2.3.3. 使用SAIL驱动TPU加载bmodel并进行推理

您可以使用如下命令下载测试图片cls.jpg：

wget \
https://sophon-file.sophon.cn/sophon-prod-s3/model/19/12/05/cls.jpg.tgz
tar -zxvf cls.jpg.tgz

然后，参考下面的程序使用SAIL进行推理：

#!/usr/bin/env python3

import cv2
import numpy as np
import sophon.sail as sail

# initialize an Engine instance using bmodel.
bmodel = sail.Engine("bmodel/compilation.bmodel", 0, sail.IOMode.SYSIO)
# graph_name is just the net_name in conversion step.
graph_name = bmodel.get_graph_names()[0]
input_tensor_name = bmodel.get_input_names(graph_name)[0]
# why transpose?
# bmodel will always be NCHW layout,
# so, if original tensorflow frozen model is formatted as NHWC,
input_data = {input_tensor_name: \
    np.expand_dims(cv2.resize(cv2.imread("cls.jpg"), (224,224)), 0)}
# do inference
outputs = bmodel.process(graph_name, input_data)