Why exp function here?

[jichilen]

g-tensorflow-models/research/object_detection/core/losses.py

Line 215 in d0ff754

w_ln = tf.exp(matched_giou)

should i use the exp function for faster r-cnn training?

g-tensorflow-models/research/object_detection/core/losses.py

Line 217 in d0ff754

return tf.reshape(weights, [-1]) * per_anchor_giou_loss

And what weigths should i choose?
0.25 in g-darknet or 10 as paper metioned about?

[nathantsoi]

Excellent question, that exp was from an experiment (since this is used in darknet), it should be removed or behind a flag. I will remove it for now.

Unfortunately I don't have a good answer for the weights, it really depends on the network. You might try even weighting then doing some search for better params