def train_dbn(X):
dbn_network = UnsupervisedDBN(
hidden_layers_structure=[337, 260, 170, 85, 45],
activation_function='relu',
learning_rate_rbm=5 * 1e-3,
n_epochs_rbm=200)
dbn_network.fit(X)
dbn_network.save(DBN_NETWORK_FILENAME)
class DBNFeatureExtractor(object):
def __init__(self, load=False):
if load:
self.bands = pickle.load(open("bands.pickle", "rb"))
self.dbn = UnsupervisedDBN.load("dbn.pickle")
self.prepca = pickle.load(open("prepca.pickle", "rb"))
self.pca = pickle.load(open("pca.pickle", "rb"))
else:
self.dbn = UnsupervisedDBN(hidden_layers_structure=[50, 50, 50],
batch_size=1024,
learning_rate_rbm=0.001,
n_epochs_rbm=5,
contrastive_divergence_iter=2,
activation_function='sigmoid')
self.bands = [
{"start": 0, "end": 150, "pca_components":30, "energy_factor": 5},
{"start": 120, "end":300, "pca_components":70, "energy_factor": 1},
{"start": 250, "end":513, "pca_components":60, "energy_factor": 0.5}
]
self.prepca = [PCA(n_components=band["pca_components"]) for band in self.bands]
self.pca = PCA(n_components=16)
def save(self):
self.dbn.save("dbn.pickle")
pickle.dump(self.bands, open("bands.pickle", "wb"))
pickle.dump(self.prepca, open("prepca.pickle", "wb"))
pickle.dump(self.pca, open("pca.pickle", "wb"))
def normalize_energy(self, banded_frames):
bands_energy = [band["energy_factor"] * np.sum(np.square(banded_frames[bandIndex]))/(band["end"] - band["start"])
for bandIndex, band in enumerate(self.bands)]
return [bf/math.sqrt(be) for bf,be in zip(banded_frames, bands_energy)]
def train(self, frames):
banded_frames = [self.prepca[bandIndex].fit_transform(frames[:,band["start"]:band["end"]])
for bandIndex, band in enumerate(self.bands)]
banded_frames = self.normalize_energy(banded_frames)
frames = np.concatenate(banded_frames, axis=1)
print(frames.shape)
self.dbn.fit(frames)
raw_features = self.dbn.transform(frames)
self.pca.fit(raw_features)
def extractFeatures(self, frames):
banded_frames = [self.prepca[bandIndex].transform(frames[:,band["start"]:band["end"]])
for bandIndex, band in enumerate(self.bands)]
banded_frames = self.normalize_energy(banded_frames)
frames = np.concatenate(banded_frames, axis=1)
raw_features = self.dbn.transform(frames)
return self.pca.transform(raw_features)