Visualizing what convnets learn confusing indexing

[jxu]

Issue Type

Documentation Bug

Source

binary

Keras Version

3

Custom Code

No

OS Platform and Distribution

No response

Python version

No response

GPU model and memory

No response

Current Behavior?

AFAIK, Keras and TF index into images as (height, width, channels) (ignoring batch size)

So why does stitched_filters have shape (width, height, 3)? Same for indexing into the ndarray.

I think this indexing may be plotting the images with x,y swapped, unless keras.utils.save_img somehow changes the numpy indexing.

Standalone code to reproduce the issue or tutorial link

https://keras.io/examples/vision/visualizing_what_convnets_learn/

The code is from https://deeplearningwithpython.io/chapters/chapter10_interpreting-what-convnets-learn/

Similar issue I opened fchollet/deep-learning-with-python-notebooks#281

Relevant log output

[dhantule]

Hi @jxu,

I think this observation is correct and the indexing is flipped, i think it should be
stitched_filters = np.zeros((height, width, 3)) and the indexing inside the loop should be

stitched_filters[
            (cropped_height + margin) * i : (cropped_height + margin) * i + cropped_height,
            (cropped_width + margin) * j : (cropped_width + margin) * j
            + cropped_width,
            :,
        ] = img

It works only because img_width == img_height, so the values are equal and no visible distortion occurs

[jxu]

It works only because img_width == img_height, so the values are equal and no visible distortion occurs

Yes, and the fact that n is used for both dimensions to make a square grid.
In my own model I had a smaller network with a layer of 32 filters, so I had to modify to use 2 dimensions for the grid.

[dhantule]

Hi @jxu, this issue was addressed in this PR, you can refer tge updated tutorial here, please feel free to close this issue.