however, that doesn't really work if we need to compute the gradients with
respect to the parameters which the layer attempted to do.  So I made the
layer's parameters non-learnable so it can continue to be inplace since the
main usecase for this layer is to be a fast replacement for bn_ during testing.

Python/Simple object detector trainer. Added upsample_limit option

Also made the bracketing phase more efficient.

amount of time.

--HG--
rename : examples/dnn_mit67_ex.cpp => examples/dnn_mnist_resnet_ex.cpp

immediately after the input layer.

introduce a configurable threshold for rendering HOGs

how much the current step size has converged.