pytorch 使用GPU报错 ->RuntimeError: CUDA out of memory. Tried to allocate 1.50 GiB

最新推荐文章于 2025-07-22 19:39:28 发布

北部湾的落日

最新推荐文章于 2025-07-22 19:39:28 发布

阅读量5.3w

点赞数 12

CC 4.0 BY-SA版权

分类专栏：深度学习 pytorch googlenet 机器学习

本文链接：https://blog.youkuaiyun.com/qq_36213248/article/details/95466392

深度学习同时被 3 个专栏收录

6 篇文章

订阅专栏

机器学习

4 篇文章

订阅专栏

pytorch

3 篇文章

订阅专栏

博主在更新PyTorch到1.1.0版本后，设置bitch_size为64时出现CUDA内存溢出，改小到2仍报错，使用复杂的GoogleNet模型时出现该问题。按参考修改代码后又出现新报错，最终通过设置x_train的requires_grad=True解决。

RuntimeError: CUDA out of memory. Tried to allocate 1.50 GiB (GPU 0; 10.92 GiB total capacity; 9.79 GiB already allocated; 539.44 MiB free; 10.28 MiB cached)

本人的pytorch的版本是1.1.0,这个是我pytorch版本更新后，我已开的bitch_size=64,出现内存不够，但是我把bitch_size改小，bitch_size=2,依然出现内存溢出报错。本文使用的模型是较为复杂的GoogleNet(使用简单的VGG是没报错),我的源码是：

# 判断gpu是否可用
if torch.cuda.is_available():
    device = 'cuda'
else:
    device = 'cpu'
# device = 'cpu'
#device = 'cpu'
device = torch.device(device) 

cnn = GoogLeNet().to(device)
optimizer = Adam(cnn.parameters(), lr=0.001, betas=(0.9, 0.999))  # 选用AdamOptimizer
loss_fn = nn.SmoothL1Loss()  # 定义损失函数

# 训练并评估模型

data = Dataset(epochs=args.EPOCHS, batch=args.BATCH, val_batch=args.BATCH)
model = Model(data)

lowest_loss = 1e5
for i in range(data.get_step()):
    cnn.train()
    x_train, y_train = data.next_train_batch()
    x_test, y_test = data.next_validation_batch()

    x_train = torch.from_numpy(x_train)
    y_train = torch.from_numpy(y_train)
    x_train = x_train.float().to(device)
    y_train = y_train.float().to(device)

    outputs = cnn(x_train) # 问题出错点

    optimizer.zero_grad()
    #print(x_train.shape,outputs.shape,y_train.shape)
    loss = loss_fn(outputs, y_train)
    loss.backward()
    optimizer.step()
    print(loss)

解决：

out of memory解决参考来源参考链接

把outputs = cnn(x_train)改为

    with torch.no_grad():
        outputs = cnn(x_train)

改完后，但是又爆出新的bugs

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn
然后又参考这个博客参考链接

修改x_train和x_train的requires_grad=True就能解决

解决后的代码为：

# 判断gpu是否可用
if torch.cuda.is_available():
    device = 'cuda'
else:
    device = 'cpu'
# device = 'cpu'
#device = 'cpu'
device = torch.device(device) 

# cnn = Net().to(device)
cnn = GoogLeNet().to(device)
optimizer = Adam(cnn.parameters(), lr=0.001, betas=(0.9, 0.999))  # 选用AdamOptimizer
loss_fn = nn.SmoothL1Loss()  # 定义损失函数

# 训练并评估模型

data = Dataset(epochs=args.EPOCHS, batch=args.BATCH, val_batch=args.BATCH)
model = Model(data)

lowest_loss = 1e5
for i in range(data.get_step()):
    cnn.train()
    x_train, y_train = data.next_train_batch()
    x_test, y_test = data.next_validation_batch()

    x_train = torch.from_numpy(x_train)
    y_train = torch.from_numpy(y_train)
    x_train = x_train.float().to(device)
    y_train = y_train.float().to(device)

    x_train = Variable(x_train, requires_grad=True)
    y_train = Variable(y_train, requires_grad=True)
    with torch.no_grad():
        outputs = cnn(x_train)

    optimizer.zero_grad()
    #print(x_train.shape,outputs.shape,y_train.shape)
    loss = loss_fn(outputs, y_train)
    loss.backward()
    optimizer.step()
    print(loss)