Belle II Software light-2406-ragdoll
theano.py
1#!/usr/bin/env python3
2
3
10
11try:
12 import theano
13 import theano.tensor
14except ImportError:
15 print("Please install theano: pip3 install theano")
16 import sys
17 sys.exit(1)
18
19import numpy
20
21
22class State:
23 """
24 State class for proper handling of parameters and data during function calls. This is a very brief theano example.
25 """
26
27 def __init__(self, x=None, y=None, params=None, cost=None, updates=None, train_function=None, eval_function=None):
28 """
29 Constructor of the State class
30 """
31 # TODO: make serializable with __getstate__(), __setstate__()
32
33 self.x = x
34
35 self.y = y
36
37
38 self.params = params
39
40 self.cost = cost
41
42
43 self.updates = updates
44
45
46 self.train_function = train_function
47
48 self.eval_function = eval_function
49
50
51def get_model(number_of_features, number_of_spectators, number_of_events, training_fraction, parameters):
52
53 x = theano.tensor.matrix('x')
54 y = theano.tensor.vector('y', dtype='float32')
55
56 # learning_rate = parameters.get('learning_rate', 0.1)
57 learning_rate = 0.1
58
59 n_in = number_of_features
60 n_out = 1
61 rng = numpy.random.RandomState(1234)
62 w_values = numpy.asarray(
63 rng.uniform(
64 low=-numpy.sqrt(6. / (n_in + n_out)),
65 high=numpy.sqrt(6. / (n_in + n_out)),
66 size=(n_in, n_out)
67 ),
68 dtype=theano.config.floatX
69 )
70
71 w_values *= 4
72 w = theano.shared(value=w_values, name='W', borrow=True)
73
74 b_values = numpy.zeros((n_out,), dtype=theano.config.floatX)
75 b = theano.shared(value=b_values, name='b', borrow=True)
76
77 activation = theano.tensor.nnet.sigmoid
78
79 output = activation(theano.tensor.dot(x, w) + b)
80
81 cost = theano.tensor.nnet.binary_crossentropy(output.T, y).mean()
82
83 params = [w, b]
84
85 grad_params = [theano.tensor.grad(cost, param) for param in params]
86
87 updates = [(param, param - learning_rate * gparam) for param, gparam in zip(params, grad_params)]
88
89 train_function = theano.function(
90 inputs=[x, y],
91 outputs=cost,
92 updates=updates
93 )
94
95 eval_function = theano.function(
96 inputs=[x],
97 outputs=output
98 )
99
100 return State(x, y, params, cost, updates, train_function, eval_function)
101
102
103def feature_importance(state):
104 """
105 Return a list containing the feature importances
106 """
107 return []
108
109
110def load(obj):
111 state = State(eval_function=obj[0])
112 return state
113
114
115def apply(state, X):
116 result = state.eval_function(X)
117 return numpy.require(result, dtype=numpy.float32, requirements=['A', 'W', 'C', 'O'])
118
119
120def begin_fit(state, Xvalid, Svalid, yvalid, wvalid, nBatches):
121 return state
122
123
124def partial_fit(state, X, S, y, w, epoch, batch):
125 avg_cost = state.train_function(X, y) / len(y)
126 print("Epoch:", f'{int(epoch):04}', "Batch:", f'{int(batch):04}', "cost=", f"{avg_cost:.9f}")
127 return True
128
129
130def end_fit(state):
131 # FIXME: this might not work as intended
132 # a better method can be found at: http://deeplearning.net/software/theano/tutorial/loading_and_saving.html
133 return [state.eval_function]
y
theano shared variable y
Definition: theano.py:35
eval_function
theano function for evaluation
Definition: theano.py:48
x
theano shared variable x
Definition: theano.py:33
def __init__(self, x=None, y=None, params=None, cost=None, updates=None, train_function=None, eval_function=None)
Definition: theano.py:27
train_function
theano function for training
Definition: theano.py:46