Tensor Flow: Weight Interpretation for Weighted Cross-Entropy

Question

The Tensorflow function tf.nn.weighted_cross_entropy_with_logits()accepts an argument pos_weight. The documentation defines it pos_weightas a “coefficient that should be used with positive examples." I suggest that this means that an pos_weightincrease increases the loss of false positives and reduces the loss of false negatives. Or do I have it back?

Answer 1

Actually, it's the other way around. Summary Documentation:

The argument is pos_weightused as a multiplier for a positive purpose.

, 5 7 , pos_weight=2, , 10 7 .

, . 5 0 . pos_weight, . , .