Analyzing train container: train_results_2019-01-06 13:07:54.558547

[rock-it-with-asher]

so in about 100,000 steps with 1e-3 learning rate and 0.5 momentum

[rock-it-with-asher]

• it seems like main cause for "big"
  loss (0.7) is rpn_bbox_loss which stand on 0.35. it seems pretty much logic since about half the objects in a generated picture are leaves , and empirically we observe that around ~0.5 of the cases the anchor choosed to contain this leaf by the rpn ( which doesn't know that it's a leaf but guesses accurately that there is an object) are rectangles which cover about half the leaf. this drives the layers which try to learn deltas for anchor "fixing" crazy ( it is harder to converge on the "right general" delta (dx,dy axes) fitting for anchors.

[rock-it-with-asher]

rpn_class_loss: 0.0153 , which is great, it means that our RPN identifies precisely (1% error) wheather there is an object in anchor or not

[rock-it-with-asher]

mrcnn_class_loss: ~0.08 . it means that it is correct to classify a given object in 92% , which is again a great achievement