In the Google's paper under section 3.4 it is said that:
However we disable backpropagation
with respect to bounding box coordinates as we have
found this to be unstable during training.
This can now be implemented using keras.backend.stop_gradient() I guess. But I am not sure how to implement it. Any ideas? @0x00b1 @yhenon @hgaiser
Thanks
In the Google's paper under section 3.4 it is said that:
This can now be implemented using keras.backend.stop_gradient() I guess. But I am not sure how to implement it. Any ideas? @0x00b1 @yhenon @hgaiser Thanks