Getting metrics on test set for multi label

Part 1 (2019)
1

I have trained a multi label vision classifier.

I now would like to get some metrics on a test set

data_test = ImageDataBunch.from_folder(
"dirwithtestdata",
valid_pct=0,
ds_tfms=get_transforms(flip_vert=False),
size=224,
num_workers=4).normalize(imagenet_stats)

I can evaluate it like this

learn.validate(data_test.train_dl, metrics=[accuracy])

which gives me

[1.5943198, tensor(0.6257)]

which I think means 62% accuracy(?)

How could I get more detailed information how it perform for each label?

2

I have not used the learn.validate() so I am not sure what those values are.

However, if you want to see real results of individual images, you can try

learn.predict(img)

This gives the prediction values based on each individual label.

3

You can do a get_preds() and look at how those were laid out. They are in order