require 'nn'
-- a typical convolutional network, with locally-normalized hidden
-- units, and L2-pooling

-- Note: the architecture of this convnet is loosely based on Pierre Sermanet's
-- work on this dataset (http://arxiv.org/abs/1204.3968). In particular
-- the use of LP-pooling (with P=2) has a very positive impact on
-- generalization. Normalization is not done exactly as proposed in
-- the paper, and low-level (first layer) features are not fed to
-- the classifier.

model = nn.Sequential()

-- stage 1 : filter bank -> squashing -> L2 pooling -> normalization
model:add(nn.SpatialConvolutionMM(nfeats, nstates[1], filtsize, filtsize))
model:add(nn.Tanh())
model:add(nn.SpatialLPPooling(nstates[1],2,poolsize,poolsize,poolsize,poolsize))
model:add(nn.SpatialSubtractiveNormalization(nstates[1], normkernel))

-- stage 2 : filter bank -> squashing -> L2 pooling -> normalization
model:add(nn.SpatialConvolutionMM(nstates[1], nstates[2], filtsize, filtsize))
model:add(nn.Tanh())
model:add(nn.SpatialLPPooling(nstates[2],2,poolsize,poolsize,poolsize,poolsize))
model:add(nn.SpatialSubtractiveNormalization(nstates[2], normkernel))

-- stage 3 : standard 2-layer neural network
model:add(nn.Reshape(nstates[2]*filtsize*filtsize))
model:add(nn.Linear(nstates[2]*filtsize*filtsize, nstates[3]))
model:add(nn.Tanh())
model:add(nn.Linear(nstates[3], noutputs))