New Code Example : Forward-Forward Algorithm for Image Classification by suvadityamuk · Pull Request #1170 · keras-team/keras-io

suvadityamuk · 2022-12-23T05:25:39Z

This PR adds a new Code Example implementation for the Forward-Forward algorithm, as introduced by Prof. Hinton in his paper at NeurIPS 2022.

Some things to note:

This is implemented completely using Keras and TensorFlow-backend code, but is an almost-direct port of this PyTorch implementation of the same algorithm here
Since the weight updates occur within the layer itself as per the algorithm, some of the code may be working against the original design of the Keras API. But such is the algorithm that looks to defy them altogether. Would appreciate any reviews and thoughts on that as well
I use TQDM and random for some cleaner visuals and simpler code.

Tagging @LukeWood @fchollet for a review. Thank you for your time!

Signed-off-by: Suvaditya Mukherjee suvadityamuk@gmail.com

Signed-off-by: Suvaditya Mukherjee <suvadityamuk@gmail.com>

suvadityamuk · 2022-12-26T05:57:52Z

Hi, @fchollet! Gentle ping for a review, thank you! Oh, and Merry Christmas 🎄 !

fchollet · 2022-12-27T09:32:08Z

Thanks for the PR! I'll take a close look in the next few days.

fchollet

Thanks for the PR. It's a great example!

fchollet · 2022-12-27T14:28:40Z

+training instead of the traditionally-used method of backpropagation, as proposed by
+[Prof. Geoffrey Hinton](https://www.cs.toronto.edu/~hinton/FFA13.pdf)
+The concept was inspired by the understanding behind [Boltzmann
+Machines](http://www.cs.toronto.edu/~fritz/absps/dbm.pdf). Backpropagation involves


Please keep all markdown links on a single line, otherwise they won't properly render

Will make this change across the file

fchollet · 2022-12-27T14:29:24Z

+[Prof. Geoffrey Hinton](https://www.cs.toronto.edu/~hinton/FFA13.pdf)
+The concept was inspired by the understanding behind [Boltzmann
+Machines](http://www.cs.toronto.edu/~fritz/absps/dbm.pdf). Backpropagation involves
+calculating loss via a cost function and propagating the error across the network. On the


I think you can find a slightly better one-line description of backprop

Sure, will update that.

fchollet · 2022-12-27T14:30:35Z

+
+The following example explores how to use the Forward-Forward algorithm to perform
+training instead of the traditionally-used method of backpropagation, as proposed by
+[Prof. Geoffrey Hinton](https://www.cs.toronto.edu/~hinton/FFA13.pdf)


"Proposed by Hinton in [The Forward-Forward Algorithm: Some Preliminary Investigations]() (2022)"

Right, will make this change

fchollet · 2022-12-27T14:31:00Z

+other hand, the FF Algorithm suggests the analogy of neurons which get "excited" based on
+looking at a certain recognized combination of an image and its correct corresponding
+label.
+This method takes certain inspiration from the biological learning process that occurs in


Please add line breaks between paragraphs

fchollet · 2022-12-27T14:31:21Z

+Date created: 2022/12/21
+Last modified: 2022/12/23
+Description: Training a Dense-layer based model using the Forward-Forward algorithm.
+


Add an Accelerator: field (either GPU or None)

fchollet · 2022-12-27T14:39:33Z

+        x = self.flatten_layer(x)
+        perm_array = tf.range(start=0, limit=x.get_shape()[0], delta=1)
+        x_pos = self.overlay_y_on_x(x, y)
+        y_numpy = y.numpy()


fchollet · 2022-12-27T14:39:44Z

+                plt.show()
+            else:
+                x = layer(x)
+        return {"FinalLoss": loss}


train_step should return average values, e.g. the output of loss_tracker.result() (where loss_tracker is a metrics.Mean instance).

Also, use snake case

Will make the change

fchollet · 2022-12-27T14:40:34Z

+        h_pos, h_neg = x_pos, x_neg
+        for idx, layer in enumerate(self.layers):
+            if idx == 0:
+                print("Input layer : No training")


Do not include any print statements or matplotlib plots in train_step. You should write a callback for these.

Right. I'll implement a callback and make use of that

fchollet · 2022-12-27T14:42:11Z

+model = FFNetwork(dims=[784, 500, 500])
+
+model.compile(
+    optimizer=keras.optimizers.Adam(learning_rate=0.03), loss="mse", run_eagerly=True


You should arrive to a model that can be run in graph mode, without requiring run_eagerly=True

Alright, I will try to do so

fchollet · 2022-12-27T14:43:09Z

+
+results = accuracy_score(preds, y_test)
+
+print(f"Accuracy score : {results*100}%")


What do you get?

The Accuracy score currently stands at ~24-27% on different runs (can set a constant Seed to enhance reproducibility). While the paper does get somewhat better results (not anywhere near SOTA though), I believe that can be achieved with more tuning and perhaps a wider network

25% accuracy on MNIST tells you that your algorithm does not work, unfortunately. A simple logistic regression does ~92% or so. A simple logistic regression after applying ~90% noise on the input (setting to 0 90% of the pixels, randomly) still does 67%. Even if it were poorly tuned, the example should achieve at least 97% if the algorithm did work as expected.

Right. Let me iterate on this algorithm for a while, see if and where I am making some errors and I will get back to you on this

DavidLandup0 · 2023-01-03T04:44:21Z

+    def forwardforward(self, x_pos, x_neg):
+        loss_list = []
+        for i in trange(self.num_epochs):
+            with tf.GradientTape() as tape:


Aren't we falling back to backprop here?
The loss is calculated, gradients are computed backwards, and then we're doing an optimizer step

This is happening locally, solely for this layer. So wouldn't qualify for backward propagation just yet.

Signed-off-by: Suvaditya Mukherjee <suvadityamuk@gmail.com>

suvadityamuk · 2023-01-08T15:30:57Z

Hi, @fchollet

I have just added a new commit with all the changes. Apologies for the delay, took some time to make sure all of your comments are addressed.
The new Test Accuracy is 97.75% as per my last training run. I have also made use of the xla.dynamic_update_slice as per your linked PR's suggestion (thanks a ton for that, by the way.)
I have also added more detailed comments and suggestions
Please let me know if any other changes are required.

Also, a very Happy New Year to you!

suvadityamuk · 2023-01-08T15:35:01Z

Also, not sure why the docker-image CI is failing. It seems due to some keras_cv-related version issue although I have not used it within this example.

fchollet · 2023-01-09T00:29:33Z

The new Test Accuracy is 97.75% as per my last training run. I have also made use of the xla.dynamic_update_slice as per your linked PR's suggestion (thanks a ton for that, by the way.)

Congrats on making it work!

The current code looks good to me. Please add the generated files. I pushed some copyedits, please pull them first.

Signed-off-by: Suvaditya Mukherjee <suvadityamuk@gmail.com>

suvadityamuk · 2023-01-09T03:00:16Z

Hi, have added the generated files and some minor factual edits. Thank you!

fchollet

Thank you for the great contribution! 👍

suvadityamuk · 2023-01-09T16:50:59Z

Happy to contribute!

michaelStettler · 2023-01-12T09:03:15Z

HI,

First of all thanks a lot for the code. It's very instructive to see how you implemented this directly in keras/tensorflow. Anyway, I wanted to try it and I just copy pasted the code but I got a:

KeyError: 'The optimizer cannot recognize variable dense_1/kernel:0. This usually means you are trying to call the optimizer to update different parts of the model separately. Please call optimizer.build(variables) with the full list of trainable variables before the training loop or use legacy optimizer
`tf.keras.optimizers.legacy.{self.class.name}.'

I've tried several thing from yesterday but none worked. Any suggestion on what that could mean on where to add this optimizer.build?

added ForwardForward py file example for PR review

4b69577

Signed-off-by: Suvaditya Mukherjee <suvadityamuk@gmail.com>

fchollet reviewed Dec 27, 2022

View reviewed changes

DavidLandup0 reviewed Jan 3, 2023

View reviewed changes

added changes as per Francois's comments

634d63d

Signed-off-by: Suvaditya Mukherjee <suvadityamuk@gmail.com>

suvadityamuk requested a review from fchollet January 8, 2023 15:33

Copyedits

5b59ced

added small final edits and generated files

0dbbfff

Signed-off-by: Suvaditya Mukherjee <suvadityamuk@gmail.com>

fchollet approved these changes Jan 9, 2023

View reviewed changes

fchollet merged commit 23b54b7 into keras-team:master Jan 9, 2023


		results = accuracy_score(preds, y_test)

		print(f"Accuracy score : {results*100}%")

Conversation

suvadityamuk commented Dec 23, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

suvadityamuk commented Dec 26, 2022

Uh oh!

fchollet commented Dec 27, 2022

Uh oh!

fchollet left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

suvadityamuk Dec 27, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

suvadityamuk Dec 28, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

suvadityamuk commented Jan 8, 2023

Uh oh!

suvadityamuk commented Jan 8, 2023

Uh oh!

fchollet commented Jan 9, 2023

Uh oh!

suvadityamuk commented Jan 9, 2023

Uh oh!

fchollet left a comment

Choose a reason for hiding this comment

Uh oh!

suvadityamuk commented Jan 9, 2023

Uh oh!

michaelStettler commented Jan 12, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

suvadityamuk commented Dec 23, 2022 •

edited

Loading

suvadityamuk Dec 27, 2022 •

edited

Loading

suvadityamuk Dec 28, 2022 •

edited

Loading