How to fix 'The kernel appears to have died. It will restart automatically" caused by pytorch

Question

I have a strange problem with Pytorch. When I use torch functions with tensors like tensor.reshape or torch.transpose, I don't have any problems; even when I created networks, it's ok. However, when I want to train network my jupyter crashed.

crashed

I find where error is but I don't know why it is there and how to fix it.

errors

I installed pytorch using conda. I have Ubuntu 18.04. I don't have cuda.

Batch size? You may run out of RAM memory or GPU memory, try to lower it down. — Szymon Maszke, Jun 25 '19 at 17:40
tried reduce batch size - Doesn't work. In keras this example work. — Elvar G, Jun 25 '19 at 19:08

score 4 · Answer 1 · answered Jun 11 '21 at 13:40

The hint for the solution is available, via the Jupyter notebook terminal (if opened it directly and not via the anaconda interface), where a more proper error code appears

OMP: Error #15: Initializing libiomp5md.dll, but found libiomp5md.dll already initialized. OMP: Hint This means that multiple copies of the OpenMP runtime have been linked into the program....

which then leads to a working solution here sklearn OMP: Error #15 when fitting models

score 3 · Answer 2 · answered Jun 25 '19 at 21:03

3

If you are on Ubuntu you may not install PyTorch just via conda.

It can be:

Conda
Pip
LibTorch
From Source

So you have multiple options.

Go to this page and select Cuda to NONE, LINUX, stable 1.1, CONDA.

conda install pytorch-cpu torchvision-cpu -c pytorch

If you have problems still, you may try also install PIP way.

pip3 install https://download.pytorch.org/whl/cpu/torch-1.1.0-cp36-cp36m-linux_x86_64.whl
pip3 install https://download.pytorch.org/whl/cpu/torchvision-0.3.0-cp36-cp36m-linux_x86_64.whl

Hopefully some of these ways will work.

answered Jun 25 '19 at 21:03

prosti

42,291
14
186
151

I don't know how to fix it, but in GoogleColab i can work and my code is work – Elvar G Jun 28 '19 at 13:17
1

in my case problem resolved when i reinstalled Pytorch and cudatoolkit via pip command instead of conda. – M. Mufti Dec 15 '21 at 06:37

score 1 · Answer 3 · answered May 01 '22 at 06:15

Simple Solution

I have solved this problem by reinstalling the torchvision, as this was the only library that was creating an issue, whenever i import that

Just reinstall that:

Directly in Jupyter notebook:

!pip3 uninstall -y torch torchvision
!pip3 install torch torchvision

or in Terminal:

pip3 uninstall -y torch torchvision
pip3 install torch torchvision

score 0 · Answer 4 · answered Jan 23 '23 at 12:12

I had sort of the same issue (on a macos 12.6.2). A kernel crash when using torchvision.

trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
                                        download=True, transform=transform)   
testset = torchvision.datasets.CIFAR10(root='./data', train=False,
                                       download=True, transform=transform)

classes = ('plane', 'car', 'bird', 'cat',
           'deer', 'dog', 'frog', 'horse', 'ship', 'truck')

batch_size = 64
trainloader = th.utils.data.DataLoader(trainset, batch_size=batch_size,
                                          shuffle=True, num_workers=0)
testloader = th.utils.data.DataLoader(testset, batch_size=batch_size,
                                         shuffle=False, num_workers=0)
dataiter = iter(trainloader)
images, labels = next(dataiter)

I solved it by reinstalling pytorch and torchvision. Run in a script (jupyter notebook or elsewhere):

!conda install --yes pytorch torchvision -c pytorch

How to fix 'The kernel appears to have died. It will restart automatically" caused by pytorch

4 Answers4

Linked