search

udacity / sdc-issue-reports

29 stars 6 forks source link

Behavioral Cloning: Keras version of Car simulator and workspace are not compatible. #1429

Closed vaidhin closed 5 years ago

vaidhin commented 5 years ago

For Behavioural Cloning project,I trained my model in workspace and opened simulator in GPU enabled work space. When I entered command python "drive.py model.h5" , I got this error "You are using Keras version b'2.0.8' , but the model was built using b'2.0.9' Segmentation fault (core dumped).

After having few discussions with mentors it was concluded that workspace is not compatible to test our trained model in the Behavioural cloning project. Fix would be appreciated.

mvirgo commented 5 years ago

Hello,

It looks like the classroom is currently using 2.0.9 on jupyter workspaces but 2.0.8 on GPU terminal workspaces. I'll flag for the workspaces team to look into. Thanks for noting!

Note that this is only an issue if you train your model outside of the workspace - the intended functionality is for you to train your network inside the workspace, and then test it there. If you already have a GPU to train your network, you should be able to run the simulator locally on your machine as well.

vaidhin commented 5 years ago

I trained my network inside workspace and tried to test with simulator inside workspace. But still there is an incompatibility issue.

mvirgo commented 5 years ago

If you trained the network from scratch in the GPU workspace and tested in the workspace, there should not be a compatibility issue - you are using the exact same environment for both. Did you use any type of pre-trained or pre-saved model?

vaidhin commented 5 years ago

Now I got the point......I havent use any pre trained or presaved models.... BUT I trained network without GPU in workspace.. then I enable GPU and tested it. Thanks for your reply.. It saved lot of my effort

mvirgo commented 5 years ago

Ah gotcha! Glad that helped. I'd suggest using the GPU to train (even outside of this particular compatibility difference) - it should be 20-30X faster depending on your network! You only want it shut off during the actual coding of the network so as to avoid using up your hours when the GPU isn't actively needed.