TensorDimension

The AIModelManager::GetModelIOTensorDim API can be called to obtain the information about the model input/output size, including the tensor dimension information, data type, buffer size, and tensor name. This data type is defined in ai_types.h.

/*
* Describe the dimensions of a tensor.
    */
    struct TensorDimension
    {
        uint32_t n; // Batch size
        uint32_t c; // Feature map channels
        uint32_t h; // Feature map height
        uint32_t w; // Feature map width
        uint32_t data_type;
        uint32_t size;
        std::string  name;
    };

If the parameters related to AIPP are not configured during model conversion, the information about the model input/output dimensions obtained by calling AIModelManager::GetModelIOTensorDim is as follows:

n, c, h, and w are those before model conversion.

The enumerated value of data_type is defined as follows:

enum DataType
{
DT_UNDEFINED = 0; // Used to indicate a DataType field has not been set.
DT_FLOAT = 1; // float type, 4 bytes
DT_FLOAT16 = 2; // fp16 type, 2 bytes
DT_INT8 = 3; // int8 type, 1 byte
DT_UINT8 = 4; // uint8 type, 1 byte
DT_INT16 = 5; // int16 type, 2 bytes
DT_UINT16 = 6; // uint16 type, 2 bytes
DT_INT32 = 7; // int32 type, 4 bytes
DT_INT64 = 8; // int64 type, 8 bytes
DT_UINT32 = 9; // unsigned int32, 4 bytes
DT_UINT64 = 10; // unsigned int64, 8 bytes
DT_BOOL = 11; // boolean type, 1 byte
DT_DOUBLE = 12; // double type, 8 bytes
}

size = n*c*h*w*Number of bytes of each data type

If AIPP parameters are set during model conversion, the information about the model input/output size obtained by AIModelManager::GetModelIOTensorDim is as follows:

n, c, h, and w are those before model conversion, instead of h (corresponding to the src_image_size_h parameter) and w (corresponding to the src_image_size_w parameter) configured in AIPP.
The value of data_type is fixed at 3, indicating the DT_INT8 type.

The parameter size indicates the size of the picture data processed by AIPP. The size varies according to the input format. For details, see Table 1.

**Table 1** Size formula
input_format	size
YUV400_U8	n * src_image_size_w * src_image_size_h
YUV420SP_U8	n * src_image_size_w * src_image_size_h * 1.5
XRGB8888_U8	n * src_image_size_w * src_image_size_h * 4
RGB888_U8	n * src_image_size_w * src_image_size_h * 3