class InnerProductLayer(Layer):
def __init__(self, K, N):
Layer.__init__(self)
# input number of neuron
self.K_ = K
# output number of neuron
self.N_ = N
self.bias_term_ = None
self.bias_multiplier_ = None
self.transpose_ = False
self.W = Blob()
self.b = Blob()
self.blobs_.append(self.W)
self.blobs_.append(self.b)
def LayerSetup(self, bottom, top):
self.M_ = bottom[0].data().shape[0]
W_shape = (self.K_, self.N_)
b_shape = (1, self.N_)
self.W.Reshape(W_shape)
self.b.Reshape(b_shape)
# Xavier
fan_in = self.W.count() / self.W.shape()[0]
fan_out = self.W.count() / self.W.shape()[1]
n = (fan_in + fan_out) / 2
scale = numpy.sqrt(3.0 / n)
#self.W.set_data(numpy.random.uniform(-scale, scale, self.W.count()) )
self.W.set_data(
numpy.random.normal(loc=0.0, scale=0.1, size=self.W.count()))
self.W.Reshape(W_shape)
def Reshape(self, bottom, top):
top_shape = (self.M_, self.N_)
top[0].Reshape(top_shape)
def type(self):
return 'InnerProduct'
def ExactNumBottomBlobs(self):
return 1
def ExactNumTopBlobs(self):
return 1
def Forward_cpu(self, bottom, top):
top[0].set_data(
numpy.matmul(bottom[0].data(), self.W.data()) + self.b.data())
def Backward_cpu(self, top, propagate_down, bottom):
self.W.set_diff(
self.W.diff() +
numpy.matmul(bottom[0].data().transpose(), top[0].diff()))
self.b.set_diff(self.b.diff() + numpy.sum(top[0].diff(), axis=0))
bottom[0].set_diff(
numpy.matmul(top[0].diff(),
self.W.data().transpose()))
