第一週：深度學習及pytorch基礎
阿新 • • 發佈：2020-07-24
一、"AI研習社"實驗過程
1、將下載好的資料集上傳到colab平臺
-
首先連線Google雲端：
import os
from google.colab import drive
drive.mount('/content/drive')
path = "/content/drive/My Drive"
os.chdir(path)
os.listdir(path)
-
將資料集上傳到雲端，執行程式碼時結果報錯：
-
解決：(解決方法錯誤，和ImageFolder有關，具體可見第2步）:
2、將訓練集和驗證集按類別分類
import os
import shutil
shutil.rmtree('/content/drive/My Drive/cat_dog/test/1')
print('delete finished')
os.mkdir('/content/drive/My Drive/cat_dog/val/cat')
os.mkdir('/content/drive/My Drive/cat_dog/val/dog')
path = r'/content/drive/My Drive/cat_dog/val'
newcat = '/content/drive/My Drive/cat_dog/val/cat'
newdog = '/content/drive/My Drive/cat_dog/val/dog'
fns = [os.path.join(root,fn) for root, dirs, files in os.walk(path) for fn in files]
for f in fns:
name1 = str(f)
if 'cat_dog/val/3/cat' in name1:
shutil.copy(f, newcat)
else:
shutil.copy(f, newdog)
print(len(fns))分類後檢視資料集發現目錄中包含了隱藏資料夾ipynb_checkpoints
['.ipynb_checkpoints', 'cat', 'dog']
{'.ipynb_checkpoints': 0, 'cat': 1, 'dog': 2}
解決：
cd /content/drive/My Drive/cat_dog/train !rm -rf .ipynb_checkpoints
3、上傳測試集資料並自定義測試集testData
from PIL import Image
test_data_dir = '/content/drive/My Drive/cat_dog/test'
class TestDS(torch.utils.data.Dataset):
def __init__(self, transform=None):
self.test_data = os.listdir(test_data_dir)
self.test_label = np.zeros(2000)
self.transform = transform
def __getitem__(self, index):
# 根據索引返回資料和對應的標籤
image = Image.open(self.test_data[index]).convert('RGB')
image = self.transform(image)
return image, self.test_label[index]
def __len__(self):
# 返回檔案資料的數目
return len(self.test_data)
# 讀取測試集
testData = TestDS(transform=vgg_format)
print(len(testData))4、載入模型：
# 載入與預訓練模型
model_vgg = models.vgg16(pretrained=True)
# 凍結模型引數
for param in model_vgg_new.parameters():
param.requires_grad = False
# 修改最後一層模型
model_vgg_new.classifier._modules['6'] = nn.Linear(4096, 2)
# 損失函式nn.CrossEntropyLoss = log_softmax() + NLLLoss()
model_vgg_new.classifier._modules['7'] = torch.nn.LogSoftmax(dim = 1)
# 修改優化器為adam
optimizer_vgg = torch.optim.Adam(model_vgg_new.classifier[6].parameters(),lr = lr)5、模型訓練並用驗證集檢查效果
# 訓練模型
def train_model(model,dataloader,size,epochs=1,optimizer=None):
model.train()
for epoch in range(epochs):
running_loss = 0.0
running_corrects = 0
count = 0
for inputs,classes in dataloader:
inputs = inputs.to(device)
classes = classes.to(device)
outputs = model(inputs)
loss = criterion(outputs,classes)
optimizer.zero_grad()
loss.backward()
optimizer.step()
_,preds = torch.max(outputs.data,1)
# statistics
running_loss += loss.data.item()
running_corrects += torch.sum(preds == classes.data)
count += len(inputs)
# print('Training: No. ', count, ' process ... total: ', size)
epoch_loss = running_loss / size
epoch_acc = running_corrects.data.item() / size
print('Loss: {:.4f} Acc: {:.4f}'.format(
epoch_loss, epoch_acc))
# 模型訓練
train_model(model_vgg_new,loader_train,size=dset_sizes['train'], epochs=1,
optimizer=optimizer_vgg)
#模型驗證
def test_model(model,dataloader,size):
model.eval()
predictions = np.zeros(size)
all_classes = np.zeros(size)
all_proba = np.zeros((size,2))
i = 0
running_loss = 0.0
running_corrects = 0
for inputs,classes in dataloader:
inputs = inputs.to(device)
classes = classes.to(device)
outputs = model(inputs)
loss = criterion(outputs,classes)
_,preds = torch.max(outputs.data,1)
# statistics
running_loss += loss.data.item()
running_corrects += torch.sum(preds == classes.data)
predictions[i:i+len(classes)] = preds.to('cpu').numpy()
all_classes[i:i+len(classes)] = classes.to('cpu').numpy()
all_proba[i:i+len(classes),:] = outputs.data.to('cpu').numpy()
i += len(classes)
# print('Testing: No. ', i, ' process ... total: ', size)
epoch_loss = running_loss / size
epoch_acc = running_corrects.data.item() / size
print('Loss: {:.4f} Acc: {:.4f}'.format(
epoch_loss, epoch_acc))
return predictions, all_proba, all_classes
predictions, all_proba, all_classes = test_model(model_vgg_new,loader_valid,size=dset_sizes['val'])
6、編寫測試程式碼：
def test_model(model,dataloader,size):
model.eval()
predictions = np.zeros(size)
all_classes = np.zeros(size)
all_proba = np.zeros((size,2))
i = 0
running_loss = 0.0
running_corrects = 0
for inputs,classes in dataloader:
inputs = inputs.to(device)
classes = classes.to(device)
outputs = model(inputs)
loss = criterion(outputs,classes)
_,preds = torch.max(outputs.data,1)
predictions[i:i+len(classes)] = preds.to('cpu').numpy()
i += len(classes)
return predictions
results = test_model(model_vgg_new,loader_test,size=2000)
print(results)
7、將測試集的輸出結果儲存：
name = []
for i in testData.test_data:
j = i[37:-4]
name.append(int(j))
print(results)
print(name)
import pandas as pd
#字典中的key值即為csv中列名
dataframe = pd.DataFrame({'name':name,'results':results})
#將DataFrame儲存為csv,index表示是否顯示行名，default=True
dataframe.to_csv("/content/sample_data/test.csv",index=False,sep=',')
8、將csv檔案按平臺要求修改，觀察結果
二、改進模型
1、修改訓練輪數epoch=3——沒什麼用，訓練輪數不宜太多，可能還會過擬合
2、使用resnet152預訓練模型
-
修改模型和最後一層：
model_resnet = models.resnet152(pretrained=True) model_resnet.fc = nn.Linear(2048, 2)
-
損失函式使用：
criterion = nn.CrossEntropyLoss()
-
修改學習率和隨機梯度下降，每隔7個epoch學習率降低
lr = 0.001
optimizer_vgg = torch.optim.Adam(model_resnet.fc.parameters(),lr = lr)
def adjust_learning_rate(optimizer, lr):
for param_group in optimizer.param_groups:
param_group['lr'] = lr
for eooch in range(epochs):
if epoch > 7:
lr = 0.0001
elif epoch > 14:
lr = 0.00001
elif epoch > 21:
lr = 0.000001
elif epoch > 28:
lr = 0.0000001
adjust_learning_rate(optimizer, lr)
-
修改 epoch = 30 並進行訓練，得到驗證集結果：
-
上傳平臺，並未得到改進：（多次嘗試epoch和lr的設定，學習率總在98%左右，未能達到99%以上）
3、構造SENet網路，訓練過程引數太多，導致colab記憶體溢位
4、借鑑網路中的神經網路結構：https://www.cnblogs.com/ansang/p/9126427.html，將其復現為PyTorch結構，也不能達到99%以上的準確率：
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(3, 32, kernel_size=(3, 3), stride=1, padding=1)
self.max_pooling2d = nn.MaxPool2d(2,stride=2)
self.conv2 = nn.Conv2d(32, 32, kernel_size=(3, 3), stride=1, padding=1)
self.conv3 = nn.Conv2d(32, 64, kernel_size=(3, 3), stride=1, padding=1)
self.conv4 = nn.Conv2d(64, 64, kernel_size=(3, 3), stride=1, padding=1)
self.fc1 = nn.Linear(28*28*64, 1024)
self.fc2 = nn.Linear(1024, 2)
self.relu = nn.ReLU(128)
self.dropout = nn.Dropout(0.4)
def forward(self, x):
x = self.relu(self.conv1(x))
x = self.max_pooling2d(x)
x = self.relu(self.conv2(x))
x = self.max_pooling2d(x)
x = self.relu(self.conv3(x))
x = self.relu(self.conv4(x))
x = self.max_pooling2d(x)
x = x.view(-1, x.shape[1]*x.shape[2]*x.shape[3])
x = self.relu(self.fc1(x))
x = self.dropout(x)
x = self.fc2(x)
return x