Mercurial > pylearn
annotate pylearn/shared/layers/sigmoidal_layer.py @ 1405:f9e4d71aa353
Add L1 and L2² costs to sigmoidal layer
| author | Pascal Lamblin <lamblinp@iro.umontreal.ca> |
|---|---|
| date | Wed, 26 Jan 2011 16:55:44 -0500 |
| parents | 912be602c3ac |
| children |
| rev | line source |
|---|---|
| 834 | 1 """ Provide the "normal" sigmoidal layers for making multi-layer perceptrons / neural nets |
| 2 | |
| 3 """ | |
| 4 import logging | |
| 5 import numpy | |
| 6 | |
| 7 import theano | |
| 8 from theano import tensor | |
|
893
912be602c3ac
fix import for the move of shared var outside of the sandbox.
Frederic Bastien <nouiz@nouiz.org>
parents:
891
diff
changeset
|
9 from theano.compile import shared, pfunc |
|
891
fc9779dcd710
some backport to python 2.4
Frederic Bastien <nouiz@nouiz.org>
parents:
834
diff
changeset
|
10 from pylearn.shared.layers.util import update_locals, add_logging |
|
fc9779dcd710
some backport to python 2.4
Frederic Bastien <nouiz@nouiz.org>
parents:
834
diff
changeset
|
11 from pylearn.shared.layers.squash import squash |
| 834 | 12 |
| 13 | |
| 14 class SigmoidalLayer(object): | |
| 15 def __init__(self, input, w, b, squash_fn, params): | |
| 16 """ | |
| 17 :param input: a symbolic tensor of shape (n_examples, n_in) | |
| 18 :param w: a symbolic weight matrix of shape (n_in, n_out) | |
| 19 :param b: symbolic bias terms of shape (n_out,) | |
| 20 :param squash: an squashing function | |
| 21 """ | |
| 22 output = squash_fn(tensor.dot(input, w) + b) | |
|
1405
f9e4d71aa353
Add L1 and L2² costs to sigmoidal layer
Pascal Lamblin <lamblinp@iro.umontreal.ca>
parents:
893
diff
changeset
|
23 l1 = abs(w).sum() |
|
f9e4d71aa353
Add L1 and L2² costs to sigmoidal layer
Pascal Lamblin <lamblinp@iro.umontreal.ca>
parents:
893
diff
changeset
|
24 l2_sqr = (w**2).sum() |
| 834 | 25 update_locals(self, locals()) |
| 26 | |
| 27 @classmethod | |
| 28 def new(cls, rng, input, n_in, n_out, squash_fn=tensor.tanh, dtype=None): | |
| 29 """Allocate a SigmoidLayer with weights to transform inputs with n_in dimensions, | |
| 30 to outputs of n_out dimensions. | |
| 31 | |
| 32 Weights are initialized randomly using rng. | |
| 33 | |
| 34 :param squash_fn: an op constructor function, or a string that has been registed as a | |
| 35 `squashing_function`. | |
| 36 | |
| 37 :param dtype: the numerical type to use for the parameters (i.e. 'float32', 'float64') | |
| 38 | |
| 39 """ | |
| 40 if dtype is None: | |
| 41 dtype = input.dtype | |
| 42 cls._debug('allocating weights and biases', n_in, n_out, dtype) | |
| 43 w = shared( | |
| 44 numpy.asarray( | |
| 45 rng.uniform(low=-2/numpy.sqrt(n_in), high=2/numpy.sqrt(n_in), | |
| 46 size=(n_in, n_out)), dtype=dtype)) | |
| 47 b = shared(numpy.asarray(numpy.zeros(n_out), dtype=dtype)) | |
| 48 return cls(input, w, b, squash(squash_fn), [w,b]) | |
| 49 | |
| 50 add_logging(SigmoidalLayer) |