Angles Comparasion Loss Function for Tensorflow in Python

Question

I have a CNN, takes in an image, outs a single value - an angle. The data set is made of (x = image, y = angle) couples.

I want the network for each image, to predict an angle.

I have found this suggestion: https://stats.stackexchange.com/a/218547 But I can't seem to understand how to translate it into a working Tensorflow in Python code.

x_CNN = tf.placeholder(tf.float32, (None, 14, 14, 3))
y_CNN = tf.placeholder(tf.int32, (None))
keep_prob_CNN = tf.placeholder(tf.float32)
one_hot_y_CNN = tf.one_hot(y_CNN, 1)
def MyCNN(x):
    # Network's architecture: In: Image, Out: Angle.
logits_CNN = MyCNN(x)

# Loss Function attempt <------------------------------
outs = tf.tanh(logits_CNN)
outc = tf.tanh(logits_CNN)
loss_operation_CNN = tf.reduce_mean(0.5 * (tf.square(tf.sin(one_hot_y_CNN) - outs) + tf.square(tf.cos(one_hot_y_CNN) - outc)))

learning_rate_placeholder_CNN = tf.placeholder(tf.float32, shape=[])
optimizer_CNN = tf.train.AdamOptimizer(learning_rate = learning_rate_placeholder_CNN)
training_operation_CNN = optimizer_CNN.minimize(loss_operation_CNN)
correct_prediction_CNN = tf.equal(tf.argmax(logits_CNN, 1), tf.argmax(one_hot_y_CNN, 1))
accuracy_operation_CNN = tf.reduce_mean(tf.cast(correct_prediction_CNN, tf.float32))

# And a working Training and testing code...

Answer 1

That is going in the right direction, but the idea is that, instead of having MyCNN produce a single angle value for each example, produce two values. So if the return value of MyCNN is currently something with shape like (None,) or (None, 1), you should change it to (None, 2) - that is, the last layer should have one more output. If you have doubts about how to do this please provide more details about the body of MyCNN.

Then you would just have:

outs = tf.tanh(logits_CNN[:, 0])
outc = tf.tanh(logits_CNN[:, 1])
out_radians = tf.atan2(outs, outc)  # This is the final angle output in radians

About the loss, I am not sure I understand your Y input. If you are trying to predict an angle, shouldn't it be a float value, and not an integer? In that case you would have:

# Example angle in radians
y_CNN = tf.placeholder(tf.float32, (None,))

# ... 

loss_operation_CNN = tf.reduce_mean(0.5 * (tf.square(tf.sin(y_CNN) - outs) +
                                           tf.square(tf.cos(y_CNN) - outc)))