Test model on train data - not corresponding answers

[kalanityL]

Hi @SullyChen

First of all thank you a lot for letting available your code that is very nice to read, plus the DataSet and pre-trained model. It makes it very handy for experimenting with it :)

Before testing on new data, I wanted to verify I installed all correctly, so I wrote a basic test file, to run the model on few samples from the training data. I was thinking it would output values very close to the ground truth. But it is not the case...

What am I missing? Or are the output from the run actually good?

I hope you can help.. :) In any case again thank you very much

K.

Images tested from the training data set: ['3252.jpg', '3253.jpg', '3254.jpg', '33255.jpg', '33256.jpg', '33257.jpg', '9501.jpg', '9502.jpg', '9503.jpg']

Corresponding values in data.txt file: 3252.jpg 2.420000 3253.jpg 2.420000 3254.jpg 2.420000

33255.jpg -37.210000 33256.jpg -37.210000 33257.jpg -37.310000

9501.jpg -17.240000 9502.jpg -17.450000 9503.jpg -17.550000

Output from the network: Predicted steering angle: -61.0882010686 degrees ie -1.06619 rad for image: 3252.jpg Predicted steering angle: -60.9035537346 degrees ie -1.06297 rad for image: 3253.jpg Predicted steering angle: -60.0010808395 degrees ie -1.04722 rad for image: 3254.jpg

Predicted steering angle: -27.5703796036 degrees ie -0.481194 rad for image: 33255.jpg Predicted steering angle: -28.3578594338 degrees ie -0.494938 rad for image: 33256.jpg Predicted steering angle: -36.3197870436 degrees ie -0.6339 rad for image: 33257.jpg

Predicted steering angle: 46.7075895259 degrees ie 0.815201 rad for image: 9501.jpg Predicted steering angle: 49.5550373342 degrees ie 0.864899 rad for image: 9502.jpg Predicted steering angle: 52.0180820961 degrees ie 0.907887 rad for image: 9503.jpg

The code for testing:

import tensorflow as tf
import scipy.misc
import model
import cv2
import glob

sess = tf.InteractiveSession()
saver = tf.train.Saver()
saver.restore(sess, "save/model.ckpt")

all_img = glob.glob("*.jpg")
all_img.sort()
for img_path in all_img:
    img = cv2.imread(img_path)
    image = scipy.misc.imresize(img, [66, 200]) / 255.0
    radian = model.y.eval(feed_dict={model.x: [image], model.keep_prob: 1.0})[0][0]
    degrees = radian * 180 / scipy.pi
    print("Predicted steering angle: " + str(degrees) + " degrees ie "+ str(radian) + " rad for image: "+img_path)

[SullyChen]

It's possible that the trained model included in the repo was trained on a different dataset (sometimes I update the dataset if I collect more data), and there was a shift in the camera or something in the new data that causes a big change in the distribution of the data. This kind of problem can be eliminated with data augmentation or other more robust training methods.