MODE = '256ary' # binary or 256ary
MODEL = 'pixel_rnn' # either pixel_cnn or pixel_rnn
DIM = 32
PIXEL_CNN_LAYERS = 4
LR = 2e-4
BATCH_SIZE = 100
N_CHANNELS = 1
HEIGHT = 28
WIDTH = 28
TIMES = ('iters', 10*500, 1000*500)
lib.print_model_settings(locals().copy())
# inputs.shape: (batch size, n channels, height, width)
if MODE=='256ary':
inputs = T.itensor4('inputs')
inputs_float = inputs.astype(theano.config.floatX) * lib.floatX(2./255)
inputs_float -= lib.floatX(0.5)
else:
inputs = T.tensor4('inputs')
inputs_float = inputs
output = lib.ops.conv2d.Conv2D(
'InputConv',
input_dim=N_CHANNELS,
output_dim=DIM,
filter_size=7,
# FOLDER_PREFIX/params: saves all checkpoint params as pkl
# FOLDER_PREFIX/samples: keeps all checkpoint samples as wav
# FOLDER_PREFIX/best: keeps the best parameters, samples, ...
if not os.path.exists(FOLDER_PREFIX):
os.makedirs(FOLDER_PREFIX)
PARAMS_PATH = os.path.join(FOLDER_PREFIX, 'params')
if not os.path.exists(PARAMS_PATH):
os.makedirs(PARAMS_PATH)
SAMPLES_PATH = os.path.join(FOLDER_PREFIX, 'samples')
if not os.path.exists(SAMPLES_PATH):
os.makedirs(SAMPLES_PATH)
BEST_PATH = os.path.join(FOLDER_PREFIX, 'best')
if not os.path.exists(BEST_PATH):
os.makedirs(BEST_PATH)
lib.print_model_settings(locals(), path=FOLDER_PREFIX, sys_arg=True)
### Import the data_feeder ###
# Handling WHICH_SET
from datasets.dataset import music_train_feed_epoch as train_feeder
from datasets.dataset import music_valid_feed_epoch as valid_feeder
from datasets.dataset import music_test_feed_epoch as test_feeder
def load_data(data_feeder):
"""
Helper function to deal with interface of different datasets.
`data_feeder` should be `train_feeder`, `valid_feeder`, or `test_feeder`.
"""
return data_feeder(WHICH_SET, BATCH_SIZE, SEQ_LEN, OVERLAP, Q_LEVELS,
Q_ZERO, Q_TYPE)
CONV_FILTER_SIZE = 3
LATENT_DIM = 128
ALPHA_ITERS = 20000
VANILLA = False
LR = 2e-4
BATCH_SIZE = 100
N_CHANNELS = 1
HEIGHT = 28
WIDTH = 28
TIMES = ('iters', 10*500, 1000*500)
# TIMES = ('seconds', 60*30, 60*60*6)
lib.print_model_settings(locals().copy())
theano_srng = RandomStreams(seed=234)
def Encoder(inputs):
if MODE=='256ary':
embedded = lib.ops.embedding.Embedding(
'Encoder.Embedding',
256,
CONV_BASE_N_FILTERS,
inputs
)
embedded = embedded.dimshuffle(0,1,4,2,3)
embedded = embedded.reshape((
embedded.shape[0],
embedded.shape[1] * embedded.shape[2],
PARAMS_PATH = os.path.join(FOLDER_PREFIX, 'params')
if not os.path.exists(PARAMS_PATH):
os.makedirs(PARAMS_PATH)
SAMPLES_PATH = os.path.join(FOLDER_PREFIX, 'samples')
if not os.path.exists(SAMPLES_PATH):
os.makedirs(SAMPLES_PATH)
BEST_PATH = os.path.join(FOLDER_PREFIX, 'best')
if not os.path.exists(BEST_PATH):
os.makedirs(BEST_PATH)
lib.print_model_settings(locals(), path=FOLDER_PREFIX, sys_arg=True)
### Creating computation graph ###
def create_wavenet_block(inp, num_dilation_layer, input_dim, output_dim, name =None):
assert name is not None
layer_out = inp
skip_contrib = []
skip_weights = lib.param(name+".parametrized_weights", lib.floatX(numpy.ones((num_dilation_layer,))))
for i in range(num_dilation_layer):
layer_out, skip_c = lib.ops.dil_conv_1D(
layer_out,
output_dim,
input_dim if i == 0 else output_dim,
2,
dilation = 2**i,
