Compound metrics

An older thread here discusses how to manage a compound loss function (in particular, being able to separately record the parts of the loss in the Recorder). The solution described there make sense to me, but I don’t think it quite meets one need I have: I want to send results not just to the Recorder but to another manager as well. (In my case, I’m using, but it just as easily be Tensorboard, etc.).

To make this grounded, suppose I have a metric callback called SomeInterestingStats that will produce multiple metrics, and a metric-consuming callback called SendToNeptune. Under the normal way of doing things, SendToNeptune would pick up the list of metric names it can expect to see from the metrics_names argument to Callback.on_train_begin:

    def on_train_begin(self, metrics_names, **kwargs):
        self.metrics_names = metrics_names

But SomeInterestingStats is going to create multiple metrics. To do this, it updates the Recorder’s list of metric names with a line of code that looks something like this:


And then proceeds to append multiple results to the end of last_metrics when it computes them.

SendToNeptune is going to see the wrong list of metric names. It will at least miss the extra metrics, or even worse, get them mixed up (if SomeInterestingStats isn’t the last metric callback in the list).

I haven’t implemented it yet, but I think there is a workaround: SendToNeptune can get its list of metrics from the Recorder, and ignore what is passed in on_training_begin. But that is kind of … ugly.

So first of all, I want to make sure I’ve got this right, and haven’t misunderstood the complex information flow through callbacks.

If I am right, then this isn’t a problem just for me, but arguably for any Callback that thinks it can use the metrics_names argument to on_train_begin. Given the (published and used) technique of expanding the list of metrics, the metrics_names argument is simply unreliable.

I think we should either:

  • Add a method for to allow callbacks to declare the name(s) they will publish (and CallbackHandler or Learner, rather than the callback, take responsibility for synchronizing with Recorder)
  • Deprecate the metrics_names argument to on_train_begin and make the workaround above the ‘official’ way to do things.

Note that Callbacks have an order attribute to allow you to make sure there are run in a certain order. We won’t do any deep change in the callback system in v1, as we are close to releasing v2.

Thanks — _order is a useful attribute.

It occurred to me later that there actually is a fix that is mostly non-invasive: let metric callbacks update metrics_names in on_training_begin. If they do that, then they don’t need to update Recorder directly. You’d leave any existing callbacks that update Recorder alone so they don’t break anything, but new callbacks could seamlessly provide info for both Recorder and any other callbacks that are listening.

Of course, depending on how soon v2 is out, this may not be a worthwhile investment either.
If it is, though, I’d be happy to submit a PR…