def train():
"""DeepSpeech2 training."""
# check if set use_gpu=True in paddlepaddle cpu version
check_cuda(args.use_gpu)
# check if paddlepaddle version is satisfied
check_version()
if args.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
train_generator = DataGenerator(vocab_filepath=args.vocab_path,
mean_std_filepath=args.mean_std_path,
augmentation_config=io.open(args.augment_conf_path, mode='r', encoding='utf8').read(),
max_duration=args.max_duration,
min_duration=args.min_duration,
specgram_type=args.specgram_type,
place=place)
dev_generator = DataGenerator(vocab_filepath=args.vocab_path,
mean_std_filepath=args.mean_std_path,
augmentation_config="{}",
specgram_type=args.specgram_type,
place=place)
train_batch_reader = train_generator.batch_reader_creator(manifest_path=args.train_manifest,
batch_size=args.batch_size,
sortagrad=args.use_sortagrad if args.init_from_pretrained_model is None else False,
shuffle_method=args.shuffle_method)
dev_batch_reader = dev_generator.batch_reader_creator(manifest_path=args.dev_manifest,
batch_size=args.batch_size,
sortagrad=False,
shuffle_method=None)
ds2_model = DeepSpeech2Model(vocab_size=train_generator.vocab_size,
num_conv_layers=args.num_conv_layers,
num_rnn_layers=args.num_rnn_layers,
rnn_layer_size=args.rnn_layer_size,
use_gru=args.use_gru,
share_rnn_weights=args.share_rnn_weights,
place=place,
init_from_pretrained_model=args.init_from_pretrained_model,
output_model_dir=args.output_model_dir)
ds2_model.train(train_batch_reader=train_batch_reader,
dev_batch_reader=dev_batch_reader,
learning_rate=args.learning_rate,
gradient_clipping=400,
batch_size=args.batch_size,
num_samples=args.num_samples,
num_epoch=args.num_epoch,
save_epoch=args.save_epoch,
num_iterations_print=args.num_iter_print,
test_off=args.test_off)
def load_model():
# check if set use_gpu=True in paddlepaddle cpu version
check_cuda(args.use_gpu)
# check if paddlepaddle version is satisfied
check_version()
if args.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
# Load model
data_generator = DataGenerator(
vocab_filepath=args.vocab_path,
mean_std_filepath=args.mean_std_path,
augmentation_config='{}',
specgram_type=args.specgram_type,
keep_transcription_text=True,
place = place,
is_training = False)
ds2_model = DeepSpeech2Model(
vocab_size=data_generator.vocab_size,
num_conv_layers=args.num_conv_layers,
num_rnn_layers=args.num_rnn_layers,
rnn_layer_size=args.rnn_layer_size,
use_gru=args.use_gru,
share_rnn_weights=args.share_rnn_weights,
place=place,
init_from_pretrained_model=args.model_path)
# decoders only accept string encoded in utf-8
vocab_list = data_generator.vocab_list
ds2_model.init_ext_scorer(args.alpha, args.beta, args.lang_model_path,
vocab_list)
return ds2_model, data_generator, vocab_list
def evaluate():
"""Evaluate on whole test data for DeepSpeech2."""
# check if set use_gpu=True in paddlepaddle cpu version
check_cuda(args.use_gpu)
# check if paddlepaddle version is satisfied
check_version()
if args.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
data_generator = DataGenerator(
vocab_filepath=args.vocab_path,
mean_std_filepath=args.mean_std_path,
augmentation_config='{}',
specgram_type=args.specgram_type,
keep_transcription_text=True,
place = place,
is_training = False)
batch_reader = data_generator.batch_reader_creator(
manifest_path=args.test_manifest,
batch_size=args.batch_size,
sortagrad=False,
shuffle_method=None)
ds2_model = DeepSpeech2Model(
vocab_size=data_generator.vocab_size,
num_conv_layers=args.num_conv_layers,
num_rnn_layers=args.num_rnn_layers,
rnn_layer_size=args.rnn_layer_size,
use_gru=args.use_gru,
share_rnn_weights=args.share_rnn_weights,
place=place,
init_from_pretrained_model=args.model_path)
# decoders only accept string encoded in utf-8
vocab_list = [chars for chars in data_generator.vocab_list]
if args.decoding_method == "ctc_beam_search":
ds2_model.init_ext_scorer(args.alpha, args.beta, args.lang_model_path,
vocab_list)
errors_func = char_errors if args.error_rate_type == 'cer' else word_errors
errors_sum, len_refs, num_ins = 0.0, 0, 0
ds2_model.logger.info("start evaluation ...")
for infer_data in batch_reader():
probs_split = ds2_model.infer_batch_probs(
infer_data=infer_data,
feeding_dict=data_generator.feeding)
if args.decoding_method == "ctc_greedy":
result_transcripts = ds2_model.decode_batch_greedy(
probs_split=probs_split,
vocab_list=vocab_list)
else:
result_transcripts = ds2_model.decode_batch_beam_search(
probs_split=probs_split,
beam_alpha=args.alpha,
beam_beta=args.beta,
beam_size=args.beam_size,
cutoff_prob=args.cutoff_prob,
cutoff_top_n=args.cutoff_top_n,
vocab_list=vocab_list,
num_processes=args.num_proc_bsearch)
target_transcripts = infer_data[1]
for target, result in zip(target_transcripts, result_transcripts):
errors, len_ref = errors_func(target, result)
errors_sum += errors
len_refs += len_ref
num_ins += 1
print("Error rate [%s] (%d/?) = %f" %
(args.error_rate_type, num_ins, errors_sum / len_refs))
print("Final error rate [%s] (%d/%d) = %f" %
(args.error_rate_type, num_ins, num_ins, errors_sum / len_refs))
ds2_model.logger.info("finish evaluation")
def evaluate():
# 检测PaddlePaddle环境
check_cuda(args.use_gpu)
check_version()
# 是否使用GPU
place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
# 获取数据生成器
data_generator = DataGenerator(vocab_filepath=args.vocab_path,
mean_std_filepath=args.mean_std_path,
augmentation_config='{}',
specgram_type=args.specgram_type,
keep_transcription_text=True,
place=place,
is_training=False)
# 获取评估数据
batch_reader = data_generator.batch_reader_creator(
manifest_path=args.test_manifest,
batch_size=args.batch_size,
shuffle_method=None)
# 获取DeepSpeech2模型，并设置为预测
ds2_model = DeepSpeech2Model(vocab_size=data_generator.vocab_size,
num_conv_layers=args.num_conv_layers,
num_rnn_layers=args.num_rnn_layers,
rnn_layer_size=args.rnn_layer_size,
use_gru=args.use_gru,
share_rnn_weights=args.share_rnn_weights,
place=place,
init_from_pretrained_model=args.model_path,
is_infer=True)
# 读取数据列表
with open(args.test_manifest, 'r', encoding='utf-8') as f_m:
test_len = len(f_m.readlines())
# 定向搜索方法的处理
if args.decoding_method == "ctc_beam_search":
ds2_model.init_ext_scorer(args.alpha, args.beta, args.lang_model_path,
data_generator.vocab_list)
# 获取评估函数，有字错率和词错率
errors_func = char_errors if args.error_rate_type == 'cer' else word_errors
errors_sum, len_refs, num_ins = 0.0, 0, 0
ds2_model.logger.info("开始评估 ...")
start = time.time()
# 开始评估
for infer_data in batch_reader():
# 获取一批的识别结果
probs_split = ds2_model.infer_batch_probs(infer_data=infer_data)
# 执行解码
if args.decoding_method == "ctc_greedy":
# 最优路径解码
result_transcripts = ds2_model.decode_batch_greedy(
probs_split=probs_split, vocab_list=data_generator.vocab_list)
else:
# 定向搜索解码
result_transcripts = ds2_model.decode_batch_beam_search(
probs_split=probs_split,
beam_alpha=args.alpha,
beam_beta=args.beta,
beam_size=args.beam_size,
cutoff_prob=args.cutoff_prob,
cutoff_top_n=args.cutoff_top_n,
vocab_list=data_generator.vocab_list,
num_processes=args.num_proc_bsearch)
target_transcripts = infer_data[1]
# 计算字错率
for target, result in zip(target_transcripts, result_transcripts):
errors, len_ref = errors_func(target, result)
errors_sum += errors
len_refs += len_ref
num_ins += 1
print("错误率：[%s] (%d/%d) = %f" %
(args.error_rate_type, num_ins, test_len, errors_sum / len_refs))
end = time.time()
print("消耗时间：%ds, 总错误率：[%s] (%d/%d) = %f" %
((end - start), args.error_rate_type, num_ins, num_ins,
errors_sum / len_refs))
ds2_model.logger.info("完成评估！")
def train():
# 检测PaddlePaddle环境
check_cuda(args.use_gpu)
check_version()
# 是否使用GPU
if args.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
# 获取训练数据生成器
train_generator = DataGenerator(vocab_filepath=args.vocab_path,
mean_std_filepath=args.mean_std_path,
augmentation_config=io.open(
args.augment_conf_path,
mode='r',
encoding='utf8').read(),
max_duration=args.max_duration,
min_duration=args.min_duration,
specgram_type=args.specgram_type,
place=place)
# 获取测试数据生成器
test_generator = DataGenerator(vocab_filepath=args.vocab_path,
mean_std_filepath=args.mean_std_path,
augmentation_config="{}",
specgram_type=args.specgram_type,
keep_transcription_text=True,
place=place,
is_training=False)
# 获取训练数据
train_batch_reader = train_generator.batch_reader_creator(
manifest_path=args.train_manifest,
batch_size=args.batch_size,
sortagrad=args.use_sortagrad
if args.init_from_pretrained_model is None else False,
shuffle_method=args.shuffle_method)
# 获取测试数据
test_batch_reader = test_generator.batch_reader_creator(
manifest_path=args.dev_manifest,
batch_size=args.batch_size,
sortagrad=False,
shuffle_method=None)
# 获取DeepSpeech2模型
ds2_model = DeepSpeech2Model(
vocab_size=train_generator.vocab_size,
num_conv_layers=args.num_conv_layers,
num_rnn_layers=args.num_rnn_layers,
rnn_layer_size=args.rnn_layer_size,
use_gru=args.use_gru,
share_rnn_weights=args.share_rnn_weights,
place=place,
init_from_pretrained_model=args.init_from_pretrained_model,
output_model_dir=args.output_model_dir,
vocab_list=test_generator.vocab_list)
# 获取训练数据数量
num_samples = get_data_len(args.train_manifest, args.max_duration,
args.min_duration)
print("[%s] 训练数据数量：%d\n" % (datetime.now(), num_samples))
# 开始训练
ds2_model.train(train_batch_reader=train_batch_reader,
dev_batch_reader=test_batch_reader,
learning_rate=args.learning_rate,
gradient_clipping=400,
batch_size=args.batch_size,
num_samples=num_samples,
num_epoch=args.num_epoch,
save_epoch=args.save_epoch,
num_iterations_print=args.num_iter_print,
test_off=args.test_off)
def infer():
"""Inference for DeepSpeech2."""
# check if set use_gpu=True in paddlepaddle cpu version
check_cuda(args.use_gpu)
# check if paddlepaddle version is satisfied
check_version()
if args.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
data_generator = DataGenerator(
vocab_filepath=args.vocab_path,
mean_std_filepath=args.mean_std_path,
augmentation_config='{}',
specgram_type=args.specgram_type,
keep_transcription_text=True,
place = place,
is_training = False)
batch_reader = data_generator.batch_reader_creator(
manifest_path=args.infer_manifest,
batch_size=args.num_samples,
sortagrad=False,
shuffle_method=None)
infer_data = next(batch_reader())
ds2_model = DeepSpeech2Model(
vocab_size=data_generator.vocab_size,
num_conv_layers=args.num_conv_layers,
num_rnn_layers=args.num_rnn_layers,
rnn_layer_size=args.rnn_layer_size,
use_gru=args.use_gru,
share_rnn_weights=args.share_rnn_weights,
place=place,
init_from_pretrained_model=args.model_path)
# decoders only accept string encoded in utf-8
vocab_list = [chars.encode("utf-8") for chars in data_generator.vocab_list]
if args.decoding_method == "ctc_greedy":
ds2_model.logger.info("start inference ...")
probs_split = ds2_model.infer_batch_probs(
infer_data=infer_data,
feeding_dict=data_generator.feeding)
result_transcripts = ds2_model.decode_batch_greedy(
probs_split=probs_split,
vocab_list=vocab_list)
else:
ds2_model.init_ext_scorer(args.alpha, args.beta, args.lang_model_path,
vocab_list)
ds2_model.logger.info("start inference ...")
probs_split= ds2_model.infer_batch_probs(
infer_data=infer_data,
feeding_dict=data_generator.feeding)
result_transcripts= ds2_model.decode_batch_beam_search(
probs_split=probs_split,
beam_alpha=args.alpha,
beam_beta=args.beta,
beam_size=args.beam_size,
cutoff_prob=args.cutoff_prob,
cutoff_top_n=args.cutoff_top_n,
vocab_list=vocab_list,
num_processes=args.num_proc_bsearch)
error_rate_func = cer if args.error_rate_type == 'cer' else wer
target_transcripts = infer_data[1]
for target, result in zip(target_transcripts, result_transcripts):
print("\nTarget Transcription: %s\nOutput Transcription: %s" %
(target, result))
print("Current error rate [%s] = %f" %
(args.error_rate_type, error_rate_func(target, result)))
ds2_model.logger.info("finish inference")
def infer(transcript_name):
"""Inference for DeepSpeech2."""
# check if set use_gpu=True in paddlepaddle cpu version
check_cuda(args.use_gpu)
# check if paddlepaddle version is satisfied
check_version()
if args.use_gpu:
place = fluid.CUDAPlace(0)
else:
place = fluid.CPUPlace()
data_generator = DataGenerator(
vocab_filepath=args.vocab_path,
mean_std_filepath=args.mean_std_path,
augmentation_config='{}',
specgram_type=args.specgram_type,
keep_transcription_text=True,
place = place,
is_training = False)
batch_reader = data_generator.batch_reader_creator(
manifest_path=args.infer_manifest,
batch_size=args.num_samples,
sortagrad=False,
shuffle_method=None)
infer_data = next(batch_reader())
ds2_model = DeepSpeech2Model(
vocab_size=data_generator.vocab_size,
num_conv_layers=args.num_conv_layers,
num_rnn_layers=args.num_rnn_layers,
rnn_layer_size=args.rnn_layer_size,
use_gru=args.use_gru,
share_rnn_weights=args.share_rnn_weights,
place=place,
init_from_pretrained_model=args.model_path)
# decoders only accept string encoded in utf-8
vocab_list = [chars.encode("utf-8") for chars in data_generator.vocab_list]
if args.decoding_method == "ctc_greedy":
ds2_model.logger.info("start inference ...")
probs_split = ds2_model.infer_batch_probs(
infer_data=infer_data,
feeding_dict=data_generator.feeding)
result_transcripts = ds2_model.decode_batch_greedy(
probs_split=probs_split,
vocab_list=vocab_list)
else:
ds2_model.init_ext_scorer(args.alpha, args.beta, args.lang_model_path,
vocab_list)
ds2_model.logger.info("start inference ...")
probs_split= ds2_model.infer_batch_probs(
infer_data=infer_data,
feeding_dict=data_generator.feeding)
result_transcripts= ds2_model.decode_batch_beam_search(
probs_split=probs_split,
beam_alpha=args.alpha,
beam_beta=args.beta,
beam_size=args.beam_size,
cutoff_prob=args.cutoff_prob,
cutoff_top_n=args.cutoff_top_n,
vocab_list=vocab_list,
num_processes=args.num_proc_bsearch)
transcription = result_transcripts[0].capitalize() + '.'
print(transcription)
with codecs.open('dataset/tap/transcription/'+transcript_name+'.txt', 'w', 'utf-8') as out_file:
out_file.write(transcription)
ds2_model.logger.info("finish inference")