Simpletransformers: Raw Outputs

[ShivamSharma1997]

Hi,

Can someone please explain the actual results we get as raw outputs and what are it's range?

The following are the results I am getting as raw output features.
[[ 2.3422227 0.04336229 -0.7625307 -2.8956182 ]
[-1.1994745 6.9426236 -1.8545047 -3.119325 ]
[ 6.8584485 -2.2693088 -3.3302417 -3.4118474 ]
[ 6.5169826 -2.4898453 -3.148984 -3.0034273 ]
[-1.9056513 0.6704526 3.7555788 -1.8298897 ]
[ 6.6204324 -1.4047363 -2.4027085 -4.5800114 ]
[-1.0606726 5.4772515 1.3901057 -4.785892 ]
[ 6.927998 -2.0280929 -3.284417 -3.7614038 ]
[ 6.141445 -1.2969164 -2.7105436 -4.2589316 ]
[ 6.8645945 -2.3979468 -2.797596 -3.737698 ]]

[chrzyki]

Heya,

assuming that this is the output of a ClassificationModel or similar you can apply the Softmax function to get the raw probabilities.

from scipy.special import softmax
import numpy

numpy.array([softmax(element) for element in your_probability_array[1]])

The resulting array then holds the probability for each label in your data set, i.e. [0] is the probability for label 0, [1] the probability for label 1, etc. with 1.00 being equal to 100% and 0.00 being equal to 0%.

[ShivamSharma1997]

Thank you so much!
Yes its from the Multi Class Classification model

[ShivamSharma1997]

I am assuming I should pass every raw output array to the softmax function separately and not the whole resulting matrix as one, right?

[ShivamSharma1997]

@chrzyki The results which I am getting with the regression model has value greater than 1 where as all my labels are between 0 and 1. Can you help please?

[chrzyki]

@ShivamSharma1997 can you show us a minimal example of your raw outputs and the code you're using for applying the softmax function?

[ShivamSharma1997]

@chrzyki The following is the code for getting the softmax values for each output. Each output in the raw_outputs list is a vector of length 25.

_, raw_outputs = model.predict(text_list)

predictions = []

for output in raw_outputs:
weights = softmax(output)
predictions.append(weights)

but I think this is working fine. I am confused about the output of the regression model, should I pass every output through the softmax function?

The predictions for the regression models are as follows:

0.4002847969532013
0.35405269265174866
0.2953510582447052
0.2686978876590729
0.4050685167312622
0.4273129999637604
0.4273129999637604
0.2625053822994232
0.4580164849758148
0.2639479637145996

[chrzyki]

@ShivamSharma1997 for every item of every list of outputs (i.e. label probabilities) you should apply the softmax function. The result should be a list of probabilities of labels for each element in text_list. Judging from your code, predictions should then be a list of list of probabilities.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[ShivamSharma1997]

Hi @chrzyki
Can you tell me a way to get the probabilities for each label for Multi-Label classification task?