几乎恒定的训练和验证准确性
Nearly Constant training and validation accuracy
我是pytorch新手，我的问题可能有点幼稚
我正在我的数据集上训练一个预训练的 VGG16 网络，它的大小接近 8 类 中的 33000 张图像，标签为 [1,2,…,8] 而我的 类 是不平衡的。我的问题是在训练期间，验证和训练准确率很低而且没有增加，我的代码有什么问题吗？
如果不是，您对改进培训有何建议？
'''
import torch
import time
import torch.nn as nn
import numpy as np
from sklearn.model_selection import train_test_split
from torch.optim import Adam
import cv2
import torchvision.models as models
from classify_dataset import Classification_dataset
from torchvision import transforms
transform = transforms.Compose([transforms.Resize((224,224)),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomVerticalFlip(p=0.5),
transforms.RandomRotation(degrees=45),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
dataset = Classification_dataset(root_dir=r'//home/arisa/Desktop/Hamid/IQA/Hamid_Dataset',
csv_file=r'/home/arisa/Desktop/Hamid/IQA/new_label.csv',transform=transform)
target = dataset.labels - 1
train_indices, test_indices = train_test_split(np.arange(target.shape[0]), stratify=target)
test_dataset = torch.utils.data.Subset(dataset, indices=test_indices)
train_dataset = torch.utils.data.Subset(dataset, indices=train_indices)
class_sample_count = np.array([len(np.where(target[train_indices] == t)[0]) for t in np.unique(target)])
weight = 1. / class_sample_count
samples_weight = np.array([weight[t] for t in target[train_indices]])
samples_weight = torch.from_numpy(samples_weight)
samples_weight = samples_weight.double()
sampler = torch.utils.data.WeightedRandomSampler(samples_weight, len(samples_weight), replacement = True)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=64,
sampler=sampler)
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=64,
shuffle=False)
for param in model.parameters():
param.requires_grad = False
num_ftrs = model.classifier[0].in_features
model.classifier = nn.Linear(num_ftrs,8)
optimizer = Adam(model.parameters(), lr = 0.0001 )
criterion = nn.CrossEntropyLoss()
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=50, gamma=0.01)
path = '/home/arisa/Desktop/Hamid/IQA/'
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print(device)
def train_model(model, train_loader,valid_loader, optimizer, criterion, scheduler=None, num_epochs=10 ):
min_valid_loss = np.inf
model.train()
start = time.time()
TrainLoss = []
model = model.to(device)
for epoch in range(num_epochs):
total = 0
correct = 0
train_loss = 0
#lr_scheduler.step()
print('Epoch {}/{}'.format(epoch+1, num_epochs))
print('-' * 10)
train_loss = 0.0
for x,y in train_loader:
x = x.to(device)
#print(y.shape)
y = y.view(y.shape[0],).to(device)
y = y.to(device)
y -= 1
out = model(x)
loss = criterion(out, y)
optimizer.zero_grad()
loss.backward()
TrainLoss.append(loss.item()* y.shape[0])
train_loss += loss.item() * y.shape[0]
_,predicted = torch.max(out.data,1)
total += y.size(0)
correct += (predicted == y).sum().item()
optimizer.step()
lr_scheduler.step()
accuracy = 100*correct/total
valid_loss = 0.0
val_loss = []
model.eval()
val_correct = 0
val_total = 0
with torch.no_grad():
for x_val, y_val in test_loader:
x_val = x_val.to(device)
y_val = y_val.view(y_val.shape[0],).to(device)
y_val -= 1
target = model(x_val)
loss = criterion(target, y_val)
valid_loss += loss.item() * y_val.shape[0]
_,predicted = torch.max(target.data,1)
val_total += y_val.size(0)
val_correct += (predicted == y_val).sum().item()
val_loss.append(loss.item()* y_val.shape[0])
val_acc = 100*val_correct / val_total
print(f'Epoch {epoch + 1} \t\t Training Loss: {train_loss / len(train_loader)} \t\t Validation Loss: {valid_loss / len(test_loader)} \t\t Train Acc:{accuracy} \t\t Validation Acc:{val_acc}')
if min_valid_loss > (valid_loss / len(test_loader)):
print(f'Validation Loss Decreased({min_valid_loss:.6f}--->{valid_loss / len(test_loader):.6f}) \t Saving The Model')
min_valid_loss = valid_loss / len(test_loader)
state = {'state_dict': model.state_dict(),'optimizer': optimizer.state_dict(),}
torch.save(state,'/home/arisa/Desktop/Hamid/IQA/checkpoint.t7')
end = time.time()
print('TRAIN TIME:')
print('%.2gs'%(end-start))
train_model(model=model, train_loader=train_loader, optimizer=optimizer, criterion=criterion, valid_loader= test_loader,num_epochs=500 )
提前致谢
这是15个epoch
的结果
Epoch 1/500
----------
Epoch 1 Training Loss: 205.63448420514916 Validation Loss: 233.89266112356475 Train Acc:39.36360386127994 Validation Acc:24.142040038131555
Epoch 2/500
----------
Epoch 2 Training Loss: 199.05699240435197 Validation Loss: 235.08799531243065 Train Acc:41.90998291820601 Validation Acc:24.27311725452812
Epoch 3/500
----------
Epoch 3 Training Loss: 199.15626737127448 Validation Loss: 236.00033430619672 Train Acc:41.1035633416756 Validation Acc:23.677311725452814
Epoch 4/500
----------
Epoch 4 Training Loss: 199.02581041173886 Validation Loss: 233.60767459869385 Train Acc:41.86628530568466 Validation Acc:24.606768350810295
Epoch 5/500
----------
Epoch 5 Training Loss: 198.61493769454472 Validation Loss: 233.7503859202067 Train Acc:41.53656695665991 Validation Acc:25.0
Epoch 6/500
----------
Epoch 6 Training Loss: 198.71323942956585 Validation Loss: 234.17176149830675 Train Acc:41.639852222619474 Validation Acc:25.369399428026693
Epoch 7/500
----------
Epoch 7 Training Loss: 199.9395153770592 Validation Loss: 234.1744423635078 Train Acc:40.98041552456998 Validation Acc:24.84509056244042
Epoch 8/500
----------
Epoch 8 Training Loss: 199.3533399020355 Validation Loss: 235.4645173188412 Train Acc:41.26643626107337 Validation Acc:24.165872259294567
Epoch 9/500
----------
Epoch 9 Training Loss: 199.6451746921249 Validation Loss: 233.33387595956975 Train Acc:40.96452548365312 Validation Acc:24.59485224022879
Epoch 10/500
----------
Epoch 10 Training Loss: 197.9305159737011 Validation Loss: 233.76405122063377 Train Acc:41.8782028363723 Validation Acc:24.6186844613918
Epoch 11/500
----------
Epoch 11 Training Loss: 199.33247244055502 Validation Loss: 234.41085289463854 Train Acc:41.59218209986891 Validation Acc:25.119161105815063
Epoch 12/500
----------
Epoch 12 Training Loss: 199.87399289874256 Validation Loss: 234.23621463775635 Train Acc:41.028085647320545 Validation Acc:24.49952335557674
Epoch 13/500
----------
Epoch 13 Training Loss: 198.85540591944292 Validation Loss: 234.33149099349976 Train Acc:41.206848607635166 Validation Acc:24.857006673021925
Epoch 14/500
----------
Epoch 14 Training Loss: 199.92641723337513 Validation Loss: 233.37722391070741 Train Acc:41.15520597465539 Validation Acc:24.988083889418494
Epoch 15/500
----------
Epoch 15 Training Loss: 197.82172771698328 Validation Loss: 234.4943131533536 Train Acc:41.69943987605768 Validation Acc:24.380362249761678
您通过
冻结您的模型
for param in model.parameters():
param.requires_grad = False
这基本上是说“不计算任何权重的任何梯度”，这相当于不更新权重 - 因此没有优化
我的问题在 model.train()。这个短语应该在训练循环中。但在我的例子中，我把它放在训练循环之外，当涉及到 model.eval() 时，模型保持在这种模式下
我是pytorch新手，我的问题可能有点幼稚 我正在我的数据集上训练一个预训练的 VGG16 网络，它的大小接近 8 类 中的 33000 张图像，标签为 [1,2,…,8] 而我的 类 是不平衡的。我的问题是在训练期间，验证和训练准确率很低而且没有增加，我的代码有什么问题吗？ 如果不是，您对改进培训有何建议？ '''
import torch
import time
import torch.nn as nn
import numpy as np
from sklearn.model_selection import train_test_split
from torch.optim import Adam
import cv2
import torchvision.models as models
from classify_dataset import Classification_dataset
from torchvision import transforms
transform = transforms.Compose([transforms.Resize((224,224)),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomVerticalFlip(p=0.5),
transforms.RandomRotation(degrees=45),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
dataset = Classification_dataset(root_dir=r'//home/arisa/Desktop/Hamid/IQA/Hamid_Dataset',
csv_file=r'/home/arisa/Desktop/Hamid/IQA/new_label.csv',transform=transform)
target = dataset.labels - 1
train_indices, test_indices = train_test_split(np.arange(target.shape[0]), stratify=target)
test_dataset = torch.utils.data.Subset(dataset, indices=test_indices)
train_dataset = torch.utils.data.Subset(dataset, indices=train_indices)
class_sample_count = np.array([len(np.where(target[train_indices] == t)[0]) for t in np.unique(target)])
weight = 1. / class_sample_count
samples_weight = np.array([weight[t] for t in target[train_indices]])
samples_weight = torch.from_numpy(samples_weight)
samples_weight = samples_weight.double()
sampler = torch.utils.data.WeightedRandomSampler(samples_weight, len(samples_weight), replacement = True)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=64,
sampler=sampler)
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=64,
shuffle=False)
for param in model.parameters():
param.requires_grad = False
num_ftrs = model.classifier[0].in_features
model.classifier = nn.Linear(num_ftrs,8)
optimizer = Adam(model.parameters(), lr = 0.0001 )
criterion = nn.CrossEntropyLoss()
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=50, gamma=0.01)
path = '/home/arisa/Desktop/Hamid/IQA/'
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print(device)
def train_model(model, train_loader,valid_loader, optimizer, criterion, scheduler=None, num_epochs=10 ):
min_valid_loss = np.inf
model.train()
start = time.time()
TrainLoss = []
model = model.to(device)
for epoch in range(num_epochs):
total = 0
correct = 0
train_loss = 0
#lr_scheduler.step()
print('Epoch {}/{}'.format(epoch+1, num_epochs))
print('-' * 10)
train_loss = 0.0
for x,y in train_loader:
x = x.to(device)
#print(y.shape)
y = y.view(y.shape[0],).to(device)
y = y.to(device)
y -= 1
out = model(x)
loss = criterion(out, y)
optimizer.zero_grad()
loss.backward()
TrainLoss.append(loss.item()* y.shape[0])
train_loss += loss.item() * y.shape[0]
_,predicted = torch.max(out.data,1)
total += y.size(0)
correct += (predicted == y).sum().item()
optimizer.step()
lr_scheduler.step()
accuracy = 100*correct/total
valid_loss = 0.0
val_loss = []
model.eval()
val_correct = 0
val_total = 0
with torch.no_grad():
for x_val, y_val in test_loader:
x_val = x_val.to(device)
y_val = y_val.view(y_val.shape[0],).to(device)
y_val -= 1
target = model(x_val)
loss = criterion(target, y_val)
valid_loss += loss.item() * y_val.shape[0]
_,predicted = torch.max(target.data,1)
val_total += y_val.size(0)
val_correct += (predicted == y_val).sum().item()
val_loss.append(loss.item()* y_val.shape[0])
val_acc = 100*val_correct / val_total
print(f'Epoch {epoch + 1} \t\t Training Loss: {train_loss / len(train_loader)} \t\t Validation Loss: {valid_loss / len(test_loader)} \t\t Train Acc:{accuracy} \t\t Validation Acc:{val_acc}')
if min_valid_loss > (valid_loss / len(test_loader)):
print(f'Validation Loss Decreased({min_valid_loss:.6f}--->{valid_loss / len(test_loader):.6f}) \t Saving The Model')
min_valid_loss = valid_loss / len(test_loader)
state = {'state_dict': model.state_dict(),'optimizer': optimizer.state_dict(),}
torch.save(state,'/home/arisa/Desktop/Hamid/IQA/checkpoint.t7')
end = time.time()
print('TRAIN TIME:')
print('%.2gs'%(end-start))
train_model(model=model, train_loader=train_loader, optimizer=optimizer, criterion=criterion, valid_loader= test_loader,num_epochs=500 )
提前致谢 这是15个epoch
的结果Epoch 1/500
----------
Epoch 1 Training Loss: 205.63448420514916 Validation Loss: 233.89266112356475 Train Acc:39.36360386127994 Validation Acc:24.142040038131555
Epoch 2/500
----------
Epoch 2 Training Loss: 199.05699240435197 Validation Loss: 235.08799531243065 Train Acc:41.90998291820601 Validation Acc:24.27311725452812
Epoch 3/500
----------
Epoch 3 Training Loss: 199.15626737127448 Validation Loss: 236.00033430619672 Train Acc:41.1035633416756 Validation Acc:23.677311725452814
Epoch 4/500
----------
Epoch 4 Training Loss: 199.02581041173886 Validation Loss: 233.60767459869385 Train Acc:41.86628530568466 Validation Acc:24.606768350810295
Epoch 5/500
----------
Epoch 5 Training Loss: 198.61493769454472 Validation Loss: 233.7503859202067 Train Acc:41.53656695665991 Validation Acc:25.0
Epoch 6/500
----------
Epoch 6 Training Loss: 198.71323942956585 Validation Loss: 234.17176149830675 Train Acc:41.639852222619474 Validation Acc:25.369399428026693
Epoch 7/500
----------
Epoch 7 Training Loss: 199.9395153770592 Validation Loss: 234.1744423635078 Train Acc:40.98041552456998 Validation Acc:24.84509056244042
Epoch 8/500
----------
Epoch 8 Training Loss: 199.3533399020355 Validation Loss: 235.4645173188412 Train Acc:41.26643626107337 Validation Acc:24.165872259294567
Epoch 9/500
----------
Epoch 9 Training Loss: 199.6451746921249 Validation Loss: 233.33387595956975 Train Acc:40.96452548365312 Validation Acc:24.59485224022879
Epoch 10/500
----------
Epoch 10 Training Loss: 197.9305159737011 Validation Loss: 233.76405122063377 Train Acc:41.8782028363723 Validation Acc:24.6186844613918
Epoch 11/500
----------
Epoch 11 Training Loss: 199.33247244055502 Validation Loss: 234.41085289463854 Train Acc:41.59218209986891 Validation Acc:25.119161105815063
Epoch 12/500
----------
Epoch 12 Training Loss: 199.87399289874256 Validation Loss: 234.23621463775635 Train Acc:41.028085647320545 Validation Acc:24.49952335557674
Epoch 13/500
----------
Epoch 13 Training Loss: 198.85540591944292 Validation Loss: 234.33149099349976 Train Acc:41.206848607635166 Validation Acc:24.857006673021925
Epoch 14/500
----------
Epoch 14 Training Loss: 199.92641723337513 Validation Loss: 233.37722391070741 Train Acc:41.15520597465539 Validation Acc:24.988083889418494
Epoch 15/500
----------
Epoch 15 Training Loss: 197.82172771698328 Validation Loss: 234.4943131533536 Train Acc:41.69943987605768 Validation Acc:24.380362249761678
您通过
冻结您的模型for param in model.parameters():
param.requires_grad = False
这基本上是说“不计算任何权重的任何梯度”，这相当于不更新权重 - 因此没有优化
我的问题在 model.train()。这个短语应该在训练循环中。但在我的例子中，我把它放在训练循环之外，当涉及到 model.eval() 时，模型保持在这种模式下