-
Notifications
You must be signed in to change notification settings - Fork 0
/
nn_af_newTest.py
228 lines (181 loc) · 6.83 KB
/
nn_af_newTest.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
import numpy as np
import math
from sklearn.preprocessing import LabelBinarizer
import pylab as plt
import arrayfire as af
#af.set_device(1)
def _relu(x, eps=1e-5): return max(eps, x)
def _d_relu(x, eps=1e-5): return 1. if x > eps else 0.0
def _sigmoid(x): return 1 / (1 + math.exp(-x))
def _tanh(x): return math.tanh(x)
def _d_tanh(x):
t = _tanh(x)
return 1 - t*t
def _d_sigmoid(x):
s = _sigmoid(x)
return s * (1 - s)
#2 sites, 50 items. To 50 sites, 2 items
def d_cost(output, target): return output - target
sigmoid = np.vectorize(_sigmoid)
d_sigmoid = np.vectorize(_d_sigmoid)
relu = np.vectorize(_relu)
d_relu = np.vectorize(_d_relu)
tanh = np.vectorize(_tanh)
d_tanh = np.vectorize(_d_tanh)
def activate(act):
"""
Returns a function and derivative tuple
:param act: 'sigmoid', 'tanh', or 'ReLU'
:return: a tuple of functions (fun, grad)
"""
if act == 'sigmoid': return (sigmoid, d_sigmoid)
if act == 'relu': return (relu, d_relu)
if act == 'tanh': return (tanh, d_tanh)
def addadadelta(nn):
# need to add error checking if weights have been already initialized
nn['E2'] = {}
nn['E2']['W'] = []
nn['E2']['b'] = []
nn['EW2'] = {}
nn['EW2']['W'] = []
nn['EW2']['b'] = []
for i in range(len(nn['weights'])):
nn['E2']['W'].append(np.zeros(nn['weights'][i].shape))
nn['E2']['b'].append(np.zeros(nn['biases'][i].shape))
nn['EW2']['W'].append(np.zeros(nn['weights'][i].shape))
nn['EW2']['b'].append(np.zeros(nn['biases'][i].shape))
def weight_matrix(seed, innum, outnum, type='glorot', layer=0):
"""
Returns randomly initialized weight matrix of appropriate dimensions
:param innum: number of neurons in the layer i
:param outnum: number of neurons in layer i+1
:return: weight matrix
"""
np.random.seed(seed)
if type == 'glorot': W = af.randu(outnum,innum) #np.random.uniform(low=-np.sqrt(6.0/(2*layer+1)), high=np.sqrt(6.0/(2*layer+1)), size=(outnum, innum))
if type == 'normal': W = af.randu(outnum,innum)# np.random.rand(outnum, innum)
return W
def nn_build(seed,layerlist, nonlin='sigmoid', eta=0.01, init='glorot'):
"""
Returns a neural network of given architecture
:param layerlist: a list of number of neurons in each layer starting with input and ending with output
:param nonlin: nonlinearity which is either 'sigmoid', 'tanh', or 'ReLU'
:return: a dictionary with neural network parameters
"""
nn = {}
nn['eta'] = eta
nn['weights'] = []
weights = []
biases = []
nn['biases'] = []
nn['nonlin'] = []
for i in range(len(layerlist) - 1):
nn['weights'].append(weight_matrix(seed,layerlist[i], layerlist[i + 1],layer=i,type=init))
nn['biases'].append(af.constant(1,1,layerlist[i + 1]) * 0.1)
nn['nonlin'].append(activate(nonlin))
#nn['weights'] = weights
#nn['biases'] = biases
#addadadelta(nn)
return nn
def forward(nn_np, data):
"""
Given a dictionary representing a feed forward neural net and an input data matrix compute the network's output and store it within the dictionary
:param nn: neural network dictionary
:param data: a numpy n by m matrix where m in the number of input units in nn
:return: the output layer activations
"""
nn = nn_np
nn['activations'] =[]
afData = data.ctypes.data, data.shape, data.dtype.char
for i in range(data.shape[0]):
nn['activations'].append(af.Array(data[i].ctypes.data,data[i].shape,data[i].dtype.char))
nn['zs'] = []
for w, s, b in map(None, nn['weights'], nn['nonlin'], nn['biases']):
#print af.transpose(nn['activations'][-1])
print nn['activations'][-1]
print ""
print w
z = af.dot(nn['activations'][-1],w)
nn['zs'].append(af.transpose(z))
nn['activations'].append(s[0](af.transpose(z)))
return nn['activations'][-1]
def test_forward():
nn = {}
nn['eta'] = 0.1
nn['weights'] = [np.array([[0.1, 0.2], [0.2, 0.5], [0.3, 0.4]]), np.array(np.array([0.1, 0.3, 0.2]))]
nn['biases'] = [np.ones(3) * 0.1, np.ones(1) * 0.1]
nn['nonlin'] = [(sigmoid, d_sigmoid)]
x = np.array([1, 0])
t = sigmoid(np.dot(nn['weights'][1], sigmoid(np.dot(nn['weights'][0], x) + nn['biases'][0]) + nn['biases'][1]))
#ipdb.set_trace()
print forward(nn, x)
print t
def average_gradient(deltas, activations):
dW = 0
for i in range(deltas.shape[1]):
dW += np.outer(deltas[:,i], activations[:,i].T)
return dW#/deltas.shape[1]
def gradient(nn, delta):
nabla_b = []
nabla_w = []
# output
dact = nn['nonlin'][-1][1]
dW = average_gradient(delta*dact(nn['zs'][-1]), nn['activations'][-2])
nabla_b.append(np.mean(delta, axis=1))
nabla_w.append(dW)
for i in range(len(nn['weights']) - 2, -1, -1):
dact = nn['nonlin'][i][1]
delta = np.dot(nn['weights'][i+1].T, delta * dact(nn['zs'][i+1]))
dW = average_gradient(delta,nn['activations'][i])
nabla_b.append(np.mean(delta*dact(nn['zs'][i]),axis=1))
nabla_w.append(dW)
return nabla_w, nabla_b
def backprop(nn, nabla_w, nabla_b):
eta = nn['eta']
for i in range(len(nn['weights'])):
nn['weights'][i] -= eta * nabla_w[-i - 1]
nn['biases'][i] -= eta * nabla_b[-i - 1]
def expand_labels(labels):
n = len(np.unique(labels))
# if n == 2:
# l = np.array([(0,1) if row==0 else (1,0) for row in labels]).T
# else:
lb = LabelBinarizer()
l = lb.fit_transform(labels).T
return l
def master_node(nn,data,labels):
nabla_w = []
nabla_b = []
for net in nn:
r = forward(net, data)
delta = d_cost(r,labels)
w,b = gradient(net, delta)
nabla_w += w
nabla_b += b
nabla_w = [x / len(nn) for x in nabla_w]
nabla_b = [x / len(nn) for x in nabla_b]
for net in nn:
backprop(net,nabla_w,nabla_b)
def minibatch_fit(nn, data, labels):
r = forward(nn, data)
dact = nn['nonlin'][-1][1]
delta = d_cost(r, labels) #* dact(nn['zs'][-1])
backprop(nn, delta)
# ipdb.set_trace()
return np.sqrt(np.sum(np.square(r - labels))) / 2
def plot_decision2D(nn, data, labels, res=0.01):
# http://stackoverflow.com/questions/19054923/plot-decision-boundary-matplotlib?answertab=votes#tab-top
x_min, x_max = data[:, 0].min() - 1, data[:, 0].max() + 1
y_min, y_max = data[:, 1].min() - 1, data[:, 1].max() + 1
xx, yy = np.meshgrid(np.arange(x_min, x_max, res),
np.arange(y_min, y_max, res))
Y = np.c_[xx.ravel(), yy.ravel()]
Z = np.array([round(forward(nn,x)[0]) for x in Y])
# Put the result into a color plot
Z = Z.reshape(xx.shape)
plt.hold(True)
plt.contourf(xx, yy, Z, cmap=plt.cm.Paired)
plt.axis('off')
# Plot also the training points
plt.scatter(data[:, 0], data[:, 1], c=labels, cmap=plt.cm.Paired)
plt.hold(False)