Mercurial > pylearn
changeset 563:16f91ca016b1
* added NStages as a stopper (moved from hpu/conv)
* added a argmax_standalone output to logistic_regression which is independent
of the targets, which was needed to compute an output independently of the
target
* fixed some import discrepancies between pylearn and pylearn_refactor (mostly
for datasets)
* added testDataset which generates sequential or random data for a given shape
| author | desjagui@atchoum.iro.umontreal.ca |
|---|---|
| date | Wed, 03 Dec 2008 17:21:05 -0500 |
| parents | 96221aa02fcb |
| children | e878003c3009 |
| files | pylearn/algorithms/logistic_regression.py pylearn/algorithms/stopper.py pylearn/datasets/MNIST.py pylearn/datasets/shapeset1.py pylearn/datasets/smallNorb.py pylearn/datasets/testDataset.py |
| diffstat | 6 files changed, 60 insertions(+), 2 deletions(-) [+] |
line wrap: on
line diff
--- a/pylearn/algorithms/logistic_regression.py Mon Dec 01 16:16:21 2008 -0500 +++ b/pylearn/algorithms/logistic_regression.py Wed Dec 03 17:21:05 2008 -0500 @@ -40,11 +40,15 @@ #here we actually build the model self.linear_output = T.dot(self.input, self.w) + self.b if 0: + # TODO: pending support for target being a sparse matrix self.softmax = nnet.softmax(self.linear_output) self._max_pr, self.argmax = T.max_and_argmax(self.linear_output) self._xent = self.target * T.log(self.softmax) else: + # TODO: when above is fixed, remove this hack (need an argmax + # which is independent of targets) + self.argmax_standalone = T.argmax(self.linear_output); (self._xent, self.softmax, self._max_pr, self.argmax) =\ nnet.crossentropy_softmax_max_and_argmax_1hot( self.linear_output, self.target)
--- a/pylearn/algorithms/stopper.py Mon Dec 01 16:16:21 2008 -0500 +++ b/pylearn/algorithms/stopper.py Wed Dec 03 17:21:05 2008 -0500 @@ -122,6 +122,16 @@ raise StopIteration +class NStages(ICML08Stopper): + """Run for a fixed number of steps, checking validation set every so + often.""" + def __init__(self, hard_limit, v_int): + ICML08Stopper.__init__(self, hard_limit, v_int, 1.0, 1.0, hard_limit) + + #TODO: could optimize next() function. Most of what's in ICML08Stopper.next() + #is not necessary + + @stopper_factory('icml08') def icml08_stopper(i_wait, v_int, min_improvement, patience, hard_limit): return ICML08Stopper(i_wait, v_int, min_improvement, patience, hard_limit)
--- a/pylearn/datasets/MNIST.py Mon Dec 01 16:16:21 2008 -0500 +++ b/pylearn/datasets/MNIST.py Wed Dec 03 17:21:05 2008 -0500 @@ -46,6 +46,7 @@ y=all_targ[ntrain+nvalid:ntrain+nvalid+ntest]) rval.n_classes = 10 + rval.img_shape = (28,28) return rval
--- a/pylearn/datasets/shapeset1.py Mon Dec 01 16:16:21 2008 -0500 +++ b/pylearn/datasets/shapeset1.py Wed Dec 03 17:21:05 2008 -0500 @@ -7,7 +7,7 @@ import os import numpy -from ..amat import AMat +from ..io.amat import AMat from .config import data_root def _head(path, n):
--- a/pylearn/datasets/smallNorb.py Mon Dec 01 16:16:21 2008 -0500 +++ b/pylearn/datasets/smallNorb.py Wed Dec 03 17:21:05 2008 -0500 @@ -1,6 +1,6 @@ import os import numpy -from ..filetensor import read +from ..io.filetensor import read from .config import data_root #Path = '/u/bergstrj/pub/data/smallnorb'
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/pylearn/datasets/testDataset.py Wed Dec 03 17:21:05 2008 -0500 @@ -0,0 +1,43 @@ +""" +Various routines to load/access MNIST data. +""" +from __future__ import absolute_import + +import os +import numpy + +from ..io.amat import AMat +from .config import data_root +from .dataset import dataset_factory, Dataset + +VALSEQ, VALRAND = range(2) + +@dataset_factory('DEBUG') +def mnist_factory(variant='', ntrain=10, nvalid=10, ntest=10, \ + nclass=2, ndim=1, dshape=None, valtype=VALSEQ): + + temp = [] + [temp.append(5) for i in range(ndim)] + dshape = temp if dshape is None else dshape + + rval = Dataset() + rval.n_classes = nclass + rval.img_shape = dshape + + dsize = numpy.prod(dshape); + + print ntrain, nvalid, ntest, nclass, dshape, valtype + + ntot = ntrain + nvalid + ntest + xdata = numpy.arange(ntot*numpy.prod(dshape)).reshape((ntot,dsize)) \ + if valtype is VALSEQ else \ + numpy.random.random((ntot,dsize)); + ydata = numpy.round(numpy.random.random(ntot)); + + rval.train = Dataset.Obj(x=xdata[0:ntrain],y=ydata[0:ntrain]) + rval.valid = Dataset.Obj(x=xdata[ntrain:ntrain+nvalid],\ + y=ydata[ntrain:ntrain+nvalid]) + rval.test = Dataset.Obj(x=xdata[ntrain+nvalid:ntrain+nvalid+ntest], + y=ydata[ntrain+nvalid:ntrain+nvalid+ntest]) + + return rval