C5W1: Jazz music generation, function "music_inference_model"

Course Q&A Deep Learning Specialization Sequence Models
1

Hello,

All previous tests are passed, and I think my code is correct. I read some topics regarding problems with dimensions, but they did not fix my problem.

tf.math.argmax (x = tf.math.argmax(x,axis=-1)) is not returning the dimensions it should and I do not understand why:

after argmax: (None, 1) , must be (None,)
after one_hot: (None, 1, 90) , must be (None, 90)

Since the “1” gets added, the function raises the following error in the line “x = RepeatVector(1)(x)”:

ValueError: Input 0 of layer repeat_vector is incompatible with the layer: expected ndim=2, found ndim=3. Full shape received: [None, 1, 90]

What might be incorrect?

Kind regards,

Sara

2

Hi, @Sara_Valiente!

Try with tf.squeeze, as it is shown in the Tensorflow documentation.

I think that tf.squeeze(output, 1) should return what you want in the one-hot encoding vector

3

Hello @alvaroramajo!

Thank you for your help; I inserted the tf.squeeze in the argamax argument instead, since I think that was the root of the problem, and now the function appears to be working well. However, I still get an error one cell after, in the comparator. Any clue why that might be happening?

I’ll paste my code for clarity.

{moderator edit - solution code removed}

Thank you in advance!

Sara

4

@Sara_Valiente please check what size does x variable have before RepeatVector(1) and try to use tf.squeeze there to get the tensor size you want

5

tf.squeeze() is not needed in this assignment.

6

The problem is you’re using the wrong parameter when you call tf.math.argmax().
The instructions say:
# Select the next value according to "out",

7

That was it! What a silly mistake. Thank you both @alvaroramajo @TMosh!!