What should the final shapes be for calibration?

[komodovaran]

It's not entirely clear to me what's going on in the cal_results function.

My classifications are of shape (n_samples, n_timesteps, n_classes), which is turned into a 1D array of shape (n_samples, ) by doing argmax on axis -1, which then passes all comparisons between y_true and y_pred.

But the logits are pickled as they are, which leads to a value error in cal_results, when calling model.fit().

I'm not sure I understand correctly. Should they be simply be unravelled, such that

[531, 21, 5, 681, ...] is compared to [1, 0, 0, 1]? Do I need to argmax the logits first? Because then the calibration fit spits out ValueError: X should be a 1d array.

But if I unravel it, softmax() throws a numpy.AxisError: axis 1 is out of bounds for array of dimension 1.

So which dimensionality is correct?

[markus93]

Hey,

sorry for the late response.

I am not sure if it is the issue, but the pickled logits file ((logits_val, y_val), (logits_test, y_test)) should contain the following:

• logits_val ( n_val_samples, n_classes)
• y_val (n_val_samples)
• logits_test (n_test_samples, n_classes)
• y_test (n_test_samples)

Let me know if You have some more questions.