Kernel dies when trying to run PyTorch code on Linux

[ccordoba12]

Please see this SO question for the details:

https://stackoverflow.com/questions/45875837/spyder-just-kernel-died-restarting-when-i-run-net-forward-in-pytorch

@andfoy, please take a look at this one.

[andfoy]

@ccordoba12 Does mixing Spyder versions from PyPi and conda might have some responsibility here? According to the package dump posted by the user, Spyder 3.1.4 and 3.2.1 are both listed:

spyder                    3.2.1                     <pip>
spyder                    3.1.4                    py27_0

[ccordoba12]

Probably, but did you try to run the OP code?

[andfoy]

The code posted by the OP works as expected:

Actually, it works if I use a deeper model such as the ResNet-101:

It works also on the GPU

[selous123]

Thank you very much for your answering.

also i run this code with the same version in my other computer.it works well!!!

I try to solve the mixing version problem. delete the redundant package 3.1.4.but still "kernel died,restarting"

first I doubt whether Spyder and pytorch have collision in the kernel.
Obviously,The problem should be in the configuration of my own computer
Sadly,there is no error message in the terminal about the kernel died

[ccordoba12]

It seems this is fixed in the latest version of PyTorch:

jupyter/notebook#2784 (comment)

[selous123]

not only it has errors in the spyder,but also in jupyter and terminal..

Problem Description

[selous123]

Just now,I update all the packages in anaconda

conda update --all

delightedly the code runs well in terminal and jupyter,but the error in the spyder also exists

[andfoy]

Well, I'm running the latest package versions on my current setup, and it works as expected. Any ideas @ccordoba12 ?

[andfoy]

Also, the user reports that on a new installation, the package works. @selous123 Have you tried to create a new conda environment on your conflicting PC?

[selous123]

everything goes well

when i use Source installation instead of "pip install..." as official website proposal to reinstall my pytorch and it works.

thank you very much

[andfoy]

@selous123 For future installations, PyTorch developers recommend to install the package using conda:

conda install pytorch torchvision -c soumith

I'm glad that it works as expected