visualizing_what_convnets_learn Image Resolution

[chrislos]

https://github.com/keras-team/keras-io/blob/313a4d15b4f858de35dac1246bfd83fbd07d1a7a/examples/vision/visualizing_what_convnets_learn.py#L29

Dear Keras team,

first of all thank you for your exceptional job. I went through fchollet's article about the techniques of visualiting convnets. Thanks to your help I finally could see what my nets are "seeing". Super interesting.

There is one underling question of understanding left for me.

So, given the fact I've trained a pix2pix net with Images with a size of 2048 x 2048. Would I have to set these dimensions as input image values in your script for a scientific correct filter-visualization of a layer "x" at a filter-index "y".

It seems that smaller image-dimensions deliver similar filter-outputs (eg. 1024px x 1024px or even 512px x 512px as Input-Dimensions), even if my net was trained on a higher resolution.

Is there a downsampling procress somwhere hidden, that I haven't found in the code so far or why do smaller image-inputs also result in filters that appear to have learned patterns?

Thanks in advance, Christian

[sachinprasadhs]

Hi @chrislos, Thanks for reporting the issue.

Yes, in the code it actually resizes the image to 180x180 for any given input format.

img_width = 180
img_height = 180

And in the below code, it is processed.

def initialize_image():
    # We start from a gray image with some random noise
    img = tf.random.uniform((1, img_width, img_height, 3))
    # ResNet50V2 expects inputs in the range [-1, +1].
    # Here we scale our random inputs to [-0.125, +0.125]
    return (img - 0.5) * 0.25

You can also provide the resolution of your desire to experiment with your model. Thanks!

[github-actions[bot]]

This issue is stale because it has been open for 14 days with no activity. It will be closed if no further activity occurs. Thank you.

[github-actions[bot]]

This issue was closed because it has been inactive for 28 days. Please reopen if you'd like to work on this further.