Test the trained network in simulation

[kargarisaac]

I modified the network and trained a new network. Now I want to use the main.py file to test it. But it wants pkl file. Why does it need data file? When I set the record=True to create a pkl file it says that I have to drive but the GUI doesn't show anything. I don't know why should I gather data. I used your data file and trained the model.

Thanks

[Iftimie]

Hi!

When testing as well as training, the neural network requires a GPS-like path. Such a path can be found in the .pkl file that has just been recorded. The pickle contains a list of states and among them is the coordinates of the car. To find the interval of states to be drawn, we need to find which recorded state is closest to the car, and start drawing the fat GPS line from that state up + Config.path_future_len

ChauffeurNet/simulator/control/car_controller/NeuralController.py

Line 34 in deddf29

path_idx = self.path.get_point_idx_close_to_car(self.vehicle, path_idx)

ChauffeurNet/simulator/util/Path.py

Line 43 in deddf29

def get_point_idx_close_to_car(self, vehicle, path_idx):

ChauffeurNet/simulator/util/Path.py

Line 65 in deddf29

selected_for_projection = self.vertices_W[:,path_idx:path_idx+ Config.path_future_len]

A good thing to do might be to decouple the train and test GPS paths, but at that time I did not cared that much about over fitting.
If you used the same data file to train the new model, just make sure that when Simulator is instantiated, the relative path to the .pkl file can be found and loaded. You don't need to record new data.
If you fork the repo and commit your modifications, maybe I could guide you better.
Hope I answered your question.

[kargarisaac]

Thank you. I will test it after the training is finished tomorrow.