Example of Dynamic Resolution Inference (ResNet-50)

Sample Code

import numpy as np
from mindx.sdk import base
from mindx.sdk.base import Tensor
import cv2 as cv

try:
    # Set the path of the image to be inferred.
    jpg_path = "./test.jpg"  
    # Read the original image.
    img = cv.imread(jpg_path)
    # Resize the image to 224 x 224. The output size format is (width, height).
    img_test1 = cv.resize(img, (224, 224))
    # Perform CSC: BGR > YUV420
    input_image = cv.cvtColor(img_test1, cv.COLOR_BGR2YUV_I420)
    
    device_id = 1  # Device ID
    base.mx_init()  # Perform global initialization and allocate device and log resources.
    input_image1 = np.reshape(input_image,((1,) + input_image.shape[:2] +  (1,)))
    
i    imgTensor = [Tensor(input_image1)]  # Convert the Image class to the Tensor class.
    model_path = "./resnet50_dynamic_image.om"  # Model path
    model_ = base.model(model_path, deviceId=device_id)  # Create a model object.
    outputs = model_.infer(imgTensor)  # Use a model to perform inference on tensor objects.
        
    label_path = "./resnet50_clsidx_to_labels.names"  # Classification label file path
    output0 = outputs[0]  # Obtain the inference result.
    output0.to_host()  # Transfer the inference result tensor to the host.
    result0_array = np.array(output0)  # Convert the inference result tensor into a NumPy array for post-processing.
    max_idx = result0_array.argmax(axis=1)[0]  # Obtain the label subscript index with the highest confidence score.
    print('label_index:', max_idx)  # Output the label subscript index with the highest confidence score.
    
    with open(label_path, 'r', encoding='utf-8') as f:   # Read classification label files.
        file = f.read()
    label_list = file.split('\n')[1:-1]  # Obtain valid classification labels.
    infer_result = label_list[max_idx]  # Obtain label names.
    print('infer result:{}'.format(infer_result))  # Print the inference result.
except Exception as e:
    print(e)

Example

For details about the inference process when a dynamic resolution model uses the DVPP class for data pre-processing, see End-to-End Inference Processing Example (ResNet-50). This sample shows the data pre-processing and inference process using OpenCV.

The example is for reference only. For details about the commands and configurations for model building, see the CANN ATC Instructions.

If AIPP CSC is configured when the ATC tool is used, ensure that the CSC corresponds to the OM model input when the OpenCV reads images.

• Example of the dynamic resolution conversion command used for model building:

  atc --model=./resnet50_v1.5.pb --framework=3 --output=resnet50_dynamic_image --output_type=FP32 --soc_version=Ascend310 --input_shape=input_tensor:1,-1,-1,3 --insert_op_conf=./aipp_resnet50_dynamic_image.aippconfig --dynamic_image_size="224,224;448,448"

• Example of dynamic resolution AIPP configuration:

  aipp_op {
      aipp_mode: static
      input_format : YUV420SP_U8
      csc_switch : true
      rbuv_swap_switch : false
      src_image_size_w : 0
      src_image_size_h : 0
      matrix_r0c0 : 256
      matrix_r0c1 : 0
      matrix_r0c2 : 359
      matrix_r1c0 : 256
      matrix_r1c1 : -88
      matrix_r1c2 : -183
      matrix_r2c0 : 256
      matrix_r2c1 : 454
      matrix_r2c2 : 0
      input_bias_0 : 0
      input_bias_1 : 128
      input_bias_2 : 128
  
      mean_chn_0 : 124
      mean_chn_1 : 117
      mean_chn_2 : 104
  }