Cifar-10图像分类竞赛_python

机器学习 Cifar图像分类竞赛

一、实验环境

PC机，Python

二、代码

#%%
import torch
import torch.nn as nn
import torch.nn.functional as F

import torchvision
import torchvision.datasets as dset
import torchvision.transforms as transforms

import torch.optim as optim
import torchvision.models as models

import PIL.Image as Image
import os
#%%
image_size = (224,224)
data_transform=transforms.Compose([
    transforms.RandomHorizontalFlip(),
    transforms.Resize(image_size),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
    ])
#%%
train_data=dset.ImageFolder(root="../input/cifar10",transform=data_transform)
# 数据集长度
totallen = len(train_data)
print('train data length:',totallen)
#%%
trainlen = int(totallen*0.95)
vallen = totallen - trainlen
train_db,val_db=torch.utils.data.random_split(train_data,[trainlen,vallen])
print('train:',len(train_db),'validation:',len(val_db))
#%%
# batch size
bs=20
# 训练集
train_loader=torch.utils.data.DataLoader(train_db,batch_size=bs, shuffle=True,num_workers=2)
# 验证集
val_loader=torch.utils.data.DataLoader(val_db,batch_size=bs, shuffle=True,num_workers=2)
#%%
def get_num_correct(out, labels):
    return out.argmax(dim=1).eq(labels).sum().item()
#%%
batch = next(iter(train_loader))
#%%
batch[1]
#%%
import torchvision.models as models
resnext101= models.resnet.resnext101_32x8d(pretrained=True)
#%%
model = resnext101
n_classes = len(train_data.classes)
model.fc = nn.Linear(2048, n_classes)
#%%
import torch.nn.init as init

for name, module in model._modules.items():
    if(name=='fc'):
        # print(module.weight.shape)
        init.kaiming_uniform_(module.weight, a=0, mode='fan_in')
#%%
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(device)
#%%
optimizer=torch.optim.SGD(model.parameters(),lr=0.01)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min',patience=1)
epoch_num = 12
model = model.to('cuda')
for epoch in range(epoch_num):
    total_loss=0
    total_correct=0
    val_correct=0
    for batch in train_loader:#GetBatch
        images,labels=batch
        images = images.to('cuda')
        labels = labels.to('cuda')
        outs=model(images)#PassBatch
        loss=F.cross_entropy(outs,labels)#CalculateLoss
        optimizer.zero_grad()
        loss.backward()#CalculateGradients
        optimizer.step()#UpdateWeights
        total_loss+=loss.item()
        total_correct+=get_num_correct(outs,labels)
    scheduler.step(total_loss)
    for batch in val_loader:
        images,labels=batch
        images = images.to('cuda')
        labels = labels.to('cuda')
        outs=model(images)
        val_correct+=get_num_correct(outs,labels)
    print("loss:",total_loss,"train_correct:",total_correct/trainlen, "val_correct:",val_correct/vallen)
#%%
torch.save(model, 'Cifar10-Resnext101_0.978.pkl')
#%%
import os
def file_name(file_dir):
    L=[]
    for root, dirs, files in os.walk(file_dir):
        for file in files:
            if os.path.splitext(file)[1] == '.png':
                L.append(os.path.join(root, file))
    L.sort()
    return L
#%%
model = torch.load('Cifar10-Resnext101_0.978.pkl')
model.eval()
#%%
test_path=file_name('/ilab/datasets/local/cifar10/test')
#%%
model.to('cpu')
pre=[]
for i in range(9000):
    filename=test_path[i]
    input_image = Image.open(filename).convert('RGB')
    input_tensor = data_transform(input_image)
    input_batch = input_tensor.unsqueeze(0) 
    output = model(input_batch)
    #print(output[0].shape)
    prob = F.softmax(output[0], dim=0)
    indexs = torch.argsort(-prob)
    #if i%100==0:
    print("i=",i," index:", indexs[0].item(), " prob: ", prob[indexs[0]])
    pre.append(indexs[0].item())
#%%
with open('/home/ilab/submission','a') as f:
    for i in range(9000):
        f.write('%04d.png %s\n'%(i+1,train_data.classes[pre[i]]))