parent
3fc94427db
commit
70a343a499
@ -0,0 +1,94 @@
|
|||||||
|
import paddle.v2 as paddle
|
||||||
|
import audio_data_utils
|
||||||
|
import argparse
|
||||||
|
from model import deep_speech2
|
||||||
|
import gzip
|
||||||
|
from itertools import groupby
|
||||||
|
|
||||||
|
parser = argparse.ArgumentParser(
|
||||||
|
description='Simpled version of DeepSpeech2 inference.')
|
||||||
|
parser.add_argument(
|
||||||
|
"--num_samples", default=10, type=int, help="Number of inference samples.")
|
||||||
|
parser.add_argument(
|
||||||
|
"--num_conv_layers", default=2, type=int, help="Convolution layer number.")
|
||||||
|
parser.add_argument(
|
||||||
|
"--num_rnn_layers", default=3, type=int, help="RNN layer number.")
|
||||||
|
parser.add_argument(
|
||||||
|
"--rnn_layer_size", default=512, type=int, help="RNN layer cell number.")
|
||||||
|
parser.add_argument(
|
||||||
|
"--use_gpu", default=True, type=bool, help="Use gpu or not.")
|
||||||
|
args = parser.parse_args()
|
||||||
|
|
||||||
|
|
||||||
|
def remove_duplicate_and_blank(id_list, blank_id):
|
||||||
|
# remove consecutive duplicate tokens
|
||||||
|
id_list = [x[0] for x in groupby(id_list)]
|
||||||
|
# remove blank
|
||||||
|
return [id for id in id_list if id != blank_id]
|
||||||
|
|
||||||
|
|
||||||
|
def max_infer():
|
||||||
|
# create network config
|
||||||
|
_, vocab_list = audio_data_utils.get_vocabulary()
|
||||||
|
dict_size = len(vocab_list)
|
||||||
|
audio_data = paddle.layer.data(
|
||||||
|
name="audio_spectrogram",
|
||||||
|
height=161,
|
||||||
|
width=1000,
|
||||||
|
type=paddle.data_type.dense_vector(161000))
|
||||||
|
text_data = paddle.layer.data(
|
||||||
|
name="transcript_text",
|
||||||
|
type=paddle.data_type.integer_value_sequence(dict_size))
|
||||||
|
_, max_id = deep_speech2(
|
||||||
|
audio_data=audio_data,
|
||||||
|
text_data=text_data,
|
||||||
|
dict_size=dict_size,
|
||||||
|
num_conv_layers=args.num_conv_layers,
|
||||||
|
num_rnn_layers=args.num_rnn_layers,
|
||||||
|
rnn_size=args.rnn_layer_size)
|
||||||
|
|
||||||
|
# load parameters
|
||||||
|
parameters = paddle.parameters.Parameters.from_tar(
|
||||||
|
gzip.open("params.tar.gz"))
|
||||||
|
|
||||||
|
# prepare infer data
|
||||||
|
feeding = {
|
||||||
|
"audio_spectrogram": 0,
|
||||||
|
"transcript_text": 1,
|
||||||
|
}
|
||||||
|
test_batch_reader = audio_data_utils.padding_batch_reader(
|
||||||
|
paddle.batch(
|
||||||
|
audio_data_utils.reader_creator(
|
||||||
|
manifest_path="./libri.manifest.test", sort_by_duration=False),
|
||||||
|
batch_size=args.num_samples),
|
||||||
|
padding=[-1, 1000])
|
||||||
|
infer_data = test_batch_reader().next()
|
||||||
|
|
||||||
|
# run inference
|
||||||
|
max_id_results = paddle.infer(
|
||||||
|
output_layer=max_id,
|
||||||
|
parameters=parameters,
|
||||||
|
input=infer_data,
|
||||||
|
field=['id'])
|
||||||
|
|
||||||
|
# postprocess
|
||||||
|
instance_length = len(max_id_results) / args.num_samples
|
||||||
|
instance_list = [
|
||||||
|
max_id_results[i:i + instance_length]
|
||||||
|
for i in xrange(0, args.num_samples)
|
||||||
|
]
|
||||||
|
for i, instance in enumerate(instance_list):
|
||||||
|
id_list = remove_duplicate_and_blank(instance, dict_size)
|
||||||
|
output_transcript = ''.join([vocab_list[id] for id in id_list])
|
||||||
|
target_transcript = ''.join([vocab_list[id] for id in infer_data[i][1]])
|
||||||
|
print("Target Transcript: %s \nOutput Transcript: %s \n" %
|
||||||
|
(target_transcript, output_transcript))
|
||||||
|
|
||||||
|
|
||||||
|
def main():
|
||||||
|
paddle.init(use_gpu=args.use_gpu, trainer_count=1)
|
||||||
|
max_infer()
|
||||||
|
|
||||||
|
|
||||||
|
if __name__ == '__main__':
|
||||||
|
main()
|
@ -0,0 +1,106 @@
|
|||||||
|
import paddle.v2 as paddle
|
||||||
|
|
||||||
|
|
||||||
|
def conv_bn_layer(input, filter_size, num_channels_in, num_channels_out, stride,
|
||||||
|
padding, act):
|
||||||
|
conv_layer = paddle.layer.img_conv(
|
||||||
|
input=input,
|
||||||
|
filter_size=filter_size,
|
||||||
|
num_channels=num_channels_in,
|
||||||
|
num_filters=num_channels_out,
|
||||||
|
stride=stride,
|
||||||
|
padding=padding,
|
||||||
|
act=paddle.activation.Linear(),
|
||||||
|
bias_attr=False)
|
||||||
|
return paddle.layer.batch_norm(input=conv_layer, act=act)
|
||||||
|
|
||||||
|
|
||||||
|
def bidirectonal_simple_rnn_bn_layer(name, input, size, act):
|
||||||
|
def __simple_rnn_step__(input):
|
||||||
|
last_state = paddle.layer.memory(name=name + "_state", size=size)
|
||||||
|
input_fc = paddle.layer.fc(
|
||||||
|
input=input,
|
||||||
|
size=size,
|
||||||
|
act=paddle.activation.Linear(),
|
||||||
|
bias_attr=False)
|
||||||
|
input_fc_bn = paddle.layer.batch_norm(
|
||||||
|
input=input_fc, act=paddle.activation.Linear())
|
||||||
|
state_fc = paddle.layer.fc(
|
||||||
|
input=last_state,
|
||||||
|
size=size,
|
||||||
|
act=paddle.activation.Linear(),
|
||||||
|
bias_attr=False)
|
||||||
|
return paddle.layer.addto(
|
||||||
|
name=name + "_state", input=[input_fc_bn, state_fc], act=act)
|
||||||
|
|
||||||
|
forward = paddle.layer.recurrent_group(
|
||||||
|
step=__simple_rnn_step__, input=input)
|
||||||
|
return forward
|
||||||
|
# argument reverse is not exposed in V2 recurrent_group
|
||||||
|
#backward = paddle.layer.recurrent_group(
|
||||||
|
|
||||||
|
|
||||||
|
#step=__simple_rnn_step__,
|
||||||
|
#input=input,
|
||||||
|
#reverse=True)
|
||||||
|
#return paddle.layer.concat(input=[forward, backward])
|
||||||
|
|
||||||
|
|
||||||
|
def conv_group(input, num_stacks):
|
||||||
|
conv = conv_bn_layer(
|
||||||
|
input=input,
|
||||||
|
filter_size=(11, 41),
|
||||||
|
num_channels_in=1,
|
||||||
|
num_channels_out=32,
|
||||||
|
stride=(3, 2),
|
||||||
|
padding=(5, 20),
|
||||||
|
act=paddle.activation.BRelu())
|
||||||
|
for i in xrange(num_stacks - 1):
|
||||||
|
conv = conv_bn_layer(
|
||||||
|
input=conv,
|
||||||
|
filter_size=(11, 21),
|
||||||
|
num_channels_in=32,
|
||||||
|
num_channels_out=32,
|
||||||
|
stride=(1, 2),
|
||||||
|
padding=(5, 10),
|
||||||
|
act=paddle.activation.BRelu())
|
||||||
|
return conv
|
||||||
|
|
||||||
|
|
||||||
|
def rnn_group(input, size, num_stacks):
|
||||||
|
output = input
|
||||||
|
for i in xrange(num_stacks):
|
||||||
|
output = bidirectonal_simple_rnn_bn_layer(
|
||||||
|
name=str(i), input=output, size=size, act=paddle.activation.BRelu())
|
||||||
|
return output
|
||||||
|
|
||||||
|
|
||||||
|
def deep_speech2(audio_data,
|
||||||
|
text_data,
|
||||||
|
dict_size,
|
||||||
|
num_conv_layers=2,
|
||||||
|
num_rnn_layers=3,
|
||||||
|
rnn_size=256):
|
||||||
|
conv_group_output = conv_group(input=audio_data, num_stacks=num_conv_layers)
|
||||||
|
conv2seq = paddle.layer.block_expand(
|
||||||
|
input=conv_group_output,
|
||||||
|
num_channels=32,
|
||||||
|
stride_x=1,
|
||||||
|
stride_y=1,
|
||||||
|
block_x=1,
|
||||||
|
block_y=21)
|
||||||
|
rnn_group_output = rnn_group(
|
||||||
|
input=conv2seq, size=rnn_size, num_stacks=num_rnn_layers)
|
||||||
|
fc = paddle.layer.fc(
|
||||||
|
input=rnn_group_output,
|
||||||
|
size=dict_size + 1,
|
||||||
|
act=paddle.activation.Linear(),
|
||||||
|
bias_attr=True)
|
||||||
|
cost = paddle.layer.warp_ctc(
|
||||||
|
input=fc,
|
||||||
|
label=text_data,
|
||||||
|
size=dict_size + 1,
|
||||||
|
blank=dict_size,
|
||||||
|
norm_by_times=True)
|
||||||
|
max_id = paddle.layer.max_id(input=fc)
|
||||||
|
return cost, max_id
|
@ -1,5 +1,5 @@
|
|||||||
pip install wget
|
pip install wget
|
||||||
pip install soundfile
|
pip install soundfile
|
||||||
|
|
||||||
# For Linux only
|
# For Ubuntu only
|
||||||
apt-get install libsndfile1
|
apt-get install libsndfile1
|
||||||
|
Loading…
Reference in new issue