search

pytorch / ignite

High-level library to help with training and evaluating neural networks in PyTorch flexibly and transparently.
https://pytorch-ignite.ai
BSD 3-Clause "New" or "Revised" License
4.51k stars 610 forks source link

[Feature request] class MaximumAbsoluteError(Metric) #292

Closed zasdfgbnm closed 5 years ago

zasdfgbnm commented 5 years ago

error = (y_pred - y).abs().max()

Would this be a good thing to add?

vfdev-5 commented 5 years ago

@zasdfgbnm could you please add more context on where such metric can be of interest ?

zasdfgbnm commented 5 years ago

@vfdev-5 This is for regression problem, it characterizes how far the worst point are deviated from ground truth.

vfdev-5 commented 5 years ago

@zasdfgbnm in general we can add something like this in the contrib module. Let's see what others think about the importance of this metric.

jasonkriss commented 5 years ago

In general, I think more metrics are definitely welcomed. That being said, for this specific one, I don't now that there is a wide enough use to justify its addition.

alykhantejani commented 5 years ago

I echo @jasonkriss sentiment, that more metrics are welcom. MAE is often used in recommender systems, and other regression tasks to measure performance.

I say go for it, might as well just put it in core as it's pretty simple to implement and I don't envisage much support being needed going forward

@zasdfgbnm can you send a PR?

zasdfgbnm commented 5 years ago

Sure, I will. @alykhantejani