How to understand Softmax layer activations for pretrained CNN?

Hello,

I was able to get softmax layer probabilities for the squeezenet network. Using

act1 = activations(net,i,'prob','OutputAs','rows');

However, Im unsure exactly what the probabilities represent, or how to identify what class they correspond with. Could you please explain what these probabilities mean? And how to get the corresponding class?

Thank you!

Answer 1

[Jon Cherrie]

How to identify what class they correspond with?

We can get this from the class names property of the output layer.

What these probabilities mean?

They are the probability that the given image is in the corresponding class.

Here's an example

net = squeezenet;
img = imread('peppers.png');
img = imresize(img,net.Layers(1).InputSize(1:2));

We can get the names of the classes from the output layer, which happens to bne the last layer for this network:

c = net.Layers(end).Classes;

Then get the activations from the softmax layer

p = activations(net,img,'prob','OutputAs','rows');

Note that these sum to 1

sum(p)

ans = single

1

We can then find the maximum probability and which class that corresponds to

[pm, i] = max(p)

pm = single

0.4172

i = 946

c(i)

ans = categorical

bell pepper

This is the same information that comes from the classify command:

[cc, pp] = classify(net,img);
cc

cc = categorical

bell pepper

isequal(pp, p)

ans = logical

1

[AK]

This is great, thank you!

Is there a way to match the classess with all the probabilities?

[Jon Cherrie]

Do you mean something like this:

class = net.Layers(end).Classes;
probability = activations(net,img,'prob','OutputAs','columns');
t = table(class,probability);
head(t)

You can then sort that by probability, e.g.,

t = sortrows(t,'probability','descend');
head(t)

[AK]

Yes! This is perfect! Thank you!

[AK]

How do i get the probability of a particular index (i) ? As in, I know the index and want the associated probability. Thanks!