Hi, ive had a go at changing the learning rate to 0.001 and ramped up the epoc’s to 20000 and i still don’t get the exact same picture; are there other parameters I need to consider to get a one to one match? getting there but not the same, no poppies around the pyramid.
i am also running this locally.
Hi Ernie,
I am not sure how the final picture was arrived at that is presented in the notebook. But in light of ongoing improvements to the code of the notebook it is probably best to see the presented final picture as illustrative rather than absolute.
Hello Ernie,
How did you get this output image after changing learning rate to 0.001 and epochs to 20000? I tried the same experiment but this is the image I end up with:
Is your alpha and beta still 10 and 40? Did you start with the content image specified in the notebook or a completely white noise image? What other hyperparameters did you change?
Regards,
Rahul
Hi,
I did originally get something like that and i noticed that i had done something wrong with the a_S, a_C and a_G in the compute_layer_style_cost… or train_step 
Did you get “All tests passed” for your compute_layer_style_cost or train_step cells, before you made this correction? What was your mistake and how did you correct it?
I’ve experimented a lot with hyperparameters but not getting anywhere close to the output image you got, or the one in the notebook.
All my cell tests pass and autograder gave 100% for my notebook submission. It’s possible I made the same mistake as you did, so could you let me know what corrections you made?
i did get 100 % but the image output looked odd… all i did was adjust the reps to 20001 and rate to 0.001. however i noticed something in the bits i mentioned above. i cannot post the answer i guess as it’s not allowed 
Ok, could you give some hints r.e. what you noticed in compute_layer_style_cost or train_step? e.g. what print statements or output to look at?
i think it was in compute_layer_style_cost, i actually cant remember and coursera has zapped my copy on the site.
a_S = tf.reshape(tf.transpose(a_S, perm=[…]), shape=[…])
may have been around here ?
That part is correct in my notebook. I have the reshape and transpose reversed, and I don’t use the “perm” parameter in tf.transpose i.e.
a_S = tf.transpose(tf.reshape(a_S, shape=[…]))
I posted a new thread r.e. this.
I understand if you are not comfortable posting your answer, but there has to be another way to debug this problem. My compute_layer_style_cost is correctly implemented, so problem is somewhere else.
