optimizer.minimize does not work with maxPooling layers

[generic-github-user]

TensorFlow.js version

Latest / 0.14.2

Browser version

Google Chrome
Version 71.0.3578.98 (Official Build) (64-bit)

Describe the problem

When using optimizer.minimize() with model.predict() to train a tf.model with a loss function, I encounter an issue. This only occurs when I use a maxPooling2D layer in a convolutional neural network with code similar to the code below. It produces this error: Cannot read property 'backend' of undefined. I'm not sure what is causing this or how to resolve it. The error does not occur when using a convolutional layer (tf.layers.conv2d()) without any pooling layers. I have not had this issue with tf.layers.averagePooling2d(), leading me to believe that this could be a possible bug.

Code to reproduce the error

loss = (pred, label) => pred.sub(label).square().mean();
optimizer = tf.train.sgd(0.001);

const input = tf.input({shape: [100, 100, 4]});
const conv = tf.layers.conv2d({
	kernelSize: 5,
	filters: 8,
	strides: 1,
	activation: 'relu',
	kernelInitializer: 'VarianceScaling'
});
const pool = tf.layers.maxPooling2d({
	poolSize: [2, 2],
	strides: [2, 2]
});
const flat = tf.layers.flatten();
const dense = tf.layers.dense({units: 10});
const output = dense.apply(flat.apply(pool.apply(conv.apply(input))));
const model = tf.model({inputs: input, outputs: output});

for (var i = 0; i < 10; i++) {
	optimizer.minimize(() =>
		loss(model.predict([tf.ones([1, 100, 100, 4])]), tf.ones([1, 10]))
	);
}

[generic-github-user]

It appears that this is in fact a bug in TensorFlow.js version 0.14 (https://stackoverflow.com/a/54495577/10940584). No error appears when the above code is run in version 0.13.3 and earlier.

[caisq]

@generic-github-user

In 0.14+, there is a change that disables backpropagation support in the Model.predict() method. You can use the Model.apply() method with the {training: true} flag to fix your code.

I.e., change

        optimizer.minimize(() =>
		loss(model.predict([tf.ones([1, 100, 100, 4])]), tf.ones([1, 10]))
	);

to

       optimizer.minimize(() =>
		loss(model.apply([tf.ones([1, 100, 100, 4])], {training: true}), tf.ones([1, 10]))
	);

[generic-github-user]

Thank you! This solved my problem, I appreciate the help.

[quantuminformation]

Initially, apply confused me as its similar name to the apply method on functions prototype.

[ichko]

This doesn't seem to work for me. Has the API been changed?
What I am trying to do is do an optim update using the optim.minimize function and a model defined with the layers API.
The optim step is not erroring out, but my loss is not going down.
I could not find a documentation or an example of using apply with this training: true parameter. Neither were I able to find an example of using a model defined with layers being updated with the core API (optimizer.minimize(...))

Can anyone help me out?
(maybe this is not a question for this thread, but this was the only issue I was able to find that contains the setup that I am trying to recreate)