A project by @dberger @jbarment @ldevelle @llenotre @gilles595
A big thank to Qarnot who supports us through this endeaviour by offering us cloud computing. If you would like to see how we interact with their platforn to launch our calculations, Here's our wrapper repository
From base directory (or model save and load is broken):
python3 Archi/train_simulator.py --sim ../../DonkeyCar/DonkeySimLinux/donkey_sim.x86_64 --model 'new_model.h5'
-
Input data
- Interaction with simulator
- From datasets
-
Preprocessing
- From raw data
-
Model training
- reward optimisation
- all hyper_parameters given in
init()
- followup of metrics (loss / accuracy)
-
Model evaluation
- Saving the model, with HyperParams
- Evaluate the model
- Multiprocessing -> Takes complete architecture in hand
- Qarnot computing -> HyperParams optim : https://github.com/ezalos/Qarnot_Wrapper
- Distance control of WS : https://github.com/ezalos/emails
Source the .env file, import Client, Server and start_server from the package.
use
c = Client()
c.request_simulator()
c.kill_sim()
Where we train the agent:
export PS="wesh" ; python3.8 srcs --sim simlaunch3000 --model 'new_model.h5' --agent DDQN
Where we run the simulator:
cd srcs/simlaunch3000
export PS="wesh" ; export SIM_PATH="/home/ezalos/Downloads/DonkeySimLinux/donkey_sim.x86_64" ; python3.8 test_server.py