Closed agude closed 8 years ago
total_loss is often larger than the sum of all the loses, for example here is the output of a training run:
total_loss
Loss stats: { mid_box_reg_loss : 0.0019056373813539 captioning_loss : 5.0455059438944 end_box_reg_loss : 0.002998169460654 end_objectness_loss : 0.038921515512466 total_loss : 10.398667634428 mid_objectness_loss : 0.11150163569537 } Average loss: 10.398667634428
Note that 10.3... is not the sum of the other losses in the losses table. I think there is a bug here: https://github.com/jcjohnson/densecap/blob/master/densecap/DenseCapModel.lua#L455-L458
losses
total_loss is added to the losses table, the table is then iterated over, and all values are added to the total_loss entry. At some point in that for loop it will look like this: total_loss = total_loss + total_loss.
total_loss = total_loss + total_loss
total_lossis often larger than the sum of all the loses, for example here is the output of a training run:Note that 10.3... is not the sum of the other losses in the
lossestable. I think there is a bug here: https://github.com/jcjohnson/densecap/blob/master/densecap/DenseCapModel.lua#L455-L458total_lossis added to thelossestable, the table is then iterated over, and all values are added to thetotal_lossentry. At some point in that for loop it will look like this:total_loss = total_loss + total_loss.