How to replace the one-hot-encoded label with a (Gaussian) distribution in the training of a CNN for classification ?

Hi there,

I am trying to train a CNN for classification using imagedatastore, in which the cross entropy loss needs to be calculated between the softmax output and a (Gaussian) distribution instead of its one-hot-encoded version of the label.

In case I need to customize Matlab functions such that my own function (that converts an original label into a distribution) runs instead of 'onehotencode', where (or in which function) does this 'onehotencode' run ?

There may be other simpler ways to achieve my goal. I would greatly appreciate if anyone could help me on this.

Best regards,

HK

Answer 1

[yanqi liu]

yes，sir，may be just compute vector，such as

class_info = [1; 2; 2; 1]

class_info = 4×1

1 2 2 1

class_num = length(unique(class_info));
Y = zeros(length(class_info), class_num);
for i = 1 : length(class_info)
    Y(i, class_info(i)) = 1;
end
Y

Y = 4×2

1 0 0 1 0 1 1 0

[hyeonjin kim]

Dear Yanqi,

Thanks very much for your thought. However, that is not the point of my question.

I am looking for the place at which the automatic one-hot-encoding takes places on the labels in a deep learning classification problem, such that I can replace the one-hot-encode function with my customized function. I do appreciate your answer in that it can be useful for writing the cusomized function of mine.

Best,

HK

[yanqi liu]

yes，it is same question i faced，may be can not make multi output in this environment，so just use label value. but in TensorFlow or Pytorch，it is easy to make multi output，and just add loss can train model

[hyeonjin kim]

Dear Yanqi, thanks very much for sharing your thoughts. I will try some more to tackle this issue, but might ultimately have to go for Pytorch, for instance, as you told.

Best !