基于CNN(LeNet)的垃圾分类(C语言实现CNN算子)
一、先使用python训练模型
具体步骤参考基于pytorch的MNIST数据集的四层CNN,测试准确率99.77%这篇文章,各种步骤我写的很详细,只需要将MNIST数据集换成垃圾分类的数据集,再调整一下参数就好了。
二、提取参数
提取模型参数
权重和偏置
我们需要提取每一个具有学习的参数的训练层的权重和偏置,我使用了两层卷积和两层全连接,就要提取两个卷积层的权重和偏置,两个全连接层的权重和偏置。
# Extraction_Parameter.py
#引入库
#引用需要用到的库
import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
#model
class CNNModel(nn.Module):
def __init__(self):
super(CNNModel, self).__init__()
# Convolution layer 1
self.conv1 = nn.Conv2d(in_channels = 3 , out_channels = 8, kernel_size = 3, stride = 1, padding = 0 )
self.relu1 = nn.ReLU()
self.maxpool1 = nn.MaxPool2d(kernel_size = 2, stride = 2)
# Convolution layer 2
self.conv2 = nn.Conv2d(in_channels =8 , out_channels = 16, kernel_size = 3, stride = 1, padding = 0 )
self.relu2 = nn.ReLU()
self.maxpool2 = nn.MaxPool2d(kernel_size = 2, stride = 2)
# Fully-Connected layer 1
self.fc1 = nn.Linear(400,40)
# Fully-Connected layer 2
self.fc2 = nn.Linear(40,4)
def forward(self, x):
# conv layer 1 的前向计算,3行代码
out = self.conv1(x)
#print(out.shape)
out = self.relu1(out)
out = self.maxpool1(out)
#print(out.shape)
# conv layer 2 的前向计算,3行代码
out = self.conv2(out)
#print(out.shape)
out = self.relu2(out)
out = self.maxpool2(out)
#print(out.shape)
#Flatten拉平操作
out = out.view(out.size(0),-1)
#print(out.shape)
#FC layer的前向计算(2行代码)
out = self.fc1(out)
out = self.fc2(out)
return F.log_softmax(out,dim = 1)
#实例化模型
network = CNNModel()
#加载模型
model_path = "model1.pth"
network.load_state_dict(torch.load(model_path, map_location = torch.device('cpu')))
#network.eval()
parm = {
}
for name,parameters in network.state_dict().items():
parm[name] = parameters.detach().numpy()
print(name, parameters)
w1 = parm['conv1.weight']
b1 = parm['conv1.bias']
w2 = parm['conv2.weight']
b2 = parm['conv2.bias']
fc1_w = parm['fc1.weight']
fc1_b = parm['fc1.bias']
fc2_w = parm['fc2.weight']
fc2_b = parm['fc2.bias']
#print(type(w1))
#print(len(w1[0]))
#print(len(w1[0][0]))
#print(len(w1[0][0][0]))
#conv1_wb
with open("parameters1_wb.h","a") as f:
print(type(w1))
#new_str1 = str(np.transpose(w1).tolist())
new_str1 = str(w1.tolist())
new_str2 = new_str1.replace('[','')
new_str3 = new_str2.replace(']','')
f.write("float conv1_weight[8][3][9] = {" + new_str3 + "};\n\n")
print("第一层卷积的权重保存成功")
f.close()
with open("parameters1_wb.h","a") as f:
print(type(b1))
#new_str1 = str(np.transpose(b1).tolist())
new_str1 = str(b1.tolist())
new_str2 = new_str1.replace('[','')
new_str3 = new_str2.replace(']','')
f.write("float conv1_bias[8] = {" + new_str3 + "};\n\n")
print("第一层卷积的偏置保存成功")
f.close()
#conv2_wb
with open("parameters1_wb.h","a") as f:
print(type(w2))
#new_str1 = str(np.transpose(w2).tolist())
new_str1 = str(w2.tolist())
new_str2 = new_str1.replace('[','')
new_str3 = new_str2.replace(']','')
f.write("float conv2_weight[16][8][9] = {" + new_str3 + "};\n\n")
print("第二层卷积的权重保存成功")
f.close()
with open("parameters1_wb.h","a") as f:
print(type(b2))
#new_str1 = str(np.transpose(b2).tolist())
new_str1 = str(b2.tolist())
new_str2 = new_str1.replace('[','')
new_str3 = new_str2.replace(']','')
f.write("float conv2_bias[16] = {" + new_str3 + "};\n\n")
print("第二层卷积的偏置保存成功")
f.close()
#fc1_wb
with open("parameters1_wb.h","a") as f:
print(type(fc1_w))
new_str1 = str(np.transpose(fc1_w).tolist())
#new_str1 = str(fc1_w.tolist())
new_str2 = new_str1.replace('[','')
new_str3 = new_str2.replace(']','')
f.write("float fc1_weight[" + str(400*40) + "] = {" + new_str3 + "};\n\n")
print("第一层全连接的权重保存成功")
f.close()
with open("parameters1_wb.h","a") as f:
print(type(fc1_b))
#new_str1 = str(np.transpose(fc1_b).tolist())
new_str1 = str(fc1_b.tolist())
new_str2 = new_str1.replace('[','')
new_str3 = new_str2.replace(']','')
f.write("float fc1_bias[40] = {" + new_str3 + "};\n\n")
print("第一层全连接的偏置保存成功")
f.close()
#fc2_wb
with open("parameters1_wb.h","a") as f:
print(type(fc2_w))
new_str1 = str(np.transpose(fc2_w).tolist())
#new_str1 = str(fc2_w.tolist())
new_str2 = new_str1.replace('[','')
new_str3 = new_str2.replace(']','')
f.write("float fc2_weight[" + str(40*4) + "] = {" + new_str3 + "};\n\n")
print("第二层全连接的权重保存成功")
f.close()
with open("parameters1_wb.h","a") as f:
print(type(fc2_b))
#new_str1 = str(np.transpose(fc2_b).tolist())
new_str1 = str(fc2_b.tolist())
new_str2 = new_str1.replace('[','')
new_str3 = new_str2.replace(']','')
f.write("float fc2_bias[4] = {" + new_str3 + "};\n\n")
print("第二层全连接的偏置保存成功")
f.close()
提取成功后会得到一个parameters1_wb.h文件,如图所示
提取图片
将测试的图片同样提取为.h文件
# Extract_Image.py
from torchvision import transforms
import torch
import numpy as np
from PIL import Image
from itertools import chain
# 在训练模型时对图片进行怎样的预处理
# 在提取图片参数时需要先进行同样的处理再提取,不然维度数据对不上
data_transform = transforms.Compose(
[transforms.ToTensor()
#transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
img = Image.open("./test/Others1.jpg")#预测图片
img = data_transform(img)
img = torch.unsqueeze(img, dim=0)
img = img.numpy()
#img = img * (1.0/255.0)
#img = img.tolist()
print(img)
#参数提取
with open("pic1.h","a") as f:
#new_str1 = str(np.transpose(img).tolist())
new_str1 = str(img.tolist())
new_str2 = new_str1.replace('[','')
new_str3 = new_str2.replace(']','')
f.write("float Others2" + "[3][28][28] = {" + new_str3 + "};\n\n")
print("图片Others1参数读取完成")
f.close()
三、编写CNN算子
在windows中实现
#include <stdio.h>
#