class MDNDataSet(SupervisedDataSet):
_cdataset = None
def __init__(self, inp, target, M, inp_transform=None, tgt_transform=None):
"""
Initialize an MDN dataset.
@param inp: int or np.array, input data or input dimension
@param target: int or np.array, target data or target dimension
@param M: number of Gaussian kernels #TODO: remove this from dataset!
@param inp_transform: set of input transformations that have been
applied to the dataset #TODO: remove
@param tgt_transform: #TODO: remove
"""
SupervisedDataSet.__init__(self, inp, target)
self.M = M
if np.isscalar(target):
self.c = target
else:
self.c = target.shape[1]
self.inp_transform=inp_transform
self.tgt_transform=tgt_transform
def __reduce__(self):
# not sure why this is overwritten in SequentialDataSet
# But we have to include additional arguments to __init__ in
# order to make deepcopy work
_, _, state, _, _ = super(SupervisedDataSet, self).__reduce__()
creator = self.__class__
args = self.indim, self.outdim, self.M, self.inp_transform, self.tgt_transform
return creator, args, state, iter([]), iter({})
def getTargets(self):
tgts = []
for k in range(self.getLength()):
tgts.append(self.getSample(k)[1])
return np.array(tgts)
def getFastDataset(self):
if not self._cdataset:
self._cdataset = SupervisedSimpleDataset(self.indim, self.c)
for k in range(self.getLength()):
sample = self.getSample(k)
self._cdataset.append(sample[0], sample[1])
return self._cdataset
def appendDataset(self, dataset):
"""
Append the given dataset.
"""
for inp, tgt in dataset:
self.addSample(inp, tgt)
def getFastDataset(self):
if not self._cdataset:
self._cdataset = SupervisedSimpleDataset(self.indim, self.c)
for k in range(self.getLength()):
sample = self.getSample(k)
self._cdataset.append(sample[0], sample[1])
return self._cdataset