C2W3 - UNQ_C3 and UNQ_C4 errors

Course Q&A Generative Adversarial Networks (GANS) Build Better Generative Adversarial Networks
1

Hello guys

Today I will ask you to help me with my mistakes :smiley:

UNQ_C3
in InjectNoise class, as nn.Parameter:

{moderator edit - solution code removed}

As forward instructions:

{moderator edit - solution code removed}

But I get the error:

AssertionError                            Traceback (most recent call last)
<ipython-input-204-8b78854b35cd> in <module>
     14 # Check that the change is per-channel
     15 assert torch.abs((inject_noise(fake_images) - fake_images).std(0)).mean() > 1e-4
---> 16 assert torch.abs((inject_noise(fake_images) - fake_images).std(1)).mean() < 1e-4
     17 assert torch.abs((inject_noise(fake_images) - fake_images).std(2)).mean() > 1e-4
     18 assert torch.abs((inject_noise(fake_images) - fake_images).std(3)).mean() > 1e-4
AssertionError:

UNQ_C4
in AdaIN class, the init code:

{moderator edit - solution code removed}

As forward instructions:

{moderator edit - solution code removed}

But i get the error:

AssertionError                            Traceback (most recent call last)
<ipython-input-203-30b30eb31fdf> in <module>
     23 test_w = torch.ones(n_test, w_channels)
     24 test_output = adain(test_input, test_w)
---> 25 assert(torch.abs(test_output[0, 0, 0, 0] - 3 / 5 + torch.sqrt(torch.tensor(9 / 8))) < 1e-4)
     26 assert(torch.abs(test_output[0, 0, 1, 0] - 3 / 5 - torch.sqrt(torch.tensor(9 / 32))) < 1e-4)
     27 print("Success!")
AssertionError:

I feel strange because I am pretty sure i have understood everything but probably not :frowning:

Do you have some clues for me?

Best regards Samir!!

2

For the InjectNoise function, your parameter looks correct, although you could have written it more simply by using torch.randn. But your noise_shape is wrong for the next part. Itā€™s the ā€œmirror imageā€ of the shape the parameters section, although maybe thatā€™s too mysterious a way to say it. How about this: one of the dimensions needs to be 1. Guess which one?

For the AdaIN section, you got the init part correct as far as I can see. But your code for the transformed_image is quite a bit different than mine. What is going on with the torch.mean divided by torch.std for the normalized image? I thought it was already normalized. They gave you that code, right?

3

Hello Paul
I really appreciate your answers and the time you have put on it, I will try to adjust my code and I hope everything will be fine, thank you.
see you next time!!!

4

Everything fine and Everything is done (thanks to you), the UNQ_C3 was a mistake of mine, because I was tired I suppose.

For the UNQ_C4 to be honest I am not very good in math, but with some logic we succeededā€¦ thank you again!!

5

Thanks for this hint. I was able to get the UNQ_C3 cell to succeed because of your hint, but I donā€™t understand why this works. Can you explain why this mirror image idea is correct? I donā€™t understand how this makes the change per channel. I guess Iā€™m having trouble thinking in 4 dimensions. (I had originally put the 1 in the n_sample parameter position.)

6

When dealing with tensors for image data in PyTorch or TensorFlow, there are two choices of the orientation of the dimensions. ā€œChannels lastā€ and ā€œchannels firstā€. In all the DLS courses, in which TensorFlow is used, they use ā€œchannels lastā€ orientation, which is the default in TF. That means you have you a batch of m samples, it is a 4D tensor with the following dimensions:

samples x height x width x channels

So if you have 100 samples, each of which is a 1024 x 2048 pixel RGB image, then the dimensions will be:

(100, 1024, 2048, 3)

The last dimension is the 3 channels giving the color value for each pixel.

But in PyTorch, the default is ā€œchannels firstā€ orientation, so the dimensions of a batch of samples will be:

samples x channels x height x width

So if all the characteristics of the batch of samples were the same as above, youā€™d have dimensions:

(100, 3, 1024, 2048)