Hard Negative Mining by Loss Ratio

Problem

Select the hardest negative examples for training.

Given negative losses, choose min(len(losses), pos_count * neg_ratio) negatives with the largest losses. Return their indices in descending loss order, breaking ties by smaller index.

Test cases

case 1

in: [[0.1,0.9,0.5,0.8],1,2]

expected: [1,3]

case 2

in: [[0.1,0.9,0.5,0.8],1,1]

expected: [1]

case 3

in: [[0.1,0.9,0.5,0.8],0,3]

expected: []

case 4

in: [[0.1,0.9,0.5,0.8],2,5]

expected: [1,3,2,0]

Starter codePython

def hard_negative_indices(neg_losses, pos_count, neg_ratio):
    # Your code here
    pass

4 test cases · run your code against them in the editor