Development Requirement Analysis

Model inference

Select an open-source classification network, for example, the Caffe ResNet-50 network and convert it into an offline model (.om file) that adapts to Ascend AI Processors.

• Ascend Tensor Compiler (ATC) for converting the ResNet-50 network into an offline model (.om file) that adapts to Ascend AI Processors.

  For details, see ATC Tool Instructions.

• API for model loading:

  acl.mdl.load_from_file_with_mem

• API for model execution:

  acl.mdl.execute

• API for model unloading:

  acl.mdl.unload

Data preprocessing (including decoding, cropping, resizing, encoding, CSC)

The ResNet-50 network takes 224 x 224 RGB images only.

Therefore, the existing 1024 x 683 .jpg images need to be decoded into the YUV420SP format, resized to 224 x 224, and then converted into the RGB format.

• APIs for decoding and resizing:
  • acl.media.dvpp_create_channel
  • acl.media.dvpp_jpeg_decode_async
  • acl.media.dvpp_vpc_resize_async
  • acl.media.dvpp_destroy_channel
• CSC:

  Specify the --insert_op_conf parameter when using the ATC (Ascend Tensor Compiler) for model conversion and save the CSC settings in the .om file for calling during model execution.

  For details, see Feature Configuration > AIPP Configuration in ATC Tool Instructions.

Data postprocessing

Post-process the model inference result. That is, call the Cast and ArgMaxD operators to identify the class label with the highest confidence score of each image from the inference result.

Call the Cast operator to convert the data type of the inference result from float32 to float16, and then call the ArgMaxD operator to identify the class label with the highest confidence score of each image from the inference result.

• Build the .json definition files of the Cast and ArgMaxD operators into an offline model (.om file) that adapts to Ascend AI Processors respectively to verify the single operators.

  For details, see ATC Tool Instructions.

• Operator execution: acl.op.execute

Data copy

Copy image from the host to the device for processing and copy the processing result to the host after processing is complete.

API for data copy: acl.rt.memcpy