autorope / donkeycar

Open source hardware and software platform to build a small scale self driving car.
http://www.donkeycar.com
MIT License
3.11k stars 1.28k forks source link

How can I train my donkeycar on jetson nano #783

Open lee-junho21 opened 3 years ago

lee-junho21 commented 3 years ago

I'm using Jetpack4.4.1 on my jetson nano. And I installed donkey car 4.1 version. After then, I could drive my donkey car with ps4 controller. But, when I try to train my donkey car, I faced a problem. On jetson nano, I tried the train code

donkey train --tub ./mycar/data --model ./models/mypilot.h5 --type=linear

Then, Error occurs. like

AttributeError: 'NoneType' object has no attribute 'DEFAULT_AI_FRAMEWORK'

So, I added DEFAULT_AI_FRAMEWORK = tensorflow in myconfig.py Then, ModuleNotFoundError: No module named 'imgaug'

I also tried this code on my window computer, but I had same problem. How can I solve this problem. I really hope to train my donkeycar...

custom-build-robots commented 3 years ago

Hi @lee-junho21, the command you are using looks wrong to me.

Try the command for example as follows: python train.py --tubs=data/tub_2_21-01-16/,data/tub_1_21-01-16/,data/tub_3_21-01-16/ --model models/mypilot.h5

For some more details how to train on a jetson nano (what is not the best way but it is possible) I have published a pdf file describing the setup on a jetson nano and how to train and run a neural net on the jetson nano. Ingmars Donkey Car – E-Book

Or visit the project in the discord channel to get some help: discord server

Best,

Ingmar

TCIII commented 3 years ago

@lee-junho21,

DEFAULT_AI_FRAMEWORK = 'tensorflow'

kory2000pr commented 2 years ago

Did you resolve the imaaug problem???

Ezward commented 1 year ago

The library we use for image augmentations does not work correctly on Jetson Nano, so we do not install it on the Nano, that is why you are getting that error. Unfortunately the maintainer of imgaug has said that he will not fix it. https://github.com/autorope/donkeycar/issues/970

We have a separate issue to completely refactor the image augmentations and we can address this issue when we do that. https://github.com/autorope/donkeycar/issues/1109