Penalize some classes more than others in cross entropy (but only for a specific class!)

[steeltrooper17 (Rohit Gajawada)]

I have three classes: small drone, large drone and bird.

I want my loss function to penalize small drone and large drone predictions equally when classifying a bird sample.
However, if the sample is actually a small drone – it should penalize large drone misclassification (small drone sample misclassified as large drone) less compared to a bird misclassification (small drone sample misclassified as bird).
Similarly, if the sample is actually a large drone – it should penalize small drone misclassification (large drone sample mis classified as small drone) less compared to a bird misclassification (large drone sample mis classified as bird).

How would I achieve this? Do I need to make a custom loss function or is this supported in pytorch?

[zshn25 (Zeeshan Khan Suri)]

You could do multi-stage binary classification.

1. Stage: 0 for bird, 1 for drone
2. Stage: only if drone in 1st stage.

Final loss would be combination of the above with additional weighting if required

[KFrank]

Hi Rohit!

@steeltrooper17

How would I achieve this?

If I understand your question, you will need a custom loss function.

I believe that your use case is addressed by this post (and its entire thread):

Create a Variation of Cross Entropy Loss with Per Pair of Classes Weights

Hi David! Cross entropy (by definition) doesn’t work this way. But you can write some other loss function that does have “per-pair” penalties. I give a couple of possibilities in the example script, below. misclassA() just weights the probability for an incorrect prediction by the per-pair weight given in your matrix D. My guess is that such a loss function won’t train very well. In my experience, the logarithmic divergence for bad predictions in cross entropy seems to be very helpfu…

Best.

K. Frank