Mercurial > pylearn
annotate mlp_factory_approach.py @ 242:ef70a665aaaf
Hmm... that was committed by Fred I think, I got lost by Mercurial I think
author | delallea@opale.iro.umontreal.ca |
---|---|
date | Fri, 30 May 2008 10:19:16 -0400 |
parents | c047238e5b3f |
children | 3156a9976183 |
rev | line source |
---|---|
225
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
1 """ |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
2 |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
3 |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
4 |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
5 This file is deprecated. I'm continuing development in hpu/models.py. |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
6 |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
7 Get that project like this: hg clone ssh://user@lgcm/../bergstrj/hpu |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
8 |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
9 |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
10 |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
11 |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
12 |
8bc16220b29a
deprecation note for mlp
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
215
diff
changeset
|
13 """ |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
14 import copy, sys |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
15 import numpy |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
16 |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
17 import theano |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
18 from theano import tensor as t |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
19 |
218
df3fae88ab46
small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
215
diff
changeset
|
20 from pylearn import dataset, nnet_ops, stopper |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
21 |
211
bd728c83faff
in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
208
diff
changeset
|
22 |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
23 def _randshape(*shape): |
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
24 return (numpy.random.rand(*shape) -0.5) * 0.001 |
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
25 |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
26 def _cache(d, key, valfn): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
27 #valfn() is only evaluated if key isn't in dictionary d |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
28 if key not in d: |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
29 d[key] = valfn() |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
30 return d[key] |
190
aa7a3ecbcc90
progress toward early stopping
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
189
diff
changeset
|
31 |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
32 class _Model(object): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
33 def __init__(self, algo, params): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
34 self.algo = algo |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
35 self.params = params |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
36 v = algo.v |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
37 self.update_fn = algo._fn([v.input, v.target] + v.params, [v.nll] + v.new_params) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
38 self._fn_cache = {} |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
39 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
40 def __copy__(self): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
41 return _Model(self.algo, [copy.copy(p) for p in params]) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
42 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
43 def update(self, input_target): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
44 """Update this model from more training data.""" |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
45 params = self.params |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
46 #TODO: why should we have to unpack target like this? |
218
df3fae88ab46
small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
215
diff
changeset
|
47 # tbm : creates problem... |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
48 for input, target in input_target: |
232 | 49 rval= self.update_fn(input, target, *params) |
212
9b57ea8c767f
previous commit was supposed to concern only one file, dataset.py, try to undo my other changes with this commit (nothing was broken though, just useless debugging prints)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
211
diff
changeset
|
50 #print rval[0] |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
51 |
218
df3fae88ab46
small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
215
diff
changeset
|
52 def __call__(self, testset, fieldnames=['output_class'],input='input',target='target'): |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
53 """Apply this model (as a function) to new data""" |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
54 #TODO: cache fn between calls |
218
df3fae88ab46
small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
215
diff
changeset
|
55 assert input == testset.fieldNames()[0] # why first one??? |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
56 assert len(testset.fieldNames()) <= 2 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
57 v = self.algo.v |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
58 outputs = [getattr(v, name) for name in fieldnames] |
218
df3fae88ab46
small debugging
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
215
diff
changeset
|
59 inputs = [v.input] + ([v.target] if target in testset else []) |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
60 inputs.extend(v.params) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
61 theano_fn = _cache(self._fn_cache, (tuple(inputs), tuple(outputs)), |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
62 lambda: self.algo._fn(inputs, outputs)) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
63 lambda_fn = lambda *args: theano_fn(*(list(args) + self.params)) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
64 return dataset.ApplyFunctionDataSet(testset, lambda_fn, fieldnames) |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
65 |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
66 class AutonameVars(object): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
67 def __init__(self, dct): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
68 for key, val in dct.items(): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
69 if type(key) is str and hasattr(val, 'name'): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
70 val.name = key |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
71 self.__dict__.update(dct) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
72 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
73 class MultiLayerPerceptron(object): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
74 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
75 def __init__(self, ninputs, nhid, nclass, lr, |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
76 l2coef=0.0, |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
77 linker='c&py', |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
78 hidden_layer=None, |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
79 early_stopper=None, |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
80 validation_portion=0.2, |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
81 V_extern=None): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
82 class V_intern(AutonameVars): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
83 def __init__(v_self, lr, l2coef, **kwargs): |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
84 lr = t.constant(lr) |
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
85 l2coef = t.constant(l2coef) |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
86 input = t.matrix() # n_examples x n_inputs |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
87 target = t.ivector() # len: n_examples |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
88 W2, b2 = t.matrix(), t.vector() |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
89 |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
90 if hidden_layer: |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
91 hid, hid_params, hid_ivals, hid_regularization = hidden_layer(input) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
92 else: |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
93 W1, b1 = t.matrix(), t.vector() |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
94 hid = t.tanh(b1 + t.dot(input, W1)) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
95 hid_params = [W1, b1] |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
96 hid_regularization = l2coef * t.sum(W1*W1) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
97 hid_ivals = lambda : [_randshape(ninputs, nhid), _randshape(nhid)] |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
98 |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
99 params = [W2, b2] + hid_params |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
100 activations = b2 + t.dot(hid, W2) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
101 nll, predictions = nnet_ops.crossentropy_softmax_1hot(activations, target) |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
102 regularization = l2coef * t.sum(W2*W2) + hid_regularization |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
103 output_class = t.argmax(activations,1) |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
104 loss_01 = t.neq(output_class, target) |
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
105 g_params = t.grad(nll + regularization, params) |
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
106 new_params = [t.sub_inplace(p, lr * gp) for p,gp in zip(params, g_params)] |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
107 self.__dict__.update(locals()); del self.self |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
108 AutonameVars.__init__(v_self, locals()) |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
109 self.nhid = nhid |
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
110 self.nclass = nclass |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
111 self.v = V_intern(**locals()) if V_extern is None else V_extern(**locals()) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
112 self.linker = linker |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
113 self.early_stopper = early_stopper if early_stopper is not None else lambda: stopper.NStages(10,1) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
114 self.validation_portion = validation_portion |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
115 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
116 def _fn(self, inputs, outputs): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
117 # Caching here would hamper multi-threaded apps |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
118 # prefer caching in _Model.__call__ |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
119 return theano.function(inputs, outputs, unpack_single=False, linker=self.linker) |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
120 |
211
bd728c83faff
in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
208
diff
changeset
|
121 def __call__(self, trainset=None, iparams=None, input='input', target='target'): |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
122 """Allocate and optionally train a model""" |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
123 if iparams is None: |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
124 iparams = [_randshape(self.nhid, self.nclass), _randshape(self.nclass)]\ |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
125 + self.v.hid_ivals() |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
126 rval = _Model(self, iparams) |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
127 if trainset: |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
128 if len(trainset) == sys.maxint: |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
129 raise NotImplementedError('Learning from infinite streams is not supported') |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
130 nval = int(self.validation_portion * len(trainset)) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
131 nmin = len(trainset) - nval |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
132 assert nmin >= 0 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
133 minset = trainset[:nmin] #real training set for minimizing loss |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
134 valset = trainset[nmin:] #validation set for early stopping |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
135 best = rval |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
136 for stp in self.early_stopper(): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
137 rval.update( |
211
bd728c83faff
in __get__, problem if the i.stop was None, i being the slice, added one line replacing None by the len(self)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
208
diff
changeset
|
138 minset.minibatches([input, target], minibatch_size=min(32, |
232 | 139 len(minset)))) |
212
9b57ea8c767f
previous commit was supposed to concern only one file, dataset.py, try to undo my other changes with this commit (nothing was broken though, just useless debugging prints)
Thierry Bertin-Mahieux <bertinmt@iro.umontreal.ca>
parents:
211
diff
changeset
|
140 #print 'mlp.__call__(), we did an update' |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
141 if stp.set_score: |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
142 stp.score = rval(valset, ['loss_01']) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
143 if (stp.score < stp.best_score): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
144 best = copy.copy(rval) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
145 rval = best |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
146 return rval |
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
147 |
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
148 |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
149 import unittest |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
150 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
151 class TestMLP(unittest.TestCase): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
152 def test0(self): |
191
e816821c1e50
added early stopping to mlp.__call__
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
190
diff
changeset
|
153 |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
154 training_set1 = dataset.ArrayDataSet(numpy.array([[0, 0, 0], |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
155 [0, 1, 1], |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
156 [1, 0, 1], |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
157 [1, 1, 1]]), |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
158 {'input':slice(2),'target':2}) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
159 training_set2 = dataset.ArrayDataSet(numpy.array([[0, 0, 0], |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
160 [0, 1, 1], |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
161 [1, 0, 0], |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
162 [1, 1, 1]]), |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
163 {'input':slice(2),'target':2}) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
164 test_data = dataset.ArrayDataSet(numpy.array([[0, 0, 0], |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
165 [0, 1, 1], |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
166 [1, 0, 0], |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
167 [1, 1, 1]]), |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
168 {'input':slice(2)}) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
169 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
170 learn_algo = MultiLayerPerceptron(2, 10, 2, .1 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
171 , linker='c&py' |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
172 , early_stopper = lambda:stopper.NStages(100,1)) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
173 |
232 | 174 model1 = learn_algo(training_set1) |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
175 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
176 model2 = learn_algo(training_set2) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
177 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
178 n_match = 0 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
179 for o1, o2 in zip(model1(test_data), model2(test_data)): |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
180 #print o1 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
181 #print o2 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
182 n_match += (o1 == o2) |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
183 |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
184 assert n_match == (numpy.sum(training_set1.fields()['target'] == |
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
185 training_set2.fields()['target'])) |
191
e816821c1e50
added early stopping to mlp.__call__
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
190
diff
changeset
|
186 |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
187 if __name__ == '__main__': |
208
bf320808919f
back to James' version
Yoshua Bengio <bengioy@iro.umontreal.ca>
parents:
207
diff
changeset
|
188 unittest.main() |
187
ebbb0e749565
added mlp_factory_approach
James Bergstra <bergstrj@iro.umontreal.ca>
parents:
diff
changeset
|
189 |