def from_spec(cls, spec, model):
spec = spec.copy()
hparams = spec['hparams']
if 'use_chars_concat' in hparams and hparams['use_chars_concat']:
raise NotImplementedError("Support for use_chars_concat has been removed")
if 'sentence_max_len' not in hparams:
hparams['sentence_max_len'] = 300
if 'use_elmo' not in hparams:
hparams['use_elmo'] = False
if 'elmo_dropout' not in hparams:
hparams['elmo_dropout'] = 0.5
if 'use_bert' not in hparams:
hparams['use_bert'] = False
if 'use_bert_only' not in hparams:
hparams['use_bert_only'] = False
if 'predict_tags' not in hparams:
hparams['predict_tags'] = False
if 'bert_transliterate' not in hparams:
hparams['bert_transliterate'] = ""
spec['hparams'] = nkutil.HParams(**hparams)
res = cls(**spec)
if use_cuda:
res.cpu()
if not hparams['use_elmo']:
res.load_state_dict(model)
else:
state = {k: v for k, v in res.state_dict().items() if k not in model}
state.update(model)
res.load_state_dict(state)
if use_cuda:
res.cuda()
return res
def from_spec(cls, spec, model):
spec = spec.copy()
hparams = spec['hparams']
if 'sentence_max_len' not in hparams:
hparams['sentence_max_len'] = 300
if 'use_elmo' not in hparams:
hparams['use_elmo'] = False
if 'elmo_dropout' not in hparams:
hparams['elmo_dropout'] = 0.5
spec['hparams'] = nkutil.HParams(**hparams)
res = cls(**spec)
if use_cuda:
res.cpu()
if not hparams['use_elmo']:
res.load_state_dict(model)
else:
state = {
k: v
for k, v in res.state_dict().items() if k not in model
}
state.update(model)
res.load_state_dict(state)
if use_cuda:
res.cuda()
return res
def make_hparams():
return nkutil.HParams(
max_len_train=0, # no length limit
max_len_dev=0, # no length limit
sentence_max_len=300,
learning_rate=0.0008,
learning_rate_warmup_steps=160,
clip_grad_norm=0., #no clipping
step_decay=True, # note that disabling step decay is not implemented
step_decay_factor=0.5,
step_decay_patience=5,
max_consecutive_decays=3, # establishes a termination criterion
partitioned=True,
num_layers_position_only=0,
num_layers=8,
d_model=1024,
num_heads=8,
d_kv=64,
d_ff=2048,
d_label_hidden=250,
d_tag_hidden=250,
tag_loss_scale=5.0,
attention_dropout=0.2,
embedding_dropout=0.0,
relu_dropout=0.1,
residual_dropout=0.2,
use_tags=False,
use_words=False,
use_chars_lstm=False,
use_elmo=False,
use_bert=False,
use_bert_only=False,
predict_tags=False,
d_char_emb=32, # A larger value may be better for use_chars_lstm
tag_emb_dropout=0.2,
word_emb_dropout=0.4,
morpho_emb_dropout=0.2,
timing_dropout=0.0,
char_lstm_input_dropout=0.2,
elmo_dropout=0.5, # Note that this semi-stacks with morpho_emb_dropout!
bert_model="bert-base-uncased",
bert_do_lower_case=True,
bert_transliterate="",
use_syntactic=False,
embed_layer=-1,
word_level="first",
freeze_embeddings=True,
syntactic_checkpoint="checkpoints/model.pt",
)
def make_hparams():
return nkutil.HParams(
max_len_train=0, # no length limit
max_len_dev=0, # no length limit
sentence_max_len=300,
learning_rate=0.0008,
learning_rate_warmup_steps=160,
clip_grad_norm=0., #no clipping
step_decay=True, # note that disabling step decay is not implemented
step_decay_factor=0.5,
step_decay_patience=5,
max_consecutive_decays=3, # establishes a termination criterion
partitioned=True,
num_layers_position_only=0,
num_layers=8,
d_model=1024,
num_heads=8,
d_kv=64,
d_ff=2048,
d_label_hidden=250,
d_tag_hidden=250,
tag_loss_scale=5.0,
attention_dropout=0.2,
embedding_dropout=0.0,
relu_dropout=0.1,
residual_dropout=0.2,
use_tags=False,
use_words=False,
use_chars_lstm=False,
use_elmo=False,
use_bert=False,
use_bert_only=False,
predict_tags=False,
d_char_emb=32, # A larger value may be better for use_chars_lstm
tag_emb_dropout=0.2,
word_emb_dropout=0.4,
morpho_emb_dropout=0.2,
timing_dropout=0.0,
char_lstm_input_dropout=0.2,
elmo_dropout=0.5, # Note that this semi-stacks with morpho_emb_dropout!
bert_model="bert-base-uncased",
bert_do_lower_case=True,
bert_transliterate="",
zero_empty=False,
metric="dot",
batch_cky=False,
label_weights=False,
no_mlp=False,
use_label_weights=False,
# Integration strategy of retrieved labels
# - soft mixes in representation space
# - hard mixes in score space
integration="hard", # ["soft", "hard"]
)
def make_hparams():
return nkutil.HParams(
max_len_train=0, # no length limit
max_len_dev=0, # no length limit
sentence_max_len=300,
learning_rate=0.0008,
learning_rate_warmup_steps=160,
clip_grad_norm=0., #no clipping
step_decay=True, # note that disabling step decay is not implemented
step_decay_factor=0.5,
step_decay_patience=5,
partitioned=True,
num_layers_position_only=0,
num_layers=8,
d_model=1024,
num_heads=8,
d_kv=64,
d_ff=2048,
d_label_hidden=250,
attention_dropout=0.2,
embedding_dropout=0.0,
relu_dropout=0.1,
residual_dropout=0.2,
use_tags=False,
use_words=False,
use_chars_lstm=False,
use_chars_concat=False,
use_elmo=False,
d_char_emb=32, # A larger value may be better for use_chars_lstm
tag_emb_dropout=0.2,
word_emb_dropout=0.4,
morpho_emb_dropout=0.2,
timing_dropout=0.0,
char_lstm_input_dropout=0.2,
elmo_dropout=0.5, # Note that this semi-stacks with morpho_emb_dropout!
task='language_modeling',
fp16=False,
vocab='/private/home/yinhanliu/BiLM_Training/vocab-2016-09-10.txt',
path=
'/checkpoint/yinhanliu/20180915/fairseq-lstm-train/layers_6/32_0.001/checkpoint_best.pt',
sample_break_mode='eos',
raw_text=False,
data='/private/home/abaevski/data/gbw',
output_dictionary_size=-1,
character_embeddings=True,
)
def make_hparams():
return nkutil.HParams(
max_len_train=0, # no length limit
max_len_dev=0, # no length limit
sentence_max_len=300,
learning_rate=0.0008,
learning_rate_warmup_steps=160,
clip_grad_norm=0., #no clipping
step_decay=True, # note that disabling step decay is not implemented
step_decay_factor=0.5,
step_decay_patience=5,
partitioned=True,
num_layers_position_only=0,
num_layers=8,
d_model=1024,
num_heads=8,
d_kv=64,
d_ff=2048,
d_label_hidden=250,
attention_dropout=0.2,
embedding_dropout=0.0,
relu_dropout=0.1,
residual_dropout=0.2,
use_tags=False,
use_words=False,
use_chars_lstm=False,
use_chars_concat=False,
use_elmo=False,
d_char_emb=32, # A larger value may be better for use_chars_lstm
tag_emb_dropout=0.2,
word_emb_dropout=0.4,
morpho_emb_dropout=0.2,
timing_dropout=0.0,
char_lstm_input_dropout=0.2,
elmo_dropout=0.5, # Note that this semi-stacks with morpho_emb_dropout!
)
def create_hparams():
hparams = nkutil.HParams(
################################
# PolyPhonic Parameters #
################################
num_classes = 1665, ##？
class2idx = "./filelists/uni_class2idx.json",
merge_cedict = "./filelists/universal_cedict.json",
saved_model_path_poly = "./save/poly_only/97.98_model.pt",
saved_model_path_structure_poly = "./save/poly_only_syntax_frozen/97.16_model.pt",
train_file = "./filelists/train_polyphonic.sent",
train_label = "./filelists/train_polyphonic.lb",
val_file = "./filelists/dev_polyphonic.sent",
val_label = "./filelists/dev_polyphonic.lb",
test_file = "./filelists/test_polyphonic.sent",
test_label = "./filelists/test_polyphonic.lb",
poly_batch_size = 32,
poly_max_length = 512,
poly_epochs = 1500,
poly_lr = 5e-5,
use_output_mask = True,
# control whether use syntax structure information in TTS
poly_use_structure = True,
tts_use_structure = True,
encoder_input_dim=[812, 512, 512],
################################
# Experiment Parameters #
################################
epochs=500,
iters_per_checkpoint=500,
seed=4321, #?
dynamic_loss_scaling=True,
fp16_run=False, ##?
distributed_run=False, ##?
dist_backend="nccl",
dist_url="tcp://localhost:54321", ## str 这个URL指定了如何初始化互相通信的进程。
cudnn_enabled=True,
cudnn_benchmark=False,
ignore_layers=['embedding.weight'],
synth_batch_size = 1,
################################
# Data Parameters #
################################
# load_mel_from_disk=False,
load_mel_from_disk=True, ##? 人工切换
pretrain_model_path_structure = './models/ch_bert_bmes_dev=93.97.pt',
saved_model_path_sandhi_structure = './save/poly_tts_CNN_syntax_frozen/96.39_model.pt',
saved_model_path_sandhi = './save/poly_tts_CNN/96.84_model.pt',
training_files='filelists/bznsyp_character_audio_text_train_filelist.txt',
validation_files='filelists/bznsyp_character_audio_text_val_filelist.txt',
mel_training_files='filelists/mel-bznsyp_character_audio_text_train_filelist.txt',
mel_validation_files='filelists/mel-bznsyp_character_audio_text_val_filelist.txt',
polyphone_dict_files = 'filelists/polyphone_dict.json',
mask_dict_files = 'filelists/polyphone_mask.json',
text_cleaners=['english_cleaners'],
################################
# Audio Parameters #
################################
max_wav_value=32768.0,
sampling_rate=16000,
filter_length=1024,
# hop_length=256,
# win_length=1024,
hop_length = 200,
win_length = 800,
n_mel_channels=80,
mel_fmin=0.0,
mel_fmax=8000.0,
signal_normalization = True,
allow_clipping_in_normalization = True,
use_lws=False,
max_abs_value = 4.,
symmetric_mels = True,
min_level_db = -100,
ref_level_db = 20,
magnitude_power = 2.,
fmin = 55,
fmax = 7600,
power = 1.5,
griffin_lim_iters = 60,
preemphasize = True,
preemphasis = 0.97,
################################
# Model Parameters #
################################
n_yinsu_symbols=len(yinsu_symbols),
n_character_symbols=len(character_symbols),
n_pinyin_symbols=1665,
character_symbols_embedding_dim=512,
yinsu_symbols_embedding_dim=512,
structure_feature_dim=300,
# Encoder parameters
encoder_kernel_size=5,
encoder_n_convolutions=3,
# encoder_embedding_dim=1836, # = 1024 + 512 + 300
# encoder_embedding_dim=1324, # = 1024 + 300
encoder_embedding_dim=512, # 512
encoder_output_dim=[512, 512, 512],
# Decoder parameters
n_frames_per_step=1, # currently only 1 is supported
decoder_rnn_dim=1024,
prenet_dim=256,
max_decoder_steps=1000,
gate_threshold=0.5,
p_attention_dropout=0.1,
p_decoder_dropout=0.1,
# Attention parameters
attention_rnn_dim=1024,
attention_dim=128,
# Location Layer parameters
attention_location_n_filters=32,
attention_location_kernel_size=31,
# Mel-post processing network parameters
postnet_embedding_dim=512,
postnet_kernel_size=5,
postnet_n_convolutions=5,
################################
# Optimization Hyperparameters #
################################
use_saved_learning_rate=False,
learning_rate=1e-4,
weight_decay=1e-6,
grad_clip_thresh=1.0,
batch_size=32,
mask_padding=True, # set model's padded outputs to padded values
)
return hparams
# hparams = {
# "epochs": 500,
# "iters_per_checkpoint": 500,
# "seed": 1234,
# "dynamic_loss_scaling": True,
# "fp16_run": False,
# "distributed_run": False,
# "dist_backend": "nccl",
# "dist_url": 'tcp://localhost:54321',
# "cudnn_enabled": True,
# "cudnn_benchmark": False,
# "ignore_layers": ['embedding.weight'],
# "load_mel_from_disk": False,
# "training_files": 'filelists/bznsyp_character_audio_text_train_filelist.txt',
# "validation_files": 'filelists/bznsyp_character_audio_text_val_filelist.txt',
# "polyphone_dict_files": 'filelists/polyphone_dict.json',
# "mask_dict_files": 'filelists/polyphone_mask.json',
# "text_cleaners": ['english_cleaners'],
# "max_wav_value": 32768.0,
# "sampling_rate": 16000,
# "filter_length": 1024,
# # "hop_length": 256,
# # "win_length": 1024,
# "hop_length": 200,
# "win_length": 800,
# "n_mel_channels": 80,
# "mel_fmin": 0.0,
# "mel_fmax": 8000.0,
# "n_yinsu_symbols": len(yinsu_symbols),
# "n_character_symbols": len(character_symbols),
# "n_pinyin_symbols": len(pinyin_symbols),
# "character_symbols_embedding_dim": 512,
# "yinsu_symbols_embedding_dim": 512,
# "encoder_kernel_size": 5,
# "encoder_n_convolutions": 3,
# "encoder_embedding_dim": 512,
# "n_frames_per_step": 1, # currently only 1 is supported
# "decoder_rnn_dim": 1024,
# "prenet_dim": 256,
# "max_decoder_steps": 1000,
# "gate_threshold": 0.5,
# "p_attention_dropout": 0.1,
# "p_decoder_dropout": 0.1,
# "attention_rnn_dim": 1024,
# "attention_dim": 128,
# "attention_location_n_filters": 32,
# "attention_location_kernel_size": 31,
# "postnet_embedding_dim": 512,
# "postnet_kernel_size": 5,
# "postnet_n_convolutions": 5,
# "use_saved_learning_rate": False,
# "learning_rate": 1e-3,
# "weight_decay": 1e-6,
# "grad_clip_thresh": 1.0,
# "batch_size": 8,
# "mask_padding": True
# }
# hparams = tf.contrib.training.HParams(
# ################################
# # Experiment Parameters #
# ################################
# epochs=500,
# iters_per_checkpoint=500,
# seed=1234,
# dynamic_loss_scaling=True,
# fp16_run=False,
# distributed_run=False,
# dist_backend="nccl",
# dist_url="tcp://localhost:54321",
# cudnn_enabled=True,
# cudnn_benchmark=False,
# ignore_layers=['embedding.weight'],
# ################################
# # Data Parameters #
# ################################
# load_mel_from_disk=False,
# training_files='filelists/bznsyp_character_audio_text_train_filelist.txt',
# validation_files='filelists/bznsyp_character_audio_text_val_filelist.txt',
# polyphone_dict_files = 'filelists/polyphone_dict.json',
# mask_dict_files = 'filelists/polyphone_mask.json',
# text_cleaners=['english_cleaners'],
# ################################
# # Audio Parameters #
# ################################
# max_wav_value=32768.0,
# sampling_rate=16000,
# filter_length=1024,
# # hop_length=256,
# # win_length=1024,
# hop_length = 200,
# win_length = 800,
# n_mel_channels=80,
# mel_fmin=0.0,
# mel_fmax=8000.0,
# ################################
# # Model Parameters #
# ################################
# n_yinsu_symbols=len(yinsu_symbols),
# n_character_symbols=len(character_symbols),
# n_pinyin_symbols=len(pinyin_symbols),
# character_symbols_embedding_dim=512,
# yinsu_symbols_embedding_dim=512,
# # Encoder parameters
# encoder_kernel_size=5,
# encoder_n_convolutions=3,
# # encoder_embedding_dim=1836, # = 1024 + 512 + 300
# # encoder_embedding_dim=1324, # = 1024 + 300
# encoder_embedding_dim=512, # = 1024 + 300 + 512
# # Decoder parameters
# n_frames_per_step=1, # currently only 1 is supported
# decoder_rnn_dim=1024,
# prenet_dim=256,
# max_decoder_steps=1000,
# gate_threshold=0.5,
# p_attention_dropout=0.1,
# p_decoder_dropout=0.1,
# # Attention parameters
# attention_rnn_dim=1024,
# attention_dim=128,
# # Location Layer parameters
# attention_location_n_filters=32,
# attention_location_kernel_size=31,
# # Mel-post processing network parameters
# postnet_embedding_dim=512,
# postnet_kernel_size=5,
# postnet_n_convolutions=5,
# ################################
# # Optimization Hyperparameters #
# ################################
# use_saved_learning_rate=False,
# learning_rate=1e-3,
# weight_decay=1e-6,
# grad_clip_thresh=1.0,
# batch_size=8,
# mask_padding=True # set model's padded outputs to padded values
# )
