How to set up a plateau detector?

[souravraha]

I am trying to detect plateaus in metric values as I train a GAN progressively. Once a plateau is reached, I would activate subsequent layers for training on bigger images.

Here is how I am thinking of implementing this:

1. Pass "None" as the 'reduction' argument for the metric.
2. Call the 'update', followed by 'compute' method.
3. Slice off the last 1000 items from the output of step 2.
4. Compute mean and std from previous output.
5. Check if abs(std/mean) is less than 0.1.
6. Implement stuff depending upon the previous result.
7. Call 'reset' method when the number of elements reaches a threshold, say 7500.

Am I missing a better way to go about doing this?