#76 Add Sliding Window to DAGMM

[WGierke]

This addresses #76.
Would be nice to have the name of the dataset in the title :)
[roc_syn extreme outliers pol 0 1 -2-14-2018-06-13-164021]
[roc_syn extreme outliers pol 0 3 -2-14-2018-06-13-164021]
[roc_syn extreme outliers pol 0 5 -2-14-2018-06-13-164021]
[roc_syn extreme outliers pol 1 -2-14-2018-06-13-164021]
[roc_synthetic combined outliers 4-dimensional-2-14-2018-06-13-164020]
[roc_synthetic combined outliers-2-14-2018-06-13-164020]
[roc_synthetic extreme outliers-2-14-2018-06-13-164019]
[roc_synthetic shift outliers-2-14-2018-06-13-164019]
[roc_synthetic trend outliers-2-14-2018-06-13-164020]
[roc_synthetic variance outliers-2-14-2018-06-13-164019]
[roc_syn variance outliers mis 0 1 -2-14-2018-06-13-164020]
[roc_syn variance outliers mis 0 3 -2-14-2018-06-13-164020]
[roc_syn variance outliers mis 0 5 -2-14-2018-06-13-164020]
[roc_syn variance outliers mis 0 8 -2-14-2018-06-13-164021]

[danthe96]

You need to take the mean of sample energy here as well.
I implemented it like this in the other branch:

train_length = len(data_loader)*self.batch_size + self.sequence_length - 1
train_energy = np.full((self.sequence_length, train_length), np.nan)
for i1, ts_batch in enumerate(data_loader):
    _, _, z, _ = self.dagmm(to_var(ts_batch).float())
    sample_energies, _ = self.dagmm.compute_energy(z, phi=train_phi, mu=train_mu, cov=train_cov,
                                                                                          size_average=False)
    for i2, sample_energy in enumerate(sample_energies):
        index = i1 * self.batch_size + i2
        window_elements = list(range(index, index + self.sequence_length, 1))
        train_energy[index % self.sequence_length, window_elements] = sample_energy.data.cpu().numpy()
self.train_energy = np.nanmean(train_energy, axis=0)

One issue here is that we have batches larger than 1, so we can't do it as we do in predict. In addition we're dropping the last batch, so the energy array isn't quite equal in length to data.

[danthe96]

Is taken care of by #83.