def create_convnet(theta, is_tr=True):
# input
a0 = T.matrix("x")
a00 = a0.reshape((a0.shape[0], 1, 28, 28))
# I: conv pool
theta1 = [theta[0], theta[1]]
sz1_pool = (2, 2)
a1 = models.create_conv_pool_tanh(a00, theta1, sz1_pool)
# II: conv pool
theta2 = [theta[2], theta[3]]
sz2_pool = (2, 2)
a2 = models.create_conv_pool_tanh(a1, theta2, sz2_pool)
# III: fc + dropout
theta3 = [theta[4], theta[5]]
if is_tr:
a3 = models.create_dropout(models.create_linear_sigmoid(a2.flatten(2), theta3), trng=trng)
else:
a3 = models.create_dropout(models.create_linear_sigmoid(a2.flatten(2), theta3), trng=None)
# IV: fc, output
theta4 = [theta[6], theta[7]]
a4 = models.create_linear(a3, theta4)
return a0, a4
def create_mlp(theta, is_tr=True):
# input layer
x = T.matrix('x')
# hidden layer I
theta1 = (theta[0], theta[1])
if is_tr:
a1 = models.create_dropout(models.create_linear_sigmoid(x, theta1), trng=trng)
else:
a1 = models.create_dropout(models.create_linear_sigmoid(x, theta1), trng=None)
# hidden layer II
theta2 = (theta[2], theta[3])
if is_tr:
a2 = models.create_dropout(models.create_linear_sigmoid(a1, theta2), trng=trng)
else:
a2 = models.create_dropout(models.create_linear_sigmoid(a1, theta2), trng=None)
# output layer
theta3 = (theta[4], theta[5])
a3 = models.create_linear_sigmoid(a2, theta3)
return x, a3