【周易AIPU 仿真】周易AIPU模拟器测试图片分类器

1、搭建模型

搭建一个分类网络，分类物体，模型结构如下:

输入: Placeholder
输出: output

训练数据集

1.1 训练

git clone https://gitee.com/Yang123321/simple-cnn.git
cd simple-cnn
python3 cnn.py

log:

step = 0        mean loss = 1.119428038597107
step = 10       mean loss = 1.0487422943115234
step = 20       mean loss = 0.9574651122093201
step = 30       mean loss = 0.8987987041473389
step = 40       mean loss = 0.8074296116828918
step = 50       mean loss = 0.7118208408355713
step = 60       mean loss = 0.5731518268585205
step = 70       mean loss = 0.42807847261428833
step = 80       mean loss = 0.25988009572029114
step = 90       mean loss = 0.5996816754341125
step = 100      mean loss = 0.20746763050556183
step = 110      mean loss = 0.08659818023443222
step = 120      mean loss = 0.027754301205277443
step = 130      mean loss = 0.009440970607101917
step = 140      mean loss = 0.005224811844527721
训练结束，保存模型到model/image_model

1.2 测试

修改cnn.py中的train = False

python3 cnn.py

log:

allImg: 150, trueImg: 150, percent: 100.0%

全部分类正确

2、模型量化

在1.2中会生成带有权重的pb文件

2.1 配置量化环境

git clone -b simpleCNN https://gitee.com/Yang123321/r329_aipu_simulator.git
cd  r329_aipu_simulator

2.2 矫正文件

三类物体，每类物体使用两帧图像作为矫正图像

cd dataset
python3 preprocess_dataset.py

dataset.npy
label.npy

2.3 cfg

量化配置文件build.cfg
模拟推理配置文件:run.cfg
生成输入测试文件(图像数据的二进制文件)

cd model
python3 gen_inputbin.py

2.4 仿真测试

aipubuild config/build.cfg

aipubuild config/run.cfg

共3类, 每类测试了一张图像，
输入为model目录下的input0.bin、input1.bin、input2.bin
测试结果如下
output_simpleCNN0.bin
output_simpleCNN1.bin
output_simpleCNN2.bin

模型直接输出的是类别，结果正确