Hi everyone,
I am going through the part 1 videos and wanted to test the lesson 2 notebook on my own dataset. I use fastai packages on my local system (Win 10). When running this cell:
data.show_batch(rows=3, figsize=(7,8))
I get the following error:
RuntimeError: cuda runtime error (30) : unknown error at …\aten\src\THC\THCGeneral.cpp:87
Anyone got an idea what to do to solve this?
And beside this error, running the cell below:
takes ages to run and does not finish running even after a long time waiting?
I tried to use Google’s Colab service as the cells above work fine there, but wasn’t familiar with how to use my own dataset in there and also wanted to know what is wrong with my own laptop.
Thanks in advance
Hi @zenlytix, thanks for the reply. Error is gone and looks like jupyter has finally found access to my GPU. Your suggestion worked perfectly. Thank you!
Just one thing I noticed right now. In task manager my GPU usage is only about one percent. Is there any way to increase this to a higher value in order to decrease the model training time? And is it even beneficial to do so?
Normally it is good to keep gpu highly utilized during model training. When moving data to gpu util will be low but should see close to peak utilization once training loop starts. Please check with nvidia-smi command to see more detailed utilization on GPU.
Thanks. I searched about the nvidia-smi command and saw a couple of articles. But all of them were about observing the amount of utilization and were missing the part on how to increase this amount. Have you any idea about it?
You can try to increase batch size. Often the model may not be “deep” enough to fully utilize the available GPU. Anyway there are a few threads on this topic on the forums like this one.