Mercurial > pylearn
comparison kernel_regression.py @ 421:e01f17be270a
Kernel regression learning algorithm
author | Yoshua Bengio <bengioy@iro.umontreal.ca> |
---|---|
date | Sat, 19 Jul 2008 10:11:22 -0400 |
parents | |
children | 32c5f87bc54e |
comparison
equal
deleted
inserted
replaced
420:040cb796f4e0 | 421:e01f17be270a |
---|---|
1 """ | |
2 Implementation of kernel regression: | |
3 """ | |
4 | |
5 from pylearn.learner import OfflineLearningAlgorithm | |
6 from theano import tensor as T | |
7 from nnet_ops import prepend_1_to_each_row | |
8 from theano.scalar import as_scalar | |
9 from common.autoname import AutoName | |
10 import theano | |
11 import numpy | |
12 | |
13 class KernelRegression(OfflineLearningAlgorithm): | |
14 """ | |
15 Implementation of kernel regression: | |
16 * the data are n (x_t,y_t) pairs and we want to estimate E[y|x] | |
17 * the predictor computes | |
18 f(x) = b + \sum_{t=1}^n \alpha_t K(x,x_t) | |
19 with free parameters b and alpha, training inputs x_t, | |
20 and kernel function K (gaussian by default). | |
21 Clearly, each prediction involves O(n) computations. | |
22 * the learner chooses b and alpha to minimize | |
23 lambda alpha' G' G alpha + \sum_{t=1}^n (f(x_t)-y_t)^2 | |
24 where G is the matrix with entries G_ij = K(x_i,x_j). | |
25 The first (L2 regularization) term is the squared L2 | |
26 norm of the primal weights w = \sum_t \alpha_t phi(x_t) | |
27 where phi is the function s.t. K(u,v)=phi(u).phi(v). | |
28 * this involves solving a linear system with (n+1,n+1) | |
29 matrix, which is an O(n^3) computation. In addition, | |
30 that linear system matrix requires O(n^2) memory. | |
31 So this learning algorithm should be used only for | |
32 small datasets. | |
33 * the linear system is | |
34 (M + lambda I_n) theta = (1, y)' | |
35 where theta = (b, alpha), I_n is the (n+1)x(n+1) matrix that is the identity | |
36 except with a 0 at (0,0), M is the matrix with G in the sub-matrix starting | |
37 at (1,1), 1's in column 0, except for a value of n at (0,0), and sum_i G_{i,j} | |
38 in the rest of row 0. | |
39 | |
40 Note that this is gives an estimate of E[y|x,training_set] that is the | |
41 same as obtained with a Gaussian process regression. The GP | |
42 regression would also provide a Bayesian Var[y|x,training_set]. | |
43 It corresponds to an assumption that f is a random variable | |
44 with Gaussian (process) prior distribution with covariance | |
45 function K. Because we assume Gaussian noise we obtain a Gaussian | |
46 posterior for f (whose mean is computed here). | |
47 | |
48 | |
49 Usage: | |
50 | |
51 kernel_regressor=KernelRegression(L2_regularizer=0.1,kernel=GaussianKernel(gamma=0.5)) | |
52 kernel_predictor=kernel_regressor(training_set) | |
53 all_results_dataset=kernel_predictor(test_set) # creates a dataset with "output" and "squared_error" field | |
54 outputs = kernel_predictor.compute_outputs(inputs) # inputs and outputs are numpy arrays | |
55 outputs, errors = kernel_predictor.compute_outputs_and_errors(inputs,targets) | |
56 errors = kernel_predictor.compute_errors(inputs,targets) | |
57 mse = kernel_predictor.compute_mse(inputs,targets) | |
58 | |
59 | |
60 | |
61 The training_set must have fields "input" and "target". | |
62 The test_set must have field "input", and needs "target" if | |
63 we want to compute the squared errors. | |
64 | |
65 The predictor parameters are obtained analytically from the training set. | |
66 Training is only done on a whole training set rather than on minibatches | |
67 (no online implementation). | |
68 | |
69 The dataset fields expected and produced by the learning algorithm and the trained model | |
70 are the following: | |
71 | |
72 - Input and output dataset fields (example-wise quantities): | |
73 | |
74 - 'input' (always expected as an input_dataset field) | |
75 - 'target' (always expected by the learning algorithm, optional for learned model) | |
76 - 'output' (always produced by learned model) | |
77 - 'squared_error' (optionally produced by learned model if 'target' is provided) | |
78 = example-wise squared error | |
79 """ | |
80 def __init__(self, kernel=None, L2_regularizer=0, gamma=1): | |
81 self.kernel = kernel | |
82 self.L2_regularizer=L2_regularizer | |
83 self.gamma = gamma # until we fix things, the kernel type is fixed, Gaussian | |
84 self.equations = KernelRegressionEquations() | |
85 | |
86 def __call__(self,trainset): | |
87 n_examples = len(trainset) | |
88 first_example = trainset[0] | |
89 n_inputs = first_example['input'].size | |
90 n_outputs = first_example['target'].size | |
91 M = numpy.zeros((n_examples+1,n_examples+1)) | |
92 Y = numpy.zeros((n_examples+1,n_outputs)) | |
93 for i in xrange(n_inputs): | |
94 M[i+1,i+1]=self.L2_regularizer | |
95 data = trainset.fields() | |
96 train_inputs = numpy.array(data['input']) | |
97 Y[0]=1 | |
98 Y[1:,:] = numpy.array(data['target']) | |
99 M,train_inputs_square=self.equations.compute_system_matrix(train_inputs,M) | |
100 theta=numpy.linalg.solve(M,Y) | |
101 return KernelPredictor(theta,self.gamma, train_inputs, train_inputs_square) | |
102 | |
103 class KernelPredictorEquations(AutoName): | |
104 train_inputs = T.matrix() # n_examples x n_inputs | |
105 train_inputs_square = T.vector() # n_examples | |
106 inputs = T.matrix() # minibatchsize x n_inputs | |
107 targets = T.matrix() # minibatchsize x n_outputs | |
108 theta = T.matrix() # (n_examples+1) x n_outputs | |
109 gamma = T.scalar() | |
110 inv_gamma2 = 1./(gamma*gamma) | |
111 b = theta[0] | |
112 alpha = theta[1:,:] | |
113 inputs_square = T.sum(inputs*inputs,axis=1) | |
114 Kx = exp(-(train_inputs_square-2*dot(inputs,train_inputs.T)+inputs_square)*inv_gamma2) | |
115 outputs = T.dot(Kx,alpha) + b # minibatchsize x n_outputs | |
116 squared_errors = T.sum(T.sqr(targets-outputs),axis=1) | |
117 | |
118 __compiled = False | |
119 @classmethod | |
120 def compile(cls,linker='c|py'): | |
121 if cls.__compiled: | |
122 return | |
123 def fn(input_vars,output_vars): | |
124 return staticmethod(theano.function(input_vars,output_vars, linker=linker)) | |
125 | |
126 cls.compute_outputs = fn([cls.inputs,cls.theta,cls.gamma,cls.train_inputs,cls.train_inputs_square],[cls.outputs]) | |
127 cls.compute_errors = fn([cls.outputs,cls.targets],[cls.squared_errors]) | |
128 | |
129 cls.__compiled = True | |
130 | |
131 def __init__(self): | |
132 self.compile() | |
133 | |
134 class KernelRegressionEquations(KernelPredictorEquations): | |
135 # P = KernelPredictorEquations | |
136 M = T.matrix() # (n_examples+1) x (n_examples+1) | |
137 inputs = T.matrix() # n_examples x n_inputs | |
138 G = M[1:,1:] | |
139 new_G = gemm(G,1.,inputs,inputs.T,1.) | |
140 M2 = T.add_inplace(M,new_G) | |
141 M2[0,0] = M.shape[0] | |
142 M2[1:,0] = 1 | |
143 M2[0,1:] = T.sum(G,axis=0) | |
144 inputs_square = T.sum(inputs*inputs,axis=1) | |
145 | |
146 __compiled = False | |
147 | |
148 @classmethod | |
149 def compile(cls,linker='c|py'): | |
150 if cls.__compiled: | |
151 return | |
152 def fn(input_vars,output_vars): | |
153 return staticmethod(theano.function(input_vars,output_vars, linker=linker)) | |
154 | |
155 cls.compute_system_matrix = fn([cls.inputs,cls.M],[cls.M2,cls.inputs_square]) | |
156 | |
157 cls.__compiled = True | |
158 | |
159 def __init__(self): | |
160 self.compile() | |
161 | |
162 class KernelPredictor(object): | |
163 """ | |
164 A kernel predictor has parameters theta (a bias vector and a weight matrix alpha) | |
165 it can use to make a non-linear prediction (according to the KernelPredictorEquations). | |
166 It can compute its output (bias + alpha * kernel(train_inputs,input) and a squared error (||output - target||^2). | |
167 """ | |
168 def __init__(self, theta, gamma, train_inputs, train_inputs_square): | |
169 self.theta=theta | |
170 self.gamma=gamma | |
171 self.train_inputs=train_inputs | |
172 self.train_inputs_square=train_inputs_square | |
173 self.equations = LinearPredictorEquations() | |
174 | |
175 def compute_outputs(self,inputs): | |
176 return self.equations.compute_outputs(inputs,self.theta,self.gamma,self.train_inputs,self.train_inputs_square) | |
177 def compute_errors(self,inputs,targets): | |
178 return self.equations.compute_errors(self.compute_outputs(inputs),targets) | |
179 def compute_outputs_and_errors(self,inputs,targets): | |
180 outputs = self.compute_outputs(inputs) | |
181 return [outputs,self.equations.compute_errors(outputs,targets)] | |
182 def compute_mse(self,inputs,targets): | |
183 errors = self.compute_errors(inputs,targets) | |
184 return numpy.sum(errors)/errors.size | |
185 | |
186 def __call__(self,dataset,output_fieldnames=None,cached_output_dataset=False): | |
187 assert dataset.hasFields(["input"]) | |
188 if output_fieldnames is None: | |
189 if dataset.hasFields(["target"]): | |
190 output_fieldnames = ["output","squared_error"] | |
191 else: | |
192 output_fieldnames = ["output"] | |
193 output_fieldnames.sort() | |
194 if output_fieldnames == ["squared_error"]: | |
195 f = self.compute_errors | |
196 elif output_fieldnames == ["output"]: | |
197 f = self.compute_outputs | |
198 elif output_fieldnames == ["output","squared_error"]: | |
199 f = self.compute_outputs_and_errors | |
200 else: | |
201 raise ValueError("unknown field(s) in output_fieldnames: "+str(output_fieldnames)) | |
202 | |
203 ds=ApplyFunctionDataSet(dataset,f,output_fieldnames) | |
204 if cached_output_dataset: | |
205 return CachedDataSet(ds) | |
206 else: | |
207 return ds | |
208 | |
209 |