Evaluate network in training=True configuration

To get help from the community, we encourage using Stack Overflow and the tensorflow.js tag.

TensorFlow.js version:

1.7.3

Browser version: Chrome

81.0.4044.113

Describe the problem or feature request:

I have trained a pix2pix network in python, and converted it to a tfjs graph model. In python, I can evaluate the network with the training=True flag, which is required in order for the batch normalization layers to use the sample mean and variance rather than the mean and variance from the training set. Is it possible to either call the model in a training configuration, or to modify the batchnorm layer to use sample statistics rather than training statistics?

There is a similar issue here: #562

[B0RJA]

The solution is to use the apply method with the training kwarg set to true. Tested with TensorFlow.js 2.0.

const output = model.apply(input, {'training': true});

[AmitMY]

Works for me too 😃 I just hope that it does not “cost” more compute to run it in training mode

Now after evaluating the time:

model.apply(tensor, {training: true})

Takes 6-8ms for me, while

model.predict(tensor)

Takes 3-4ms to me (50%!)

While it is not a lot, (assuming that model inference takes 40ms~) It can make the difference from 24 to 25fps 😃