def _transform_multi_channel_data(self, X, y):
# Data partitioning
parted_X, parted_y = self._partition_data(
X=X, y=y, partition_size=self.window_size)
transposed_X = np.transpose(parted_X, [0, 2, 1])
converted_X = np.reshape(transposed_X,
(transposed_X.shape[0], transposed_X.shape[1],
1, transposed_X.shape[2]))
# Create view converter
view_converter = DefaultViewConverter(shape=self.sample_shape,
axes=('b', 0, 1, 'c'))
# Convert data into a design matrix
view_converted_X = view_converter.topo_view_to_design_mat(converted_X)
assert np.all(converted_X == view_converter.design_mat_to_topo_view(
view_converted_X))
# Format the target into proper format
sum_y = np.sum(parted_y, axis=1)
sum_y[sum_y > 0] = 1
one_hot_formatter = OneHotFormatter(max_labels=self.n_classes)
hot_y = one_hot_formatter.format(sum_y)
return view_converted_X, hot_y, view_converter
class ConditionalGeneratorTestCase(unittest.TestCase):
def setUp(self):
self.noise_dim = 10
self.num_labels = 10
self.condition_dtype = 'uint8'
self.condition_space = VectorSpace(dim=self.num_labels,
dtype=self.condition_dtype)
self.condition_formatter = OneHotFormatter(self.num_labels,
dtype=self.condition_dtype)
self.condition_distribution = OneHotDistribution(self.condition_space)
# TODO this nvis stuff is dirty. The ConditionalGenerator should handle it
self.mlp_nvis = self.noise_dim + self.num_labels
self.mlp_nout = 1
# Set up model
self.mlp = MLP(nvis=self.mlp_nvis,
layers=[Linear(self.mlp_nout, 'out', irange=0.1)])
self.G = ConditionalGenerator(
input_condition_space=self.condition_space,
condition_distribution=self.condition_distribution,
noise_dim=self.noise_dim,
mlp=self.mlp)
def test_conditional_generator_input_setup(self):
"""Check that conditional generator correctly sets up composite
input layer."""
# Feedforward: We want the net to ignore the noise and simply
# convert the one-hot vector to a number
weights = np.concatenate([
np.zeros((self.mlp_nout, self.noise_dim)),
np.array(range(self.num_labels)).reshape(
(1, -1)).repeat(self.mlp_nout, axis=0)
],
axis=1).T.astype(theano.config.floatX)
self.mlp.layers[0].set_weights(weights)
inp = (T.matrix(), T.matrix(dtype=self.condition_dtype))
f = theano.function(inp, self.G.mlp.fprop(inp))
assert_array_equal(
f(
np.random.rand(self.num_labels,
self.noise_dim).astype(theano.config.floatX),
self.condition_formatter.format(
np.array(range(self.num_labels)))),
np.array(range(self.num_labels)).reshape(self.num_labels, 1))
def test_sample_noise(self):
"""Test barebones noise sampling."""
n = T.iscalar()
cond_inp = self.condition_distribution.sample(n)
sample_and_noise = theano.function([n],
self.G.sample_and_noise(
cond_inp, all_g_layers=True)[1])
print sample_and_noise(15)
def test_dtype_errors():
# Try to call theano_expr with a bad label dtype.
raised = False
fmt = OneHotFormatter(max_labels=50)
try:
fmt.theano_expr(theano.tensor.vector(dtype=theano.config.floatX))
except TypeError:
raised = True
assert raised
# Try to call format with a bad label dtype.
raised = False
try:
fmt.format(numpy.zeros(10, dtype='float64'))
except TypeError:
raised = True
assert raised
def test_dtype_errors():
# Try to call theano_expr with a bad label dtype.
raised = False
fmt = OneHotFormatter(max_labels=50)
try:
fmt.theano_expr(theano.tensor.vector(dtype=theano.config.floatX))
except TypeError:
raised = True
assert raised
# Try to call format with a bad label dtype.
raised = False
try:
fmt.format(numpy.zeros(10, dtype='float64'))
except TypeError:
raised = True
assert raised
def check_one_hot_formatter(seed, max_labels, dtype, ncases):
rng = numpy.random.RandomState(seed)
fmt = OneHotFormatter(max_labels=max_labels, dtype=dtype)
integer_labels = rng.random_integers(0, max_labels - 1, size=ncases)
one_hot_labels = fmt.format(integer_labels)
assert len(zip(*one_hot_labels.nonzero())) == ncases
for case, label in enumerate(integer_labels):
assert one_hot_labels[case, label] == 1
def format_targets(y):
# matlab has one-based indexing and one-based labels
# have to convert to zero-based labels so subtract 1...
y = y - 1
# we need only a 1d-array of integers
# squeeze in case of 2 dimensions, make sure it is still 1d in case of
# a single number (can happen for test runs with just one trial)
y = np.atleast_1d(y.squeeze())
y = y.astype(int)
target_formatter = OneHotFormatter(4)
y = target_formatter.format(y)
return y
def test_bad_arguments():
# Make sure an invalid max_labels raises an error.
raised = False
try:
fmt = OneHotFormatter(max_labels=-10)
except ValueError:
raised = True
assert raised
raised = False
try:
fmt = OneHotFormatter(max_labels='10')
except ValueError:
raised = True
assert raised
# Make sure an invalid dtype identifier raises an error.
raised = False
try:
fmt = OneHotFormatter(max_labels=10, dtype='invalid')
except TypeError:
raised = True
assert raised
# Make sure an invalid ndim raises an error for format().
fmt = OneHotFormatter(max_labels=10)
raised = False
try:
fmt.format(numpy.zeros((2, 3), dtype='int32'))
except ValueError:
raised = True
assert raised
# Make sure an invalid ndim raises an error for theano_expr().
raised = False
try:
fmt.theano_expr(theano.tensor.imatrix())
except ValueError:
raised = True
assert raised
def test_bad_arguments():
# Make sure an invalid max_labels raises an error.
raised = False
try:
fmt = OneHotFormatter(max_labels=-10)
except ValueError:
raised = True
assert raised
raised = False
try:
fmt = OneHotFormatter(max_labels='10')
except ValueError:
raised = True
assert raised
# Make sure an invalid dtype identifier raises an error.
raised = False
try:
fmt = OneHotFormatter(max_labels=10, dtype='invalid')
except TypeError:
raised = True
assert raised
# Make sure an invalid ndim raises an error for format().
fmt = OneHotFormatter(max_labels=10)
raised = False
try:
fmt.format(numpy.zeros((2, 3, 4), dtype='int32'))
except ValueError:
raised = True
assert raised
# Make sure an invalid ndim raises an error for theano_expr().
raised = False
try:
fmt.theano_expr(theano.tensor.itensor3())
except ValueError:
raised = True
assert raised
def generate_datasets(inputs):
targets = np.zeros(inputs.shape[0]).astype('int')
targets[::2] = 1 # every second target is class 1 others class 0
inputs[targets == 1] = inputs[targets == 1] + 1
target_formatter = OneHotFormatter(2)
targets_one_hot = target_formatter.format(targets)
train_set = VolumetricDenseDesignMatrix(topo_view=inputs[0:50],
y=targets_one_hot[0:50], axes=('b', 0, 1, 2, 'c'))
valid_set = VolumetricDenseDesignMatrix(topo_view=inputs[50:75],
y=targets_one_hot[50:75], axes=('b', 0, 1, 2, 'c'))
test_set = VolumetricDenseDesignMatrix(topo_view=inputs[75:100],
y=targets_one_hot[75:100], axes=('b', 0, 1, 2, 'c'))
return train_set, valid_set, test_set
def _transform_single_channel_data(self, X, y):
windowed_X = np.reshape(X, (-1, self.window_size))
windowed_y = np.reshape(y, (-1, self.window_size))
# Format the target into proper format
sum_y = np.sum(windowed_y, axis=1)
sum_y[sum_y > 0] = 1
# Duplicate the labels for all channels
dup_y = np.tile(sum_y, self.n_channels)
one_hot_formatter = OneHotFormatter(max_labels=self.n_classes)
hot_y = one_hot_formatter.format(dup_y)
return windowed_X, hot_y, None
def test_one_hot_formatter_simple():
def check_one_hot_formatter(seed, max_labels, dtype, ncases):
rng = numpy.random.RandomState(seed)
fmt = OneHotFormatter(max_labels=max_labels, dtype=dtype)
integer_labels = rng.random_integers(0, max_labels - 1, size=ncases)
one_hot_labels = fmt.format(integer_labels)
assert len(list(zip(*one_hot_labels.nonzero()))) == ncases
for case, label in enumerate(integer_labels):
assert one_hot_labels[case, label] == 1
rng = numpy.random.RandomState(0)
for seed, dtype in enumerate(all_types):
yield (check_one_hot_formatter, seed, rng.random_integers(1, 30), dtype, rng.random_integers(1, 100))
fmt = OneHotFormatter(max_labels=10)
assert fmt.format(numpy.zeros((1, 1), dtype="uint8")).shape == (1, 1, 10)
def test_one_hot_formatter_simple():
def check_one_hot_formatter(seed, max_labels, dtype, ncases):
rng = numpy.random.RandomState(seed)
fmt = OneHotFormatter(max_labels=max_labels, dtype=dtype)
integer_labels = rng.random_integers(0, max_labels - 1, size=ncases)
one_hot_labels = fmt.format(integer_labels)
assert len(list(zip(*one_hot_labels.nonzero()))) == ncases
for case, label in enumerate(integer_labels):
assert one_hot_labels[case, label] == 1
rng = numpy.random.RandomState(0)
for seed, dtype in enumerate(all_types):
yield (check_one_hot_formatter, seed, rng.random_integers(1, 30),
dtype, rng.random_integers(1, 100))
fmt = OneHotFormatter(max_labels=10)
assert fmt.format(numpy.zeros((1, 1), dtype='uint8')).shape == (1, 1, 10)
def _transform_single_channel_data(self, X, y):
windowed_X = np.reshape(X, (-1, self.window_size))
windowed_y = np.reshape(y, (-1, self.window_size))
# Format the target into proper format
sum_y = np.sum(windowed_y, axis=1)
sum_y[sum_y > 0] = 1
# Duplicate the labels for all channels
dup_y = np.tile(sum_y, self.n_channels)
one_hot_formatter = OneHotFormatter(max_labels=self.n_classes)
hot_y = one_hot_formatter.format(dup_y)
return windowed_X, hot_y, None
class ConditionalGeneratorTestCase(unittest.TestCase):
def setUp(self):
self.noise_dim = 10
self.num_labels = 10
self.condition_dtype = 'uint8'
self.condition_space = VectorSpace(dim=self.num_labels, dtype=self.condition_dtype)
self.condition_formatter = OneHotFormatter(self.num_labels, dtype=self.condition_dtype)
self.condition_distribution = OneHotDistribution(self.condition_space)
# TODO this nvis stuff is dirty. The ConditionalGenerator should handle it
self.mlp_nvis = self.noise_dim + self.num_labels
self.mlp_nout = 1
# Set up model
self.mlp = MLP(nvis=self.mlp_nvis, layers=[Linear(self.mlp_nout, 'out', irange=0.1)])
self.G = ConditionalGenerator(input_condition_space=self.condition_space,
condition_distribution=self.condition_distribution,
noise_dim=self.noise_dim,
mlp=self.mlp)
def test_conditional_generator_input_setup(self):
"""Check that conditional generator correctly sets up composite
input layer."""
# Feedforward: We want the net to ignore the noise and simply
# convert the one-hot vector to a number
weights = np.concatenate([np.zeros((self.mlp_nout, self.noise_dim)),
np.array(range(self.num_labels)).reshape((1, -1)).repeat(self.mlp_nout, axis=0)],
axis=1).T.astype(theano.config.floatX)
self.mlp.layers[0].set_weights(weights)
inp = (T.matrix(), T.matrix(dtype=self.condition_dtype))
f = theano.function(inp, self.G.mlp.fprop(inp))
assert_array_equal(
f(np.random.rand(self.num_labels, self.noise_dim).astype(theano.config.floatX),
self.condition_formatter.format(np.array(range(self.num_labels)))),
np.array(range(self.num_labels)).reshape(self.num_labels, 1))
def test_sample_noise(self):
"""Test barebones noise sampling."""
n = T.iscalar()
cond_inp = self.condition_distribution.sample(n)
sample_and_noise = theano.function([n], self.G.sample_and_noise(cond_inp, all_g_layers=True)[1])
print sample_and_noise(15)
def generate_datasets(inputs):
targets = np.zeros(inputs.shape[0]).astype('int')
targets[::2] = 1 # every second target is class 1 others class 0
inputs[targets == 1] = inputs[targets == 1] + 1
target_formatter = OneHotFormatter(2)
targets_one_hot = target_formatter.format(targets)
train_set = VolumetricDenseDesignMatrix(topo_view=inputs[0:50],
y=targets_one_hot[0:50],
axes=('b', 0, 1, 2, 'c'))
valid_set = VolumetricDenseDesignMatrix(topo_view=inputs[50:75],
y=targets_one_hot[50:75],
axes=('b', 0, 1, 2, 'c'))
test_set = VolumetricDenseDesignMatrix(topo_view=inputs[75:100],
y=targets_one_hot[75:100],
axes=('b', 0, 1, 2, 'c'))
return train_set, valid_set, test_set
def check_one_hot_formatter(seed, max_labels, dtype, ncases, nmultis):
rng = numpy.random.RandomState(seed)
fmt = OneHotFormatter(max_labels=max_labels, dtype=dtype)
integer_labels = rng.random_integers(0, max_labels - 1, size=ncases * nmultis).reshape(ncases, nmultis)
one_hot_labels = fmt.format(integer_labels, mode="merge")
# n_ones was expected to be equal to ncases * nmultis if integer_labels
# do not contain duplicated tags. (i.e., those labels like
# [1, 2, 2, 3, 5, 6].) Because that we are not depreciating this kind
# of duplicated labels, which allows different cases belong to
# different number of classes, and those duplicated tags will only
# activate one neuron in the k-hot representation, we need to use
# numpy.unique() here to eliminate those duplications while counting
# "1"s in the final k-hot representation.
n_ones = numpy.concatenate([numpy.unique(l) for l in integer_labels])
assert len(list(zip(*one_hot_labels.nonzero()))) == len(n_ones)
for case, label in enumerate(integer_labels):
assert numpy.sum(one_hot_labels[case, label]) == nmultis
def check_one_hot_formatter(seed, max_labels, dtype, ncases, nmultis):
rng = numpy.random.RandomState(seed)
fmt = OneHotFormatter(max_labels=max_labels, dtype=dtype)
integer_labels = rng.random_integers(
0, max_labels - 1, size=ncases*nmultis
).reshape(ncases, nmultis)
one_hot_labels = fmt.format(integer_labels, mode='merge')
# n_ones was expected to be equal to ncases * nmultis if integer_labels
# do not contain duplicated tags. (i.e., those labels like
# [1, 2, 2, 3, 5, 6].) Because that we are not depreciating this kind
# of duplicated labels, which allows different cases belong to
# different number of classes, and those duplicated tags will only
# activate one neuron in the k-hot representation, we need to use
# numpy.unique() here to eliminate those duplications while counting
# "1"s in the final k-hot representation.
n_ones = numpy.concatenate([numpy.unique(l) for l in integer_labels])
assert len(list(zip(*one_hot_labels.nonzero()))) == len(n_ones)
for case, label in enumerate(integer_labels):
assert numpy.sum(one_hot_labels[case, label]) == nmultis
def _transform_multi_channel_data(self, X, y):
# Data partitioning
parted_X, parted_y = self._partition_data(X=X, y=y, partition_size=self.window_size)
transposed_X = np.transpose(parted_X, [0, 2, 1])
converted_X = np.reshape(transposed_X, (transposed_X.shape[0],
transposed_X.shape[1],
1,
transposed_X.shape[2]))
# Create view converter
view_converter = DefaultViewConverter(shape=self.sample_shape,
axes=('b', 0, 1, 'c'))
# Convert data into a design matrix
view_converted_X = view_converter.topo_view_to_design_mat(converted_X)
assert np.all(converted_X == view_converter.design_mat_to_topo_view(view_converted_X))
# Format the target into proper format
sum_y = np.sum(parted_y, axis=1)
sum_y[sum_y > 0] = 1
one_hot_formatter = OneHotFormatter(max_labels=self.n_classes)
hot_y = one_hot_formatter.format(sum_y)
return view_converted_X, hot_y, view_converter
# Samples per condition
sample_cols = 5
# Generate conditional information
conditional_batch = model.generator.condition_space.make_theano_batch()
formatter = OneHotFormatter(rows,
dtype=model.generator.condition_space.dtype)
conditional = formatter.theano_expr(conditional_batch, mode='concatenate')
# Now sample from generator
# For some reason format_as from VectorSpace is not working right
topo_samples_batch = model.generator.sample(conditional)
topo_sample_f = theano.function([conditional], topo_samples_batch)
conditional_data = formatter.format(np.concatenate([np.repeat(i, sample_cols) for i in range(rows)])
.reshape((rows * sample_cols, 1)),
mode='concatenate')
topo_samples = topo_sample_f(conditional_data)
samples = dataset.get_design_matrix(topo_samples)
dataset.axes = ['b', 0, 1, 'c']
dataset.view_converter.axes = ['b', 0, 1, 'c']
topo_samples = dataset.get_topological_view(samples)
pv = PatchViewer(grid_shape=(rows, sample_cols + 1), patch_shape=(32,32),
is_color=True)
scale = np.abs(samples).max()
X = dataset.X
topo = dataset.get_topological_view()
index = 0
def __init__(self, which_set, onehot_dtype='uint8',
center=False, rescale=False, gcn=None,
start=None, stop=None, axes=('b', 0, 1, 'c'),
toronto_prepro=False, preprocessor=None):
"""Modified version of the CIFAR10 constructor which creates Y
as one-hot vectors rather than simple indexes. This is super
hacky. Sorry, Guido.."""
# note: there is no such thing as the cifar10 validation set;
# pylearn1 defined one but really it should be user-configurable
# (as it is here)
self.axes = axes
# we define here:
dtype = 'uint8'
ntrain = 50000
nvalid = 0 # artefact, we won't use it
ntest = 10000
# we also expose the following details:
self.img_shape = (3, 32, 32)
self.img_size = numpy.prod(self.img_shape)
self.n_classes = 10
self.label_names = ['airplane', 'automobile', 'bird', 'cat', 'deer',
'dog', 'frog', 'horse', 'ship', 'truck']
# prepare loading
fnames = ['data_batch_%i' % i for i in range(1, 6)]
datasets = {}
datapath = os.path.join(
string_utils.preprocess('${PYLEARN2_DATA_PATH}'),
'cifar10', 'cifar-10-batches-py')
for name in fnames + ['test_batch']:
fname = os.path.join(datapath, name)
if not os.path.exists(fname):
raise IOError(fname + " was not found. You probably need to "
"download the CIFAR-10 dataset by using the "
"download script in "
"pylearn2/scripts/datasets/download_cifar10.sh "
"or manually from "
"http://www.cs.utoronto.ca/~kriz/cifar.html")
datasets[name] = cache.datasetCache.cache_file(fname)
lenx = numpy.ceil((ntrain + nvalid) / 10000.) * 10000
x = numpy.zeros((lenx, self.img_size), dtype=dtype)
y = numpy.zeros((lenx, 1), dtype=dtype)
# load train data
nloaded = 0
for i, fname in enumerate(fnames):
_logger.info('loading file %s' % datasets[fname])
data = serial.load(datasets[fname])
x[i * 10000:(i + 1) * 10000, :] = data['data']
y[i * 10000:(i + 1) * 10000, 0] = data['labels']
nloaded += 10000
if nloaded >= ntrain + nvalid + ntest:
break
# load test data
_logger.info('loading file %s' % datasets['test_batch'])
data = serial.load(datasets['test_batch'])
# process this data
Xs = {'train': x[0:ntrain],
'test': data['data'][0:ntest]}
Ys = {'train': y[0:ntrain],
'test': data['labels'][0:ntest]}
X = numpy.cast['float32'](Xs[which_set])
y = Ys[which_set]
if isinstance(y, list):
y = numpy.asarray(y).astype(dtype)
if which_set == 'test':
assert y.shape[0] == 10000
y = y.reshape((y.shape[0], 1))
formatter = OneHotFormatter(self.n_classes, dtype=onehot_dtype)
y = formatter.format(y, mode='concatenate')
if center:
X -= 127.5
self.center = center
if rescale:
X /= 127.5
self.rescale = rescale
if toronto_prepro:
assert not center
assert not gcn
X = X / 255.
if which_set == 'test':
other = CIFAR10(which_set='train')
oX = other.X
oX /= 255.
X = X - oX.mean(axis=0)
else:
X = X - X.mean(axis=0)
self.toronto_prepro = toronto_prepro
self.gcn = gcn
if gcn is not None:
gcn = float(gcn)
X = global_contrast_normalize(X, scale=gcn)
if start is not None:
# This needs to come after the prepro so that it doesn't
# change the pixel means computed above for toronto_prepro
assert start >= 0
assert stop > start
assert stop <= X.shape[0]
X = X[start:stop, :]
y = y[start:stop, :]
assert X.shape[0] == y.shape[0]
if which_set == 'test':
assert X.shape[0] == 10000
view_converter = dense_design_matrix.DefaultViewConverter((32, 32, 3),
axes)
super(CIFAR10, self).__init__(X=X, y=y, view_converter=view_converter,
)#y_labels=self.n_classes)
assert not contains_nan(self.X)
if preprocessor:
preprocessor.apply(self)
# Another hack: rename 'targets' to match model expectations
space, (X_source, y_source) = self.data_specs
self.data_specs = (space, (X_source, 'condition'))
def __init__(
self,
path,
name='', # optional name
# selectors
subjects='all', # optional selector (list) or 'all'
trial_types='all', # optional selector (list) or 'all'
trial_numbers='all', # optional selector (list) or 'all'
conditions='all', # optional selector (list) or 'all'
partitioner=None,
channel_filter=NoChannelFilter(
), # optional channel filter, default: keep all
channel_names=None, # optional channel names (for metadata)
label_map=None, # optional conversion of labels
remove_dc_offset=False, # optional subtraction of channel mean, usually done already earlier
resample=None, # optional down-sampling
# optional sub-sequences selection
start_sample=0,
stop_sample=None, # optional for selection of sub-sequences
# optional signal filter to by applied before spitting the signal
signal_filter=None,
# windowing parameters
frame_size=-1,
hop_size=-1, # values > 0 will lead to windowing
hop_fraction=None, # alternative to specifying absolute hop_size
# optional spectrum parameters, n_fft = 0 keeps raw data
n_fft=0,
n_freq_bins=None,
spectrum_log_amplitude=False,
spectrum_normalization_mode=None,
include_phase=False,
flatten_channels=False,
layout='tf', # (0,1)-axes layout tf=time x features or ft=features x time
save_matrix_path=None,
keep_metadata=False,
):
'''
Constructor
'''
# save params
self.params = locals().copy()
del self.params['self']
# print self.params
# TODO: get the whole filtering into an extra class
datafiles_metadata, metadb = load_datafiles_metadata(path)
# print datafiles_metadata
def apply_filters(filters, node):
if isinstance(node, dict):
filtered = []
keepkeys = filters[0]
for key, value in node.items():
if keepkeys == 'all' or key in keepkeys:
filtered.extend(apply_filters(filters[1:], value))
return filtered
else:
return node # [node]
# keep only files that match the metadata filters
self.datafiles = apply_filters(
[subjects, trial_types, trial_numbers, conditions],
datafiles_metadata)
# copy metadata for retained files
self.metadb = {}
for datafile in self.datafiles:
self.metadb[datafile] = metadb[datafile]
# print self.datafiles
# print self.metadb
self.name = name
if partitioner is not None:
self.datafiles = partitioner.get_partition(self.name, self.metadb)
self.include_phase = include_phase
self.spectrum_normalization_mode = spectrum_normalization_mode
self.spectrum_log_amplitude = spectrum_log_amplitude
self.sequence_partitions = [
] # used to keep track of original sequences
# metadata: [subject, trial_no, stimulus, channel, start, ]
self.metadata = []
sequences = []
labels = []
n_sequences = 0
if frame_size > 0 and hop_size == -1 and hop_fraction is not None:
hop_size = np.ceil(frame_size / hop_fraction)
for i in xrange(len(self.datafiles)):
with log_timing(log,
'loading data from {}'.format(self.datafiles[i])):
# save start of next sequence
self.sequence_partitions.append(n_sequences)
data, metadata = load(os.path.join(path, self.datafiles[i]))
label = metadata['label']
if label_map is not None:
label = label_map[label]
multi_channel_frames = []
# process 1 channel at a time
for channel in xrange(data.shape[1]):
# filter channels
if not channel_filter.keep_channel(channel):
continue
samples = data[:, channel]
# subtract channel mean
if remove_dc_offset:
samples -= samples.mean()
# down-sample if requested
if resample is not None and resample[0] != resample[1]:
samples = librosa.resample(samples, resample[0],
resample[1])
# apply optional signal filter after down-sampling -> requires lower order
if signal_filter is not None:
samples = signal_filter.process(samples)
# get sub-sequence in resampled space
# log.info('using samples {}..{} of {}'.format(start_sample,stop_sample, samples.shape))
samples = samples[start_sample:stop_sample]
if n_fft is not None and n_fft > 0: # Optionally:
### frequency spectrum branch ###
# transform to spectogram
hop_length = n_fft / 4
'''
from http://theremin.ucsd.edu/~bmcfee/librosadoc/librosa.html
>>> # Get a power spectrogram from a waveform y
>>> S = np.abs(librosa.stft(y)) ** 2
>>> log_S = librosa.logamplitude(S)
'''
S = librosa.core.stft(samples,
n_fft=n_fft,
hop_length=hop_length)
# mag = np.abs(S) # magnitude spectrum
mag = np.abs(S)**2 # power spectrum
# include phase information if requested
if self.include_phase:
# phase = np.unwrap(np.angle(S))
phase = np.angle(S)
# Optionally: cut off high bands
if n_freq_bins is not None:
mag = mag[0:n_freq_bins, :]
if self.include_phase:
phase = phase[0:n_freq_bins, :]
if self.spectrum_log_amplitude:
mag = librosa.logamplitude(mag)
s = mag # for normalization
'''
NOTE on normalization:
It depends on the structure of a neural network and (even more)
on the properties of data. There is no best normalization algorithm
because if there would be one, it would be used everywhere by default...
In theory, there is no requirement for the data to be normalized at all.
This is a purely practical thing because in practice convergence could
take forever if your input is spread out too much. The simplest would be
to just normalize it by scaling your data to (-1,1) (or (0,1) depending
on activation function), and in most cases it does work. If your
algorithm converges well, then this is your answer. If not, there are
too many possible problems and methods to outline here without knowing
the actual data.
'''
## normalize to mean 0, std 1
if self.spectrum_normalization_mode == 'mean0_std1':
# s = preprocessing.scale(s, axis=0);
mean = np.mean(s)
std = np.std(s)
s = (s - mean) / std
## normalize by linear transform to [0,1]
elif self.spectrum_normalization_mode == 'linear_0_1':
s = s / np.max(s)
## normalize by linear transform to [-1,1]
elif self.spectrum_normalization_mode == 'linear_-1_1':
s = -1 + 2 * (s - np.min(s)) / (np.max(s) -
np.min(s))
elif self.spectrum_normalization_mode is not None:
raise ValueError(
'unsupported spectrum normalization mode {}'.
format(self.spectrum_normalization_mode))
#print s.mean(axis=0)
#print s.std(axis=0)
# include phase information if requested
if self.include_phase:
# normalize phase to [-1.1]
phase = phase / np.pi
s = np.vstack([s, phase])
# transpose to fit pylearn2 layout
s = np.transpose(s)
# print s.shape
### end of frequency spectrum branch ###
else:
### raw waveform branch ###
# normalize to max amplitude 1
s = librosa.util.normalize(samples)
# add 2nd data dimension
s = s.reshape(s.shape[0], 1)
# print s.shape
### end of raw waveform branch ###
s = np.asfarray(s, dtype='float32')
if frame_size > 0 and hop_size > 0:
s = s.copy(
) # FIXME: THIS IS NECESSARY IN MultiChannelEEGSequencesDataset - OTHERWISE, THE FOLLOWING OP DOES NOT WORK!!!!
frames = frame(s,
frame_length=frame_size,
hop_length=hop_size)
else:
frames = s
del s
# print frames.shape
if flatten_channels:
# add artificial channel dimension
frames = frames.reshape(
(frames.shape[0], frames.shape[1], frames.shape[2],
1))
# print frames.shape
sequences.append(frames)
# increment counter by new number of frames
n_sequences += frames.shape[0]
if keep_metadata:
# determine channel name
channel_name = None
if channel_names is not None:
channel_name = channel_names[channel]
elif 'channels' in metadata:
channel_name = metadata['channels'][channel]
self.metadata.append({
'subject':
metadata['subject'], # subject
'trial_type':
metadata['trial_type'], # trial_type
'trial_no':
metadata['trial_no'], # trial_no
'condition':
metadata['condition'], # condition
'channel':
channel, # channel
'channel_name':
channel_name,
'start':
self.sequence_partitions[-1], # start
'stop':
n_sequences # stop
})
for _ in xrange(frames.shape[0]):
labels.append(label)
else:
multi_channel_frames.append(frames)
### end of channel iteration ###
if not flatten_channels:
# turn list into array
multi_channel_frames = np.asfarray(multi_channel_frames,
dtype='float32')
# [channels x frames x time x freq] -> cb01
# [channels x frames x time x 1] -> cb0.
# move channel dimension to end
multi_channel_frames = np.rollaxis(multi_channel_frames, 0,
4)
# print multi_channel_frames.shape
# log.debug(multi_channel_frames.shape)
sequences.append(multi_channel_frames)
# increment counter by new number of frames
n_sequences += multi_channel_frames.shape[0]
if keep_metadata:
self.metadata.append({
'subject':
metadata['subject'], # subject
'trial_type':
metadata['trial_type'], # trial_type
'trial_no':
metadata['trial_no'], # trial_no
'condition':
metadata['condition'], # condition
'channel':
'all', # channel
'start':
self.sequence_partitions[-1], # start
'stop':
n_sequences # stop
})
for _ in xrange(multi_channel_frames.shape[0]):
labels.append(label)
### end of datafile iteration ###
# turn into numpy arrays
sequences = np.vstack(sequences)
# print sequences.shape;
labels = np.hstack(labels)
# one_hot_y = one_hot(labels)
one_hot_formatter = OneHotFormatter(labels.max() + 1) # FIXME!
one_hot_y = one_hot_formatter.format(labels)
self.labels = labels
if layout == 'ft': # swap axes to (batch, feature, time, channels)
sequences = sequences.swapaxes(1, 2)
log.debug('final dataset shape: {} (b,0,1,c)'.format(sequences.shape))
super(MultiChannelEEGDataset, self).__init__(topo_view=sequences,
y=one_hot_y,
axes=['b', 0, 1, 'c'])
log.info(
'generated dataset "{}" with shape X={}={} y={} labels={} '.format(
self.name, self.X.shape, sequences.shape, self.y.shape,
self.labels.shape))
if save_matrix_path is not None:
matrix = DenseDesignMatrix(topo_view=sequences,
y=one_hot_y,
axes=['b', 0, 1, 'c'])
with log_timing(
log,
'saving DenseDesignMatrix to {}'.format(save_matrix_path)):
serial.save(save_matrix_path, matrix)
def svd_accuracy(file_name, ec, kwargs, folds=10, max_svs=10, max_init=15):
"""
Classify data based on svd features.
"""
kwargs['condense'] = False
ds = ec.ECoG(file_name, which_set='train', **kwargs)
n_classes = int(np.around(ds.y.max() + 1))
max_svs = min(max_svs, n_classes)
init_list = np.arange(0, n_classes - max_svs + 1)
init_list = init_list[init_list < max_init]
nsvs_list = np.arange(1, max_svs + 1)
pa = np.inf * np.ones((folds, len(nsvs_list), len(init_list)))
ma = np.inf * np.ones((folds, len(nsvs_list), len(init_list)))
va = np.inf * np.ones((folds, len(nsvs_list), len(init_list)))
u_s = np.zeros((folds, n_classes, n_classes))
s_s = np.zeros((folds, n_classes))
v_s = np.zeros((folds, n_classes, ds.X.shape[1]))
ohf = OneHotFormatter(n_classes)
for fold in range(folds):
kwargs_copy = copy.deepcopy(kwargs)
print('fold: {}'.format(fold))
ds = ec.ECoG(file_name,
which_set='train',
fold=fold,
center=False,
**kwargs_copy)
# CV
ts = ds.get_test_set()
vs = ds.get_valid_set()
train_X = np.concatenate((ds.X, vs.X), axis=0)
train_mean = train_X.mean(axis=0)
train_X = train_X - train_mean
train_y = np.concatenate((ds.y, vs.y), axis=0)
test_X = ts.X - train_mean
test_y = ts.y
y_oh = ohf.format(train_y, mode='concatenate')
c_yx = (y_oh - y_oh.mean(axis=0)).T.dot(train_X) / train_X.shape[0]
u, s, v = np.linalg.svd(c_yx, full_matrices=False)
u_s[fold] = u
s_s[fold] = s
v_s[fold] = v
for ii, n_svs in enumerate(nsvs_list):
for jj, sv_init in enumerate(init_list):
vp = v[sv_init:sv_init + n_svs]
train_proj = train_X.dot(vp.T)
test_proj = test_X.dot(vp.T)
cl = LR(solver='lbfgs',
multi_class='multinomial').fit(train_proj,
train_y.ravel())
y_hat = cl.predict(test_proj)
p_results = []
m_results = []
v_results = []
for y, yh in zip(test_y.ravel(), y_hat.ravel()):
pr = place_equiv(y, yh)
if pr is not None:
p_results.append(pr)
mr = manner_equiv(y, yh)
if mr is not None:
m_results.append(mr)
vr = vowel_equiv(y, yh)
if vr is not None:
v_results.append(vr)
pa[fold, ii, jj] = np.array(p_results).mean()
ma[fold, ii, jj] = np.array(m_results).mean()
va[fold, ii, jj] = np.array(v_results).mean()
return pa, ma, va, u_s, s_s, v_s, init_list, nsvs_list
def load_data(self):
# Get the directory of the patient data
patient_dir = os.path.join(self.data_dir, self.patient_id)
# Load metadata about dataset form MAT file
metadata_fname = os.path.join(patient_dir, 'trainset.mat')
metadata_mat = loadmat(metadata_fname)
# Get number of seizures
self.n_seizures = metadata_mat.get('ictals').size
# Get detail of the segment
self.sampling_rate = metadata_mat['sampling_rate'][0][0]
self.segment_sec = metadata_mat['segment_sec'][0][0]
self.segment_samples = self.sampling_rate * self.segment_sec
self.preictal_samples = 0
self.nonictal_samples = 0
# Examples of indexing through MAT file
# mat['nonictals'][i][0]['filename'][0][0][0][j][0]
# mat['nonictals'][i][0]['idx'][0][0][0][j][0]
# mat['nonictals'][i][0]['n_segments'][0][0][0][0]
# Balanced classes
if self.which_set == 'train' or self.which_set == 'valid_train':
if self.which_set == 'train':
select_idx = np.setdiff1d(
range(metadata_mat['preictals'].size),
np.asarray([
self.leave_out_seizure_idx_valid,
self.leave_out_seizure_idx_test
]))
else:
select_idx = np.asarray([self.leave_out_seizure_idx_valid])
X = None
y = None
for i in select_idx:
print '====== Seizure', i, '======'
# Non-ictal data
temp_nonictal_X = self.load_segment(part='nonictals',
seizure_idx=i,
metadata_mat=metadata_mat,
patient_dir=patient_dir)
# Pre-ictal
temp_preictal_X = self.load_segment(part='preictals',
seizure_idx=i,
metadata_mat=metadata_mat,
patient_dir=patient_dir)
# Concatenate preictal and nonictal data
temp_X = np.concatenate((temp_preictal_X, temp_nonictal_X),
axis=0)
temp_y = np.zeros(temp_X.shape[0], dtype=int)
temp_y[range(temp_preictal_X.shape[0])] = 1
# Sanity check
# if not (temp_preictal_X.shape[0] == temp_nonictal_X.shape[0]):
# raise Exception('Unbalanced classes.')
print 'Preictal samples: {0}, Nonictal samples: {1}'.format(
temp_preictal_X.shape[0], temp_nonictal_X.shape[0])
if not np.all(
np.arange(temp_preictal_X.shape[0]) == np.where(temp_y)
[0]):
raise Exception(
'There is a mismatch between the number of preictal data and labels.'
)
self.preictal_samples = self.preictal_samples + temp_preictal_X.shape[
0]
self.nonictal_samples = self.nonictal_samples + temp_nonictal_X.shape[
0]
if not (X is None) and not (y is None):
X = np.concatenate((X, temp_X), axis=0)
y = np.append(y, temp_y)
else:
X = temp_X
y = temp_y
# Unbalanced classes
elif self.which_set == 'valid' or self.which_set == 'test':
if self.which_set == 'valid':
select_idx = self.leave_out_seizure_idx_valid
else:
select_idx = self.leave_out_seizure_idx_test
print '====== Seizure', select_idx, '======'
# Get metadata of all blocks
block_df = pd.read_table(os.path.join(patient_dir,
'block_metadata.txt'),
sep='\t')
# Get block index of the selected seizure
select_sz_fname = metadata_mat['preictals'][select_idx][0][
'filename'][0][0][0][0][0]
block_idx = np.where(block_df.filename == select_sz_fname)[0][0]
n_padded_block = 2
start_block_idx = block_idx - n_padded_block
end_block_idx = block_idx + n_padded_block + 1
if start_block_idx < 0:
start_block_idx = 0
if end_block_idx > block_df.shape[0]:
end_block_idx = block_df.shape[0]
select_block_idx = np.arange(start_block_idx, end_block_idx)
filenames = block_df.filename[select_block_idx].values
X = None
y = None
y_select_idx = None
ictal_labels = None
for b_idx, fname in enumerate(filenames):
# Name of the MAT files that store EEG data
data_fname = fname.replace('.data', '.mat')
# Name of the MAT file that stores indices of flat (i.e., false) segments
fname_flat = fname.replace('.data',
'_flat_signal_segment_idx.mat')
# Get all good indices (i.e., remove segments of flat signals)
flat_mat = loadmat(os.path.join(patient_dir, fname_flat))
flat_idx = np.empty(0, dtype=int)
for j in range(flat_mat['flat_signal_segment_idx'].shape[0]):
flat_idx = np.append(
flat_idx,
np.squeeze(flat_mat['flat_signal_segment_idx'][j][0]))
flat_idx = flat_idx - 1 # Change from MATLAB to python index system
data_mat = loadmat(os.path.join(patient_dir, data_fname))
if data_mat['signals'].shape[1] != block_df.samples[
select_block_idx[b_idx]]:
raise Exception(
'There is a mismatch between the number samples specified in the metadata and '
'the provided signal data')
n_segments = np.ceil(data_mat['signals'].shape[1] /
(self.segment_samples * 1.0))
all_idx = np.arange(n_segments, dtype=int)
good_idx = np.setdiff1d(all_idx, flat_idx)
# Get indicies of scalp EEG channels
elec_names = np.asarray(
[ename[0][0] for ename in data_mat['elec_names']])
scalp_channels_idx = np.empty(0, dtype=int)
for ch in self.scalp_channel_labels:
scalp_channels_idx = np.append(
scalp_channels_idx,
np.where(elec_names == ch)[0][0])
print 'Load', self.which_set, 'data from', fname
if good_idx.size > 0:
temp_X = None
for idx in range(good_idx.size):
g_idx = good_idx[idx]
start_sample_idx = np.uint32(
g_idx) * self.segment_samples
end_sample_idx = np.uint32(g_idx +
1) * self.segment_samples
if end_sample_idx > data_mat['signals'].shape[1]:
# Zero-padding if the window size is not compatible
extra = end_sample_idx - data_mat['signals'].shape[
1]
assert (data_mat['signals'].shape[1] +
extra) % self.segment_samples == 0
if extra > 0:
data_mat['signals'] = np.concatenate(
(data_mat['signals'],
np.zeros(
(data_mat['signals'].shape[0], extra),
dtype=float)),
axis=1)
assert data_mat['signals'].shape[
1] % self.segment_samples == 0
temp_sample_idx = np.arange(start_sample_idx,
end_sample_idx)
if not (temp_X is None):
temp = data_mat['signals'][:, temp_sample_idx]
temp_X = np.concatenate(
(temp_X,
np.asarray([temp[scalp_channels_idx, :]])),
axis=0)
else:
temp = data_mat['signals'][:, temp_sample_idx]
temp_X = np.asarray([temp[scalp_channels_idx, :]])
# If this record contains preictal data, get preictal labels
temp_preictal_meta_idx = -1
temp_preictal_fname_idx = -1
for preictal_meta_idx, preictal_meta in enumerate(
metadata_mat['preictals']):
for preictal_fname_idx, preictal_fname in enumerate(
preictal_meta[0]['filename'][0][0][0]):
if preictal_fname == fname:
temp_preictal_meta_idx = preictal_meta_idx
temp_preictal_fname_idx = preictal_fname_idx
break
if temp_preictal_meta_idx != -1 and temp_preictal_fname_idx != -1:
# Preictal indices
preictal_idx = metadata_mat['preictals'][
temp_preictal_meta_idx][0]['idx'][0][0][0][
temp_preictal_fname_idx][0]
preictal_idx = preictal_idx - 1 # Change from MATLAB to python index system
temp_y = np.zeros(n_segments, dtype=int)
temp_y[preictal_idx] = 1
# Sanity check
if not (preictal_idx.size == np.intersect1d(
good_idx, preictal_idx).size):
raise Exception(
'Good indices and preictal indices are mismatch.'
)
# Remove segment of flat signals from labels
temp_y = temp_y[good_idx]
self.preictal_samples = self.preictal_samples + preictal_idx.size
self.nonictal_samples = self.nonictal_samples + (
temp_y.size - preictal_idx.size)
else:
temp_y = np.zeros(temp_X.shape[0], dtype=int)
self.nonictal_samples = self.nonictal_samples + temp_y.size
# If this record contains preictal data of the leave-out-seizure index, get preictal labels
if temp_preictal_meta_idx == select_idx:
temp_y_select_idx = temp_y
else:
temp_y_select_idx = np.zeros(temp_X.shape[0],
dtype=int)
# If this record contains ictal data, get ictal labels
temp_ictal_meta_idx = -1
temp_ictal_fname_idx = -1
for ictal_meta_idx, ictal_meta in enumerate(
metadata_mat['ictals']):
for ictal_fname_idx, ictal_fname in enumerate(
ictal_meta[0]['filename'][0][0][0]):
if ictal_fname == fname:
temp_ictal_meta_idx = ictal_meta_idx
temp_ictal_fname_idx = ictal_fname_idx
break
if temp_ictal_meta_idx != -1 and temp_ictal_fname_idx != -1:
# Ictal indices
ictal_idx = metadata_mat['ictals'][
temp_ictal_meta_idx][0]['idx'][0][0][0][
temp_ictal_fname_idx][0]
ictal_idx = ictal_idx - 1 # Change from MATLAB to python index system
temp_ictal_labels = np.zeros(n_segments, dtype=int)
temp_ictal_labels[ictal_idx] = 1
# Sanity check
if not (ictal_idx.size == np.intersect1d(
good_idx, ictal_idx).size):
raise Exception(
'Good indices and ictal indices are mismatch.')
# Remove segment of flat signals from labels
temp_ictal_labels = temp_ictal_labels[good_idx]
else:
temp_ictal_labels = np.zeros(temp_X.shape[0],
dtype=int)
# Sanity check
if not (temp_X.shape[0] == temp_y.size):
raise Exception(
'Number of feature data and labels are not equal.')
if not (temp_X.shape[0] == temp_ictal_labels.size):
raise Exception(
'Number of feature data and labels are not equal.')
if not (X is None) and not (y is None) and not (
ictal_labels is None):
X = np.concatenate((X, temp_X), axis=0)
y = np.append(y, temp_y)
y_select_idx = np.append(y_select_idx,
temp_y_select_idx)
ictal_labels = np.append(ictal_labels,
temp_ictal_labels)
else:
X = temp_X
y = temp_y
y_select_idx = temp_y_select_idx
ictal_labels = temp_ictal_labels
else:
print 'There is no good segment for during this seizure'
# Store preictal labels that are from the leave-out-seizure index (use for compute accuracy)
# Note: this property will exist when which_set=='valid' or which_set=='test'
# as there is no need for ictal to be imported.
self.y_select_idx = y_select_idx
# Sanity check
if np.where(y_select_idx == 1)[0].size > np.where(y == 1)[0].size:
raise Exception(
'There is an error in collecting preictal labels only from the leave-out-seizure index.'
)
elif np.where(y_select_idx == 1)[0].size == np.where(
y == 1)[0].size:
print 'There is only one preictal periods, and this period is from the leave-out-seizure index.'
if not np.all(
np.where(y_select_idx == 1)[0] == np.where(y == 1)[0]):
raise Exception(
'There is a mismatch between y_select_idx and y.')
elif np.where(y_select_idx == 1)[0].size < np.where(
y == 1)[0].size:
print 'There are more than one preictal periods.'
if not np.all(
np.intersect1d(
np.where(y == 1)[0],
np.where(y_select_idx == 1)[0]) == np.where(
y_select_idx == 1)[0]):
raise Exception(
'There is a mismatch between y_select_idx and y in the preictal labels of the leave-out-seizure index.'
)
# Store ictal labels
# Note: this property will exist when which_set=='valid' or which_set=='test'
# as there is no need for ictal to be imported.
self.ictal_labels = ictal_labels
else:
raise Exception('Invalid dataset selection')
X = np.transpose(X, [0, 2, 1])
one_hot_formatter = OneHotFormatter(max_labels=2)
y = one_hot_formatter.format(y)
# Sanity check
if not (X.shape[0] == self.preictal_samples + self.nonictal_samples):
raise Exception(
'There is a mismatch in the number of training samples.')
if not (np.where(np.argmax(y, axis=1) == 1)[0].size
== self.preictal_samples):
raise Exception(
'There is a mismatch in the number of preictal samples and its labels.'
)
if not (X.shape[0] == y.shape[0]):
raise Exception(
'There is a mismatch in the number of training samples and its labels.'
)
return X, y
class IndexSpace(Space):
"""
A space representing indices, for example MNIST labels (0-10) or the
indices of words in a dictionary for NLP tasks. A single space can
contain multiple indices, for example the word indices of an n-gram.
IndexSpaces can be converted to VectorSpaces in two ways: Either the
labels are converted into one-hot vectors which are then concatenated,
or they are converted into a single vector where 1s indicate labels
present i.e. for 4 possible labels we have [0, 2] -> [1 0 1 0] or
[0, 2] -> [1 0 0 0 0 0 1 0].
"""
def __init__(self, max_labels, dim, **kwargs):
"""
Initialize an IndexSpace.
Parameters
----------
max_labels : int
The number of possible classes/labels. This means that
all labels should be < max_labels. Example: For MNIST
there are 10 numbers and hence max_labels = 10.
dim : int
The number of indices in one space e.g. for MNIST there is
one target label and hence dim = 1. If we have an n-gram
of word indices as input to a neurel net language model, dim = n.
kwargs: passes on to superclass constructor
"""
super(IndexSpace, self).__init__(**kwargs)
self.max_labels = max_labels
self.dim = dim
self.formatter = OneHotFormatter(self.max_labels)
def __str__(self):
"""
Return a string representation.
"""
return '%(classname)s(dim=%(dim)s, max_labels=%(max_labels)s)' % \
dict(classname=self.__class__.__name__,
dim=self.dim,
max_labels=self.max_labels)
def __eq__(self, other):
return (type(self) == type(other) and
self.max_labels == other.max_labels and
self.dim == other.dim)
def __ne__(self, other):
return (not self == other)
@functools.wraps(Space.get_total_dimension)
def get_total_dimension(self):
return self.dim
@functools.wraps(Space.np_format_as)
def np_format_as(self, batch, space):
if isinstance(space, VectorSpace):
if self.max_labels == space.dim:
rval = self.formatter.format(batch, sparse=space.sparse,
mode='merge')
elif self.dim * self.max_labels == space.dim:
rval = self.formatter.format(batch, sparse=space.sparse,
mode='concatenate')
else:
raise ValueError("Can't convert %s to %s"
% (self, space))
return rval
else:
raise ValueError("Can't convert %s to %s"
% (self, space))
@functools.wraps(Space._format_as)
def _format_as(self, batch, space):
"""
Supports formatting to a VectorSpace where indices are represented
by ones in a binary vector.
"""
if isinstance(space, VectorSpace):
if self.max_labels == space.dim:
rval = self.formatter.theano_expr(batch, sparse=space.sparse,
mode='merge')
elif self.dim * self.max_labels == space.dim:
rval = self.formatter.theano_expr(batch, sparse=space.sparse,
mode='concatenate')
else:
raise ValueError("Can't convert %s to %s"
% (self, space))
return rval
else:
raise ValueError("Can't convert %s to %s"
% (self, space))
@functools.wraps(Space.make_theano_batch)
def make_theano_batch(self, name=None, dtype=None, batch_size=None):
if batch_size == 1:
rval = T.lrow(name=name)
else:
rval = T.lmatrix(name=name)
return rval
@functools.wraps(Space.batch_size)
def batch_size(self, batch):
self.validate(batch)
return batch.shape[0]
@functools.wraps(Space.np_batch_size)
def np_batch_size(self, batch):
self.np_validate(batch)
return batch.shape[0]
@functools.wraps(Space._validate)
def _validate(self, batch):
"""
.. todo::
WRITEME
"""
if not isinstance(batch, theano.gof.Variable):
raise TypeError("IndexSpace batch should be a theano Variable, "
"got " + str(type(batch)))
if not isinstance(batch.type, (theano.tensor.TensorType,
CudaNdarrayType)):
raise TypeError("VectorSpace batch should be TensorType or "
"CudaNdarrayType, got "+str(batch.type))
if batch.ndim != 2:
raise ValueError('IndexSpace batches must be 2D, got %d '
'dimensions' % batch.ndim)
for val in get_debug_values(batch):
self.np_validate(val)
@functools.wraps(Space._np_validate)
def _np_validate(self, batch):
# Use the 'CudaNdarray' string to avoid importing theano.sandbox.cuda
# when it is not available
if (not isinstance(batch, np.ndarray)
and str(type(batch)) != "<type 'CudaNdarray'>"):
raise TypeError("The value of a IndexSpace batch should be a "
"numpy.ndarray, or CudaNdarray, but is %s."
% str(type(batch)))
if batch.ndim != 2:
raise ValueError("The value of a IndexSpace batch must be "
"2D, got %d dimensions for %s." % (batch.ndim,
batch))
if batch.shape[1] != self.dim:
raise ValueError("The width of a IndexSpace batch must match "
"with the space's dimension, but batch has shape "
"%s and dim = %d." % (str(batch.shape), self.dim))
def load_data(self):
# Get the directory of the patient data
patient_dir = os.path.join(self.data_dir, self.patient_id)
# Load metadata about dataset form MAT file
metadata_fname = os.path.join(
patient_dir, 'trainset_' + str(self.preictal_sec) + '.mat')
metadata_mat = loadmat(metadata_fname)
# Get number of seizures
self.n_seizures = metadata_mat.get('ictals').size
# Get detail of the segment
self.sampling_rate = metadata_mat['sampling_rate'][0][0]
self.segment_sec = metadata_mat['segment_sec'][0][0]
self.segment_samples = self.sampling_rate * self.segment_sec
# Get the number blocks to extend from the withheld seizure
self.n_extended_blocks_test = metadata_mat['n_extended_blocks_test'][
0][0]
self.preictal_samples = 0
self.nonictal_samples = 0
self.nan_non_flat_samples = 0
# Examples of indexing through MAT file
# mat['nonictals'][i][0]['filename'][0][0][0][j][0]
# mat['nonictals'][i][0]['idx'][0][0][0][j][0]
# mat['nonictals'][i][0]['n_segments'][0][0][0][0]
# Load shuffle data
if self.which_set == 'train' or self.which_set == 'valid_train':
if self.which_set == 'train':
select_idx = np.setdiff1d(
range(metadata_mat['preictals'].size),
np.asarray([
self.leave_out_seizure_idx_valid,
self.leave_out_seizure_idx_test
]))
else:
select_idx = np.asarray([self.leave_out_seizure_idx_valid])
X = None
y = None
if self.use_all_nonictals:
temp_preictal_X = None
for i in select_idx:
print '====== Seizure', i, '======'
# Pre-ictal
temp_X = self.load_feature(
part='preictals',
list_features=self.list_features,
seizure_idx=i,
metadata_mat=metadata_mat,
patient_dir=patient_dir)
if not (temp_preictal_X is None):
temp_preictal_X = np.concatenate(
(temp_preictal_X, temp_X), axis=1)
else:
temp_preictal_X = temp_X
self.preictal_samples = temp_preictal_X.shape[1]
# Non-ictal data
temp_nonictal_X = self.load_feature(
part='nonictals_all',
list_features=self.list_features,
seizure_idx=self.leave_out_seizure_idx_test,
metadata_mat=metadata_mat,
patient_dir=patient_dir)
X = np.concatenate((temp_preictal_X, temp_nonictal_X), axis=1)
y = np.zeros(X.shape[1], dtype=int)
y[range(self.preictal_samples)] = 1
self.nonictal_samples = temp_nonictal_X.shape[1]
print 'Preictal samples: {0}, Nonictal samples: {1}'.format(
self.preictal_samples, self.nonictal_samples)
if not np.all(
np.arange(self.preictal_samples) == np.where(y)[0]):
raise Exception(
'There is a mismatch between the number of preictal data and labels.'
)
else:
for i in select_idx:
print '====== Seizure', i, '======'
# Non-ictal data
temp_nonictal_X = self.load_feature(
part='nonictals',
list_features=self.list_features,
seizure_idx=i,
metadata_mat=metadata_mat,
patient_dir=patient_dir)
# Pre-ictal
temp_preictal_X = self.load_feature(
part='preictals',
list_features=self.list_features,
seizure_idx=i,
metadata_mat=metadata_mat,
patient_dir=patient_dir)
# Concatenate preictal and nonictal data
temp_X = np.concatenate((temp_preictal_X, temp_nonictal_X),
axis=1)
temp_y = np.zeros(temp_X.shape[1], dtype=int)
temp_y[range(temp_preictal_X.shape[1])] = 1
# Sanity check
# if not (temp_preictal_X.shape[1] == temp_nonictal_X.shape[1]):
# raise Exception('Unbalanced classes.')
print 'Preictal samples: {0}, Nonictal samples: {1}'.format(
temp_preictal_X.shape[1], temp_nonictal_X.shape[1])
if not np.all(
np.arange(temp_preictal_X.shape[1]) == np.where(
temp_y)[0]):
raise Exception(
'There is a mismatch between the number of preictal data and labels.'
)
self.preictal_samples = self.preictal_samples + temp_preictal_X.shape[
1]
self.nonictal_samples = self.nonictal_samples + temp_nonictal_X.shape[
1]
if not (X is None) and not (y is None):
X = np.concatenate((X, temp_X), axis=1)
y = np.append(y, temp_y)
else:
X = temp_X
y = temp_y
# Load continuous data
elif self.which_set == 'valid' or self.which_set == 'test':
if self.which_set == 'valid':
select_idx = self.leave_out_seizure_idx_valid
else:
select_idx = self.leave_out_seizure_idx_test
print '====== Seizure', select_idx, '======'
# Get metadata of all blocks
block_df = pd.read_table(os.path.join(patient_dir,
'block_metadata.txt'),
sep='\t')
# Get block index of the selected seizure
select_sz_fname = metadata_mat['preictals'][select_idx][0][
'filename'][0][0][0][0][0]
block_idx = np.where(block_df.filename == select_sz_fname)[0][0]
start_block_idx = block_idx - self.n_extended_blocks_test
end_block_idx = block_idx + self.n_extended_blocks_test + 1
if start_block_idx < 0:
start_block_idx = 0
if end_block_idx > block_df.shape[0]:
end_block_idx = block_df.shape[0]
select_block_idx = np.arange(start_block_idx, end_block_idx)
filenames = block_df.filename[select_block_idx].values
X = None
y = None
y_label_all = None
ictal_labels = None
for b_idx, fname in enumerate(filenames):
# Name of the MAT file that stores indices of flat (i.e., false) segments
fname_flat = fname.replace('.data',
'_flat_signal_segment_idx.mat')
# Get all good indices (i.e., remove segments of flat signals)
flat_mat = loadmat(os.path.join(patient_dir, fname_flat))
flat_idx = np.empty(0, dtype=int)
for j in range(flat_mat['flat_signal_segment_idx'].shape[0]):
flat_idx = np.append(
flat_idx,
np.squeeze(flat_mat['flat_signal_segment_idx'][j][0]))
flat_idx = flat_idx - 1 # Change from MATLAB to python index system
n_segments = np.ceil(
block_df.samples[select_block_idx[b_idx]] /
(self.segment_samples * 1.0))
all_idx = np.arange(n_segments, dtype=int)
good_idx = np.setdiff1d(all_idx, flat_idx)
print 'Load', self.which_set, 'data from', fname
if good_idx.size > 0:
# Features with shape [n_features, n_samples]
temp_X = self.load_list_feature(
list_features=self.list_features,
sample_idx=good_idx,
fname=fname,
patient_dir=patient_dir)
# If this record contains preictal data in the withheld seizures, get preictal labels
temp_y_withheld = self.get_labels(
label_type='preictals',
filename=fname,
good_idx=good_idx,
metadata_mat=metadata_mat,
n_all_segments=n_segments,
n_data_segments=temp_X.shape[1],
select_meta_idx=select_idx)
# If this record contains preictal data in the selected seizures, get preictal labels
temp_y_select = self.get_labels(
label_type='preictals',
filename=fname,
good_idx=good_idx,
metadata_mat=metadata_mat,
n_all_segments=n_segments,
n_data_segments=temp_X.shape[1])
# If this record contains preictal data in all seizures, get preictal labels
temp_y_rm = self.get_labels(
label_type='all_preictals',
filename=fname,
good_idx=good_idx,
metadata_mat=metadata_mat,
n_all_segments=n_segments,
n_data_segments=temp_X.shape[1])
tmp_preictal_withheld_idx = np.where(
temp_y_withheld == 1)[0]
tmp_preictal_select_idx = np.where(temp_y_select == 1)[0]
tmp_preictal_rm_idx = np.where(temp_y_rm == 1)[0]
tmp_preictal_select_idx = np.setdiff1d(
tmp_preictal_select_idx, tmp_preictal_withheld_idx)
tmp_preictal_rm_idx = np.setdiff1d(
tmp_preictal_rm_idx, tmp_preictal_withheld_idx)
tmp_preictal_rm_idx = np.setdiff1d(
tmp_preictal_rm_idx, tmp_preictal_select_idx)
self.preictal_samples = self.preictal_samples + np.where(
temp_y_withheld == 1)[0].size
self.nonictal_samples = self.nonictal_samples + np.where(
temp_y_withheld == 0)[0].size
if tmp_preictal_withheld_idx.size > 0:
print ' Load preictal data from the withheld seizure from this file.'
print ' Size:', tmp_preictal_withheld_idx.size, tmp_preictal_withheld_idx
if tmp_preictal_select_idx.size > 0:
print ' Load preictal data from selected seizures in addition to the withheld seizure from this file.'
print ' Size:', tmp_preictal_select_idx.size, tmp_preictal_select_idx
if tmp_preictal_rm_idx.size > 0:
print ' Load preictal data from removed seizures in addition to the withheld seizure from this file.'
print ' Size:', tmp_preictal_rm_idx.size, tmp_preictal_rm_idx
# Sanity check
if np.intersect1d(tmp_preictal_withheld_idx,
tmp_preictal_select_idx).size > 0:
raise Exception(
'There is an overlapped of the labels between the withheld seizures, and the selected seizures.'
)
if np.intersect1d(tmp_preictal_select_idx,
tmp_preictal_rm_idx).size > 0:
raise Exception(
'There is an overlapped of the labels between the selected seizures, and the removed seizures.'
)
if np.intersect1d(tmp_preictal_withheld_idx,
tmp_preictal_rm_idx).size > 0:
raise Exception(
'There is an overlapped of the labels between the withheld seizures, and the removed seizures.'
)
temp_y_all = np.zeros(temp_X.shape[1], dtype=int)
temp_y_all[
tmp_preictal_withheld_idx] = 1 # Labels for the withheld seizure
temp_y_all[
tmp_preictal_select_idx] = 2 # Labels for the selected seizure (that is not from withheld seizures)
temp_y_all[
tmp_preictal_rm_idx] = 3 # Labels for the removed seizure (that is not from withheld seizures)
# If this record contains ictal data, get ictal labels
temp_ictal_labels = self.get_labels(
label_type='all_ictals',
filename=fname,
good_idx=good_idx,
metadata_mat=metadata_mat,
n_all_segments=n_segments,
n_data_segments=temp_X.shape[1])
tmp_ictal_idx = np.where(temp_ictal_labels == 1)[0]
if tmp_ictal_idx.size > 0:
print ' Ictal label:', tmp_ictal_idx.size, tmp_ictal_idx
# Dealing with NaN features after filtering out flat segment which occurs due to noise in the data,
# not from flat segments
nan_sample_idx = np.where(np.isnan(np.sum(temp_X, 0)))[0]
nan_feature_idx = np.where(np.isnan(np.sum(temp_X, 1)))[0]
if nan_sample_idx.size > 0 or nan_feature_idx.size > 0:
print self.which_set, 'contains NaN at:'
print ' sample_idx:', good_idx[
nan_sample_idx], ' feature_idx:', nan_feature_idx
print ' shape before remove NaN:', temp_X.shape
tmp_preictal_idx = np.where(temp_y_withheld == 1)[0]
tmp_nonictal_idx = np.where(temp_y_withheld == 0)[0]
nan_preictal_sample_idx = np.intersect1d(
tmp_preictal_idx, nan_sample_idx)
nan_nonictal_sample_idx = np.intersect1d(
tmp_nonictal_idx, nan_sample_idx)
if nan_preictal_sample_idx.size > 0:
print ' NaN are in preictal index:', good_idx[
nan_preictal_sample_idx]
if nan_nonictal_sample_idx.size > 0:
print ' NaN are in nonictal index:', good_idx[
nan_nonictal_sample_idx]
all_idx = np.arange(temp_X.shape[1])
good_idx_1 = np.setdiff1d(all_idx, nan_sample_idx)
temp_X = temp_X[:, good_idx_1]
temp_y_all = temp_y_all[good_idx_1]
temp_y_withheld = temp_y_withheld[good_idx_1]
temp_ictal_labels = temp_ictal_labels[good_idx_1]
print ' shape before remove NaN:', temp_X.shape
self.nan_non_flat_samples = self.nan_non_flat_samples + nan_sample_idx.size
# Sanity check
tmp_nan_sample_idx = np.where(np.isnan(np.sum(temp_X,
0)))[0]
if tmp_nan_sample_idx.size > 0:
raise Exception('There is an error in removing NaN')
if not (temp_X.shape[1] == temp_y_all.size):
raise Exception(
'Number of feature data and labels [temp_y_all] are not equal.'
)
if not (temp_X.shape[1] == temp_y_withheld.size):
raise Exception(
'Number of feature data and labels [temp_y_withheld] are not equal.'
)
if not (temp_X.shape[1] == temp_ictal_labels.size):
raise Exception(
'Number of feature data and labels [ictal_labels] are not equal.'
)
if not (X is None) and not (y is None) and not (
ictal_labels is None):
X = np.concatenate((X, temp_X), axis=1)
y = np.append(y, temp_y_withheld)
y_label_all = np.append(y_label_all, temp_y_all)
ictal_labels = np.append(ictal_labels,
temp_ictal_labels)
else:
X = temp_X
y = temp_y_withheld
y_label_all = temp_y_all
ictal_labels = temp_ictal_labels
else:
print 'There is no good segment for during this seizure'
# Store preictal labels that are from the withheld index (use for compute accuracy), selected seizure index,
# and removed seizure index.
# Note: this property will exist when which_set=='valid' or which_set=='test'
# as there is no need for ictal to be imported.
self.y_label_all = y_label_all
# Sanity check
if np.where(y == 1)[0].size > np.where(y_label_all > 0)[0].size:
raise Exception(
'There is an error in collecting preictal labels only from the leave-out-seizure index.'
)
if np.where(y == 1)[0].size == np.where(y_label_all == 1)[0].size:
print 'There is only one preictal periods, and this period is from the leave-out-seizure index.'
if not np.all(
np.where(y == 1)[0] == np.where(y_label_all == 1)[0]):
raise Exception(
'There is a mismatch between y and y_label_all.')
if np.where(y == 1)[0].size < np.where(y_label_all > 0)[0].size:
print 'There are more than one preictal periods.'
if not np.all(
np.where(y == 1)[0] == np.where(y_label_all == 1)[0]):
raise Exception(
'There is a mismatch between y_select_idx and y in the preictal labels of the leave-out-seizure index.'
)
# Store ictal labels
# Note: this property will exist when which_set=='valid' or which_set=='test'
# as there is no need for ictal to be imported.
self.ictal_labels = ictal_labels
else:
raise Exception('Invalid dataset selection')
print 'There are {0} samples that have been removed in addition to the flat signal as due to NaN.'.format(
self.nan_non_flat_samples)
X = np.transpose(X, [1, 0])
one_hot_formatter = OneHotFormatter(max_labels=2)
y = one_hot_formatter.format(y)
# Sanity check
# Note: We ignore nan_non_flat_samples if we load shuffle data as we specify the labels after the NaN have been removed
# In contrast to loading continuous data, we specify the labels before removing NaN, so we have to remove the NaN samples for checking
if self.which_set == 'train' or self.which_set == 'valid_train':
if not (X.shape[0]
== self.preictal_samples + self.nonictal_samples):
raise Exception(
'There is a mismatch in the number of training samples ({0} != {1}).'
.format(X.shape[0],
self.preictal_samples + self.nonictal_samples))
if not (np.where(np.argmax(y, axis=1) == 1)[0].size
== self.preictal_samples):
raise Exception(
'There is a mismatch in the number of preictal samples and its labels ({0} != {1}).'
.format(
np.where(np.argmax(y, axis=1) == 1)[0].size,
self.preictal_samples))
if not (X.shape[0] == y.shape[0]):
raise Exception(
'There is a mismatch in the number of training samples and its labels ({0} != {1}).'
.format(X.shape[0], y.shape[0]))
elif self.which_set == 'valid' or self.which_set == 'test':
if not (X.shape[0] == self.preictal_samples +
self.nonictal_samples - self.nan_non_flat_samples):
raise Exception(
'There is a mismatch in the number of training samples ({0} != {1}).'
.format(
X.shape[0], self.preictal_samples +
self.nonictal_samples - self.nan_non_flat_samples))
if not ((np.where(np.argmax(y, axis=1) == 1)[0].size +
np.where(np.argmax(y, axis=1) == 0)[0].size)
== self.preictal_samples + self.nonictal_samples -
self.nan_non_flat_samples):
raise Exception(
'There is a mismatch in the number of samples and its labels ({0} != {1}).'
.format(
np.where(np.argmax(y, axis=1) == 1)[0].size +
np.where(np.argmax(y, axis=1) == 0)[0].size,
self.preictal_samples))
if not (X.shape[0] == y.shape[0]):
raise Exception(
'There is a mismatch in the number of training samples and its labels ({0} != {1}).'
.format(X.shape[0], y.shape[0]))
return X, y
def load_data(self, which_set, sample_size_second, batch_size, scaler_path):
raw_data, raw_labels, channel_labels, \
seizure_range_idx, seizure_range_second, seizure_seconds, \
n_channels, sample_size, sampling_rate = self.load_source_data(sample_size_second)
self.channel_labels = channel_labels
self.seizure_seconds_src = seizure_seconds
self.sampling_rate = sampling_rate
self.raw_data = raw_data
# Generate seiuzre index (rounded to be divided by the sampling rate)
seizure_round_sample_idx = np.empty(seizure_range_second.size, dtype=object)
for r in range(seizure_range_second.size):
start_idx = seizure_range_second[r][0] * sampling_rate
end_idx = seizure_range_second[r][-1] * sampling_rate
seizure_round_sample_idx[r] = np.arange(start_idx, end_idx)
# Generate non-seizure index
non_seizure_round_sample_idx = np.arange(raw_data.shape[1])
for s_idx in seizure_round_sample_idx:
non_seizure_round_sample_idx = np.setdiff1d(non_seizure_round_sample_idx,
s_idx)
# Partition non-seizure data into segments
# Then randomly choose for training, cv and test sets
n_segments = 10
segment_size = non_seizure_round_sample_idx.size / n_segments
segment_size = segment_size - (segment_size % sampling_rate)
segment_idx = np.empty(n_segments, dtype=object)
for i in range(n_segments):
start_segment_idx = i * segment_size
end_segment_idx = (i+1) * segment_size
if end_segment_idx > non_seizure_round_sample_idx.size:
end_segment_idx = non_seizure_round_sample_idx.size
segment_idx[i] = np.arange(start_segment_idx, end_segment_idx)
# Select test seizure index
test_seizure_idx = self.leave_one_out_seizure
np.random.seed(test_seizure_idx)
# Leave-one-out cross-validation - seizure
n_seizures = seizure_range_idx.shape[0]
rest_seizure_idx = np.setdiff1d(np.arange(n_seizures), test_seizure_idx)
perm_rest_seizure_idx = np.random.permutation(rest_seizure_idx)
train_seizure_idx = perm_rest_seizure_idx
cv_seizure_idx = perm_rest_seizure_idx
# Leave-one-out cross-validation - non-seizure
n_train_segments = int(n_segments * 0.6)
n_cv_segments = int(n_segments * 0.2)
non_seizure_segment_idx = np.arange(n_segments)
perm_non_seizure_segment_idx = np.random.permutation(non_seizure_segment_idx)
train_sample_segments = perm_non_seizure_segment_idx[:n_train_segments]
cv_sample_segments = perm_non_seizure_segment_idx[n_train_segments:n_train_segments+n_cv_segments]
test_sample_segments = perm_non_seizure_segment_idx[n_train_segments+n_cv_segments:]
train_sample_idx = np.empty(0, dtype=int)
for s in train_sample_segments:
train_sample_idx = np.append(train_sample_idx, segment_idx[s])
cv_sample_idx = np.empty(0, dtype=int)
for s in cv_sample_segments:
cv_sample_idx = np.append(cv_sample_idx, segment_idx[s])
test_sample_idx = np.empty(0, dtype=int)
for s in test_sample_segments:
test_sample_idx = np.append(test_sample_idx, segment_idx[s])
print 'Segment index for train, cv and test sets:', \
train_sample_segments, cv_sample_segments, test_sample_segments
print 'Seizure index for train, cv and test sets:', \
train_seizure_idx, cv_seizure_idx, [test_seizure_idx]
if which_set == 'train':
print("Loading training data...")
data = raw_data[:,non_seizure_round_sample_idx[train_sample_idx]]
labels = raw_labels[non_seizure_round_sample_idx[train_sample_idx]]
select_seizure = train_seizure_idx
elif which_set == 'valid':
print("Loading validation data...")
data = raw_data[:,non_seizure_round_sample_idx[cv_sample_idx]]
labels = raw_labels[non_seizure_round_sample_idx[cv_sample_idx]]
select_seizure = cv_seizure_idx
elif which_set == 'test':
print("Loading test data...")
data = raw_data[:,non_seizure_round_sample_idx[test_sample_idx]]
labels = raw_labels[non_seizure_round_sample_idx[test_sample_idx]]
select_seizure = [test_seizure_idx]
elif which_set == 'all':
print("Loading all data...")
data = raw_data
labels = raw_labels
select_seizure = []
else:
raise('Invalid set.')
# Add seizure data
for sz in select_seizure:
data = np.concatenate((data, raw_data[:, seizure_round_sample_idx[sz]]), axis=1)
labels = np.concatenate((labels, raw_labels[seizure_round_sample_idx[sz]]), axis=1)
# No filtering
# Preprocessing
if which_set == 'train':
scaler = preprocessing.StandardScaler()
scaler = scaler.fit(data.transpose())
with open(scaler_path, 'w') as f:
pickle.dump(scaler, f)
data = scaler.transform(data.transpose()).transpose()
else:
with open(scaler_path) as f:
scaler = pickle.load(f)
data = scaler.transform(data.transpose()).transpose()
# Input transformation
X = np.reshape(data, (-1, sample_size))
y = np.reshape(labels, (-1, sample_size))
y = np.sum(y, 1).transpose()
y[y > 0] = 1
print 'Seizure index after transform:', np.where(y)[0]
self.seizure_seconds = np.where(y)[0]
# Duplicate the labels for all channels
y = np.tile(y, n_channels)
# Format the target into proper format
n_classes = 2
one_hot_formatter = OneHotFormatter(max_labels=n_classes)
y = one_hot_formatter.format(y)
# Check batch size
cut_off = X.shape[0] % batch_size
if cut_off > 0:
X = X[:-cut_off,:]
y = y[:-cut_off,:]
return X, y, n_channels, sample_size
def _build_frames_w_phn(dataset, subset, wav_seqs, seqs_to_phns,
in_samples, out_samples, shift,
win_width, shuffle):
#import pdb; pdb.set_trace()
norm_seqs = utils.standardize(wav_seqs)
#norm_seqs = utils.normalize(wav_seqs)
frame_len = in_samples + out_samples
overlap = frame_len - shift
samples = []
seqs_phn_info = []
seqs_phn_shift = []
# CAUTION!: I am using here reduced phone set
# we can also try using the full set but we must store phn+1
# because 0 no more refers to 'h#' (no speech)
for ind in range(len(norm_seqs)):
#import pdb; pdb.set_trace()
wav_seq = norm_seqs[ind]
phn_seq = seqs_to_phns[ind]
phn_start_end = dataset.__dict__[subset+"_phn"][phn_seq[0]:phn_seq[1]]
# create a matrix with consecutive windows
# phones are padded by h#, because each window will be shifted once
# the first phone samples has passed
phones = np.append(phn_start_end[:,2].astype('int16'),
np.zeros((1,),dtype='int16'))
# phones = np.append(phn_start_end[:,2],
# np.zeros((1,)))
phn_windows = segment_axis(phones, win_width, win_width-1)
# array that has endings of each phone
phn_ends = phn_start_end[:,1]
# extend the last phone till the end, this is not wrong as long as the
# last phone is no speech phone (h#)
phn_ends[-1] = wav_seq.shape[0]-1
# create a mapping from each sample to phn_window
phn_win_shift = np.zeros_like(wav_seq,dtype='int16')
phn_win_shift[phn_ends] = 1
phn_win = phn_win_shift.cumsum(dtype='int16')
# minor correction!
phn_win[-1] = phn_win[-2]
# Segment samples into frames
samples.append(segment_axis(wav_seq, frame_len, overlap))
# for phones we care only about one value to mark the start of a new window.
# the start of a phone window in a frame is when all samples of previous
# phone hav passed, so we use 'min' function to choose the current phone
# of the frame
phn_frames = segment_axis(phn_win, frame_len, overlap).min(axis=1)
# replace the window index with the window itself
win_frames = phn_windows[phn_frames]
seqs_phn_info.append(win_frames)
#import pdb; pdb.set_trace()
# create a window shift for each frame
shift_frames_aux = np.roll(phn_frames,1)
shift_frames_aux[0] = 0
shift_frames = phn_frames - shift_frames_aux
# to mark the ending of the sequence - countering the first correction!
shift_frames[-1] = 1
seqs_phn_shift.append(shift_frames)
#import pdb; pdb.set_trace()
#import pdb; pdb.set_trace()
# stack all data in one matrix, each row is a frame
samples_data = np.vstack(samples[:])
phn_data = np.vstack(seqs_phn_info[:])
shift_data = np.hstack(seqs_phn_shift[:])
#convert phone data to one-hot
from pylearn2.format.target_format import OneHotFormatter
fmt = OneHotFormatter(max_labels=39, dtype='float32')
phn_data = fmt.format(phn_data)
phn_data = phn_data.reshape(phn_data.shape[0],
phn_data.shape[1]*phn_data.shape[2])
full_data = np.hstack([samples_data[:,:in_samples], phn_data, #input
samples_data[:,in_samples:], #out1
shift_data.reshape(shift_data.shape[0],1)]) #out2
if shuffle:
np.random.seed(123)
full_data = np.random.permutation(full_data)
data_x = full_data[:,:in_samples+win_width*39]
data_y1 = full_data[:,in_samples+win_width*39:-1]
data_y2 = full_data[:,-1]
print 'Done'
print 'There are %d examples in %s set'%(data_x.shape[0],subset)
print "--------------"
print 'data_x.shape', data_x.shape
print 'data_y1.shape', data_y1.shape
return utils.shared_dataset(data_x), \
utils.shared_dataset(data_y1),\
utils.shared_dataset(data_y2)
import numpy
from pylearn2_timit.timitlpc import TIMITlpc
from pylearn2.space import CompositeSpace, VectorSpace, IndexSpace
from pylearn2.format.target_format import OneHotFormatter
valid = TIMITlpc("valid", frame_length=160, overlap=159, start=10, stop=11)
valid._iter_data_specs = (CompositeSpace((IndexSpace(dim=3,max_labels=61), VectorSpace(dim=10),)), ('phones', 'lpc_features'))
formatter = OneHotFormatter(max_labels=62)
f = lambda x: formatter.format(numpy.asarray(x, dtype=int), mode='merge')
#valid._iter_convert = [f, None]
it = valid.iterator(mode='random_uniform', batch_size=100, num_batches=100)
def load_data(self, which_set, sample_size_second, batch_size,
scaler_path):
raw_data, raw_labels, channel_labels, \
seizure_range_idx, seizure_range_second, seizure_seconds, \
n_channels, sample_size, sampling_rate = self.load_source_data(sample_size_second)
self.channel_labels = channel_labels
self.seizure_seconds_src = seizure_seconds
self.sampling_rate = sampling_rate
self.raw_data = raw_data
# Generate seiuzre index (rounded to be divided by the sampling rate)
seizure_round_sample_idx = np.empty(seizure_range_second.size,
dtype=object)
for r in range(seizure_range_second.size):
start_idx = seizure_range_second[r][0] * sampling_rate
end_idx = seizure_range_second[r][-1] * sampling_rate
seizure_round_sample_idx[r] = np.arange(start_idx, end_idx)
# Generate non-seizure index
non_seizure_round_sample_idx = np.arange(raw_data.shape[1])
for s_idx in seizure_round_sample_idx:
non_seizure_round_sample_idx = np.setdiff1d(
non_seizure_round_sample_idx, s_idx)
# Partition non-seizure data into segments
# Then randomly choose for training, cv and test sets
n_segments = 10
segment_size = non_seizure_round_sample_idx.size / n_segments
segment_size = segment_size - (segment_size % sampling_rate)
segment_idx = np.empty(n_segments, dtype=object)
for i in range(n_segments):
start_segment_idx = i * segment_size
end_segment_idx = (i + 1) * segment_size
if end_segment_idx > non_seizure_round_sample_idx.size:
end_segment_idx = non_seizure_round_sample_idx.size
segment_idx[i] = np.arange(start_segment_idx, end_segment_idx)
# Select test seizure index
test_seizure_idx = self.leave_one_out_seizure
np.random.seed(test_seizure_idx)
# Leave-one-out cross-validation - seizure
n_seizures = seizure_range_idx.shape[0]
rest_seizure_idx = np.setdiff1d(np.arange(n_seizures),
test_seizure_idx)
perm_rest_seizure_idx = np.random.permutation(rest_seizure_idx)
train_seizure_idx = perm_rest_seizure_idx
cv_seizure_idx = perm_rest_seizure_idx
# Leave-one-out cross-validation - non-seizure
n_train_segments = int(n_segments * 0.6)
n_cv_segments = int(n_segments * 0.2)
non_seizure_segment_idx = np.arange(n_segments)
perm_non_seizure_segment_idx = np.random.permutation(
non_seizure_segment_idx)
train_sample_segments = perm_non_seizure_segment_idx[:n_train_segments]
cv_sample_segments = perm_non_seizure_segment_idx[
n_train_segments:n_train_segments + n_cv_segments]
test_sample_segments = perm_non_seizure_segment_idx[n_train_segments +
n_cv_segments:]
train_sample_idx = np.empty(0, dtype=int)
for s in train_sample_segments:
train_sample_idx = np.append(train_sample_idx, segment_idx[s])
cv_sample_idx = np.empty(0, dtype=int)
for s in cv_sample_segments:
cv_sample_idx = np.append(cv_sample_idx, segment_idx[s])
test_sample_idx = np.empty(0, dtype=int)
for s in test_sample_segments:
test_sample_idx = np.append(test_sample_idx, segment_idx[s])
print 'Segment index for train, cv and test sets:', \
train_sample_segments, cv_sample_segments, test_sample_segments
print 'Seizure index for train, cv and test sets:', \
train_seizure_idx, cv_seizure_idx, [test_seizure_idx]
if which_set == 'train':
print("Loading training data...")
data = raw_data[:, non_seizure_round_sample_idx[train_sample_idx]]
labels = raw_labels[non_seizure_round_sample_idx[train_sample_idx]]
select_seizure = train_seizure_idx
elif which_set == 'valid':
print("Loading validation data...")
data = raw_data[:, non_seizure_round_sample_idx[cv_sample_idx]]
labels = raw_labels[non_seizure_round_sample_idx[cv_sample_idx]]
select_seizure = cv_seizure_idx
elif which_set == 'test':
print("Loading test data...")
data = raw_data[:, non_seizure_round_sample_idx[test_sample_idx]]
labels = raw_labels[non_seizure_round_sample_idx[test_sample_idx]]
select_seizure = [test_seizure_idx]
elif which_set == 'all':
print("Loading all data...")
data = raw_data
labels = raw_labels
select_seizure = []
else:
raise ('Invalid set.')
# Add seizure data
for sz in select_seizure:
data = np.concatenate(
(data, raw_data[:, seizure_round_sample_idx[sz]]), axis=1)
labels = np.concatenate(
(labels, raw_labels[seizure_round_sample_idx[sz]]), axis=1)
# No filtering
# Preprocessing
if which_set == 'train':
scaler = preprocessing.StandardScaler()
scaler = scaler.fit(data.transpose())
with open(scaler_path, 'w') as f:
pickle.dump(scaler, f)
data = scaler.transform(data.transpose()).transpose()
else:
with open(scaler_path) as f:
scaler = pickle.load(f)
data = scaler.transform(data.transpose()).transpose()
# Input transformation
X = np.reshape(data, (-1, sample_size))
y = np.reshape(labels, (-1, sample_size))
y = np.sum(y, 1).transpose()
y[y > 0] = 1
print 'Seizure index after transform:', np.where(y)[0]
self.seizure_seconds = np.where(y)[0]
# Duplicate the labels for all channels
y = np.tile(y, n_channels)
# Format the target into proper format
n_classes = 2
one_hot_formatter = OneHotFormatter(max_labels=n_classes)
y = one_hot_formatter.format(y)
# Check batch size
cut_off = X.shape[0] % batch_size
if cut_off > 0:
X = X[:-cut_off, :]
y = y[:-cut_off, :]
return X, y, n_channels, sample_size
def __init__(self,
which_set,
onehot_dtype='uint8',
center=False,
rescale=False,
gcn=None,
start=None,
stop=None,
axes=('b', 0, 1, 'c'),
toronto_prepro=False,
preprocessor=None):
"""Modified version of the CIFAR10 constructor which creates Y
as one-hot vectors rather than simple indexes. This is super
hacky. Sorry, Guido.."""
# note: there is no such thing as the cifar10 validation set;
# pylearn1 defined one but really it should be user-configurable
# (as it is here)
self.axes = axes
# we define here:
dtype = 'uint8'
ntrain = 50000
nvalid = 0 # artefact, we won't use it
ntest = 10000
# we also expose the following details:
self.img_shape = (3, 32, 32)
self.img_size = numpy.prod(self.img_shape)
self.n_classes = 10
self.label_names = [
'airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog',
'horse', 'ship', 'truck'
]
# prepare loading
fnames = ['data_batch_%i' % i for i in range(1, 6)]
datasets = {}
datapath = os.path.join(
string_utils.preprocess('${PYLEARN2_DATA_PATH}'), 'cifar10',
'cifar-10-batches-py')
for name in fnames + ['test_batch']:
fname = os.path.join(datapath, name)
if not os.path.exists(fname):
raise IOError(fname + " was not found. You probably need to "
"download the CIFAR-10 dataset by using the "
"download script in "
"pylearn2/scripts/datasets/download_cifar10.sh "
"or manually from "
"http://www.cs.utoronto.ca/~kriz/cifar.html")
datasets[name] = cache.datasetCache.cache_file(fname)
lenx = numpy.ceil((ntrain + nvalid) / 10000.) * 10000
x = numpy.zeros((lenx, self.img_size), dtype=dtype)
y = numpy.zeros((lenx, 1), dtype=dtype)
# load train data
nloaded = 0
for i, fname in enumerate(fnames):
_logger.info('loading file %s' % datasets[fname])
data = serial.load(datasets[fname])
x[i * 10000:(i + 1) * 10000, :] = data['data']
y[i * 10000:(i + 1) * 10000, 0] = data['labels']
nloaded += 10000
if nloaded >= ntrain + nvalid + ntest:
break
# load test data
_logger.info('loading file %s' % datasets['test_batch'])
data = serial.load(datasets['test_batch'])
# process this data
Xs = {'train': x[0:ntrain], 'test': data['data'][0:ntest]}
Ys = {'train': y[0:ntrain], 'test': data['labels'][0:ntest]}
X = numpy.cast['float32'](Xs[which_set])
y = Ys[which_set]
if isinstance(y, list):
y = numpy.asarray(y).astype(dtype)
if which_set == 'test':
assert y.shape[0] == 10000
y = y.reshape((y.shape[0], 1))
formatter = OneHotFormatter(self.n_classes, dtype=onehot_dtype)
y = formatter.format(y, mode='concatenate')
if center:
X -= 127.5
self.center = center
if rescale:
X /= 127.5
self.rescale = rescale
if toronto_prepro:
assert not center
assert not gcn
X = X / 255.
if which_set == 'test':
other = CIFAR10(which_set='train')
oX = other.X
oX /= 255.
X = X - oX.mean(axis=0)
else:
X = X - X.mean(axis=0)
self.toronto_prepro = toronto_prepro
self.gcn = gcn
if gcn is not None:
gcn = float(gcn)
X = global_contrast_normalize(X, scale=gcn)
if start is not None:
# This needs to come after the prepro so that it doesn't
# change the pixel means computed above for toronto_prepro
assert start >= 0
assert stop > start
assert stop <= X.shape[0]
X = X[start:stop, :]
y = y[start:stop, :]
assert X.shape[0] == y.shape[0]
if which_set == 'test':
assert X.shape[0] == 10000
view_converter = dense_design_matrix.DefaultViewConverter((32, 32, 3),
axes)
super(CIFAR10, self).__init__(
X=X,
y=y,
view_converter=view_converter,
) #y_labels=self.n_classes)
assert not contains_nan(self.X)
if preprocessor:
preprocessor.apply(self)
# Another hack: rename 'targets' to match model expectations
space, (X_source, y_source) = self.data_specs
self.data_specs = (space, (X_source, 'condition'))
def __init__(self,
db, # data source
name = '', # optional name
selectors = dict(),
partitioner = None,
meta_sources = [], # optional sources other than 'features' and 'targets' from metadata
channel_filter = NoChannelFilter(), # optional channel filter, default: keep all
channel_names = None, # optional channel names (for metadata)
label_attribute = 'label', # metadata attribute to be used as label
label_map = None, # optional conversion of labels
use_targets = True, # use targets if provides, otherwise labels are used
remove_dc_offset = False, # optional subtraction of channel mean, usually done already earlier
resample = None, # optional down-sampling
normalize = True, # normalize to max=1
# optional sub-sequences selection
start_sample = 0,
stop_sample = None, # optional for selection of sub-sequences
zero_padding = True, # if True (default) trials that are too short will be padded with
# otherwise they will rejected.
# optional signal filter to by applied before splitting the signal
signal_filter = None,
trial_processors = [], # optional processing of the trials
target_processor = None, # optional processing of the targets, e.g. zero-padding
transformers = [], # optional transformations of the dataset
layout='tf', # (0,1)-axes layout tf=time x features or ft=features x time
debug=False,
):
'''
Constructor
'''
# save params
self.params = locals().copy()
del self.params['self']
# print self.params
self.name = name
self.debug = debug
metadb = DatasetMetaDB(db.metadata, selectors.keys())
if partitioner is not None:
pass # FIXME
selected_trial_ids = metadb.select(selectors)
log.info('selectors: {}'.format(selectors))
log.info('selected trials: {}'.format(selected_trial_ids))
if normalize:
log.info('Data will be normalized to max amplitude 1 per channel (normalize=True).')
trials = list()
labels = list()
targets = list()
meta = list()
if stop_sample == 'auto-min':
stop_sample = np.min([db.data[trial_i].shape[-1] for trial_i in selected_trial_ids])
log.info('Using minimum trial length. stop_sample={}'.format(stop_sample))
elif stop_sample == 'auto-max':
stop_sample = np.max([db.data[trial_i].shape[-1] for trial_i in selected_trial_ids])
log.info('Using maximum trial length. stop_sample={}'.format(stop_sample))
for trial_i in selected_trial_ids:
trial_meta = db.metadata[trial_i]
if use_targets:
if targets is None:
target = None
else:
target = db.targets[trial_i]
assert not np.isnan(np.sum(target))
if target_processor is not None:
target = target_processor.process(target, trial_meta)
assert not np.isnan(np.sum(target))
else:
# get and process label
label = db.metadata[trial_i][label_attribute]
if label_map is not None:
label = label_map[label]
processed_trial = []
trial = db.data[trial_i]
if np.isnan(np.sum(trial)):
print trial_i, trial
assert not np.isnan(np.sum(trial))
rejected = False # flag for trial rejection
trial = np.atleast_2d(trial)
# process 1 channel at a time
for channel in xrange(trial.shape[0]):
# filter channels
if not channel_filter.keep_channel(channel):
continue
samples = trial[channel, :]
# subtract channel mean
if remove_dc_offset:
samples -= samples.mean()
# down-sample if requested
if resample is not None and resample[0] != resample[1]:
samples = librosa.resample(samples, resample[0], resample[1], res_type='sinc_best')
# apply optional signal filter after down-sampling -> requires lower order
if signal_filter is not None:
samples = signal_filter.process(samples)
# get sub-sequence in resampled space
# log.info('using samples {}..{} of {}'.format(start_sample,stop_sample, samples.shape))
if stop_sample is not None and stop_sample > len(samples):
if zero_padding:
tmp = np.zeros(stop_sample)
tmp[:len(samples)] = samples
samples = tmp
else:
rejected = True
break # stop processing this trial
s = samples[start_sample:stop_sample]
# TODO optional channel processing
# normalize to max amplitude 1
if normalize:
s = librosa.util.normalize(s)
# add 2nd data dimension
s = s.reshape(s.shape[0], 1)
# print s.shape
s = np.asfarray(s, dtype=theano.config.floatX)
processed_trial.append(s)
### end of channel iteration ###
if rejected:
continue # next trial
processed_trial = np.asfarray([processed_trial], dtype=theano.config.floatX)
# processed_trial = processed_trial.reshape((1, processed_trial.shape))
processed_trial = np.rollaxis(processed_trial, 1, 4)
# optional (external) trial processing, e.g. windowing
# trials will be in b01c format with tf layout for 01-axes
for trial_processor in trial_processors:
processed_trial = trial_processor.process(processed_trial, trial_meta)
trials.append(processed_trial)
for k in range(len(processed_trial)):
meta.append(trial_meta)
if use_targets:
targets.append(target)
else:
labels.append(label)
### end of datafile iteration ###
# turn into numpy arrays
self.trials = np.vstack(trials)
assert not np.isnan(np.sum(self.trials))
# prepare targets / labels
if use_targets:
self.targets = np.vstack(targets)
assert not np.isnan(np.sum(self.targets))
else:
labels = np.hstack(labels)
if label_map is None:
one_hot_formatter = OneHotFormatter(max(labels) + 1)
else:
one_hot_formatter = OneHotFormatter(max(label_map.values()) + 1)
one_hot_y = one_hot_formatter.format(labels)
self.targets = one_hot_y
self.metadata = meta
if layout == 'ft': # swap axes to (batch, feature, time, channels)
self.trials = self.trials.swapaxes(1, 2)
# transform after finalizing the data structure
for transformer in transformers:
self.trials, self.targets = transformer.process(self.trials, self.targets)
self.trials = np.asarray(self.trials, dtype=theano.config.floatX)
log.debug('final dataset shape: {} (b,0,1,c)'.format(self.trials.shape))
# super(EEGEpochsDataset, self).__init__(topo_view=self.trials, y=self.targets, axes=['b', 0, 1, 'c'])
self.X = self.trials.reshape(self.trials.shape[0], np.prod(self.trials.shape[1:]))
self.y = self.targets
log.info('generated dataset "{}" with shape X={}={} y={} targets={} '.
format(self.name, self.X.shape, self.trials.shape, self.y.shape, self.targets.shape))
# determine data specs
features_space = Conv2DSpace(
shape=[self.trials.shape[1], self.trials.shape[2]],
num_channels=self.trials.shape[3]
)
features_source = 'features'
targets_space = VectorSpace(dim=self.targets.shape[-1])
targets_source = 'targets'
space_components = [features_space, targets_space]
source_components = [features_source, targets_source]
# additional support for meta information
self.meta_maps = dict()
for meta_source in meta_sources:
self.meta_maps[meta_source] = sorted(list(set([m[meta_source] for m in self.metadata])))
space_components.extend([VectorSpace(dim=1)])
source_components.extend([meta_source])
log.info('Generated meta-source "{}" with value map: {}'
.format(meta_source, self.meta_maps[meta_source]))
space = CompositeSpace(space_components)
source = tuple(source_components)
self.data_specs = (space, source)
log.debug('data specs: {}'.format(self.data_specs))
def __init__(self,
path,
name = '', # optional name
# selectors
subjects='all', # optional selector (list) or 'all'
trial_types='all', # optional selector (list) or 'all'
trial_numbers='all', # optional selector (list) or 'all'
conditions='all', # optional selector (list) or 'all'
partitioner = None,
channel_filter = NoChannelFilter(), # optional channel filter, default: keep all
channel_names = None, # optional channel names (for metadata)
label_map = None, # optional conversion of labels
remove_dc_offset = False, # optional subtraction of channel mean, usually done already earlier
resample = None, # optional down-sampling
# optional sub-sequences selection
start_sample = 0,
stop_sample = None, # optional for selection of sub-sequences
# optional signal filter to by applied before spitting the signal
signal_filter = None,
# windowing parameters
frame_size = -1,
hop_size = -1, # values > 0 will lead to windowing
hop_fraction = None, # alternative to specifying absolute hop_size
# optional spectrum parameters, n_fft = 0 keeps raw data
n_fft = 0,
n_freq_bins = None,
spectrum_log_amplitude = False,
spectrum_normalization_mode = None,
include_phase = False,
flatten_channels=False,
layout='tf', # (0,1)-axes layout tf=time x features or ft=features x time
save_matrix_path = None,
keep_metadata = False,
):
'''
Constructor
'''
# save params
self.params = locals().copy()
del self.params['self']
# print self.params
# TODO: get the whole filtering into an extra class
datafiles_metadata, metadb = load_datafiles_metadata(path)
# print datafiles_metadata
def apply_filters(filters, node):
if isinstance(node, dict):
filtered = []
keepkeys = filters[0]
for key, value in node.items():
if keepkeys == 'all' or key in keepkeys:
filtered.extend(apply_filters(filters[1:], value))
return filtered
else:
return node # [node]
# keep only files that match the metadata filters
self.datafiles = apply_filters([subjects,trial_types,trial_numbers,conditions], datafiles_metadata)
# copy metadata for retained files
self.metadb = {}
for datafile in self.datafiles:
self.metadb[datafile] = metadb[datafile]
# print self.datafiles
# print self.metadb
self.name = name
if partitioner is not None:
self.datafiles = partitioner.get_partition(self.name, self.metadb)
self.include_phase = include_phase
self.spectrum_normalization_mode = spectrum_normalization_mode
self.spectrum_log_amplitude = spectrum_log_amplitude
self.sequence_partitions = [] # used to keep track of original sequences
# metadata: [subject, trial_no, stimulus, channel, start, ]
self.metadata = []
sequences = []
labels = []
n_sequences = 0
if frame_size > 0 and hop_size == -1 and hop_fraction is not None:
hop_size = np.ceil(frame_size / hop_fraction)
for i in xrange(len(self.datafiles)):
with log_timing(log, 'loading data from {}'.format(self.datafiles[i])):
# save start of next sequence
self.sequence_partitions.append(n_sequences)
data, metadata = load(os.path.join(path, self.datafiles[i]))
label = metadata['label']
if label_map is not None:
label = label_map[label]
multi_channel_frames = []
# process 1 channel at a time
for channel in xrange(data.shape[1]):
# filter channels
if not channel_filter.keep_channel(channel):
continue
samples = data[:, channel]
# subtract channel mean
if remove_dc_offset:
samples -= samples.mean()
# down-sample if requested
if resample is not None and resample[0] != resample[1]:
samples = librosa.resample(samples, resample[0], resample[1])
# apply optional signal filter after down-sampling -> requires lower order
if signal_filter is not None:
samples = signal_filter.process(samples)
# get sub-sequence in resampled space
# log.info('using samples {}..{} of {}'.format(start_sample,stop_sample, samples.shape))
samples = samples[start_sample:stop_sample]
if n_fft is not None and n_fft > 0: # Optionally:
### frequency spectrum branch ###
# transform to spectogram
hop_length = n_fft / 4;
'''
from http://theremin.ucsd.edu/~bmcfee/librosadoc/librosa.html
>>> # Get a power spectrogram from a waveform y
>>> S = np.abs(librosa.stft(y)) ** 2
>>> log_S = librosa.logamplitude(S)
'''
S = librosa.core.stft(samples, n_fft=n_fft, hop_length=hop_length)
# mag = np.abs(S) # magnitude spectrum
mag = np.abs(S)**2 # power spectrum
# include phase information if requested
if self.include_phase:
# phase = np.unwrap(np.angle(S))
phase = np.angle(S)
# Optionally: cut off high bands
if n_freq_bins is not None:
mag = mag[0:n_freq_bins, :]
if self.include_phase:
phase = phase[0:n_freq_bins, :]
if self.spectrum_log_amplitude:
mag = librosa.logamplitude(mag)
s = mag # for normalization
'''
NOTE on normalization:
It depends on the structure of a neural network and (even more)
on the properties of data. There is no best normalization algorithm
because if there would be one, it would be used everywhere by default...
In theory, there is no requirement for the data to be normalized at all.
This is a purely practical thing because in practice convergence could
take forever if your input is spread out too much. The simplest would be
to just normalize it by scaling your data to (-1,1) (or (0,1) depending
on activation function), and in most cases it does work. If your
algorithm converges well, then this is your answer. If not, there are
too many possible problems and methods to outline here without knowing
the actual data.
'''
## normalize to mean 0, std 1
if self.spectrum_normalization_mode == 'mean0_std1':
# s = preprocessing.scale(s, axis=0);
mean = np.mean(s)
std = np.std(s)
s = (s - mean) / std
## normalize by linear transform to [0,1]
elif self.spectrum_normalization_mode == 'linear_0_1':
s = s / np.max(s)
## normalize by linear transform to [-1,1]
elif self.spectrum_normalization_mode == 'linear_-1_1':
s = -1 + 2 * (s - np.min(s)) / (np.max(s) - np.min(s))
elif self.spectrum_normalization_mode is not None:
raise ValueError(
'unsupported spectrum normalization mode {}'.format(
self.spectrum_normalization_mode)
)
#print s.mean(axis=0)
#print s.std(axis=0)
# include phase information if requested
if self.include_phase:
# normalize phase to [-1.1]
phase = phase / np.pi
s = np.vstack([s, phase])
# transpose to fit pylearn2 layout
s = np.transpose(s)
# print s.shape
### end of frequency spectrum branch ###
else:
### raw waveform branch ###
# normalize to max amplitude 1
s = librosa.util.normalize(samples)
# add 2nd data dimension
s = s.reshape(s.shape[0], 1)
# print s.shape
### end of raw waveform branch ###
s = np.asfarray(s, dtype='float32')
if frame_size > 0 and hop_size > 0:
s = s.copy() # FIXME: THIS IS NECESSARY IN MultiChannelEEGSequencesDataset - OTHERWISE, THE FOLLOWING OP DOES NOT WORK!!!!
frames = frame(s, frame_length=frame_size, hop_length=hop_size)
else:
frames = s
del s
# print frames.shape
if flatten_channels:
# add artificial channel dimension
frames = frames.reshape((frames.shape[0], frames.shape[1], frames.shape[2], 1))
# print frames.shape
sequences.append(frames)
# increment counter by new number of frames
n_sequences += frames.shape[0]
if keep_metadata:
# determine channel name
channel_name = None
if channel_names is not None:
channel_name = channel_names[channel]
elif 'channels' in metadata:
channel_name = metadata['channels'][channel]
self.metadata.append({
'subject' : metadata['subject'], # subject
'trial_type': metadata['trial_type'], # trial_type
'trial_no' : metadata['trial_no'], # trial_no
'condition' : metadata['condition'], # condition
'channel' : channel, # channel
'channel_name' : channel_name,
'start' : self.sequence_partitions[-1], # start
'stop' : n_sequences # stop
})
for _ in xrange(frames.shape[0]):
labels.append(label)
else:
multi_channel_frames.append(frames)
### end of channel iteration ###
if not flatten_channels:
# turn list into array
multi_channel_frames = np.asfarray(multi_channel_frames, dtype='float32')
# [channels x frames x time x freq] -> cb01
# [channels x frames x time x 1] -> cb0.
# move channel dimension to end
multi_channel_frames = np.rollaxis(multi_channel_frames, 0, 4)
# print multi_channel_frames.shape
# log.debug(multi_channel_frames.shape)
sequences.append(multi_channel_frames)
# increment counter by new number of frames
n_sequences += multi_channel_frames.shape[0]
if keep_metadata:
self.metadata.append({
'subject' : metadata['subject'], # subject
'trial_type': metadata['trial_type'], # trial_type
'trial_no' : metadata['trial_no'], # trial_no
'condition' : metadata['condition'], # condition
'channel' : 'all', # channel
'start' : self.sequence_partitions[-1], # start
'stop' : n_sequences # stop
})
for _ in xrange(multi_channel_frames.shape[0]):
labels.append(label)
### end of datafile iteration ###
# turn into numpy arrays
sequences = np.vstack(sequences)
# print sequences.shape;
labels = np.hstack(labels)
# one_hot_y = one_hot(labels)
one_hot_formatter = OneHotFormatter(labels.max() + 1) # FIXME!
one_hot_y = one_hot_formatter.format(labels)
self.labels = labels
if layout == 'ft': # swap axes to (batch, feature, time, channels)
sequences = sequences.swapaxes(1, 2)
log.debug('final dataset shape: {} (b,0,1,c)'.format(sequences.shape))
super(MultiChannelEEGDataset, self).__init__(topo_view=sequences, y=one_hot_y, axes=['b', 0, 1, 'c'])
log.info('generated dataset "{}" with shape X={}={} y={} labels={} '.
format(self.name, self.X.shape, sequences.shape, self.y.shape, self.labels.shape))
if save_matrix_path is not None:
matrix = DenseDesignMatrix(topo_view=sequences, y=one_hot_y, axes=['b', 0, 1, 'c'])
with log_timing(log, 'saving DenseDesignMatrix to {}'.format(save_matrix_path)):
serial.save(save_matrix_path, matrix)
class IndexSpace(Space):
"""
A space representing indices, for example MNIST labels (0-10) or the
indices of words in a dictionary for NLP tasks. A single space can
contain multiple indices, for example the word indices of an n-gram.
IndexSpaces can be converted to VectorSpaces in two ways: Either the
labels are converted into one-hot vectors which are then concatenated,
or they are converted into a single vector where 1s indicate labels
present i.e. for 4 possible labels we have [0, 2] -> [1 0 1 0] or
[0, 2] -> [1 0 0 0 0 0 1 0].
"""
def __init__(self, max_labels, dim, **kwargs):
"""
Initialize an IndexSpace.
Parameters
----------
max_labels : int
The number of possible classes/labels. This means that
all labels should be < max_labels. Example: For MNIST
there are 10 numbers and hence max_labels = 10.
dim : int
The number of indices in one space e.g. for MNIST there is
one target label and hence dim = 1. If we have an n-gram
of word indices as input to a neurel net language model, dim = n.
kwargs: passes on to superclass constructor
"""
super(IndexSpace, self).__init__(**kwargs)
self.max_labels = max_labels
self.dim = dim
self.formatter = OneHotFormatter(self.max_labels)
def __str__(self):
"""
Return a string representation.
"""
return '%(classname)s(dim=%(dim)s, max_labels=%(max_labels)s' % \
dict(classname=self.__class__.__name__,
dim=self.dim,
max_labels=self.max_labels)
@functools.wraps(Space.get_total_dimension)
def get_total_dimension(self):
return self.dim
@functools.wraps(Space.np_format_as)
def np_format_as(self, batch, space):
if isinstance(space, VectorSpace):
if self.max_labels == space.dim:
rval = self.formatter.format(batch, sparse=space.sparse,
mode='merge')
elif self.dim * self.max_labels == space.dim:
rval = self.formatter.format(batch, sparse=space.sparse,
mode='concatenate')
else:
raise ValueError("Can't convert IndexSpace to"
"VectorSpace (%d labels to %d dimensions)"
% (self.dim, space.dim))
return rval
else:
raise ValueError("Can't convert IndexSpace to %(space)s"
% (space.__class__.__name__))
@functools.wraps(Space._format_as)
def _format_as(self, batch, space):
"""
Supports formatting to a VectorSpace where indices are represented
by ones in a binary vector.
"""
if isinstance(space, VectorSpace):
if self.max_labels == space.dim:
rval = self.formatter.theano_expr(batch, sparse=space.sparse,
mode='merge')
elif self.dim * self.max_labels == space.dim:
rval = self.formatter.theano_expr(batch, sparse=space.sparse,
mode='concatenate')
else:
raise ValueError("Can't convert IndexSpace to"
"VectorSpace (%d labels to %d dimensions)"
% (self.dim, space.dim))
return rval
else:
raise ValueError("Can't convert IndexSpace to %(space)s"
% (space.__class__.__name__))
@functools.wraps(Space.make_theano_batch)
def make_theano_batch(self, name=None, dtype=None, batch_size=None):
if batch_size == 1:
rval = T.lrow(name=name)
else:
rval = T.lmatrix(name=name)
return rval
@functools.wraps(Space.batch_size)
def batch_size(self, batch):
self.validate(batch)
return batch.shape[0]
@functools.wraps(Space.np_batch_size)
def np_batch_size(self, batch):
self.np_validate(batch)
return batch.shape[0]
@functools.wraps(Space._validate)
def _validate(self, batch):
"""
.. todo::
WRITEME
"""
if not isinstance(batch, theano.gof.Variable):
raise TypeError("IndexSpace batch should be a theano Variable, "
"got " + str(type(batch)))
if not isinstance(batch.type, (theano.tensor.TensorType,
CudaNdarrayType)):
raise TypeError("VectorSpace batch should be TensorType or "
"CudaNdarrayType, got "+str(batch.type))
if batch.ndim != 2:
raise ValueError('IndexSpace batches must be 2D, got %d '
'dimensions' % batch.ndim)
for val in get_debug_values(batch):
self.np_validate(val)
@functools.wraps(Space._np_validate)
def _np_validate(self, batch):
# Use the 'CudaNdarray' string to avoid importing theano.sandbox.cuda
# when it is not available
if (not isinstance(batch, np.ndarray)
and str(type(batch)) != "<type 'CudaNdarray'>"):
raise TypeError("The value of a IndexSpace batch should be a "
"numpy.ndarray, or CudaNdarray, but is %s."
% str(type(batch)))
if batch.ndim != 2:
raise ValueError("The value of a IndexSpace batch must be "
"2D, got %d dimensions for %s." % (batch.ndim,
batch))
if batch.shape[1] != self.dim:
raise ValueError("The width of a IndexSpace batch must match "
"with the space's dimension, but batch has shape "
"%s and dim = %d." % (str(batch.shape), self.dim))
def __init__(
self,
db, # data source
name='', # optional name
selectors=dict(),
partitioner=None,
meta_sources=[], # optional sources other than 'features' and 'targets' from metadata
channel_filter=NoChannelFilter(
), # optional channel filter, default: keep all
channel_names=None, # optional channel names (for metadata)
label_attribute='label', # metadata attribute to be used as label
label_map=None, # optional conversion of labels
use_targets=True, # use targets if provides, otherwise labels are used
remove_dc_offset=False, # optional subtraction of channel mean, usually done already earlier
resample=None, # optional down-sampling
normalize=True, # normalize to max=1
# optional sub-sequences selection
start_sample=0,
stop_sample=None, # optional for selection of sub-sequences
zero_padding=True, # if True (default) trials that are too short will be padded with
# otherwise they will rejected.
# optional signal filter to by applied before splitting the signal
signal_filter=None,
trial_processors=[], # optional processing of the trials
target_processor=None, # optional processing of the targets, e.g. zero-padding
transformers=[], # optional transformations of the dataset
layout='tf', # (0,1)-axes layout tf=time x features or ft=features x time
debug=False,
):
'''
Constructor
'''
# save params
self.params = locals().copy()
del self.params['self']
# print self.params
self.name = name
self.debug = debug
metadb = DatasetMetaDB(db.metadata, selectors.keys())
if partitioner is not None:
pass # FIXME
selected_trial_ids = metadb.select(selectors)
log.info('selectors: {}'.format(selectors))
log.info('selected trials: {}'.format(selected_trial_ids))
if normalize:
log.info(
'Data will be normalized to max amplitude 1 per channel (normalize=True).'
)
trials = list()
labels = list()
targets = list()
meta = list()
if stop_sample == 'auto-min':
stop_sample = np.min(
[db.data[trial_i].shape[-1] for trial_i in selected_trial_ids])
log.info('Using minimum trial length. stop_sample={}'.format(
stop_sample))
elif stop_sample == 'auto-max':
stop_sample = np.max(
[db.data[trial_i].shape[-1] for trial_i in selected_trial_ids])
log.info('Using maximum trial length. stop_sample={}'.format(
stop_sample))
for trial_i in selected_trial_ids:
trial_meta = db.metadata[trial_i]
if use_targets:
if targets is None:
target = None
else:
target = db.targets[trial_i]
assert not np.isnan(np.sum(target))
if target_processor is not None:
target = target_processor.process(target, trial_meta)
assert not np.isnan(np.sum(target))
else:
# get and process label
label = db.metadata[trial_i][label_attribute]
if label_map is not None:
label = label_map[label]
processed_trial = []
trial = db.data[trial_i]
if np.isnan(np.sum(trial)):
print trial_i, trial
assert not np.isnan(np.sum(trial))
rejected = False # flag for trial rejection
trial = np.atleast_2d(trial)
# process 1 channel at a time
for channel in xrange(trial.shape[0]):
# filter channels
if not channel_filter.keep_channel(channel):
continue
samples = trial[channel, :]
# subtract channel mean
if remove_dc_offset:
samples -= samples.mean()
# down-sample if requested
if resample is not None and resample[0] != resample[1]:
samples = librosa.resample(samples,
resample[0],
resample[1],
res_type='sinc_best')
# apply optional signal filter after down-sampling -> requires lower order
if signal_filter is not None:
samples = signal_filter.process(samples)
# get sub-sequence in resampled space
# log.info('using samples {}..{} of {}'.format(start_sample,stop_sample, samples.shape))
if stop_sample is not None and stop_sample > len(samples):
if zero_padding:
tmp = np.zeros(stop_sample)
tmp[:len(samples)] = samples
samples = tmp
else:
rejected = True
break # stop processing this trial
s = samples[start_sample:stop_sample]
# TODO optional channel processing
# normalize to max amplitude 1
if normalize:
s = librosa.util.normalize(s)
# add 2nd data dimension
s = s.reshape(s.shape[0], 1)
# print s.shape
s = np.asfarray(s, dtype=theano.config.floatX)
processed_trial.append(s)
### end of channel iteration ###
if rejected:
continue # next trial
processed_trial = np.asfarray([processed_trial],
dtype=theano.config.floatX)
# processed_trial = processed_trial.reshape((1, processed_trial.shape))
processed_trial = np.rollaxis(processed_trial, 1, 4)
# optional (external) trial processing, e.g. windowing
# trials will be in b01c format with tf layout for 01-axes
for trial_processor in trial_processors:
processed_trial = trial_processor.process(
processed_trial, trial_meta)
trials.append(processed_trial)
for k in range(len(processed_trial)):
meta.append(trial_meta)
if use_targets:
targets.append(target)
else:
labels.append(label)
### end of datafile iteration ###
# turn into numpy arrays
self.trials = np.vstack(trials)
assert not np.isnan(np.sum(self.trials))
# prepare targets / labels
if use_targets:
self.targets = np.vstack(targets)
assert not np.isnan(np.sum(self.targets))
else:
labels = np.hstack(labels)
if label_map is None:
one_hot_formatter = OneHotFormatter(max(labels) + 1)
else:
one_hot_formatter = OneHotFormatter(
max(label_map.values()) + 1)
one_hot_y = one_hot_formatter.format(labels)
self.targets = one_hot_y
self.metadata = meta
if layout == 'ft': # swap axes to (batch, feature, time, channels)
self.trials = self.trials.swapaxes(1, 2)
# transform after finalizing the data structure
for transformer in transformers:
self.trials, self.targets = transformer.process(
self.trials, self.targets)
self.trials = np.asarray(self.trials, dtype=theano.config.floatX)
log.debug('final dataset shape: {} (b,0,1,c)'.format(
self.trials.shape))
# super(EEGEpochsDataset, self).__init__(topo_view=self.trials, y=self.targets, axes=['b', 0, 1, 'c'])
self.X = self.trials.reshape(self.trials.shape[0],
np.prod(self.trials.shape[1:]))
self.y = self.targets
log.info('generated dataset "{}" with shape X={}={} y={} targets={} '.
format(self.name, self.X.shape, self.trials.shape,
self.y.shape, self.targets.shape))
# determine data specs
features_space = Conv2DSpace(
shape=[self.trials.shape[1], self.trials.shape[2]],
num_channels=self.trials.shape[3])
features_source = 'features'
targets_space = VectorSpace(dim=self.targets.shape[-1])
targets_source = 'targets'
space_components = [features_space, targets_space]
source_components = [features_source, targets_source]
# additional support for meta information
self.meta_maps = dict()
for meta_source in meta_sources:
self.meta_maps[meta_source] = sorted(
list(set([m[meta_source] for m in self.metadata])))
space_components.extend([VectorSpace(dim=1)])
source_components.extend([meta_source])
log.info('Generated meta-source "{}" with value map: {}'.format(
meta_source, self.meta_maps[meta_source]))
space = CompositeSpace(space_components)
source = tuple(source_components)
self.data_specs = (space, source)
log.debug('data specs: {}'.format(self.data_specs))
def load_data(self):
# Get the directory of the patient data
patient_dir = os.path.join(self.data_dir, self.patient_id)
# Load metadata about dataset form MAT file
metadata_fname = os.path.join(patient_dir, 'trainset_' + str(self.preictal_sec) + '.mat')
metadata_mat = loadmat(metadata_fname)
# Get number of seizures
self.n_seizures = metadata_mat.get('ictals').size
# Get detail of the segment
self.sampling_rate = metadata_mat['sampling_rate'][0][0]
self.segment_sec = metadata_mat['segment_sec'][0][0]
self.segment_samples = self.sampling_rate * self.segment_sec
# Get the number blocks to extend from the withheld seizure
self.n_extended_blocks_test = metadata_mat['n_extended_blocks_test'][0][0]
self.preictal_samples = 0
self.nonictal_samples = 0
self.nan_non_flat_samples = 0
# Examples of indexing through MAT file
# mat['nonictals'][i][0]['filename'][0][0][0][j][0]
# mat['nonictals'][i][0]['idx'][0][0][0][j][0]
# mat['nonictals'][i][0]['n_segments'][0][0][0][0]
# Load shuffle data
if self.which_set == 'train' or self.which_set == 'valid_train':
if self.which_set == 'train':
select_idx = np.setdiff1d(range(metadata_mat['preictals'].size),
np.asarray([self.leave_out_seizure_idx_valid,
self.leave_out_seizure_idx_test]))
else:
select_idx = np.asarray([self.leave_out_seizure_idx_valid])
X = None
y = None
if self.use_all_nonictals:
temp_preictal_X = None
for i in select_idx:
print '====== Seizure', i, '======'
# Pre-ictal
temp_X = self.load_feature(part='preictals',
list_features=self.list_features,
seizure_idx=i,
metadata_mat=metadata_mat,
patient_dir=patient_dir)
if not (temp_preictal_X is None):
temp_preictal_X = np.concatenate((temp_preictal_X, temp_X), axis=1)
else:
temp_preictal_X = temp_X
self.preictal_samples = temp_preictal_X.shape[1]
# Non-ictal data
temp_nonictal_X = self.load_feature(part='nonictals_all',
list_features=self.list_features,
seizure_idx=self.leave_out_seizure_idx_test,
metadata_mat=metadata_mat,
patient_dir=patient_dir)
X = np.concatenate((temp_preictal_X, temp_nonictal_X), axis=1)
y = np.zeros(X.shape[1], dtype=int)
y[range(self.preictal_samples)] = 1
self.nonictal_samples = temp_nonictal_X.shape[1]
print 'Preictal samples: {0}, Nonictal samples: {1}'.format(self.preictal_samples,
self.nonictal_samples)
if not np.all(np.arange(self.preictal_samples) == np.where(y)[0]):
raise Exception('There is a mismatch between the number of preictal data and labels.')
else:
for i in select_idx:
print '====== Seizure', i, '======'
# Non-ictal data
temp_nonictal_X = self.load_feature(part='nonictals',
list_features=self.list_features,
seizure_idx=i,
metadata_mat=metadata_mat,
patient_dir=patient_dir)
# Pre-ictal
temp_preictal_X = self.load_feature(part='preictals',
list_features=self.list_features,
seizure_idx=i,
metadata_mat=metadata_mat,
patient_dir=patient_dir)
# Concatenate preictal and nonictal data
temp_X = np.concatenate((temp_preictal_X, temp_nonictal_X), axis=1)
temp_y = np.zeros(temp_X.shape[1], dtype=int)
temp_y[range(temp_preictal_X.shape[1])] = 1
# Sanity check
# if not (temp_preictal_X.shape[1] == temp_nonictal_X.shape[1]):
# raise Exception('Unbalanced classes.')
print 'Preictal samples: {0}, Nonictal samples: {1}'.format(temp_preictal_X.shape[1],
temp_nonictal_X.shape[1])
if not np.all(np.arange(temp_preictal_X.shape[1]) == np.where(temp_y)[0]):
raise Exception('There is a mismatch between the number of preictal data and labels.')
self.preictal_samples = self.preictal_samples + temp_preictal_X.shape[1]
self.nonictal_samples = self.nonictal_samples + temp_nonictal_X.shape[1]
if not (X is None) and not (y is None):
X = np.concatenate((X, temp_X), axis=1)
y = np.append(y, temp_y)
else:
X = temp_X
y = temp_y
# Load continuous data
elif self.which_set == 'valid' or self.which_set == 'test':
if self.which_set == 'valid':
select_idx = self.leave_out_seizure_idx_valid
else:
select_idx = self.leave_out_seizure_idx_test
print '====== Seizure', select_idx, '======'
# Get metadata of all blocks
block_df = pd.read_table(os.path.join(patient_dir, 'block_metadata.txt'), sep='\t')
# Get block index of the selected seizure
select_sz_fname = metadata_mat['preictals'][select_idx][0]['filename'][0][0][0][0][0]
block_idx = np.where(block_df.filename == select_sz_fname)[0][0]
start_block_idx = block_idx - self.n_extended_blocks_test
end_block_idx = block_idx + self.n_extended_blocks_test + 1
if start_block_idx < 0:
start_block_idx = 0
if end_block_idx > block_df.shape[0]:
end_block_idx = block_df.shape[0]
select_block_idx = np.arange(start_block_idx, end_block_idx)
filenames = block_df.filename[select_block_idx].values
X = None
y = None
y_label_all = None
ictal_labels = None
for b_idx, fname in enumerate(filenames):
# Name of the MAT file that stores indices of flat (i.e., false) segments
fname_flat = fname.replace('.data', '_flat_signal_segment_idx.mat')
# Get all good indices (i.e., remove segments of flat signals)
flat_mat = loadmat(os.path.join(patient_dir, fname_flat))
flat_idx = np.empty(0, dtype=int)
for j in range(flat_mat['flat_signal_segment_idx'].shape[0]):
flat_idx = np.append(flat_idx, np.squeeze(flat_mat['flat_signal_segment_idx'][j][0]))
flat_idx = flat_idx - 1 # Change from MATLAB to python index system
n_segments = np.ceil(block_df.samples[select_block_idx[b_idx]] / (self.segment_samples * 1.0))
all_idx = np.arange(n_segments, dtype=int)
good_idx = np.setdiff1d(all_idx, flat_idx)
print 'Load', self.which_set, 'data from', fname
if good_idx.size > 0:
# Features with shape [n_features, n_samples]
temp_X = self.load_list_feature(list_features=self.list_features,
sample_idx=good_idx,
fname=fname,
patient_dir=patient_dir)
# If this record contains preictal data in the withheld seizures, get preictal labels
temp_y_withheld = self.get_labels(label_type='preictals',
filename=fname,
good_idx=good_idx,
metadata_mat=metadata_mat,
n_all_segments=n_segments,
n_data_segments=temp_X.shape[1],
select_meta_idx=select_idx)
# If this record contains preictal data in the selected seizures, get preictal labels
temp_y_select = self.get_labels(label_type='preictals',
filename=fname,
good_idx=good_idx,
metadata_mat=metadata_mat,
n_all_segments=n_segments,
n_data_segments=temp_X.shape[1])
# If this record contains preictal data in all seizures, get preictal labels
temp_y_rm = self.get_labels(label_type='all_preictals',
filename=fname,
good_idx=good_idx,
metadata_mat=metadata_mat,
n_all_segments=n_segments,
n_data_segments=temp_X.shape[1])
tmp_preictal_withheld_idx = np.where(temp_y_withheld == 1)[0]
tmp_preictal_select_idx = np.where(temp_y_select == 1)[0]
tmp_preictal_rm_idx = np.where(temp_y_rm == 1)[0]
tmp_preictal_select_idx = np.setdiff1d(tmp_preictal_select_idx, tmp_preictal_withheld_idx)
tmp_preictal_rm_idx = np.setdiff1d(tmp_preictal_rm_idx, tmp_preictal_withheld_idx)
tmp_preictal_rm_idx = np.setdiff1d(tmp_preictal_rm_idx, tmp_preictal_select_idx)
self.preictal_samples = self.preictal_samples + np.where(temp_y_withheld == 1)[0].size
self.nonictal_samples = self.nonictal_samples + np.where(temp_y_withheld == 0)[0].size
if tmp_preictal_withheld_idx.size > 0:
print ' Load preictal data from the withheld seizure from this file.'
print ' Size:', tmp_preictal_withheld_idx.size, tmp_preictal_withheld_idx
if tmp_preictal_select_idx.size > 0:
print ' Load preictal data from selected seizures in addition to the withheld seizure from this file.'
print ' Size:', tmp_preictal_select_idx.size, tmp_preictal_select_idx
if tmp_preictal_rm_idx.size > 0:
print ' Load preictal data from removed seizures in addition to the withheld seizure from this file.'
print ' Size:', tmp_preictal_rm_idx.size, tmp_preictal_rm_idx
# Sanity check
if np.intersect1d(tmp_preictal_withheld_idx, tmp_preictal_select_idx).size > 0:
raise Exception('There is an overlapped of the labels between the withheld seizures, and the selected seizures.')
if np.intersect1d(tmp_preictal_select_idx, tmp_preictal_rm_idx).size > 0:
raise Exception('There is an overlapped of the labels between the selected seizures, and the removed seizures.')
if np.intersect1d(tmp_preictal_withheld_idx, tmp_preictal_rm_idx).size > 0:
raise Exception('There is an overlapped of the labels between the withheld seizures, and the removed seizures.')
temp_y_all = np.zeros(temp_X.shape[1], dtype=int)
temp_y_all[tmp_preictal_withheld_idx] = 1 # Labels for the withheld seizure
temp_y_all[tmp_preictal_select_idx] = 2 # Labels for the selected seizure (that is not from withheld seizures)
temp_y_all[tmp_preictal_rm_idx] = 3 # Labels for the removed seizure (that is not from withheld seizures)
# If this record contains ictal data, get ictal labels
temp_ictal_labels = self.get_labels(label_type='all_ictals',
filename=fname,
good_idx=good_idx,
metadata_mat=metadata_mat,
n_all_segments=n_segments,
n_data_segments=temp_X.shape[1])
tmp_ictal_idx = np.where(temp_ictal_labels == 1)[0]
if tmp_ictal_idx.size > 0:
print ' Ictal label:', tmp_ictal_idx.size, tmp_ictal_idx
# Dealing with NaN features after filtering out flat segment which occurs due to noise in the data,
# not from flat segments
nan_sample_idx = np.where(np.isnan(np.sum(temp_X, 0)))[0]
nan_feature_idx = np.where(np.isnan(np.sum(temp_X, 1)))[0]
if nan_sample_idx.size > 0 or nan_feature_idx.size > 0:
print self.which_set, 'contains NaN at:'
print ' sample_idx:', good_idx[nan_sample_idx], ' feature_idx:', nan_feature_idx
print ' shape before remove NaN:', temp_X.shape
tmp_preictal_idx = np.where(temp_y_withheld == 1)[0]
tmp_nonictal_idx = np.where(temp_y_withheld == 0)[0]
nan_preictal_sample_idx = np.intersect1d(tmp_preictal_idx, nan_sample_idx)
nan_nonictal_sample_idx = np.intersect1d(tmp_nonictal_idx, nan_sample_idx)
if nan_preictal_sample_idx.size > 0:
print ' NaN are in preictal index:', good_idx[nan_preictal_sample_idx]
if nan_nonictal_sample_idx.size > 0:
print ' NaN are in nonictal index:', good_idx[nan_nonictal_sample_idx]
all_idx = np.arange(temp_X.shape[1])
good_idx_1 = np.setdiff1d(all_idx, nan_sample_idx)
temp_X = temp_X[:, good_idx_1]
temp_y_all = temp_y_all[good_idx_1]
temp_y_withheld = temp_y_withheld[good_idx_1]
temp_ictal_labels = temp_ictal_labels[good_idx_1]
print ' shape before remove NaN:', temp_X.shape
self.nan_non_flat_samples = self.nan_non_flat_samples + nan_sample_idx.size
# Sanity check
tmp_nan_sample_idx = np.where(np.isnan(np.sum(temp_X, 0)))[0]
if tmp_nan_sample_idx.size > 0:
raise Exception('There is an error in removing NaN')
if not (temp_X.shape[1] == temp_y_all.size):
raise Exception('Number of feature data and labels [temp_y_all] are not equal.')
if not (temp_X.shape[1] == temp_y_withheld.size):
raise Exception('Number of feature data and labels [temp_y_withheld] are not equal.')
if not (temp_X.shape[1] == temp_ictal_labels.size):
raise Exception('Number of feature data and labels [ictal_labels] are not equal.')
if not (X is None) and not (y is None) and not (ictal_labels is None):
X = np.concatenate((X, temp_X), axis=1)
y = np.append(y, temp_y_withheld)
y_label_all = np.append(y_label_all, temp_y_all)
ictal_labels = np.append(ictal_labels, temp_ictal_labels)
else:
X = temp_X
y = temp_y_withheld
y_label_all = temp_y_all
ictal_labels = temp_ictal_labels
else:
print 'There is no good segment for during this seizure'
# Store preictal labels that are from the withheld index (use for compute accuracy), selected seizure index,
# and removed seizure index.
# Note: this property will exist when which_set=='valid' or which_set=='test'
# as there is no need for ictal to be imported.
self.y_label_all = y_label_all
# Sanity check
if np.where(y == 1)[0].size > np.where(y_label_all > 0)[0].size:
raise Exception('There is an error in collecting preictal labels only from the leave-out-seizure index.')
if np.where(y == 1)[0].size == np.where(y_label_all == 1)[0].size:
print 'There is only one preictal periods, and this period is from the leave-out-seizure index.'
if not np.all(np.where(y == 1)[0] == np.where(y_label_all == 1)[0]):
raise Exception('There is a mismatch between y and y_label_all.')
if np.where(y == 1)[0].size < np.where(y_label_all > 0)[0].size:
print 'There are more than one preictal periods.'
if not np.all(np.where(y == 1)[0] == np.where(y_label_all == 1)[0]):
raise Exception('There is a mismatch between y_select_idx and y in the preictal labels of the leave-out-seizure index.')
# Store ictal labels
# Note: this property will exist when which_set=='valid' or which_set=='test'
# as there is no need for ictal to be imported.
self.ictal_labels = ictal_labels
else:
raise Exception('Invalid dataset selection')
print 'There are {0} samples that have been removed in addition to the flat signal as due to NaN.'.format(self.nan_non_flat_samples)
X = np.transpose(X, [1, 0])
one_hot_formatter = OneHotFormatter(max_labels=2)
y = one_hot_formatter.format(y)
# Sanity check
# Note: We ignore nan_non_flat_samples if we load shuffle data as we specify the labels after the NaN have been removed
# In contrast to loading continuous data, we specify the labels before removing NaN, so we have to remove the NaN samples for checking
if self.which_set == 'train' or self.which_set == 'valid_train':
if not (X.shape[0] == self.preictal_samples + self.nonictal_samples):
raise Exception('There is a mismatch in the number of training samples ({0} != {1}).'.format(X.shape[0],
self.preictal_samples + self.nonictal_samples))
if not (np.where(np.argmax(y, axis=1) == 1)[0].size == self.preictal_samples):
raise Exception('There is a mismatch in the number of preictal samples and its labels ({0} != {1}).'.format(np.where(np.argmax(y, axis=1) == 1)[0].size,
self.preictal_samples))
if not (X.shape[0] == y.shape[0]):
raise Exception('There is a mismatch in the number of training samples and its labels ({0} != {1}).'.format(X.shape[0],
y.shape[0]))
elif self.which_set == 'valid' or self.which_set == 'test':
if not (X.shape[0] == self.preictal_samples + self.nonictal_samples - self.nan_non_flat_samples):
raise Exception('There is a mismatch in the number of training samples ({0} != {1}).'.format(X.shape[0],
self.preictal_samples + self.nonictal_samples - self.nan_non_flat_samples))
if not ((np.where(np.argmax(y, axis=1) == 1)[0].size + np.where(np.argmax(y, axis=1) == 0)[0].size) ==
self.preictal_samples + self.nonictal_samples - self.nan_non_flat_samples):
raise Exception('There is a mismatch in the number of samples and its labels ({0} != {1}).'.format(np.where(np.argmax(y, axis=1) == 1)[0].size + np.where(np.argmax(y, axis=1) == 0)[0].size,
self.preictal_samples))
if not (X.shape[0] == y.shape[0]):
raise Exception('There is a mismatch in the number of training samples and its labels ({0} != {1}).'.format(X.shape[0],
y.shape[0]))
return X, y
def __init__(self,
path, suffix='', # required data file parameters
subjects='all', # optional selector (list) or 'all'
start_sample = 0,
stop_sample = None, # optional for selection of sub-sequences
frame_size = -1,
hop_size = -1, # values > 0 will lead to windowing
label_mode='tempo',
name = '', # optional name
n_fft = 0,
n_freq_bins = None,
save_matrix_path = None,
channels = None,
resample = None,
stimulus_id_filter = None,
keep_metadata = False,
spectrum_log_amplitude = False,
spectrum_normalization_mode = None,
include_phase = False,
layout = 'tf' # 2D axes layout tf=time x features or ft= features x time
):
'''
Constructor
'''
# save params
self.params = locals().copy()
del self.params['self']
# print self.params
self.name = name;
self.include_phase = include_phase;
self.spectrum_normalization_mode = spectrum_normalization_mode;
self.spectrum_log_amplitude = spectrum_log_amplitude;
self.datafiles = [];
subject_paths = glob.glob(os.path.join(path, 'Sub*'));
for path in subject_paths:
dataset_filename = os.path.join(path, 'dataset'+suffix+'.pklz');
if os.path.isfile(dataset_filename):
log.debug('addding {}'.format(dataset_filename));
self.datafiles.append(dataset_filename);
else:
log.warn('file does not exists {}'.format(dataset_filename));
self.datafiles.sort();
if subjects == 'all':
subjects = np.arange(0,len(self.datafiles));
assert subjects is not None and len(subjects) > 0;
self.label_mode = label_mode;
self.label_converter = LabelConverter();
if stimulus_id_filter is None:
stimulus_id_filter = [];
self.stimulus_id_filter = stimulus_id_filter;
self.subject_partitions = []; # used to keep track of original subjects
self.sequence_partitions = []; # used to keep track of original sequences
self.trial_partitions = []; # keeps track of original trials
# metadata: [subject, trial_no, stimulus, channel, start, ]
self.metadata = [];
sequences = [];
labels = [];
n_sequences = 0;
last_raw_label = -1;
for i in xrange(len(self.datafiles)):
if i in subjects:
with log_timing(log, 'loading data from {}'.format(self.datafiles[i])):
self.subject_partitions.append(n_sequences); # save start of next subject
subject_sequences, subject_labels, channel_meta = load(self.datafiles[i]);
subject_trial_no = -1;
for j in xrange(len(subject_sequences)):
l = subject_labels[j]; # get raw label
if l in stimulus_id_filter:
# log.debug('skipping stimulus {}'.format(l));
continue;
c = channel_meta[j][0];
if channels is not None and not c in channels: # apply optional channel filter
log.debug('skipping channel {}'.format(c));
continue;
self.sequence_partitions.append(n_sequences); # save start of next sequence
if l != last_raw_label: # if raw label changed...
self.trial_partitions.append(n_sequences); # ...save start of next trial
subject_trial_no += 1; # increment subject_trial_no counter
last_raw_label = l;
l = self.label_converter.get_label(l[0], self.label_mode); # convert to label_mode view
s = subject_sequences[j];
s = s[start_sample:stop_sample]; # get sub-sequence in original space
# down-sample if requested
if resample is not None and resample[0] != resample[1]:
s = librosa.resample(s, resample[0], resample[1]);
if n_fft is not None and n_fft > 0: # Optionally:
# transform to spectogram
hop_length = n_fft / 4;
'''
from http://theremin.ucsd.edu/~bmcfee/librosadoc/librosa.html
>>> # Get a power spectrogram from a waveform y
>>> S = np.abs(librosa.stft(y)) ** 2
>>> log_S = librosa.logamplitude(S)
'''
# s = np.abs(librosa.core.stft(s,
# n_fft=n_fft,
# hop_length=hop_length)
# )**2;
S = librosa.core.stft(s, n_fft=n_fft, hop_length=hop_length);
# mag = np.abs(S); # magnitude spectrum
mag = np.abs(S)**2; # power spectrum
# phase = np.unwrap(np.angle(S));
phase = np.angle(S);
if n_freq_bins is not None: # Optionally:
mag = mag[0:n_freq_bins, :]; # cut off high bands
phase = phase[0:n_freq_bins, :];
if self.spectrum_log_amplitude:
mag = librosa.logamplitude(mag);
s = mag; # for normalization
'''
NOTE on normalization:
It depends on the structure of a neural network and (even more)
on the properties of data. There is no best normalization algorithm
because if there would be one, it would be used everywhere by default...
In theory, there is no requirement for the data to be normalized at all.
This is a purely practical thing because in practice convergence could
take forever if your input is spread out too much. The simplest would be
to just normalize it by scaling your data to (-1,1) (or (0,1) depending
on activation function), and in most cases it does work. If your
algorithm converges well, then this is your answer. If not, there are
too many possible problems and methods to outline here without knowing
the actual data.
'''
## normalize to mean 0, std 1
if self.spectrum_normalization_mode == 'mean0_std1':
# s = preprocessing.scale(s, axis=0);
mean = np.mean(s);
std = np.std(s);
s = (s - mean) / std;
## normalize by linear transform to [0,1]
elif self.spectrum_normalization_mode == 'linear_0_1':
s = s / np.max(s);
## normalize by linear transform to [-1,1]
elif self.spectrum_normalization_mode == 'linear_-1_1':
s = -1 + 2 * (s - np.min(s)) / (np.max(s) - np.min(s));
elif self.spectrum_normalization_mode is not None:
raise ValueError(
'unsupported spectrum normalization mode {}'.format(
self.spectrum_normalization_mode)
);
#print s.mean(axis=0)
#print s.std(axis=0)
# include phase information if requested
if self.include_phase:
# normalize phase to [-1.1]
phase = phase / np.pi
s = np.vstack([s, phase]);
# transpose to fit pylearn2 layout
s = np.transpose(s);
else:
# normalize to max amplitude 1
s = librosa.util.normalize(s);
s = np.asfarray(s, dtype='float32');
if frame_size > 0 and hop_size > 0:
s, l = self._split_sequence(s, l, frame_size, hop_size);
# print s.shape
n_sequences += len(s);
sequences.append(s);
labels.extend(l);
if keep_metadata:
self.metadata.append({
'subject' : i, # subject
'trial_no' : subject_trial_no, # trial_no
'stimulus' : last_raw_label[0], # stimulus
'channel' : c, # channel
'start' : self.sequence_partitions[-1], # start
'stop' : n_sequences # stop
});
# turn into numpy arrays
sequences = np.vstack(sequences);
print sequences.shape;
labels = np.hstack(labels);
# one_hot_y = one_hot(labels)
one_hot_formatter = OneHotFormatter(labels.max() + 1)
one_hot_y = one_hot_formatter.format(labels)
self.labels = labels; # save for later
if n_fft > 0:
sequences = np.array([sequences]);
# re-arrange dimensions
sequences = sequences.swapaxes(0,1).swapaxes(1,2).swapaxes(2,3);
if layout == 'ft':
sequences = sequences.swapaxes(1,2)
log.debug('final dataset shape: {} (b,0,1,c)'.format(sequences.shape));
print 'final dataset shape: {} (b,0,1,c)'.format(sequences.shape)
super(EEGDataset, self).__init__(topo_view=sequences, y=one_hot_y, axes=['b', 0, 1, 'c']);
else:
# if layout == 'ft':
# sequences = sequences.swapaxes(1,2)
super(EEGDataset, self).__init__(X=sequences, y=one_hot_y, axes=['b', 0, 1, 'c']);
log.debug('generated dataset "{}" with shape X={} y={} labels={} '.format(self.name, self.X.shape, self.y.shape, self.labels.shape));
if save_matrix_path is not None:
matrix = DenseDesignMatrix(X=sequences, y=one_hot_y);
with log_timing(log, 'saving DenseDesignMatrix to {}'.format(save_matrix_path)):
serial.save(save_matrix_path, matrix);