search

SullyChen / Autopilot-TensorFlow

A TensorFlow implementation of this Nvidia paper: https://arxiv.org/pdf/1604.07316.pdf with some changes
MIT License
1.25k stars 425 forks source link

Unexpected train result? #12

Closed weijia-yu closed 7 years ago

weijia-yu commented 7 years ago

Thanks for your sample code, I tried this but couldn't get expected result. I am using wheel rotation degree to train the data, this is my sample train data:

22.jpg 0 23.jpg 0 24.jpg 0 25.jpg 0 26.jpg 0 27.jpg 0 28.jpg 0 29.jpg 0 30.jpg 0 31.jpg 0 32.jpg -3.851463 33.jpg -6.549802 34.jpg -11.0871 35.jpg -14.10705 36.jpg -18.85992 37.jpg -21.56775 38.jpg -20.37253 39.jpg -17.5227 40.jpg -13.01936 41.jpg -9.757766 42.jpg -5.45346 43.jpg -2.566003 44.jpg 0 45.jpg 0 46.jpg 0 47.jpg 0 48.jpg 0 49.jpg 0 50.jpg 0 51.jpg 0 52.jpg 0 53.jpg 0 54.jpg 1.5773

mostly it is 0, but when I eval the data, it outputs:

0.848186871676 0.247460795198 0.585712200104 0.410212570454 0.451668243538 0.933786110727 0.905919685965 0.66789506391 0.497750729456 0.352852214917 0.391191160626 0.959048846033 0.931835024502 0.174691773018 -0.0430651931555 -0.260673075776 -0.258237713131 -0.620764197316 -0.750217252903 -0.715262639247 -0.710381775394 -0.733180519957 -0.473712732446 -0.605725828311 -0.814070557019

There is no 0s and when it is supposed to be 8 degree, it is like 0.x degree, did I feed the right parameter? Should I feed the steer wheel's rotation? Thank you

SullyChen commented 7 years ago

I converted the degree values to radians before feeding it in to the model. That is the issue.

weijia-yu commented 7 years ago

@SullyChen Thanks for the fast reply! Correct me if I am wrong, the code takes the degree train data and in the code it converts into radians. When running run.py, it then converts back to degree.

weijia-yu commented 7 years ago

The first data block is fed into train.py, the second data block is got by writing new line of degrees degrees = model.y.eval(feed_dict={model.x: [image], model.keep_prob: 1.0})[0][0] * 180 / scipy.pi

Another question I have is do you use radius instead of 1/radius mentioned in the paper?

SullyChen commented 7 years ago

Oh that's true. I thought you had modified my code, I didn't expect that. Also are you aware that the first few hundred images are very noisy dark images in a parking lot? Those images won't be classified well regardless of how well you train the model. There are no clear road markings and the picture quality is poor. Try it with images much later in the dataset.

As for your other question, the paper uses 1/radius where radius is the radius of the car's turn. This is roughly proportional to the angle of the steering wheel, so essentially I do use the same 1/radius as the paper.

weijia-yu commented 7 years ago

Can I ask why you didnt use max_pool?

SullyChen commented 7 years ago

I didn't use max_pool because the Nvidia paper this model is based off of didn't use any max pooling.