LeNet5 is a convolutional network architecture described in several
publications, notably in [LeCun, Bottou, Bengio and Haffner 1998]:
"gradient-based learning applied to document recognition", Proc IEEE,
Nov 1998. The paper is also available at
|220.127.116.11.16.0. (new-lenet5 image-height image-width ki0 kj0 si0 sj0 ki1 kj1 si1 sj1 hid output-size net-param)
create a new instance of net-cscscf implementing a LeNet-5 type
convolutional neural net. This network has regular sigmoid units on the
output, not an extra RBF layer as described in the Proc. IEEE paper. The
network has 6 feature maps at the first layer and 16 feature maps at the
second layer with a connection matrix between feature maps as described
in the paper. Arguments:
<image-height> <image-width>: height and width of input image
<ki0> <kj0>: height and with of convolutional kernel, first layer.
<si0> <sj0>: subsampling ratio of subsampling layer, second layer.
<ki1> <kj1>: height and with of convolutional kernel, third layer.
<si1> <sj1>: subsampling ratio of subsampling layer, fourth layer.
<hid>: number of hidden units, fifth layer
<output-size>: number of output units
<net-param>: idx1-ddparam that will hold the trainable parameters
of the network
(setq p (new idx1-ddparam 0 0.1 0.02 0.02 80000))
(setq z (new-lenet5 32 32 5 5 2 2 5 5 2 2 120 10 p))