Mercurial > pylearn

annotate external/wrap_libsvm.py @ 517:716c04512dbe

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
init
author James Bergstra <bergstrj@iro.umontreal.ca>
date Wed, 12 Nov 2008 10:54:38 -0500
parents
children
rev   line source
517
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
1 """Run an experiment using libsvm.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
2 """
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
3 import numpy
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
4 from ..datasets import dataset_from_descr
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
5
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
6 # libsvm currently has no python installation instructions/convention.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
7 #
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
8 # This module uses a specific convention for libsvm's installation.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
9 # I base this on installing libsvm-2.88.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
10 # To install libsvm's python module, do three things:
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
11 # 1. Build libsvm (run make in both the root dir and the python subdir).
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
12 # 2. touch a '__init__.py' file in the python subdir
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
13 # 3. add a symbolic link to a PYTHONPATH location that looks like this:
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
14 # libsvm -> <your root path>/libsvm-2.88/python/
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
15 #
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
16 # That is the sort of thing that this module expects from 'import libsvm'
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
17
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
18 import libsvm
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
19
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
20 def score_01(x, y, model):
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
21 assert len(x) == len(y)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
22 size = len(x)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
23 errors = 0
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
24 for i in range(size):
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
25 prediction = model.predict(x[i])
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
26 #probability = model.predict_probability
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
27 if (y[i] != prediction):
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
28 errors = errors + 1
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
29 return float(errors)/size
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
30
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
31 #this is the dbdict experiment interface... if you happen to use dbdict
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
32 class State(object):
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
33 #TODO: parametrize to get all the kernel types, not hardcode for RBF
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
34 dataset = 'MNIST_1k'
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
35 C = 10.0
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
36 kernel = 'RBF'
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
37 # rel_gamma is related to the procedure Jerome used. He mentioned why in
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
38 # quadratic_neurons/neuropaper/draft3.pdf.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
39 rel_gamma = 1.0
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
40
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
41 def __init__(self, **kwargs):
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
42 for k, v in kwargs:
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
43 setattr(self, k, type(getattr(self, k))(v))
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
44
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
45
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
46 def dbdict_run_svm_experiment(state, channel=lambda *args, **kwargs:None):
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
47 """Parameters are described in state, and returned in state.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
48
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
49 :param state: object instance to store parameters and return values
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
50 :param channel: not used
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
51
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
52 :returns: None
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
53
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
54 This is the kind of function that dbdict-run can use.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
55
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
56 """
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
57 ((train_x, train_y), (valid_x, valid_y), (test_x, test_y)) = dataset_from_descr(state.dataset)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
58
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
59 #libsvm needs stuff in int32 on a 32bit machine
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
60 #TODO: test this on a 64bit machine
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
61 train_y = numpy.asarray(train_y, dtype='int32')
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
62 valid_y = numpy.asarray(valid_y, dtype='int32')
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
63 test_y = numpy.asarray(test_y, dtype='int32')
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
64 problem = svm.svm_problem(train_y, train_x);
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
65
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
66 gamma0 = 0.5 / numpy.sum(numpy.var(train_x, axis=0))
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
67
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
68 param = svm.svm_parameter(C=state.C,
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
69 kernel_type=getattr(svm, state.kernel),
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
70 gamma=state.rel_gamma * gamma0)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
71
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
72 model = svm.svm_model(problem, param) #this is the expensive part
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
73
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
74 state.train_01 = score_01(train_x, train_y, model)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
75 state.valid_01 = score_01(valid_x, valid_y, model)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
76 state.test_01 = score_01(test_x, test_y, model)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
77
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
78 state.n_train = len(train_y)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
79 state.n_valid = len(valid_y)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
80 state.n_test = len(test_y)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
81
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
82 def run_svm_experiment(**kwargs):
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
83 """Python-friendly interface to dbdict_run_svm_experiment
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
84
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
85 Parameters are used to construct a `State` instance, which is returned after running
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
86 `dbdict_run_svm_experiment` on it.
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
87
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
88 .. code-block:: python
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
89 results = run_svm_experiment(dataset='MNIST_1k', C=100.0, rel_gamma=0.01)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
90 print results.n_train
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
91 # 1000
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
92 print results.valid_01, results.test_01
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
93 # 0.14, 0.10 #.. or something...
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
94
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
95 """
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
96 state = State(**kwargs)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
97 state_run_svm_experiment(state)
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
98 return state
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff changeset
99