Camvid - Accuracy

Part 1 (2019)
1

Hi everyone

I am not able to achieve nearly the same accuracy on Camvid as shown in Lesson 3. I am running the notebook ‘lesson3-camvid.ipynb’ without any changes to the NB on Salamander, but even in the basic training I have the validation accuracy diverging from training:

epoch train_loss valid_loss acc_camvid
1 1.528252 4.490077 0.233769
2 0.904275 3.284046 0.016770
3 0.763168 2.912649 0.270058
4 0.655638 4.700518 0.100856
5 0.593151 3.962085 0.235185

Any ideas what could be going wrong?

Thanks
Ron

2

Can you link to your notebook ?

1 Like
3

I am rerunning on Colab and don’t have the same issue. Really weird. Maybe something to do with the version of the library…

4

i am trying to do a binary dataset in the gcu and my acc is all over the place, including 0 sometimes

5

Try to kill the kernal and start again. I had some similar issues. But couldn’t find a way to reproduce it.

1 Like
6

Ron, I got the same 0.2-0.3 accuracy at one point after playing around. I blew away the files in data folder and created them from scratch and all was fine after that. Not sure if you have tried - but blowing away your data and re-downloading it (or at least re-unzipping it) from scratch may help?

1 Like
7

Very interesting! I wonder what could be the cause… All I know is when I switched to Colab the problem went away.

8

@ronrich had the same issue and got away with it when i rebooted my machine…https://forums.fast.ai/t/general-course-chat/24987/225?u=raghavab1992