神经网络cifar10 简单预测

# -*- coding: utf-8 -*- """ Created on Thu Apr 8 10:24:08 2021 @author: nkliu """ ############加载torch & 去除随机性#### import torch import random import numpy as np import os import math from skimage import io from torch import nn seed = 0 random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False ############### labels={"airplane":0,"automobile":1,"bird":2,"cat":3,"deer":4,"dog":5,"frog":6,"horse":7,"ship":8,"truck":9} def load_img(file,img_name): label = torch.tensor([labels[img_name[img_name.find("_") + 1:img_name.find(".")]]]) img = torch.from_numpy( np.expand_dims(np.transpose(io.imread(file + "\\" + img_name).astype(np.float32),(2,1,0)),0)) return label,img def load_imgs(file,img_list): datas=[] for img_name in img_list: datas.append(load_img(file,img_name)) return datas def load_data(file,batch_num): data=[] data_name=os.listdir(file) random.shuffle(data_name) data_number=len(data_name) len_per_batch=math.ceil(data_number/batch_num) extra_num=data_number-len_per_batch*batch_num index=[] #将多余的数据随机分到各个batch中 for i in range(extra_num): index.append(1) for i in range(batch_num-extra_num): index.append(0) random.shuffle(index) #将img分配到batch中 start=0 end=start+batch_num+index[0] for i in range(batch_num): batch=data_name[start:end] start=end if(i!=(batch_num-1)): end=end+batch_num+index[i] data.append(load_imgs(file,batch)) return data ###################数据加载########## # def load_data(): # # N是批大小；D是输入维度 # # H是隐藏层维度；D_out是输出维度 # N, D_in, H, D_out = 64, 1000, 100, 10 # 产生随机输入和输出张量 # x = torch.randn(N, D_in) # y = torch.randn(N, D_out) # return N, D_in, H, D_out, x, y ####################模型定义########## def define_model(H): # 使用torch.nn.Sequential定义序列化模型 # 一般模型定义方法 class Net(torch.nn.Module): def __init__(self,H): super(Net,self).__init__() self.conv1 = nn.Conv2d(3, 10, 3, padding=2) self.conv2 = nn.Conv2d(10, 30, 3, padding=2) self.pool = nn.MaxPool2d(2, 2) self.linear = nn.Linear(2430,10) def forward(self, x): x=self.pool(nn.functional.relu(self.conv1(x))) x=self.pool(nn.functional.relu(self.conv2(x))) x=x.view(-1,2430) x=self.linear(nn.functional.relu(x)) return x net = Net(H) return net ################损失函数定义######### def define_loss(): Loss = torch.nn.CrossEntropyLoss() return Loss ##############优化器定义############# def define_optimizer(): learning_rate = 1e-5 optimizer = torch.optim.Adam(net.parameters(), lr=learning_rate) return optimizer ###################模型训练######### def train(file, H, batch_num): net = define_model(H) Loss = torch.nn.CrossEntropyLoss() learning_rate = 1e-5 optimizer = torch.optim.Adam(net.parameters(), lr=learning_rate) datas=load_data(file,batch_num) for t in range(100): for k in range(batch_num): data_in_batch=datas[k] for j in range(len(data_in_batch)): x=data_in_batch[j][1] y=data_in_batch[j][0] y_pred = net(x) # 前向传播：通过像模型输入x计算预测的y loss = Loss(y_pred, y) # 计算loss optimizer.zero_grad() # 在反向传播之前，使用optimizer将它要更新的所有张量的梯度清零 loss.backward() # 反向传播：根据模型的参数计算loss的梯度 optimizer.step() # 调用Optimizer的step函数使它所有参数更新 if((t+1)%10==0): print("第{}次, 损失函数为 {}".format(t + 1, loss.item())) y_pred = net(x) # 前向传播：通过像模型输入x计算预测的y loss = Loss(y_pred, y) # 计算loss print("训练完成, 损失函数为 {}".format(loss.item())) ## 查看网络参数 #for name, parameters in net.named_parameters(): # print(name, ':', parameters.size(), ',', parameters.mean().item()) net_path = 'net.pth' torch.save(net.state_dict(), net_path) return net_path ###################模型测试######### def test(file, net_path,H): net = define_model(H) net.load_state_dict(torch.load(net_path)) # 再加载模型参数 test_datas=load_imgs(file,os.listdir(file)) total=0 correct=0 with torch.no_grad(): for data in test_datas: label,img = data outputs = net(img) _, pre = torch.max(outputs.data, 1) total += 1 correct += (pre == label).item() print("accuracy:{}".format(100 * correct / total)) if __name__ == '__main__': H=20 batch_num=10 net_path = train("cifar\\train", H,10) test("cifar\\test", net_path,H)