Improve metrics docs: 2 types of API

[vfdev-5]

🚀 Feature

Following #1538 it would be nice to explicitly mention here that there are 2 types of API :

• using metric with Engine
• using metric's reset/update/compute

and give examples.

[ydcjeff]

@vfdev-5 I am familiar with .attach() method. For reset/update/compute, the correct way is like this (computing epoch-wise)?

precision = Precision()
precision.attach(engine, 'precision')
@engine.on(Events.EPOCH_STARTED)
def reset():
    precision.reset()

def process_fn(e, b):
    ...
    precision.update((pred, target))

@engine.on(Events.EPOCH_COMPLETED)
def compute():
    precision.compute()

[sdesrozis]

We have introduced MetricUsage to do this. See

ignite/ignite/metrics/metric.py

Line 52 in 5ceacbb

class EpochWise(MetricUsage):

The doc is here : https://pytorch.org/ignite/metrics.html#metrics-and-its-usages

The relevant PR is #979

HTH

[vfdev-5]

@ydcjeff yes, the idea is to show how to use reset/update/compute API without relation to the engine:

precision = Precision()

precision.update((pred1, target1))
precision.update((pred2, target2))
precision.update((pred3, target3))

print("Precision: ", precision.compute())

precision.reset()
precision.update((pred1, target1))
precision.update((pred2, target2))
precision.update((pred3, target3))
print("Precision: ", precision.compute())

[vfdev-5]

@ydcjeff would you like to send a PR to fix this issue ?

[ydcjeff]

I can only send in next weeks