我在数据科学和 pytorch 方面没有那么丰富的经验，并且我在实现至少任何东西方面都遇到了问题（目前我正在为分割任务制作一个神经网络）。存在某种内存问题，尽管它没有任何意义 - 每个纪元占用的内存都比上升时要少得多
import torch
from torch import nn
from torch.autograd import Variable
from torch.nn import Linear, ReLU6, CrossEntropyLoss, Sequential, Conv2d, MaxPool2d, Module, Softmax, Softplus ,BatchNorm2d, Dropout, ConvTranspose2d
import torch.nn.functional as F
from torch.nn import LeakyReLU,Tanh
from torch.optim import Adam, SGD
import numpy as np
import cv2 as cv
def train(epoch,model,criterion, x_train, y_train, loss_val):
model.train()
tr_loss = 0
# getting the training set
x_train, y_train = Variable(x_train), Variable(y_train)
# converting the data into GPU format
# clearing the Gradients of the model parameters
optimizer.zero_grad()
# prediction for training and validation set
output_train = model(x_train)
# computing the training and validation loss
loss_train = criterion(output_train, y_train)
train_losses.append(loss_train)
# computing the updated weights of all the model parameters
loss_train.backward()
optimizer.step()
tr_loss = loss_train.item()
return loss_train
# printing the validation loss
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(3, 96, (3,3), padding=1)
self.conv11= nn.Conv2d(96, 96, (3,3), padding=1)
self.conv12= nn.Conv2d(96, 96, (3,3), padding=1)
self.pool = nn.MaxPool2d((2,2), 2)
self.conv2 = nn.Conv2d(96, 192, (3,3), padding=1)
self.conv21 = nn.Conv2d(192, 192, (3,3), padding=1)
self.conv22 = nn.Conv2d(192, 192, (3,3), padding=1)
self.b = BatchNorm2d(96)
self.b1 = BatchNorm2d(192)
self.b2 = BatchNorm2d(384)
self.conv3 = nn.Conv2d(192,384,(3,3), padding=1)
self.conv31= nn.Conv2d(384,384,(3,3), padding=1)
self.conv32= nn.Conv2d(384,384,(3,3), padding=1)
self.lin1 = nn.Linear(384*16*16, 256*2*2, 1)
self.lin2 = nn.Linear(256*2*2, 16*16, 1)
self.uppool = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False)
self.upconv1= nn.ConvTranspose2d(385,192,(3,3), padding=1)
self.upconv11=nn.ConvTranspose2d(192,32,(3,3), padding=1)
self.upconv12=nn.ConvTranspose2d(32,1,(3,3), padding=1)
self.upconv2= nn.ConvTranspose2d(193,96,(3,3), padding=1)
self.upconv21= nn.ConvTranspose2d(96,16,(3,3), padding=1)
self.upconv22= nn.ConvTranspose2d(16,1,(3,3), padding=1)
self.upconv3= nn.ConvTranspose2d(97,16,(3,3), padding=1)
self.upconv4= nn.ConvTranspose2d(16,8,(3,3), padding=1)
self.upconv6= nn.ConvTranspose2d(8,1,(3,3), padding=1)
def forward(self, x):
m=Tanh()
x1=self.b(m(self.conv12(m(self.conv11(m(self.conv1(x)))))))
x = self.pool(x1)
x2=self.b1(m(self.conv22(m(self.conv21(m(self.conv2(x)))))))
x = self.pool(x2)
x3=self.b2(m(self.conv32(m(self.conv31(m(self.conv3(x)))))))
x=self.pool(x3)
x = x.view(-1, 16*16*384)
x = m(self.lin1(x))
x = m(self.lin2(x))
x = x.view(1, 1, 16, 16)
x=torch.cat((x,self.pool(x3)),1)
x = self.uppool(m(self.upconv12(m(self.upconv11(m(self.upconv1(x)))))))
x=torch.cat((x,self.pool(x2)),1)
x = self.uppool(m(self.upconv22(m(self.upconv21(m(self.upconv2(x)))))))
x=torch.cat((x,self.pool(x1)),1)
x = (self.uppool(m(self.upconv3(x))))
x = (m(self.upconv4(x)))
l=Softplus()
x= l(self.upconv6(x))
return x
train_data=[]
for path in range(1000):
n="".join(["0" for i in range(5-len(str(path)))])+str(path)
paths="00000\\"+n+".png"
train_data.append(cv.imread(paths))
for path in range(2000,3000):
n="".join(["0" for i in range(5-len(str(path)))])+str(path)
paths="02000\\"+n+".png"
train_data.append(cv.imread(paths))
train_output=[]
for path in range(1,2001):
n="outputs\\"+str(path)+".jpg"
train_output.append(cv.imread(n))
data=torch.from_numpy((np.array(train_data,dtype=float).reshape(2000,3,128,128)/255)).reshape(2000,3,128,128)
data_cuda=torch.tensor(data.to('cuda'), dtype=torch.float32)
output=torch.from_numpy(np.array(train_output,dtype=float).reshape(2000,3,128,128))[:,2].view(2000,1,128,128)*2
output_cuda=torch.tensor(output.to('cuda'),dtype=torch.float32)
model=Net()
optimizer = Adam(model.parameters(), lr=0.1)
criterion = nn.BCEWithLogitsLoss()
if torch.cuda.is_available():
model = model.cuda()
criterion = criterion.cuda()
print(model)
epochs=3
n_epochs = 1
train_losses = []
val_losses = []
for epoch in range(n_epochs):
loss_train=0
for i in range(data.shape[0]):
loss_train1=train(epoch,model,criterion,data_cuda[i].reshape(1,3,128,128),output_cuda[i].reshape(1,1,128,128),train_losses)
loss_train+=loss_train1
print('Epoch : ',epoch+1, '\t', 'loss :', loss_train/data.shape[0])
with torch.no_grad():
torch.save(model.state_dict(), "C:\\Users\\jugof\\Desktop\\Python\\pytorch_models")
a=np.array(model(data_cuda).to('cpu').numpy())*255
cv.imshow('',a.reshape(128,128))
cv.waitKey(0)"""
这是错误：
> PS C:\Users\jugof\Desktop\Python> & C:/Users/jugof/anaconda3/python.exe c:/Users/jugof/Desktop/Python/3d_visual_effect1.py
c:/Users/jugof/Desktop/Python/3d_visual_effect1.py:98: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
data_cuda=torch.tensor(data.to('cuda'), dtype=torch.float32)
c:/Users/jugof/Desktop/Python/3d_visual_effect1.py:101: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
output_cuda=torch.tensor(output.to('cuda'),dtype=torch.float32)
Epoch : 1 loss : tensor(0.6933, device='cuda:0', grad_fn=<DivBackward0>)
Traceback (most recent call last):
File "c:/Users/jugof/Desktop/Python/3d_visual_effect1.py", line 120, in <module>
a=np.array(model(data_cuda).to('cpu').numpy())*255
File "C:\Users\jugof\anaconda3\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "c:/Users/jugof/Desktop/Python/3d_visual_effect1.py", line 62, in forward
x1=self.b(m(self.conv12(m(self.conv11(m(self.conv1(x)))))))
File "C:\Users\jugof\anaconda3\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "C:\Users\jugof\anaconda3\lib\site-packages\torch\nn\modules\conv.py", line 399, in forward
return self._conv_forward(input, self.weight, self.bias)
File "C:\Users\jugof\anaconda3\lib\site-packages\torch\nn\modules\conv.py", line 395, in _conv_forward
return F.conv2d(input, weight, bias, self.stride,
RuntimeError: CUDA out of memory. Tried to allocate 11.72 GiB (GPU 0; 6.00 GiB total capacity; 2.07 GiB already allocated; 1.55 GiB free; 2.62 GiB reserved in total by PyTorch)
I feed a numpy array (an image) of 128*128 shape and recieve another of the same shape, it's a segmentation model(again)
我使用的是 Flickr-Faces-HQ 数据集（FFHQ）并使用了下采样的 128*128 标签 - 我使用了 00000、01000 和 02000 文件，并且 opencv haarscascades_eye 收到了掩码
问题是你的loss_train列表，它存储了从实验开始以来的所有损失。如果您输入的损失只是浮点数，那不会是问题，但由于您没有在函数中返回浮点数train，因此您实际上是在存储损失张量，并将所有计算图嵌入其中。事实上，张量保留了参与其计算的所有张量的指针，只要指针存在，分配的内存就无法释放。
所以基本上，你保留所有时期的所有张量并防止 pytorch 清理它们；这就像（故意的）内存泄漏
nvidia-smi -l 1您可以通过在开始实验后运行来轻松监控此类问题。您将看到内存使用量线性增长，直到您的 GPU 内存耗尽（“nvidia-smi 是在 GPU 上执行操作时使用的一个很好的工具”）。
为了防止这种情况发生，只需将函数的最后一行替换train为return loss_train.item()，内存问题就会消失
| 归档时间： |
|
| 查看次数： |
3100 次 |
| 最近记录： |