Mercurial > pylearn

annotate gradient_learner.py @ 22:b6b36f65664f

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Created virtual sub-classes of DataSet: {Finite{Length,Width},Sliceable}DataSet, removed .field ability from LookupList (because of setattr problems), removed fieldNames() from DataSet (but is in FiniteWidthDataSet, where it makes sense), and added hasFields() instead. Fixed problems in asarray, and tested previous functionality in _test_dataset.py, but not yet new functionality.
author bengioy@esprit.iro.umontreal.ca
date Mon, 07 Apr 2008 20:44:37 -0400
parents 266c68cb6136
children 526e192b0699
rev   line source
13
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
1
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
2 from learner import *
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
3 from tensor import *
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
4 import gradient
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
5 from compile import Function
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
6
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
7 class GradientLearner(Learner):
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
8 """
14
5ede27026e05 Working on gradient_based_learner
bengioy@bengiomac.local
parents: 13
diff changeset
9 Base class for gradient-based optimization of a training criterion
13
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
10 that can consist in two parts, an additive part over examples, and
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
11 an example-independent part (usually called the regularizer).
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
12 The user provides a Theano formula that maps the fields of a training example
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
13 and parameters to output fields (for the use function), one of which must be a cost
14
5ede27026e05 Working on gradient_based_learner
bengioy@bengiomac.local
parents: 13
diff changeset
14 that is the training criterion to be minimized. Subclasses implement
20
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
15 a training strategy that uses the Theano formula to compute gradients and
14
5ede27026e05 Working on gradient_based_learner
bengioy@bengiomac.local
parents: 13
diff changeset
16 to compute outputs in the update method.
5ede27026e05 Working on gradient_based_learner
bengioy@bengiomac.local
parents: 13
diff changeset
17 The inputs, parameters, and outputs are lists of Theano tensors,
13
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
18 while the example_wise_cost and regularization_term are Theano tensors.
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
19 The user can specify a regularization coefficient that multiplies the regularization term.
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
20 The training algorithm looks for parameters that minimize
20
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
21 regularization_coefficient * regularization_term(parameters) +
13
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
22 sum_{inputs in training_set} example_wise_cost(inputs,parameters)
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
23 i.e. the regularization_term should not depend on the inputs, only on the parameters.
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
24 The learned function can map a subset of inputs to a subset of outputs (as long as the inputs subset
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
25 includes all the inputs required in the Theano expression for the selected outputs).
20
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
26 It is assumed that all the inputs are provided in the training set (as dataset fields
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
27 with the corresponding name), but not necessarily when using the learned function.
13
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
28 """
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
29 def __init__(self, inputs, parameters, outputs, example_wise_cost, regularization_term,
20
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
30 regularization_coefficient = astensor(1.0)):
13
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
31 self.inputs = inputs
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
32 self.outputs = outputs
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
33 self.parameters = parameters
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
34 self.example_wise_cost = example_wise_cost
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
35 self.regularization_term = regularization_term
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
36 self.regularization_coefficient = regularization_coefficient
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
37 self.parameters_example_wise_gradient = gradient.grad(example_wise_cost, parameters)
20
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
38 self.parameters_regularization_gradient = gradient.grad(self.regularization_coefficient * regularization_term, parameters)
14
5ede27026e05 Working on gradient_based_learner
bengioy@bengiomac.local
parents: 13
diff changeset
39 if example_wise_cost not in outputs:
5ede27026e05 Working on gradient_based_learner
bengioy@bengiomac.local
parents: 13
diff changeset
40 outputs.append(example_wise_cost)
5ede27026e05 Working on gradient_based_learner
bengioy@bengiomac.local
parents: 13
diff changeset
41 if regularization_term not in outputs:
5ede27026e05 Working on gradient_based_learner
bengioy@bengiomac.local
parents: 13
diff changeset
42 outputs.append(regularization_term)
5ede27026e05 Working on gradient_based_learner
bengioy@bengiomac.local
parents: 13
diff changeset
43 self.example_wise_gradient_fn = Function(inputs + parameters,
5ede27026e05 Working on gradient_based_learner
bengioy@bengiomac.local
parents: 13
diff changeset
44 [self.parameters_example_wise_gradient + self.parameters_regularization_gradient])
20
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
45 self.use_functions = {frozenset([input.name for input in inputs]+[output.name for output in outputs])
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
46 : Function(inputs, outputs)}
13
633453635d51 Starting to work on gradient_based_learner.py
bengioy@bengiomac.local
parents:
diff changeset
47
20
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
48 def use(self,input_dataset,output_fields=None,copy_inputs=True):
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
49 # obtain the function that maps the desired inputs to desired outputs
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
50 input_fields = input_dataset.fieldNames()
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
51 if output_fields is None: output_fields = [output.name for output in outputs]
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
52 # handle special case of inputs that are directly copied into outputs
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
53
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
54 use_function_key = input_fields+output_fields
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
55 if not self.use_functions.has_key(use_function_key):
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
56 self.use_function[use_function_key]=Function(input_fields,output_fields)
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
57 use_function = self.use_functions[use_function_key]
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
58 # return a virtual dataset that computes the outputs on demand
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
59 return input_dataset.apply_function(use_function,input_fields,output_fields,copy_inputs,accept_minibatches=???)
266c68cb6136 Minor editions, plus adding untested ApplyFunctionDataset for GradientLearner in the works.
bengioy@bengiomac.local
parents: 14
diff changeset
60