大神，求教 ，我用pytorch cnn識別貓狗圖片，為什么loss忽大忽小？

import os;

import torch.utils.data as data;

import numpy as np;

import torch;

import torchvision.transforms as transforms;

from PIL import Image



data_transform=transforms.Compose([

    transforms.ToTensor()

]);



IMAGE_H = 200;

IMAGE_W = 200;



class DogCatDataset(data.Dataset):

    def __init__(self,mode,dir):

        self.mode=mode;

        self.list_img=[];

        self.list_label=[];

        self.data_size=0;

        self.transform=data_transform;



        if self.mode=='train':

            dir=dir+'/train/';

            for file in os.listdir(dir):

                self.list_img.append(dir+file);

                self.data_size+=1;

                name=file.split(sep='.');

                if name[0]=='cat':

                    self.list_label.append(0);

                else:

                    self.list_label.append(1);

        elif self.mode=='test':

            dir=dir+'/test/';

            for file in os.listdir(dir):

                self.list_img.append(dir+file);

                self.data_size+=1;

                self.list_label.append(2);

        else:

            return print('undefined dataset');







    def __getitem__(self, item):

        if self.mode=='train':

            img = Image.open(self.list_img[item])  # 打開圖片

            img = img.resize((IMAGE_H, IMAGE_W))  # 將圖片resize成統一大小

            img = np.array(img)[:, :, :3]  # 資料轉換成numpy陣列形式

            label = self.list_label[item]  # 獲取image對應的label

            return self.transform(img), torch.LongTensor([label])

        elif self.mode=='test':

            img=Image.open(self.list_img[item]);

            img=img.resize((IMAGE_W,IMAGE_H));

            img=np.array(img)[:,:,:3];

            return self.transform(img);

        else:

            print('None');





    def __len__(self):

        return self.data_size;

import torch;

import torch.nn as nn;

import torch.utils.data;

import torch.nn.functional as F;



class Net(nn.Module):

    def __init__(self):

        super(Net,self).__init__();

        self.conv1 = torch.nn.Conv2d(3, 16, 3, padding=1)

        self.conv2 = torch.nn.Conv2d(16, 16, 3, padding=1)



        self.fc1 = nn.Linear(50 * 50 * 16, 128)

        self.fc2 = nn.Linear(128, 64)

        self.fc3 = nn.Linear(64, 2)



    def forward(self,x):

        x = self.conv1(x)



        x = F.relu(x)

        x = F.max_pool2d(x, 2)



        x = self.conv2(x)

        x = F.relu(x)

        x = F.max_pool2d(x, 2)



        x = x.view(x.size()[0], -1)

        x = F.relu(self.fc1(x))

        x = F.relu(self.fc2(x))

        x = self.fc3(x)



        return F.softmax(x, dim=1)

    # x=F.relu(self.fc2(x));

        # x=F.relu(self.fc3(x));

        # x=F.relu(self.fc4(x));

        # x=self.fc5(x);



        #return  F.softmax(x,dim=1);

from DogCatImageDataset import DogCatDataset as DVCD;

from torch.utils.data import DataLoader as DataLoader;

from TrainNet import Net;

import torch;

from torch.autograd import  Variable;

import  torch.nn as nn;

import torchvision.transforms as transforms;

import matplotlib.pyplot as plt;



dataset_dir='./data/';

model_cp='./model/';

workers=10;

batch_size=50;

lr=0.0001;

epochs=10;



def train():

    transform_train=transforms.Compose([

        transforms.Normalize(std=(0.5,0.5,0.5),mean=(0.5,0.5,0.5))

    ]);

    datafile=DVCD('train',dataset_dir);

    dataloader=DataLoader(datafile,batch_size=batch_size,shuffle=True,num_workers=workers);

    print('Dataset loaded! length of train set is {0}'.format(len(datafile)));



    model=Net();



    optimizer=torch.optim.SGD(model.parameters(),lr=lr,momentum=0.9);

    creiterion=torch.nn.CrossEntropyLoss();

    cnt=0;





    losses=[];

    for i in range(epochs):

        #model.train();

        print('epochs:{0}'.format(i));

        for j,(img,label) in enumerate(dataloader):

            img,label=Variable(img),Variable(label);



            optimizer.zero_grad();

            out=model(img);

            loss=creiterion(out,label.squeeze());

            loss.backward();

            optimizer.step();

            #train_loss=loss/batch_size;

            if j % 10 ==0:

                losses.append(loss.float());

                print('[epochs - {0} - {1}/{2}]loss:{3}'.format(i,j,len(datafile),loss.float()));



                plt.clf();

                plt.plot(losses);

                plt.pause(0.01);

            #cnt+=1;

            #print('Frame {0},train_loss:{1}'.format(cnt*batch_size,loss/batch_size));



    torch.save(model.state_dict(),'{0}model.pth'.format(model_cp));



if __name__=='__main__':

    train();

uj5u.com熱心網友回復：

uj5u.com熱心網友回復：

uj5u.com熱心網友回復：

uj5u.com熱心網友回復：

標籤：人工智能技術

上一篇：2019年后的商業新形態，我們該如何自處？

下一篇：敏捷開發是否可以不寫檔案？