# set log level to debug
tf.sg_verbosity(10)
#
# hyper parameters
#
batch_size = 16 # total batch size
#
# inputs
#
# corpus input tensor
data = SpeechCorpus(batch_size=batch_size * tf.sg_gpus())
# mfcc feature of audio
inputs = tf.split(data.mfcc, tf.sg_gpus(), axis=0)
# target sentence label
labels = tf.split(data.label, tf.sg_gpus(), axis=0)
# sequence length except zero-padding
seq_len = []
for input_ in inputs:
seq_len.append(
tf.not_equal(input_.sg_sum(axis=2), 0.).sg_int().sg_sum(axis=1))
# parallel loss tower
@tf.sg_parallel
import sugartensor as tf
__author__ = '*****@*****.**'
# set log level to debug
tf.sg_verbosity(10)
# batch size
batch_size = 128
# MNIST input tensor ( batch size should be adjusted for multiple GPUS )
data = tf.sg_data.Mnist(batch_size=batch_size * tf.sg_gpus())
# split inputs for each GPU tower
inputs = tf.split(data.train.image, tf.sg_gpus(), axis=0)
labels = tf.split(data.train.label, tf.sg_gpus(), axis=0)
# simple wrapping function with decorator for parallel training
@tf.sg_parallel
def get_loss(opt):
# conv layers
with tf.sg_context(name='convs', act='relu', bn=True):
conv = (opt.input[opt.gpu_index].sg_conv(
dim=16, name='conv1').sg_pool().sg_conv(
dim=32,
name='conv2').sg_pool().sg_conv(dim=32,
name='conv3').sg_pool())
# fc layers
# set log level to debug
tf.sg_verbosity(10)
#
# hyper parameters
#
batch_size = 16 # total batch size
#
# inputs
#
# corpus input tensor
data = SpeechCorpus(batch_size=batch_size * tf.sg_gpus())
# mfcc feature of audio
inputs = tf.split(data.mfcc, tf.sg_gpus(), axis=0)
# target sentence label
labels = tf.split(data.label, tf.sg_gpus(), axis=0)
# sequence length except zero-padding
seq_len = []
for input_ in inputs:
seq_len.append(tf.not_equal(input_.sg_sum(axis=2), 0.).sg_int().sg_sum(axis=1))
# parallel loss tower
@tf.sg_parallel
def get_loss(opt):
# set log level to debug tf.sg_verbosity(10) # # hyper parameters # batch_size = 1 # total batch size # # inputs # # corpus input tensor data = SpeechCorpus(batch_size=batch_size * tf.sg_gpus()) # mfcc feature of audio x = data.mfcc # target sentence label y = data.label seq_len = tf.not_equal(x.sg_sum(axis=2), 0.).sg_int().sg_sum(axis=1) # encode audio feature logit = get_logit(x, voca_size=voca_size) # CTC loss loss = logit.sg_ctc(target=y, seq_len=seq_len) decoded_sequence, _ = tf.nn.ctc_beam_search_decoder(