Understanding numpy array shapes

Question

I'm having a lot of problems to understand numpy array shape.

If have a numpy array with shape (1, 12, 12, 512), does it means?

I have 512 arrays of shape (1, 12, 12).
I have one array of shape (12, 12, 512)

What does it means?

By the way, this numpy array is a feature map from an encoder CNN model.

If I use tf.keras.layers.GlobalAveragePooling2D with that feature map, I get a Tensor with shape (1, 512) that is what it makes me doubt.

To show a simplest array and try to understand it, I have used this code:

x = tf.constant([[1., 2., 3.],
                 [4., 5., 6.]])
y = tf.reshape(x, [1, 2, 3, 1])

x array is:

array([[1., 2., 3.],
       [4., 5., 6.]], dtype=float32)

y array is:

array([[[[1.],
         [2.],
         [3.]],

        [[4.],
         [5.],
         [6.]]]], dtype=float32)

So I think is the second option, but I'm not sure.

Ivan · Accepted Answer · 2020-12-08 08:49:29Z

2

Both those two statements are correct. However, ultimately it comes down to what your data object represents. If the first axis is the batch, it's more likely to be 1 batch containing 512 12x12-feature maps.

Edit: Which is the case, since you mention Tensorflow. The library works this with format: (batch, height, width, channels). Where the channels are last, contrary to how PyTorch operates.

tf.keras.layers.GlobalAveragePooling2D will, for each batch, average each one of your 12x12 feature maps into a single point. Since you have 512 feature maps, this will leave you with a shape of (1, 512). "Global" means it will average the whole feature map.

edited Dec 8, 2020 at 8:49

answered Dec 8, 2020 at 8:44

Ivan

41.3k9 gold badges78 silver badges120 bronze badges

Sign up to request clarification or add additional context in comments.

2 Comments

VansFannel Over a year ago

Yes, you are right, because the output_channels for the last Conv2D layer is 512, so I have 512 feature maps of size 12x12 but, if you check the code that I have added to my question, it seems that the second option is the right one. Maybe because I don't understand every thing completely. Thanks a lot for your answer.

Ivan Over a year ago

In your example, y is essentially a single batch of 1 2x3 feature map, and in this case would be the correct format. Glad I could help!

Collectives™ on Stack Overflow

Understanding numpy array shapes

1 Answer 1

2 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

2 Comments

Your Answer

Sign up or log in

Post as a guest