def gen_phone(mdl):
X = theano.tensor.dmatrix('X')
P = theano.tensor.dmatrix('P')
y = mdl.fprop([X,P])
predict = theano.function([X, P], y)
resolution = 1600
step = 64
b = 1.019
n_channels = 64
D_multi = np.r_[tuple(gammatone_matrix(b, fc, resolution, step) for
fc in erb_space(150, 8000, n_channels))]
phones = np.load('test_phones_1600.npy')
X = np.asmatrix((len(phones),np.zeros(1536)))
phone_code = np.asmatrix((len(phones),np.zeros(3*62)))
for pi, p in enumerate(phones):
phone_code[pi,[p[0], p[1]+62, p[2]+2*62]] = 1 # one-hot encoding
out = np.zeros(1600 + 200*(len(phones)-1))
step = 200
for k in range(1,len(phones)):
idx = range(k*step, k*step+1600)
X[k] = predict(X[k-1], phone_code[k])
out[idx] += np.dot(X[k], D_multi)
out_scaled = np.asarray(out/max(abs(out)), dtype='float32')
return out_scaled
def __init__(self, which_set, frame_length, overlap=0,
frames_per_example=1, start=0, stop=None,
audio_only=False, n_prev_phones=0, n_next_phones=0,
samples_to_predict=1, filter_fn=None,
rng=_default_seed, b=1.019, step=8, n_channels=50):
"""
Parameters
----------
which_set : str
Either "train", "valid" or "test"
frame_length : int
Number of acoustic samples contained in a frame
overlap : int, optional
Number of overlapping acoustic samples for two consecutive frames.
Defaults to 0, meaning frames don't overlap.
frames_per_example : int, optional
Number of frames in a training example. Defaults to 1.
start : int, optional
Starting index of the sequences to use. Defaults to 0.
stop : int, optional
Ending index of the sequences to use. Defaults to `None`, meaning
sequences are selected all the way to the end of the array.
audio_only : bool, optional
Whether to load only the raw audio and no auxiliary information.
Defaults to `False`.
rng : object, optional
A random number generator used for picking random indices into the
design matrix when choosing minibatches.
"""
self.frame_length = frame_length
self.overlap = overlap
self.frames_per_example = frames_per_example
self.offset = self.frame_length - self.overlap
self.audio_only = audio_only
self.n_prev_phones = n_prev_phones
self.n_next_phones = n_next_phones
self.samples_to_predict = samples_to_predict
self.b = b
self.step = step
self.n_channels = n_channels
# Initializing the dictionary
self.D = numpy.r_[tuple(gammatone_matrix(self.b, fc,
self.frame_length,
self.step) for fc in
erb_space(150, 8000,
self.n_channels))]
#self.coder = SparseCoder(dictionary=self.D, transform_n_nonzero_coefs=None, transform_alpha=1, transform_algorithm='omp')
# RNG initialization
if hasattr(rng, 'random_integers'):
self.rng = rng
else:
self.rng = numpy.random.RandomState(rng)
# Load data from disk
self._load_data(which_set)
# Standardize data
for i, sequence in enumerate(self.raw_wav):
self.raw_wav[i] = self.scaler.transform(sequence)
if filter_fn is not None:
filter_fn = eval(filter_fn)
indexes = filter_fn(self.speaker_info_list[self.speaker_id])
self.raw_wav = self.raw_wav[indexes]
if not self.audio_only:
self.phones = self.phones[indexes]
self.phonemes = self.phonemes[indexes]
self.words = self.words[indexes]
# Slice data
if stop is not None:
self.raw_wav = self.raw_wav[start:stop]
if not self.audio_only:
self.phones = self.phones[start:stop]
self.phonemes = self.phonemes[start:stop]
self.words = self.words[start:stop]
else:
self.raw_wav = self.raw_wav[start:]
if not self.audio_only:
self.phones = self.phones[start:]
self.phonemes = self.phonemes[start:]
self.words = self.words[start:]
examples_per_sequence = [0]
for sequence_id, samples_sequence in enumerate(self.raw_wav):
if not self.audio_only:
# Phones segmentation
phones_sequence = self.phones[sequence_id]
phones_segmented_sequence = segment_axis(phones_sequence,frame_length,overlap)
phones_mode = numpy.concatenate([scipy.stats.mode(phones_segmented_sequence[k-self.n_prev_phones:k+self.n_next_phones+1],axis=1)[0].T for k in range(self.n_prev_phones,len(phones_segmented_sequence)-self.n_next_phones)])
self.phones[sequence_id] = numpy.asarray(phones_mode, dtype=numpy.int16)
# Phonemes segmentation
phonemes_sequence = self.phonemes[sequence_id]
phonemes_segmented_sequence = segment_axis(phonemes_sequence,
frame_length,
overlap)
if self.n_next_phones == 0:
self.phonemes[sequence_id] = phonemes_segmented_sequence[self.n_prev_phones:]
else:
self.phonemes[sequence_id] = phonemes_segmented_sequence[self.n_prev_phones:-self.n_next_phones]
# Words segmentation
words_sequence = self.words[sequence_id]
words_segmented_sequence = segment_axis(words_sequence,
frame_length,
overlap)
if self.n_next_phones == 0:
self.words[sequence_id] = words_segmented_sequence[self.n_prev_phones:]
else:
self.words[sequence_id] = words_segmented_sequence[self.n_prev_phones:-self.n_next_phones]
if self.n_next_phones == 0:
self.raw_wav[sequence_id] = self.raw_wav[sequence_id][self.n_prev_phones:]
else:
self.raw_wav[sequence_id] = self.raw_wav[sequence_id][self.n_prev_phones:-self.n_next_phones]
# TODO: change me
# Generate features/targets/phones/phonemes/words map
num_frames = self.raw_wav[sequence_id].shape[0]
num_examples = num_frames - self.frames_per_example
examples_per_sequence.append(num_examples)
self.cumulative_example_indexes = numpy.cumsum(examples_per_sequence)
self.samples_sequences = self.raw_wav
#numpy.save('/home/jfsantos/data/%s_sparse_frames.npy' % which_set, self.samples_sequences)
if not self.audio_only:
self.phones_sequences = self.phones
self.phonemes_sequences = self.phonemes
self.words_sequences = self.words
self.num_examples = self.cumulative_example_indexes[-1]
# DataSpecs
features_space = VectorSpace(
dim=self.D.shape[0] * self.frames_per_example
)
features_source = 'features'
def features_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(self.samples_sequences[sequence_index][example_index:example_index + self.frames_per_example].todense())
return rval
targets_space = VectorSpace(dim=self.D.shape[0])
targets_source = 'targets'
def targets_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(self.samples_sequences[sequence_index][example_index + self.frames_per_example].todense())
return rval
space_components = [features_space, targets_space]
source_components = [features_source, targets_source]
map_fn_components = [features_map_fn, targets_map_fn]
batch_components = [None, None]
if not self.audio_only:
num_phones = numpy.max([numpy.max(sequence) for sequence
in self.phones]) + 1
phones_space = IndexSpace(max_labels=num_phones, dim=1+self.n_prev_phones+self.n_next_phones,
dtype=str(self.phones_sequences[0].dtype))
phones_source = 'phones'
def phones_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(self.phones_sequences[sequence_index][example_index].ravel())
return rval
num_phonemes = numpy.max([numpy.max(sequence) for sequence
in self.phonemes]) + 1
phonemes_space = IndexSpace(max_labels=num_phonemes, dim=1,
dtype=str(self.phonemes_sequences[0].dtype))
phonemes_source = 'phonemes'
def phonemes_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(self.phonemes_sequences[sequence_index][example_index
+ self.frames_per_example].ravel())
return rval
num_words = numpy.max([numpy.max(sequence) for sequence
in self.words]) + 1
words_space = IndexSpace(max_labels=num_words, dim=1,
dtype=str(self.words_sequences[0].dtype))
words_source = 'words'
def words_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(self.words_sequences[sequence_index][example_index
+ self.frames_per_example].ravel())
return rval
space_components.extend([phones_space, phonemes_space,
words_space])
source_components.extend([phones_source, phonemes_source,
words_source])
map_fn_components.extend([phones_map_fn, phonemes_map_fn,
words_map_fn])
batch_components.extend([None, None, None])
space = CompositeSpace(space_components)
source = tuple(source_components)
self.data_specs = (space, source)
self.map_functions = tuple(map_fn_components)
self.batch_buffers = batch_components
# Defaults for iterators
self._iter_mode = resolve_iterator_class('shuffled_sequential')
self._iter_data_specs = (CompositeSpace((features_space,
targets_space)),
(features_source, targets_source))
def __init__(
self,
which_set,
frame_length,
overlap=0.5,
frames_per_example=1,
start=0,
stop=None,
audio_only=True,
n_prev_phones=0,
n_next_phones=0,
samples_to_predict=1,
filter_fn=None,
rng=_default_seed,
b=1.019,
step=64,
n_channels=64,
):
"""
Parameters
----------
which_set : str
Either "train", "valid" or "test"
frame_length : int
Number of acoustic samples contained in a frame
overlap : int, optional
Number of overlapping acoustic samples for two consecutive frames.
Defaults to 0, meaning frames don't overlap.
frames_per_example : int, optional
Number of frames in a training example. Defaults to 1.
start : int, optional
Starting index of the sequences to use. Defaults to 0.
stop : int, optional
Ending index of the sequences to use. Defaults to `None`, meaning
sequences are selected all the way to the end of the array.
audio_only : bool, optional
Whether to load only the raw audio and no auxiliary information.
Defaults to `False`.
rng : object, optional
A random number generator used for picking random indices into the
design matrix when choosing minibatches.
"""
self.frame_length = frame_length
if overlap < 1.0:
self.overlap = overlap * frame_length
else:
self.overlap = overlap
self.frames_per_example = frames_per_example
self.offset = self.frame_length - self.overlap
self.audio_only = audio_only
self.n_prev_phones = n_prev_phones
self.n_next_phones = n_next_phones
self.samples_to_predict = samples_to_predict
self.b = b
self.step = step
self.n_channels = n_channels
print "Frame length %d, overlap %d" % (self.frame_length, self.overlap)
# Initializing the dictionary
self.D = numpy.r_[
tuple(
gammatone_matrix(self.b, fc, self.frame_length, self.step)
for fc in erb_space(150, 8000, self.n_channels)
)
]
print "Using dictionary with shape", self.D.shape
self.coder = SparseCoder(
dictionary=self.D, transform_n_nonzero_coefs=None, transform_alpha=None, transform_algorithm="omp"
)
# RNG initialization
if hasattr(rng, "random_integers"):
self.rng = rng
else:
self.rng = numpy.random.RandomState(rng)
# Load data from disk
self._load_data(which_set)
examples_per_sequence = [0]
for sequence_id, samples_sequence in enumerate(self.raw_wav):
print "Sentence %d/%d" % (sequence_id, len(self.raw_wav))
X = segment_axis(samples_sequence, frame_length, overlap, end="pad")
X = numpy.hanning(self.frame_length) * X
self.raw_wav[sequence_id] = scipy.sparse.csr_matrix(self.coder.transform(X))
# TODO: change me
# Generate features/targets/phones/phonemes/words map
num_frames = self.raw_wav[sequence_id].shape[0]
num_examples = num_frames - self.frames_per_example
examples_per_sequence.append(num_examples)
self.cumulative_example_indexes = numpy.cumsum(examples_per_sequence)
self.samples_sequences = self.raw_wav
numpy.save("%s_sparse_frames.npy" % which_set, self.samples_sequences)
self.num_examples = self.cumulative_example_indexes[-1]
# DataSpecs
features_space = VectorSpace(dim=self.D.shape[0] * self.frames_per_example)
features_source = "features"
def features_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(
self.samples_sequences[sequence_index][
example_index : example_index + self.frames_per_example
].todense()
)
return rval
targets_space = VectorSpace(dim=self.D.shape[0])
targets_source = "targets"
def targets_map_fn(indexes):
rval = []
for sequence_index, example_index in self._fetch_index(indexes):
rval.append(self.samples_sequences[sequence_index][example_index + self.frames_per_example].todense())
return rval
space_components = [features_space, targets_space]
source_components = [features_source, targets_source]
map_fn_components = [features_map_fn, targets_map_fn]
batch_components = [None, None]
space = CompositeSpace(space_components)
source = tuple(source_components)
self.data_specs = (space, source)
self.map_functions = tuple(map_fn_components)
self.batch_buffers = batch_components
# Defaults for iterators
self._iter_mode = resolve_iterator_class("shuffled_sequential")
self._iter_data_specs = (CompositeSpace((features_space, targets_space)), (features_source, targets_source))