2.3.2. 使用 bmnett 将 mobilenet 编译为 bmodel

我们已经在官网上传了一份 tensorflow mobilenet 的模型，直接下载即可：

wget \
https://sophon-file.sophon.cn/sophon-prod-s3/model/19/05/28/mobilenetv1_tf.tar.gz
tar -zxvf mobilenetv1_tf.tar.gz

然后，使用下面的脚本将模型编译为 bmodel：

#!/usr/bin/env python3
import bmnett

model_path = "mobilenetv1.pb"  # path of tensorflow frozen model to be converted.
outdir = "bmodel/"             # path of the generated bmodel.
target = "BM1684"              # targeted TPU platform, BM1684 or BM1682.
input_names = ["input"]        # input operation names.
output_names = ["MobilenetV1/Predictions/Reshape_1"]  # output operation names.
shapes = [(1, 224, 224, 3)]    # input shapes.
net_name = "mobilenetv1"       # name of the generated bmodel.

bmnett.compile(model_path, outdir, target, \
 input_names, output_names, shapes=shapes, net_name=net_name)

运行结束之后，在上面脚本指定的${outdir}目录下，会生成 compilation.bmodel.