Question on training the model

[xubo92]

Hello @felipecode @juaxix @nsubiron @dosovits @marcgpuig ,

I'm currently working on training the model you provided in paper in order to reproduce the result. However, I found the loss is hard to decrease. It always around certain level, up and down again and again. I have tried different lr and optimizer but seems not work.

I check the label, AKA the first three columns in "target" dataset (steer, acc, brake) . I found the steer angle is too small most of time. Maybe that is the reason for hard training since the loss calculated every time is also small and the corresponding gradient is small too?

I want to know if you have done any pre-processing on the label? or is there any tricks for training?

Very thankful for your work and hope you can help me with these questions.

[felipecode]

Hey lvlvlvlvlv,
Yes the steering is mostly small.
I think for your case, the trick is to balance the data. Make sure that every minibatch covers a good range of steering values. Other potential problems:

Regularization, if it is too big it wont converge.
Optimizer, i used adam optimizer ( Not ideal , but should maybe help in your case)
Mini Batch size. Are you using a reasonably big one ? I was using 120 for training.

Cheers.

[xubo92]

@felipecode Thanks for your tips. I am trying them.
By the way, are you training the model to learn target as a continuous-value regression problem or a classification problem? Is the regression problem more difficult to find suitable params? I am thinking if there are any ways to divide the steer value into several range and turn it into a classification problem.