per class AP, and metrics

[benlten]

Unless I am misreading the code return, the function:
ap_per_class in (https://github.com/ultralytics/yolov5/blob/master/utils/metrics.py)

returns the mean AP and not a list of APs.

Goal
I am trying to plot how each individual class is doing, or rather the precision, AP50, and recall per class label as would be shown in a confusion matrix.

Currently I see wandb show the average per P, R, AP50 etc - but I'd like to log a dict of those per individual class performance to analyze the misclassification errors.

If someone has gone through the same steps to save data out, please let me know.

[glenn-jocher]

@bencten confusion matrices and per-class statistics are automatically reported by test.py on final epoch, or when called directly (i.e. python test.py):

See #1474 for confusion matrix:

test.py verbose (per class) output is enabled automatically for small datasets, or can be forced manually with python test.py --verbose. See #1997

[benlten]

@glenn-jocher
There may be a bug. I was tracing through the code to the metrics method for a while.
In my csv dump, I kept seeing absurdly high numbers as if something was being cached coming from when test.py returned.
class_name 6 1 1.0 1.0 0.9954166666666666 0.9954166666666667, no matter what.
I'll keep an eye on this and present a small reproducible test set if it occurs again.

[glenn-jocher]

@bencten thanks for the info! If you spot any reproducible bugs, please raise a new issue using the 🐛 Bug Report template with a minimum reproducible example to help us better understand and diagnose your problem. Thank you!