Results from first notebook dramatically different

I have searched the forums for others with the same issue but didn’t find anything.
I have run the first notebook without any changes, and my results are around 84% (0.16 error rate)
Even with the res50, this only goes down by less than a percent.
As this is 4-5x worse than expected, without any changes, I feel that something is wrong with either my setup, the notebook, the code or the data.
I am using the gradient setup on paperspace with the fastAI ‘default’. This did run MUCH quicker than Jeremy’s demo.
I feel that I need to understand what is causing this, else there would be no point in continuing the course.

Hi !

Could you gave us more information to help you debug your issue ? For example screenshots of your notebook, or even better a gist of your notebook would be great.

The Notebook is the exact one used by Jeremy in the course.
When I get some time I will run it again and post the results.

Here are my results using resnet34:


I’m using the P5000 Gradient using the fastAI setup as per the docs

BASE CONTAINER: Paperspace + Fast.AI 1.0 (V3)

BASE CONTAINER URL: paperspace/fastai:1.0-CUDA9.2-base-3.0-v1.0.6

The results above are from a fresh download of the first notebook, with no changes.

Ok, I seemed to have fixed this issue.
It appears it was a data issue. When I originally untared the data, only 10% were actually saved.
I deleted the data, and redownloaded and now I’m getting much better results!

So the key is to have enough data to train!


I was going to ask you to show the data, congrats for solving it yourself !

1 Like