Question regarding same results for 'micro' average type

[sarthakpati]

Hi,

Thank you for the amazing work you have done with this.

I wanted to ask the reasoning for this behavior I am seeing. The "micro" average type for many classification metrics is giving identical results, but "macro", "weighted", or "none" do not. Here are the steps to reproduce this:

# create and activate python virtual environment
pip install torchmetrics==0.5.1 # I am seeing this on the latest torchmetrics as well, just that code needs to be changed
pip install torch

The actual script:

# file: metrics_test.py
import torch
import torchmetrics as tm

ground_truths=torch. Tensor([0,1,1,0,1,2]).long()
predictions=torch. Tensor([0,1,0,1,1,1]).long()

for average_type in ["micro", "macro", "weighted", "none"]:
    print("accuracy_"+average_type, tm.Accuracy(num_classes=3, average=average_type)(predictions, ground_truths))
    print("precision_"+average_type, tm.Precision(num_classes=3, average=average_type)(predictions, ground_truths))
    print("recall_"+average_type, tm.Recall(num_classes=3, average=average_type)(predictions, ground_truths))
    print("f1_"+average_type, tm.F1(num_classes=3, average=average_type)(predictions, ground_truths)) # make this tm.F1Score for newer versions of torchmetrics

The output:

> python metrics_test.py
accuracy_micro tensor(0.5000)
precision_micro tensor(0.5000)
recall_micro tensor(0.5000)
f1_micro tensor(0.5000)
accuracy_macro tensor(0.3889)
precision_macro tensor(0.3333)
recall_macro tensor(0.3889)
f1_macro tensor(0.3571)
accuracy_weighted tensor(0.5000)
precision_weighted tensor(0.4167)
recall_weighted tensor(0.5000)
f1_weighted tensor(0.4524)
accuracy_none tensor([0.5000, 0.6667, 0.0000])
precision_none tensor([0.5000, 0.5000, 0.0000])
recall_none tensor([0.5000, 0.6667, 0.0000])
f1_none tensor([0.5000, 0.5714, 0.0000])

Any insight into this would be really helpful, thank you!

Cheers,
Sarthak