alexstaravoitau / behavioral-cloning Goto Github PK

Hey! Awesome code, read you blog post and found your data augmentation methods awesome.

I checked out the data.py file, and noticed you were using np.append to add a new image to the feature set. It is much faster to first use Python lists, and then convert them numpy arrays later. It can be from 10 -30 times faster.

So you could do something like:

x = []
y = []
...
x.append(image)
y.append(angle)
...
x = np.asarray(x)
y = np.asarray(y)

Hi, I am interested in that how do you decided the steer adjustment(in your code,0.25) for left and right image?Thanks in advanced!

Hi, thank you for the code. I have the following issue:

Command used:
python drive.py model.json

Error:
Using Theano backend.
Can not use cuDNN on context None: Disabled by dnn.enabled flag
Mapped name None to device cuda: GeForce GTX TITAN X (0000:01:00.0)
(13106) wsgi starting up on http://0.0.0.0:4567
(13106) accepted ('127.0.0.1', 53040)
127.0.0.1 - - [06/Sep/2018 15:54:56] "GET / HTTP/1.1" 404 366 0.006205

Only change I made is in my theanorc file to disable the dnn flag.

Please let me know how exactly to run this code.

Hi,
I want to know your final result value of loss and val loss to compare it with my result
thank s.

Hi Alex, I really appreciate your efforts in sharing your code and thoughts! This is pretty helpful for understanding the whole pipeline to implement behavioral cloning!

Currently, we are trying to replicate the behavioral cloning with the same model on jetson racecar to make it run indoors but it didn't work well.

I have a few questions in the data collection section in https://navoshta.com/end-to-end-deep-learning/

This would contain both driving in “smooth” mode (staying right in the middle of the road for the whole lap), and “recovery” mode (letting the car drive off center and then interfering to steer it back in the middle).

1. What's the ratio of the data between "smooth" mode and "recovery" mode?
2. In the "recovery" mode, when do you start to log the data (by data I mean everything that is going to be fed into the network like frames, steering angle and so on) ?
   2.1) start to log data at the moment when you drive the car off the center and then interfering to steer it back in the middle. Or
   2.2) drive the car off the center, and then start to log data at the moment when you steer it back in the middle?

Look forward to your answers.
Thanks!

Hi Sir,
where can we download the training data? Can you give some tips?

Thanks

Can u write about how to start up your code fully in anaconda ...