IterationMetric discussion

[vfdev-5]

🚀 Feature

I'd like to discuss a bit about your PR on IterationMetric (https://github.com/Project-MONAI/MONAI/pull/1488). As far as I understand, the idea is to compute metric value on each iteration and "all gather" everything on epoch ended ? However, final list contains all metric values for a single epoch: [m1, m2, ..., mN]. @Nic-Ma, let's discuss here then.

[Nic-Ma]

Hi @vfdev-5 ,

Thanks for raising the discussion here. IterationMetric is the base class for other metrics, every metric will make a list. I don't quite understand your question, could you please share more details about where is wrong?

Thanks.

[vfdev-5]

Hi @Nic-Ma , it was not a real question but just my understanding of the feature you introduced...

I can understand to use-case but I'm hesitating a little about storing all values in RAM without a high memory consumption... I agree that things may depend on what kind of processing is intended after gathering all values: [m1, m2, ..., mN]. In the simplest case, we can think that we just write each value to the disk in a separate file. In a more complicated, all values into a single file (either appending each time or once on event completed).

@sdesrozis what are your thoughts about that ?

PS: BTW, we have also https://github.com/pytorch/ignite/blob/702d3a2eed412118651004d6c7c03739ac33dfd4/ignite/contrib/handlers/stores.py#L6

which stores any output data during the epoch (and does not all_gather)

[Nic-Ma]

Hi @vfdev-5 ,

Here is the following PR to consume the metrics output: https://github.com/Project-MONAI/MONAI/pull/1497 It's still under review. The idea is to save metric details into engine.state, then other handlers can save it to file or visualize somewhere.

Thanks.

[vfdev-5]

Thanks for the link @Nic-Ma !

[sdesrozis]

@Nic-Ma In your implementation, the IterationMetric could be used without inheritance, right ?

def my_algorithm(y, y_pred):
   # ... do something

my_metric = IterationMetric(metric_fn=my_algorithm, 
                            output_transform=output_transform, 
                            device=device)

Somehow the functional and algorithmic parts are split.

I think it is a good idea to reuse existing metrics.

[Nic-Ma]

Hi @sdesrozis ,

Yes, that's the expected behavior, you can use an algorithm(function or callable class) as arg or define subclass.

Thanks.

[vfdev-5]

@sdesrozis I think we can try to implement this as new SampleWise metric usage. Would you like to work on that ? Maybe, it could also open a discussion on how to better generalize MetricUsage structure and cover the issue with RunningAverage and Frequency metrics...