc, size_t index)

format_info = acl.mdl.get_output_format(model_desc, index)

Obtains the format of an output in a model.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input to be obtained, starting from 0.

format_info: int, format of the specified input.

• –1: ACL_FORMAT_UNDEFINED (unknown format, default value)

• 0: ACL_FORMAT_NCHW, in NCHW format
• 1: ACL_FORMAT_NHWC, in NHWC format
• 2: ACL_FORMAT_ND, which indicates that all formats are supported. Exercise cautions when using operators except for single-input operators that are processed inherently such as square and tanh.
• 3: ACL_FORMAT_NC1HWC0, Huawei proprietary 5D data format
• 4: ACL_FORMAT_FRACTAL_Z, convolution weight format

None

None

aclDataType aclmdlGetInputDataType(const aclmdlDesc *modelDesc, size_t index)

data_type = acl.mdl.get_input_data_type(model_desc, index)

Obtains the data type of an input in a model.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input to be obtained, starting from 0.

data_type: int, data type of the specified input.

• –1: ACL_DT_UNDEFINED, unknown data type (default)
• 0: ACL_FLOAT
• 1: ACL_FLOAT16
• 2: ACL_INT8
• 3: ACL_INT32
• 4: ACL_UINT8
• 6: ACL_INT16
• 7: ACL_UINT16
• 8: ACL_UINT32
• 9: ACL_INT64
• 10: ACL_UINT64
• 11: ACL_DOUBLE
• 12: ACL_BOOL

None

None

aclDataType aclmdlGetOutputDataType(const aclmdlDesc *modelDesc, size_t index)

data_type = acl.mdl.get_output_data_type(model_desc, index)

Obtains the data type of an input in a model.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the input to be obtained, starting from 0.

ata_type: int, data type of the specified input.

• –1: ACL_DT_UNDEFINED, unknown data type (default)
• 0: ACL_FLOAT
• 1: ACL_FLOAT16
• 2: ACL_INT8
• 3: ACL_INT32
• 4: ACL_UINT8
• 6: ACL_INT16
• 7: ACL_UINT16
• 8: ACL_UINT32
• 9: ACL_INT64
• 10: ACL_UINT64
• 11: ACL_DOUBLE
• 12: ACL_BOOL

None

None

aclError aclmdlGetInputIndexByName(const aclmdlDesc *modelDesc, const char *name, size_t *index)

index, ret = acl.mdl.get_input_index_by_name(model_desc, name)

Obtains the index of an input using the input name in a model.

model_desc: int, pointer to the data of the aclmdlDesc type.

name: str, input name.

index: int, index of the input to be obtained.

ret: int, error code.

• 0 indicates success.
• Other values indicate failure.

None

For dynamic batch and image size, the input name is fixed to ascend_mbatch_shape_data. For dynamic AIPP, the input name is fixed to ascend_dynamic_aipp_data.

aclError aclmdlGetOutputIndexByName(const aclmdlDesc *modelDesc, const char *name, size_t *index)

index, ret = acl.mdl.get_output_index_by_name(model_desc, name)

Obtains the index of the output by the output name in the model.

model_desc: int, pointer to the data of the aclmdlDesc type.

name: str, input name.

index: int, index of the input to be obtained.

ret: int, error code.

• 0 indicates success.
• Other values indicate failure.

None

The name is fixed to data.

aclError aclmdlGetDynamicBatch(const aclmdlDesc *modelDesc, aclmdlBatch *batch)

batch, ret = acl.mdl.get_dynamic_batch(model_desc)

Obtains the dynamic batch size information supported by a model.

model_desc: int, pointer to the data of the aclmdlDesc type.

batch: dict, dynamic batch information supported by the obtained model

• batch = {

  "batchCount": int, # Number of batch choices supported by the model

  "batch": [xx, xx,] # Specific batch choices supported by the model

  }

  The maximum number of batches supported by the model is 128.

• If batchCount is 0, dynamic batch size is not supported. The batch size of the model applies.

ret: int, error code.

• 0 indicates success.
• Other values indicate failure.

None

None

aclError aclmdlGetDynamicHW(const aclmdlDesc *modelDesc, size_t index, aclmdlHW *hw)

hw_info, ret = acl.mdl.get_dynamic_hw(model_desc, index)

Obtains the dynamic size information supported by a model.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: Reserved. The value is fixed to –1.

hw_info: dict, width and height supported by the obtained model.

• hw_info= {

  "hwCount": int, # Number of image size choices supported by the model

  "hw": [xx, xx,] # Supported dynamic image size choices. In each array, element 0 indicates the height, and element 1 indicates the width.*/

  }

  The maximum number of HW is 128.

• If hwCount is 0, dynamic image size is not supported. The image size of the model applies.

ret: int, error code.

• 0 indicates success.
• Other values indicate failure.

None

None

aclErroraclmdlGetCurOutputDims(const aclmdlDesc *modelDesc, size_t index, aclIODims *dims)

dims, ret = acl.mdl.get_cur_output_dims(model_desc, index)

Obtains actual tensor dimension information of an output based on the model description.

model_desc: int, pointer to the data of the aclmdlDesc type.

index: int, sequence number of the output whose Dims is to be obtained, indexing starting from 0.

dims: dict, actual dimension information.

• dims= {

  "name": str, # Tensor name

  "dimCount": int, # Number of dimensions in the shape of a tensor

  "dims": [xx, xx,] # Shape dimension information of the tensor

  }

ret: int, error code.

• 0 indicates success.
• Other values indicate failure.

None

None

aclError aclmdlGetFirstAippInfo(uint32_t modelId, size_t index, aclAippInfo *aippinfo)

aipp_info, ret = acl.mdl.get_first_aipp_info(model_id, index)

Obtains the AIPP (static or dynamic) configuration of a model.

Static AIPP provides the following functions. They are executed in the exact sequence. For details, see ATC Tool Instructions.

Cropping -> CSC -> Mean subtraction/normalization -> Padding

model_id: int, model ID.

acl.mdl.load_from_file, acl.mdl.load_from_mem, acl.mdl.load_from_file_with_mem, and acl.mdl.load_from_mem_with_mem are called to load the model. After the model is successfully loaded, the model ID is returned.

index: int, sequence number of the specified input to the model

aipp_info: dict. Static AIPP configuration.

None

None