Trainers
Trainers are responsable for create and train a neural network model to be used by an Agent.
Trainers only appear on the App Training Config Screen if there are a description file in json format, along with the trainer module inside the game/trainers/
directory. The description file should be named the same as the module file.
The trained models have to be stored on game/models/
folder, if not, the APP won't be able to detect it. Also the name of the trained model have to start with the Agent model_key
defined on the Agent class.
Directory structure
game/ trainers/ new_trainer.py new_trainer.json models/ model_name.h5
JSON File example
{ "name": "Name of the trainer", "short_description": "short description", "description": "full description" }
APP Requirements
To work properly with the app the trainer must implement a function named start
which will receive as argument a function log
. The log function receive as argument a string which will be shown on the log screen in the app.
It's also required to implement the variables:
- kwargs
- dictionary, all kwargs informed by the user in the "Variables" on the Training Config Screen.
- kill_training
- boolean, flag setted to True when the user press the "Cancel" button on the Training Screen.
Example
kwargs = None kill_training = False def start(log): log('Starting traning...') for kw, value in kwargs.items(): if kw == 'quantity_games': quantity_games = eval(value) # ...