y = speech["target"]
vocabulary = speech["vocabulary"]
vocabulary_size = speech["vocabulary_size"]
reconstruct = speech["reconstruct"]
fs = speech["sample_rate"]
X = np.array([x.astype(theano.config.floatX) for x in X])
y = np.array([yy.astype(theano.config.floatX) for yy in y])
minibatch_size = 20
n_epochs = 20000 # Used way at the bottom in the training loop!
checkpoint_every_n = 500
# Was 300 for handwriting
cut_len = 31 # Used way at the bottom in the training loop!
random_state = np.random.RandomState(1999)
train_itr = list_iterator([X, y], minibatch_size, axis=1, stop_index=80,
randomize=True, make_mask=True)
valid_itr = list_iterator([X, y], minibatch_size, axis=1, start_index=80,
make_mask=True)
X_mb, X_mb_mask, c_mb, c_mb_mask = next(train_itr)
train_itr.reset()
input_dim = X_mb.shape[-1]
n_bins = 10
n_kernels = 32
conv_size1 = 11
conv_size2 = 5
deconv_size1 = 5
deconv_size2 = 11
n_hid = 512
att_size = 10
vocabulary = speech["vocabulary"]
vocabulary_size = speech["vocabulary_size"]
reconstruct = speech["reconstruct"]
fs = speech["sample_rate"]
X = np.array([x.astype(theano.config.floatX) for x in X])
y = np.array([yy.astype(theano.config.floatX) for yy in y])
minibatch_size = 1
n_epochs = 200 # Used way at the bottom in the training loop!
checkpoint_every_n = 10
cut_len = 41 # Used way at the bottom in the training loop!
random_state = np.random.RandomState(1999)
train_itr = list_iterator([X, y],
minibatch_size,
axis=1,
stop_index=105,
randomize=True,
make_mask=True)
valid_itr = list_iterator([X, y],
minibatch_size,
axis=1,
start_index=105 - minibatch_size,
randomize=True,
make_mask=True)
X_mb, X_mb_mask, c_mb, c_mb_mask = next(train_itr)
train_itr.reset()
n_hid = 256
att_size = 10
n_proj = 256
n_v_proj = 5
X1 = np.array([x.astype(theano.config.floatX) for x in X1])
X2 = np.array([x.astype(theano.config.floatX) for x in X2])
X = np.array(
[np.hstack((x1[:, None], x2[:, None])) for x1, x2 in zip(X1, X2)])
y = np.array([yy.astype(theano.config.floatX) for yy in y])
minibatch_size = 20
n_epochs = 1000 # Used way at the bottom in the training loop!
checkpoint_every_n = 100
# Was 300
cut_len = 21 # Used way at the bottom in the training loop!
random_state = np.random.RandomState(1999)
train_itr = list_iterator([X, y],
minibatch_size,
axis=1,
stop_index=80,
make_mask=True)
valid_itr = list_iterator([X, y],
minibatch_size,
axis=1,
start_index=80,
make_mask=True)
X_mb, X_mb_mask, c_mb, c_mb_mask = next(train_itr)
train_itr.reset()
n_hid = 256
att_size = 10
n_proj = 1024
n_softmax1 = X1_size
n_softmax2 = X2_size
#speech = fetch_fruitspeech_spectrogram()
mnist = fetch_binarized_mnist()
X = mnist["data"]
train_indices = mnist["train_indices"]
valid_indices = mnist["valid_indices"]
X = np.array([x.astype(theano.config.floatX) for x in X])
minibatch_size = 16
n_epochs = 10000 # Used way at the bottom in the training loop!
checkpoint_every_n = 10
n_bins = 1
random_state = np.random.RandomState(1999)
# bit weird but for MNIST this will return 28, 1, 28
train_itr = list_iterator([X], minibatch_size, axis=1,
stop_index=train_indices[-1] + 1, randomize=True,
make_mask=True)
valid_itr = list_iterator([X], minibatch_size, axis=1,
start_index=valid_indices[0],
stop_index=valid_indices[-1] + 1,
randomize=True, make_mask=True)
X_mb, X_mb_mask = next(train_itr)
train_itr.reset()
desc = "Speech generation"
parser = argparse.ArgumentParser(description=desc)
parser.add_argument('-s', '--sample',
help='Sample from a checkpoint file',
default=None,
required=False)
def restricted_int(x):
import theano
import sys
from kdllib import fetch_fruitspeech_spectrogram, run_loop, list_iterator
speech = fetch_fruitspeech_spectrogram()
data = speech["data"]
# 10 classes
X_train = data[0] / 10.
X_train = X_train[None].astype("float32")
y_train = X_train.astype("float32")
minibatch_size = 1
# Make easy iterators
data = [(X_train),]
target = [(y_train),]
train_itr = list_iterator([data, target], minibatch_size, axis=0, stop_index=1)
valid_itr = list_iterator([data, target], minibatch_size, axis=0, stop_index=1)
X_mb, y_mb = train_itr.next()
train_itr.reset()
# set recursion limit so pickle doesn't error
sys.setrecursionlimit(40000)
random_state = np.random.RandomState(1999)
n_epochs = 200
# theano land tensor4 for 4 dimensions
input_var = tensor.tensor4('X')
target_var = tensor.tensor4('y')
outchan = y_train.shape[0]
inchan = X_train.shape[0]
mnist = fetch_binarized_mnist()
X = mnist["data"]
train_indices = mnist["train_indices"]
valid_indices = mnist["valid_indices"]
X = np.array([x.astype(theano.config.floatX) for x in X])
minibatch_size = 16
n_epochs = 10000 # Used way at the bottom in the training loop!
checkpoint_every_n = 10
n_bins = 1
random_state = np.random.RandomState(1999)
# bit weird but for MNIST this will return 28, 1, 28
train_itr = list_iterator([X],
minibatch_size,
axis=1,
stop_index=train_indices[-1] + 1,
randomize=True,
make_mask=True)
valid_itr = list_iterator([X],
minibatch_size,
axis=1,
start_index=valid_indices[0],
stop_index=valid_indices[-1] + 1,
randomize=True,
make_mask=True)
X_mb, X_mb_mask = next(train_itr)
train_itr.reset()
desc = "Speech generation"
parser = argparse.ArgumentParser(description=desc)
parser.add_argument('-s',
speech = fetch_fruitspeech_spectrogram()
data = speech["data"]
# 10 classes
X_train = data[0] / 10.
X_train = X_train[None].astype("float32")
y_train = X_train.astype("float32")
minibatch_size = 1
# Make easy iterators
data = [
(X_train),
]
target = [
(y_train),
]
train_itr = list_iterator([data, target], minibatch_size, axis=0, stop_index=1)
valid_itr = list_iterator([data, target], minibatch_size, axis=0, stop_index=1)
X_mb, y_mb = train_itr.next()
train_itr.reset()
# set recursion limit so pickle doesn't error
sys.setrecursionlimit(40000)
random_state = np.random.RandomState(1999)
n_epochs = 200
# theano land tensor4 for 4 dimensions
input_var = tensor.tensor4('X')
target_var = tensor.tensor4('y')
outchan = y_train.shape[0]
inchan = X_train.shape[0]
