The accuracy between “learn.recorder.plot_metrics()” and munually computed from "plot_confusion_matrix()" are NOT THE SAME

[jazzylee]

Recently I‘ve begun to use tsai,but I got confused by the accuracy.
When I use the "learn.recorder.plot_metrics()" method,a figure containing accuracy is shown,which peforms very well.
But when I draw the confusion matrix using "plot_confusion_matrix()" and compute the accuracy manually,the result is not the same as the above way.
So how can I get the accuracy and why the two ways mentioned above are different?

[vrodriguezf]

Hi @jazzylee , can you show an example?

[jazzylee]

Hi,vrodriguezf,thanks for your replying!😁
I tried to apply InceptionTime in my project and found the issue.Here's some core screenshots below.
accuracy from acc figure:

accuracy calculated from confusion matrix:

[oguiza]

Hi @jazzylee,
I'm sorry about the late reply.
I've tried to replicate this issue, and have tried different settings, but I couldn't. I always get the same result. This is the code that I've used:

from tsai.all import *

dsid = 'ECG200'
X, y, splits = get_UCR_data(dsid, split_data=False)
tfms = [None, TSClassification()]
dls = get_ts_dls(X, y, splits=splits, tfms=tfms)
learn = ts_learner(dls, TST, metrics=accuracy, cbs=ShowGraph())
learn.fit_one_cycle(10, 1e-4)
interp = ClassificationInterpretation.from_learner(learn)
interp.plot_confusion_matrix()
print(f'from train: {learn.recorder.values[-1][-1]:.3f} from interp: {np.trace(interp.confusion_matrix()) / interp.confusion_matrix().sum():.3f}')


# output
# from train: 0.730 from interp: 0.730

Do you still get these conflicting results?