Suppose if we have a dataset consisting of 5 classes than how to train it for one vs all classification?

I was trying to do but i was unable to do it using fast ai library , if anyone knows please help!!

You could use the output from the sigmoid layer (=last layer) and calculate a softmax from it and then take the argmax?
(And test the case when two units output a similar high value.)

Thanks for suggestion
But currently i am using resnet34 architecture so how to make the change which you are telling?