Hi there,
Keeping to the required model architecture, number of epochs (3) and learning rate (0.001), the average structural similarity index I can achieve is 0.87. Any suggestions on how to improve this? Below is my do_salience function, as well as the compile and fit operations.
[Removed code from moderator]
Have a look on this post maybe you are facing the same issue SSI
Thank you for the response! I solved it in the end…turned out to be a simple syntax error.
I had also failed submission, except that my scores were much lower around 0.03 on average. It turned out it was because I did not take the absolute values of the gradients (for each color channel) before doing tf.reduce_sum() to turn it into the grayscale_tensor.