Mercurial > pylearn
changeset 748:863e34a3d01e
New option 'corruption_pattern' to alter the way inputs are corrupted
| author | Olivier Delalleau <delallea@iro> |
|---|---|
| date | Tue, 02 Jun 2009 13:50:43 -0400 |
| parents | 6117969dd37f |
| children | 84d22b7d835a |
| files | pylearn/algorithms/sandbox/DAA_inputs_groups.py |
| diffstat | 1 files changed, 27 insertions(+), 3 deletions(-) [+] |
line wrap: on
line diff
--- a/pylearn/algorithms/sandbox/DAA_inputs_groups.py Tue Jun 02 11:31:50 2009 -0400 +++ b/pylearn/algorithms/sandbox/DAA_inputs_groups.py Tue Jun 02 13:50:43 2009 -0400 @@ -48,6 +48,7 @@ regularize = False, tie_weights = False, hid_fn = 'sigmoid_act', reconstruction_cost_function=cost.cross_entropy, interface = True, ignore_missing=None, reconstruct_missing=False, + corruption_pattern=None, **init): """ :param regularize: WRITEME @@ -67,6 +68,12 @@ auxilary ones (that should never contain missing values). In fact, in the current implementation, auxiliary inputs cannot be used when this option is True. + :param corruption_pattern: if not None, may specify a particular way to + corrupt the input with missing values. Valid choices are: + - 'by_pair': consider that features are given as pairs, and corrupt + (or not) the whole pair instead of considering them independently. + Elements in a pair are not consecutive, instead they are assumed to + be at distance (total number of features / 2) of each other. :param reconstruct_missing: if True, then the reconstruction cost on missing inputs will be backpropagated. Otherwise, it will not. :todo: Default noise level for all daa levels @@ -91,6 +98,7 @@ self.interface = interface self.ignore_missing = ignore_missing self.reconstruct_missing = reconstruct_missing + self.corruption_pattern = corruption_pattern assert hid_fn in ('sigmoid_act','tanh_act','softsign_act') self.hid_fn = eval(hid_fn) @@ -137,7 +145,8 @@ self.noisy_input = self.corrupt_input() if self.auxinput is not None: self.noisy_idx_list , self.noisy_auxinput = \ - scannoise(self.idx_list,self.auxinput,self.noise_level,self.noise_level_group) + scannoise(self.idx_list, self.auxinput,self.noise_level, + self.noise_level_group) self.noise = ScratchPad() self.clean = ScratchPad() @@ -270,7 +279,19 @@ self.validate = theano.Method(listin, [self.clean.cost, self.clean.rec]) def corrupt_input(self): - return self.random.binomial(T.shape(self.input), 1, 1 - self.noise_level) * self.input + if self.corruption_pattern is None: + mask = self.random.binomial(T.shape(self.input), 1, 1 - self.noise_level) + elif self.corruption_pattern == 'by_pair': + shape = T.shape(self.input) + scale = numpy.ones(2) + scale[1] = 2 + shape = shape / scale + mask = self.random.binomial(shape, 1, 1 - self.noise_level) + mask = T.hstack((mask, mask)) + else: + raise ValueError('Unknown value for corruption_pattern: %s' + % self.corruption_pattern) + return mask * self.input def reconstruction_costs(self, rec): if (self.input is not None) and (self.auxinput is not None): @@ -348,6 +369,7 @@ reconstruction_cost_function=cost.cross_entropy, n_out = 2, target = None, debugmethod = False, totalupdatebool=False, ignore_missing=None, reconstruct_missing=False, + corruption_pattern=None, **init): super(StackedDAAig, self).__init__() @@ -375,6 +397,7 @@ self.totalupdatebool = totalupdatebool self.ignore_missing = ignore_missing self.reconstruct_missing = reconstruct_missing + self.corruption_pattern = corruption_pattern # init for model construction inputprec = input @@ -441,7 +464,8 @@ False, self.tie_weights, self.hid_fn, self.reconstruction_cost_function,False] dict_params = dict(ignore_missing = self.ignore_missing, - reconstruct_missing = self.reconstruct_missing) + reconstruct_missing = self.reconstruct_missing, + corruption_pattern = self.corruption_pattern) print '\tLayer init= ', i+1 self.daaig[i] = DAAig(*param, **dict_params)