def __init__(self, nnModel=None, nnTestData=None):
"""
Define the model structure
"""
self.nnModel = nnModel
if (nnModel is None):
l1 = NNLayer(3,
activation=NNLayer.A_Linear,
layerType=NNLayer.T_Input)
l2 = NNLayer(2,
activation=NNLayer.A_Linear,
layerType=NNLayer.T_Hidden)
l3 = NNLayer(1,
activation=NNLayer.A_Linear,
layerType=NNLayer.T_Output)
c1 = NNConnection(l1, l2)
c2 = NNConnection(l2, l3)
#c1 = NNConnection(l1,l3)
self.nnModel = NNModel()
self.nnModel.addConnection(c1)
self.nnModel.addConnection(c2)
"""
Create the test data using the number of nodes in the InputLayer for the col
dimension of the X input
"""
# test data
self.m = 100
self.n = nnModel.getInputLayer().getNodeCnt()
self.x = rng.rand(self.m, self.n)
# the larger
self.w = np.linspace(0, 3, self.n) + 1.0
self.b = 3.5
#err = rng.randn(m)
z = np.dot(self.x, np.transpose(self.w))
self.y = np.asmatrix(z + self.b).reshape(self.m, 1)
else:
if (nnTestData is None):
raise ValueError(
"TestData must be supplied when not using a default nnModel value"
)
# transfer test data to gradient tester
self.m = nnTestData.m
self.n = nnTestData.n
self.x = nnTestData.x
self.w = nnTestData.w
self.b = nnTestData.b
self.y = nnTestData.y
X = iris.data Y = iris.target Yd = pd.get_dummies(Y) Y = Yd.as_matrix() # Test data m = X.shape[0] n = X.shape[1] numClasses = Y.shape[1] # Always make sure the input and output layers and their connections # are the first and last in the list lIn = NNLayer(n,activation=NNLayer.A_Tanh, layerType=NNLayer.T_Input) lH1 = NNLayer(20,activation=NNLayer.A_Tanh, layerType=NNLayer.T_Hidden) lOut = NNLayer(numClasses,activation=NNLayer.A_Softmax, layerType=NNLayer.T_Output) c1 = NNConnection(lIn,lH1) c2 = NNConnection(lH1,lOut) #c1 = NNConnection(l1,l3) nnModel = NNModel() nnModel.addConnection(c1) nnModel.addConnection(c2) # Build the controller and train the model nnController = NNController(nnModel) #yh = nnModel.forwardPropigation(X, Y) #cost = nnModel.getCost(Y) # Train the model
sys.path.append('C:/Users/tschlosser/Google Drive/Python/NNModel')
import numpy as np
import matplotlib.pyplot as plt
rng = np.random
from NNLayer import NNLayer
from NNConnection import NNConnection
from NNModel import NNModel
from NNController import NNController
import NNFunctions as nnf
# Always make sure the input and output layers and their connections
# are the first and last in the list
lIn = NNLayer(2, layerType=NNLayer.T_Input)
lOut = NNLayer(1,activation=NNLayer.A_Sigmoid, layerType=NNLayer.T_Output)
c1 = NNConnection(lIn,lOut)
#c1 = NNConnection(l1,l3)
nnModel = NNModel()
nnModel.addConnection(c1)
# Test data
m = 200
n = lIn.getNodeCnt()
x = 10*(rng.rand(m,n) - 0.5)
# the larger
w = np.linspace(-6,6,n) + 1.0
b = -3.5
err = 0 #10*rng.randn(m)
z = np.asmatrix(np.dot(x,np.transpose(w)) + b + err).reshape(m,1)
# from reaching zero and the cost baseline tends to be related to the noise
# value
x, y, coef = datasets.make_regression(n_samples=m,
n_features=n,
n_informative=1,
noise=10,
coef=True,
random_state=None)
y = np.asmatrix(y).reshape(m, 1)
# Always make sure the input and output layers and their connections
# are the first and last in the list
l1 = NNLayer(n, activation=NNLayer.A_Linear, layerType=NNLayer.T_Input)
l2 = NNLayer(3, activation=NNLayer.A_Linear, layerType=NNLayer.T_Hidden)
l3 = NNLayer(1, activation=NNLayer.A_Linear, layerType=NNLayer.T_Output)
c1 = NNConnection(l1, l2)
c2 = NNConnection(l2, l3)
#c1 = NNConnection(l1,l3)
nnModel = NNModel()
nnModel.addConnection(c1)
nnModel.addConnection(c2)
# save initial params
nnParams = nnModel.getModelParams()
nnModel.setModelParams(nnParams)
'''
Run SG
'''
nnController = NNController(nnModel)
begTime = dt.datetime.now()
results = nnController.trainModelSGD(x,