def get_seq_mnist_streams(hidden_dim, batch_size=100, drop_prob=0.5):
permutation = np.random.randint(0, 784, size=(784, ))
train_set, valid_set, test_set = load_data('mnist.pkl.gz')
train_x = train_set[0].reshape((50000 / batch_size, batch_size, 784))
train_x = np.swapaxes(train_x, 2, 1)
train_x = train_x[:, :, :, np.newaxis]
# Now the dimension is num_batches x 784 x batch_size x 1
train_y = (np.zeros(train_set[0].shape) - 1)
# label for each time-step is -1 and for the last one is the real label
train_y[:, -1] = train_set[1]
train_y = train_y.reshape((50000 / batch_size, batch_size, 784))
train_y = np.swapaxes(train_y, 2, 1)
train_y = train_y[:, :, :, np.newaxis]
# Now the dimension is num_batches x 784 x batch_size x 1
valid_x = valid_set[0].reshape((10000 / batch_size, batch_size, 784))
valid_x = np.swapaxes(valid_x, 2, 1)
valid_x = valid_x[:, :, :, np.newaxis]
# Now the dimension is num_batches x 784 x batch_size x 1
valid_y = (np.zeros(valid_set[0].shape) - 1)
# label for each time-step is -1 and for the last one is the real label
valid_y[:, -1] = valid_set[1]
valid_y = valid_y.reshape((10000 / batch_size, batch_size, 784))
valid_y = np.swapaxes(valid_y, 2, 1)
valid_y = valid_y[:, :, :, np.newaxis]
# Now the dimension is num_batches x 784 x batch_size x 1
train_x = train_x[:, permutation]
valid_x = valid_x[:, permutation]
train = IterableDataset({
'x': train_x.astype(floatX),
'y': train_y[:, -1, :, 0].astype('int32')
})
train_stream = DataStream(train)
train_stream = SampleDrops(train_stream, drop_prob, hidden_dim, False)
train_stream.sources = ('y', 'x', 'drops')
train_stream.get_epoch_iterator().next()
valid = IterableDataset({
'x': valid_x.astype(floatX),
'y': valid_y[:, -1, :, 0].astype('int32')
})
valid_stream = DataStream(valid)
valid_stream = SampleDrops(valid_stream, drop_prob, hidden_dim, True)
valid_stream.sources = ('y', 'x', 'drops')
return train_stream, valid_stream
def get_seq_mnist_streams(hidden_dim, batch_size=100, drop_prob=0.5):
permutation = np.random.randint(0, 784, size=(784,))
train_set, valid_set, test_set = load_data('mnist.pkl.gz')
train_x = train_set[0].reshape((50000 / batch_size, batch_size, 784))
train_x = np.swapaxes(train_x, 2, 1)
train_x = train_x[:, :, :, np.newaxis]
# Now the dimension is num_batches x 784 x batch_size x 1
train_y = (np.zeros(train_set[0].shape) - 1)
# label for each time-step is -1 and for the last one is the real label
train_y[:, -1] = train_set[1]
train_y = train_y.reshape((50000 / batch_size, batch_size, 784))
train_y = np.swapaxes(train_y, 2, 1)
train_y = train_y[:, :, :, np.newaxis]
# Now the dimension is num_batches x 784 x batch_size x 1
valid_x = valid_set[0].reshape((10000 / batch_size, batch_size, 784))
valid_x = np.swapaxes(valid_x, 2, 1)
valid_x = valid_x[:, :, :, np.newaxis]
# Now the dimension is num_batches x 784 x batch_size x 1
valid_y = (np.zeros(valid_set[0].shape) - 1)
# label for each time-step is -1 and for the last one is the real label
valid_y[:, -1] = valid_set[1]
valid_y = valid_y.reshape((10000 / batch_size, batch_size, 784))
valid_y = np.swapaxes(valid_y, 2, 1)
valid_y = valid_y[:, :, :, np.newaxis]
# Now the dimension is num_batches x 784 x batch_size x 1
train_x = train_x[:, permutation]
valid_x = valid_x[:, permutation]
train = IterableDataset({'x': train_x.astype(floatX),
'y': train_y[:, -1, :, 0].astype('int32')})
train_stream = DataStream(train)
train_stream = SampleDrops(train_stream, drop_prob, hidden_dim, False)
train_stream.sources = ('y', 'x', 'drops')
train_stream.get_epoch_iterator().next()
valid = IterableDataset({'x': valid_x.astype(floatX),
'y': valid_y[:, -1, :, 0].astype('int32')})
valid_stream = DataStream(valid)
valid_stream = SampleDrops(valid_stream, drop_prob, hidden_dim, True)
valid_stream.sources = ('y', 'x', 'drops')
return train_stream, valid_stream
def test_ngram_stream_error_on_multiple_sources():
# Check that NGram accepts only data streams with one source
sentences = [list(numpy.random.randint(10, size=sentence_length))
for sentence_length in [3, 5, 7]]
stream = DataStream(IterableDataset(sentences))
stream.sources = ('1', '2')
assert_raises(ValueError, NGrams, 4, stream)
def test_ngram_stream_error_on_multiple_sources():
# Check that NGram accepts only data streams with one source
sentences = [
list(numpy.random.randint(10, size=sentence_length))
for sentence_length in [3, 5, 7]
]
stream = DataStream(IterableDataset(sentences))
stream.sources = ('1', '2')
assert_raises(ValueError, NGrams, 4, stream)
def prep_dataset(dataset):
dataset = dataset[:(len(dataset) - (len(dataset) % (seq_len * batch_size)))]
dataset = dataset.reshape(batch_size, -1, seq_len).transpose((1, 0, 2))
stream = DataStream(IndexableDataset(indexables=OrderedDict([
('data', dataset)])),
iteration_scheme=SequentialExampleScheme(dataset.shape[0]))
stream = Transpose(stream, [(1, 0)])
stream = SampleDropsNPWord(
stream, z_prob_states, z_prob_cells, drop_prob_igates,
layer_size, num_layers, False, stoch_depth, share_mask,
gaussian_drop, alphabetsize)
stream.sources = ('data',) * 3 + stream.sources + ('zoneouts_states', 'zoneouts_cells', 'zoneouts_igates')
return (stream,)
def prep_dataset(dataset):
dataset = dataset[:(len(dataset) - (len(dataset) %
(seq_len * batch_size)))]
dataset = dataset.reshape(batch_size, -1, seq_len).transpose((1, 0, 2))
stream = DataStream(
IndexableDataset(indexables=OrderedDict([('data', dataset)])),
iteration_scheme=SequentialExampleScheme(dataset.shape[0]))
stream = Transpose(stream, [(1, 0)])
stream = SampleDropsNPWord(stream, z_prob_states, z_prob_cells,
drop_prob_igates, layer_size, num_layers,
False, stoch_depth, share_mask,
gaussian_drop, alphabetsize)
stream.sources = ('data', ) * 3 + stream.sources + (
'zoneouts_states', 'zoneouts_cells', 'zoneouts_igates')
return (stream, )
def test_sources_setter(self):
stream = DataStream(self.dataset)
stream.sources = ('features', )
assert_equal(stream.sources, ('features', ))
def test_sources_setter(self):
stream = DataStream(self.dataset)
stream.sources = ('features',)
assert_equal(stream.sources, ('features',))