Converted to keras core external attention

[anas-rz]

Update imports
Remove TensorFlow Addons Dependency
Convert TensorFlow ops in functional component to keras_core ops

Attached git-diff
Colab Notebook
eanet_keras_core.diff.txt

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[fchollet]

Thanks for the PR!

[fchollet]

Since the example is TF-only, start by setting the backend using os

[fchollet]

Every TF call except this one can be converted to ops. This will make the example less TF dependent.

[fchollet]

I believe extract patches can be implemented by a convolution, no? It's a convolution with identity kernel. Maybe we could do that and make the example backend-agnostic.

Alternatively, maybe we can add a cross-backend extract_patches function in ops.image (implemented using convolution for other backends). Actually, this is a better solution, since this method comes up a lot.

[anas-rz]

I can create a PR for it.
Another workaround: For this example can I use frameworks like einops for creating patches?

[fchollet]

Prefer using a conv.

[anas-rz]

I have found this easy way to extract patches. Can you please elaborate/refer some example of using conv. As far as I understand, conv uses reduce sum of a filter. Is there any 'do nothing' filter for convolution that can be used for upsampling?

[fchollet]

Your way is fine. Conv is likely faster on GPU though. It could be worth trying both if we add the op to the API (just for this example we can keep your way since it's concise).

Is there any 'do nothing' filter for convolution that can be used for upsampling?

"Do nothing" is just an identity kernel. There is no "sum" operation, it's a matmul. It turns into a sum if your kernel is all ones.

[anas-rz]

Here's the initial 'code' version I have got for extracting patches and it is working. Thank you for guidance. Will create a PR for adding this to cross-backend ops.

import tensorflow as tf
input_image = tf.random.uniform((1, 224, 224, 3))
patch_h, patch_w = 14, 14
channels_in = 3
out_dim = 588
kernel = tf.eye(patch_h * patch_w * channels_in, out_dim)
kernel = tf.reshape(kernel, (14, 14, 3, 588))
out_conv_layer = tf.nn.conv2d(input_image, kernel, 14, 'VALID')
out_tf = tf.image.extract_patches(input_image, [1, 14, 14, 1], [1, 14, 14, 1], [1, 1, 1, 1], "VALID")
check = tf.equal(out_conv_layer, out_tf)
print(np.unique(check))```

[fchollet]

Thanks! extract_patches will be a great addition to the ops.image API.

Should we wait for this feature before revisiting this example?

[anas-rz]

Sure. That'll be better.

[fchollet]

LGTM, thank you!