class MLP:
def __init__(self, input, label, n_in, n_hidden, n_out, rng=None):
self.x = input
self.y = label
if rng is None:
rng = numpy.random.RandomState(1234)
# construct hidden_layer
self.hidden_layer = HiddenLayer(input=self.x,
n_in=n_in,
n_out=n_hidden,
rng=rng,
activation=tanh)
# construct log_layer
self.log_layer = LogisticRegression(input=self.hidden_layer.output,
label=self.y,
n_in=n_hidden,
n_out=n_out)
def train(self):
# forward hidden_layer
layer_input = self.hidden_layer.forward()
self.log_layer.train(input=layer_input)
# backward hidden_layer
self.hidden_layer.backward(prev_layer=self.log_layer)
def predict(self, x):
x = self.hidden_layer.output(input=x)
return self.log_layer.predict(x)
class CNN:
def __init__(self,
N,
label,
n_hidden,
n_out,
image_size,
channel,
n_kernels,
kernel_sizes,
pool_sizes,
rng=None,
activation=ReLU):
if rng is None:
rng = numpy.random.RandomState(1234)
self.N = N
self.n_hidden = n_hidden
self.n_kernels = n_kernels
self.pool_sizes = pool_sizes
self.conv_layers = []
self.conv_sizes = []
# construct 1st conv_layer
conv_layer0 = ConvPoolLayer(N, image_size, channel, n_kernels[0],
kernel_sizes[0], pool_sizes[0], rng,
activation)
self.conv_layers.append(conv_layer0)
conv_size = [
(image_size[0] - kernel_sizes[0][0] + 1) / pool_sizes[0][0],
(image_size[1] - kernel_sizes[0][1] + 1) / pool_sizes[0][1]
]
self.conv_sizes.append(conv_size)
# construct 2nd conv_layer
conv_layer1 = ConvPoolLayer(N, conv_size, n_kernels[0], n_kernels[1],
kernel_sizes[1], pool_sizes[1], rng,
activation)
self.conv_layers.append(conv_layer1)
conv_size = [
(conv_size[0] - kernel_sizes[1][0] + 1) / pool_sizes[1][0],
(conv_size[1] - kernel_sizes[1][0] + 1) / pool_sizes[1][1]
]
self.conv_sizes.append(conv_size)
# construct hidden_layer
self.hidden_layer = HiddenLayer(
None, n_kernels[-1] * conv_size[0] * conv_size[1], n_hidden, None,
None, rng, activation)
# construct log_layer
self.log_layer = LogisticRegression(None, label, n_hidden, n_out)
# def train(self, epochs, learning_rate, input=None):
def train(self, epochs, learning_rate, input, test_input=None):
for epoch in xrange(epochs):
if (epoch + 1) % 5 == 0:
print 'iter = %d/%d' % (epoch + 1, epochs)
print
print '------------------'
print 'TEST PROCESSING...'
print self.predict(test_input)
print '------------------'
print
# forward first conv layer
pooled_X = self.conv_layers[0].forward(input=input)
# forward second conv layer
pooled_X = self.conv_layers[1].forward(input=pooled_X)
# flatten input
layer_input = self.flatten(pooled_X)
# forward hidden layer
layer_input = self.hidden_layer.forward(input=layer_input)
# forward & backward logistic layer
self.log_layer.train(lr=learning_rate, input=layer_input)
# backward hidden layer
self.hidden_layer.backward(prev_layer=self.log_layer,
lr=learning_rate)
flatten_size = self.n_kernels[-1] * self.conv_sizes[-1][
0] * self.conv_sizes[-1][1]
delta_flatten = numpy.zeros((self.N, flatten_size))
for n in xrange(self.N):
for i in xrange(flatten_size):
for j in xrange(self.n_hidden):
delta_flatten[n][i] += self.hidden_layer.W[i][
j] * self.hidden_layer.d_y[n][j]
# unflatten delta
delta = numpy.zeros(
(len(delta_flatten), self.n_kernels[-1],
self.conv_sizes[-1][0], self.conv_sizes[-1][1]))
for n in xrange(len(delta)):
index = 0
for k in xrange(self.n_kernels[-1]):
for i in xrange(self.conv_sizes[-1][0]):
for j in xrange(self.conv_sizes[-1][1]):
delta[n][k][i][j] = delta_flatten[n][index]
index += 1
# backward second conv layer
delta = self.conv_layers[1].backward(delta, self.conv_sizes[1],
learning_rate)
# backward first conv layer
self.conv_layers[0].backward(delta, self.conv_sizes[0],
learning_rate)
def flatten(self, input):
flatten_size = self.n_kernels[-1] * self.conv_sizes[-1][
0] * self.conv_sizes[-1][1]
flattened_input = numpy.zeros((len(input), flatten_size))
for n in xrange(len(flattened_input)):
index = 0
for k in xrange(self.n_kernels[-1]):
for i in xrange(self.conv_sizes[-1][0]):
for j in xrange(self.conv_sizes[-1][1]):
flattened_input[n][index] = input[n][k][i][j]
index += 1
# print flattened_input
return flattened_input
def predict(self, x):
pooled_X = self.conv_layers[0].forward(input=x)
pooled_X = self.conv_layers[1].forward(input=pooled_X)
layer_input = self.flatten(pooled_X)
x = self.hidden_layer.output(input=layer_input)
return self.log_layer.predict(x)