Remove extra activation layer in image_classification_from_scratch.py

[SuryanarayanaY]

The image_classification_from_scratch.py tutorial has one extra relu activation layer which is being applied on the layer that already gone through relu activation which makes one code of line redundant. Though it may not affect output it unnecessarily creates confusion for the users and also consumes some resources if used for training.Hence I am proposing to remove the unwanted line of code.

Fixes #1119

[divyashreepathihalli]

Thank you for the PR. It is recommended to remove the first call at line 238 instead of the one in the loop.

[SuryanarayanaY]

Actually line no.240 previous_block_activation = x (where x = layers.Activation("relu")(x) defined at line no.238 ) stored under variable previous_block_activation as residual and used it later at Line No.254.

keras-io/examples/vision/image_classification_from_scratch.py

Lines 238 to 256 in 8c48c9f

	x = layers.Activation("relu")(x)
	previous_block_activation = x # Set aside residual
	for size in [256, 512, 728]:
	x = layers.Activation("relu")(x)
	x = layers.SeparableConv2D(size, 3, padding="same")(x)
	x = layers.BatchNormalization()(x)
	x = layers.Activation("relu")(x)
	x = layers.SeparableConv2D(size, 3, padding="same")(x)
	x = layers.BatchNormalization()(x)
	x = layers.MaxPooling2D(3, strides=2, padding="same")(x)
	# Project residual
	residual = layers.Conv2D(size, 1, strides=2, padding="same")(
	previous_block_activation
	)

[divyashreepathihalli]

@SuryanarayanaY in that case this change will break the implementation. I will close this PR. Thanks!

[SuryanarayanaY]

@divyashreepathihalli ,

I am thinking in perspective of user request for removing extra layer. But here issue seems more than that. Here author used activation layer output as residual. I think it's better to pass residuals without activation. In that case removing the first activation call at line no.138 itself makes sense. Like the amendment in #1398

If the author intention is to pass activation layer output as residual then earlier proposal is the solution. Anyways passing already activated output to same activation layer again gives same result.Hence the proposed changes will not break. I have tested the code with earlier proposal and it executes fine as per attached gist. Correct me if I am missing anything here.

I think what we need to pay attention more here is whether we have to pass residuals as with relu activation or without activation. If we need to pass residuals without going through activation then we need to remove line no.238 as proposed by you.

Please suggest.

[SuryanarayanaY]

@divyashreepathihalli ,

Can I go ahead with the proposal you referred earlier in this comment ?

[divyashreepathihalli]

@divyashreepathihalli ,

Can I go ahead with the proposal you referred earlier in this comment ?

Sure, please go ahead and update the changes.

[SuryanarayanaY]

@divyashreepathihalli , Done the changes.May please review.

[gbaned]

Hi @divyashreepathihalli, Can you please review this PR ? Thank you!

[gbaned]

Hi @fchollet, Can you please review this PR ? Thank you!

[sachinprasadhs]

Fines are outdated, can you please make the changes to the latest file https://keras.io/examples/vision/image_classification_from_scratch/

[github-actions]

This PR 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]

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