So I come up with an idea to use those powerful cloud servers to help us train faster. Particularly with AWS, we can use spot instance to get a powerful server at a cheap rate to train the model. In order to streamline the whole process to create the instance, upload the tubs and download the final model, I have created the Donkey Car Console.
Donkey Car Console is a web-based application running directly on the Pi that can do the followings:
- A few simple clicks to upload the tubs, train the model using the instance type of your choice and download the model back to the Pi. Normally a training completes within 10 minutes
- Manage tubs and model using a GUI. Delete useless tubs and assign remarks to good tubs or model.
Best of all, donkey car console is open source and it is FREE for you to use it. You just need to pay for the EC2 instance hour you used, which is around US$0.1-0.2 for each training assuming each training complete within 10 mins (Disclaimer: I don't work for AWS). I have written a detail instructions on how to use it on github. Feel free to try it and give me some feedback.
* Note: Donkey Car Console requires you to create a AWS key. Take good care of the key and set a budget alarm. You can even disable the key after you finished using it. There are bad guys who steal AWS key to make money.
How about Google Colab?
After I have developed the Donkey Car Console, some other folks suggest that we can also use Google Colab to train the model. Since Google Colab is free and equipped with GPU, it could be a very good alternative to train the model on the cloud. Personally I haven't tried it yet but I have seen people using it. If you are interested to know how to train Donkey car using Google Colab, let me know and I would see if I can put together some examples to show how it works.