def __init__(self, dim_list, eta = 0.1):
"""
Constructor for network.
Params:
dim_list: a list of the number of dimension for each layer.
eta: learning rate for each gradient descent step
"""
depth = len(dim_list)
self.depth = depth
self.dim_list = dim_list
self.eta = eta
# 1. Initiate each layer: output, partial_output and weight,
# although partial_output is useless for the input layer, similarly
# weight and bias are useless for the output layer.
#
# 2. Partial_weight is an internal variable and will not be stored in
# a layer.
#
self.layers = [ {'output':Vector.fromIterable(0 for i in xrange(dim_list[l])),
'partial_output':Vector.fromIterable(0 for i in xrange(dim_list[l])),
'weight':Matrix.fromRandom(dim_list[l + 1], dim_list[l]),
'bias':Vector.fromRandom(dim_list[l + 1])}
for l in xrange(depth - 1) ]
# output layer
self.layers.append({'output':Vector.fromList([0] * dim_list[depth - 1]),
'partial_output':Vector.fromList([0] * dim_list[depth - 1]),
'weight': None, 'bias': None})
def main(mnist_path):
puttime('start loading')
network = feedforward_network.FeedForwardNetwork(
dim_list = [2, 2, 2],
eta = 0.5
)
network.layers[0]['weight'] = Matrix(2, 2, [.15, .20, .25, .30])
network.layers[0]['bias'] = Vector.fromList([.35, .35])
network.layers[1]['weight'] = Matrix(2, 2, [.40, .45, .50, .55])
network.layers[1]['bias'] = Vector.fromList([.60, .60])
x = Vector([.05, .10])
y = Vector([.01, .99])
def generator(x, y):
yield (x, y)
# start training
puttime('start training')
network.train(generator(x, y), puttime, limit = 1)
# testing
out = network.inference(x)
print x, '->', out
# start training
puttime('start training')
network.train(generator(x, y), puttime, limit = 1)
# testing
out = network.inference(x)
print x, '->', out
def inference(self, vector):
self._forward(vector)
output = self.layers[self.depth - 1]['output']
maxpos, maxval = 0, 0
for i in xrange(len(output)):
if maxval < output[i]:
maxpos, maxval = i, output[i]
rtn = Vector.fromList([0] * len(output))
rtn[maxpos] = 1
return rtn
