From 1cfb26afc519a634194a4eaa02f9b0969f5c2cbf Mon Sep 17 00:00:00 2001
From: jmwang66 <wangjiaming.wjm@alibaba-inc.com>
Date: 星期三, 09 八月 2023 16:48:21 +0800
Subject: [PATCH] Merge pull request #790 from alibaba-damo-academy/dev_wjm_modelscope
---
funasr/datasets/small_datasets/sequence_iter_factory.py | 3
funasr/utils/prepare_data.py | 41 ++-
funasr/bin/build_trainer.py | 640 +++++++++++++++++++++++++++++++++++++++++++++++++---
3 files changed, 626 insertions(+), 58 deletions(-)
diff --git a/funasr/bin/build_trainer.py b/funasr/bin/build_trainer.py
index e7f28ed..52aa509 100644
--- a/funasr/bin/build_trainer.py
+++ b/funasr/bin/build_trainer.py
@@ -1,6 +1,35 @@
-import os
+# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
+# MIT License (https://opensource.org/licenses/MIT)
+import argparse
+import logging
+import os
+import sys
+from io import BytesIO
+
+import torch
import yaml
+
+from funasr.build_utils.build_args import build_args
+from funasr.build_utils.build_dataloader import build_dataloader
+from funasr.build_utils.build_distributed import build_distributed
+from funasr.build_utils.build_model import build_model
+from funasr.build_utils.build_optimizer import build_optimizer
+from funasr.build_utils.build_scheduler import build_scheduler
+from funasr.build_utils.build_trainer import build_trainer as build_trainer_modelscope
+from funasr.modules.lora.utils import mark_only_lora_as_trainable
+from funasr.text.phoneme_tokenizer import g2p_choices
+from funasr.torch_utils.load_pretrained_model import load_pretrained_model
+from funasr.torch_utils.model_summary import model_summary
+from funasr.torch_utils.pytorch_version import pytorch_cudnn_version
+from funasr.torch_utils.set_all_random_seed import set_all_random_seed
+from funasr.utils.nested_dict_action import NestedDictAction
+from funasr.utils.prepare_data import prepare_data
+from funasr.utils.types import int_or_none
+from funasr.utils.types import str2bool
+from funasr.utils.types import str_or_none
+from funasr.utils.yaml_no_alias_safe_dump import yaml_no_alias_safe_dump
+
def update_dct(fin_configs, root):
if root == {}:
@@ -17,26 +46,468 @@
return fin_configs
-def parse_args(mode):
- if mode == "asr":
- from funasr.tasks.asr import ASRTask as ASRTask
- elif mode == "paraformer":
- from funasr.tasks.asr import ASRTaskParaformer as ASRTask
- elif mode == "paraformer_streaming":
- from funasr.tasks.asr import ASRTaskParaformer as ASRTask
- elif mode == "paraformer_vad_punc":
- from funasr.tasks.asr import ASRTaskParaformer as ASRTask
- elif mode == "uniasr":
- from funasr.tasks.asr import ASRTaskUniASR as ASRTask
- elif mode == "mfcca":
- from funasr.tasks.asr import ASRTaskMFCCA as ASRTask
- elif mode == "tp":
- from funasr.tasks.asr import ASRTaskAligner as ASRTask
- else:
- raise ValueError("Unknown mode: {}".format(mode))
- parser = ASRTask.get_parser()
- args = parser.parse_args()
- return args, ASRTask
+def get_parser():
+ parser = argparse.ArgumentParser(
+ description="FunASR Common Training Parser",
+ )
+
+ # common configuration
+ parser.add_argument("--output_dir", help="model save path")
+ parser.add_argument(
+ "--ngpu",
+ type=int,
+ default=0,
+ help="The number of gpus. 0 indicates CPU mode",
+ )
+ parser.add_argument("--seed", type=int, default=0, help="Random seed")
+ parser.add_argument("--task_name", type=str, default="asr", help="Name for different tasks")
+
+ # ddp related
+ parser.add_argument(
+ "--dist_backend",
+ default="nccl",
+ type=str,
+ help="distributed backend",
+ )
+ parser.add_argument(
+ "--dist_init_method",
+ type=str,
+ default="env://",
+ help='if init_method="env://", env values of "MASTER_PORT", "MASTER_ADDR", '
+ '"WORLD_SIZE", and "RANK" are referred.',
+ )
+ parser.add_argument(
+ "--dist_world_size",
+ type=int,
+ default=1,
+ help="number of nodes for distributed training",
+ )
+ parser.add_argument(
+ "--dist_rank",
+ type=int,
+ default=None,
+ help="node rank for distributed training",
+ )
+ parser.add_argument(
+ "--local_rank",
+ type=int,
+ default=None,
+ help="local rank for distributed training",
+ )
+ parser.add_argument(
+ "--dist_master_addr",
+ default=None,
+ type=str_or_none,
+ help="The master address for distributed training. "
+ "This value is used when dist_init_method == 'env://'",
+ )
+ parser.add_argument(
+ "--dist_master_port",
+ default=None,
+ type=int_or_none,
+ help="The master port for distributed training"
+ "This value is used when dist_init_method == 'env://'",
+ )
+ parser.add_argument(
+ "--dist_launcher",
+ default=None,
+ type=str_or_none,
+ choices=["slurm", "mpi", None],
+ help="The launcher type for distributed training",
+ )
+ parser.add_argument(
+ "--multiprocessing_distributed",
+ default=True,
+ type=str2bool,
+ help="Use multi-processing distributed training to launch "
+ "N processes per node, which has N GPUs. This is the "
+ "fastest way to use PyTorch for either single node or "
+ "multi node data parallel training",
+ )
+ parser.add_argument(
+ "--unused_parameters",
+ type=str2bool,
+ default=False,
+ help="Whether to use the find_unused_parameters in "
+ "torch.nn.parallel.DistributedDataParallel ",
+ )
+ parser.add_argument(
+ "--gpu_id",
+ type=int,
+ default=0,
+ help="local gpu id.",
+ )
+
+ # cudnn related
+ parser.add_argument(
+ "--cudnn_enabled",
+ type=str2bool,
+ default=torch.backends.cudnn.enabled,
+ help="Enable CUDNN",
+ )
+ parser.add_argument(
+ "--cudnn_benchmark",
+ type=str2bool,
+ default=torch.backends.cudnn.benchmark,
+ help="Enable cudnn-benchmark mode",
+ )
+ parser.add_argument(
+ "--cudnn_deterministic",
+ type=str2bool,
+ default=True,
+ help="Enable cudnn-deterministic mode",
+ )
+
+ # trainer related
+ parser.add_argument(
+ "--max_epoch",
+ type=int,
+ default=40,
+ help="The maximum number epoch to train",
+ )
+ parser.add_argument(
+ "--max_update",
+ type=int,
+ default=sys.maxsize,
+ help="The maximum number update step to train",
+ )
+ parser.add_argument(
+ "--batch_interval",
+ type=int,
+ default=10000,
+ help="The batch interval for saving model.",
+ )
+ parser.add_argument(
+ "--patience",
+ type=int_or_none,
+ default=None,
+ help="Number of epochs to wait without improvement "
+ "before stopping the training",
+ )
+ parser.add_argument(
+ "--val_scheduler_criterion",
+ type=str,
+ nargs=2,
+ default=("valid", "loss"),
+ help="The criterion used for the value given to the lr scheduler. "
+ 'Give a pair referring the phase, "train" or "valid",'
+ 'and the criterion name. The mode specifying "min" or "max" can '
+ "be changed by --scheduler_conf",
+ )
+ parser.add_argument(
+ "--early_stopping_criterion",
+ type=str,
+ nargs=3,
+ default=("valid", "loss", "min"),
+ help="The criterion used for judging of early stopping. "
+ 'Give a pair referring the phase, "train" or "valid",'
+ 'the criterion name and the mode, "min" or "max", e.g. "acc,max".',
+ )
+ parser.add_argument(
+ "--best_model_criterion",
+ nargs="+",
+ default=[
+ ("train", "loss", "min"),
+ ("valid", "loss", "min"),
+ ("train", "acc", "max"),
+ ("valid", "acc", "max"),
+ ],
+ help="The criterion used for judging of the best model. "
+ 'Give a pair referring the phase, "train" or "valid",'
+ 'the criterion name, and the mode, "min" or "max", e.g. "acc,max".',
+ )
+ parser.add_argument(
+ "--keep_nbest_models",
+ type=int,
+ nargs="+",
+ default=[10],
+ help="Remove previous snapshots excluding the n-best scored epochs",
+ )
+ parser.add_argument(
+ "--nbest_averaging_interval",
+ type=int,
+ default=0,
+ help="The epoch interval to apply model averaging and save nbest models",
+ )
+ parser.add_argument(
+ "--grad_clip",
+ type=float,
+ default=5.0,
+ help="Gradient norm threshold to clip",
+ )
+ parser.add_argument(
+ "--grad_clip_type",
+ type=float,
+ default=2.0,
+ help="The type of the used p-norm for gradient clip. Can be inf",
+ )
+ parser.add_argument(
+ "--grad_noise",
+ type=str2bool,
+ default=False,
+ help="The flag to switch to use noise injection to "
+ "gradients during training",
+ )
+ parser.add_argument(
+ "--accum_grad",
+ type=int,
+ default=1,
+ help="The number of gradient accumulation",
+ )
+ parser.add_argument(
+ "--resume",
+ type=str2bool,
+ default=False,
+ help="Enable resuming if checkpoint is existing",
+ )
+ parser.add_argument(
+ "--train_dtype",
+ default="float32",
+ choices=["float16", "float32", "float64"],
+ help="Data type for training.",
+ )
+ parser.add_argument(
+ "--use_amp",
+ type=str2bool,
+ default=False,
+ help="Enable Automatic Mixed Precision. This feature requires pytorch>=1.6",
+ )
+ parser.add_argument(
+ "--log_interval",
+ default=None,
+ help="Show the logs every the number iterations in each epochs at the "
+ "training phase. If None is given, it is decided according the number "
+ "of training samples automatically .",
+ )
+ parser.add_argument(
+ "--use_tensorboard",
+ type=str2bool,
+ default=True,
+ help="Enable tensorboard logging",
+ )
+
+ # pretrained model related
+ parser.add_argument(
+ "--init_param",
+ type=str,
+ action="append",
+ default=[],
+ help="Specify the file path used for initialization of parameters. "
+ "The format is '<file_path>:<src_key>:<dst_key>:<exclude_keys>', "
+ "where file_path is the model file path, "
+ "src_key specifies the key of model states to be used in the model file, "
+ "dst_key specifies the attribute of the model to be initialized, "
+ "and exclude_keys excludes keys of model states for the initialization."
+ "e.g.\n"
+ " # Load all parameters"
+ " --init_param some/where/model.pb\n"
+ " # Load only decoder parameters"
+ " --init_param some/where/model.pb:decoder:decoder\n"
+ " # Load only decoder parameters excluding decoder.embed"
+ " --init_param some/where/model.pb:decoder:decoder:decoder.embed\n"
+ " --init_param some/where/model.pb:decoder:decoder:decoder.embed\n",
+ )
+ parser.add_argument(
+ "--ignore_init_mismatch",
+ type=str2bool,
+ default=False,
+ help="Ignore size mismatch when loading pre-trained model",
+ )
+ parser.add_argument(
+ "--freeze_param",
+ type=str,
+ default=[],
+ action="append",
+ help="Freeze parameters",
+ )
+
+ # dataset related
+ parser.add_argument(
+ "--dataset_type",
+ type=str,
+ default="small",
+ help="whether to use dataloader for large dataset",
+ )
+ parser.add_argument(
+ "--dataset_conf",
+ action=NestedDictAction,
+ default=dict(),
+ help=f"The keyword arguments for dataset",
+ )
+ parser.add_argument(
+ "--data_dir",
+ type=str,
+ default=None,
+ help="root path of data",
+ )
+ parser.add_argument(
+ "--train_set",
+ type=str,
+ default="train",
+ help="train dataset",
+ )
+ parser.add_argument(
+ "--valid_set",
+ type=str,
+ default="validation",
+ help="dev dataset",
+ )
+ parser.add_argument(
+ "--data_file_names",
+ type=str,
+ default="wav.scp,text",
+ help="input data files",
+ )
+ parser.add_argument(
+ "--speed_perturb",
+ type=float,
+ nargs="+",
+ default=None,
+ help="speed perturb",
+ )
+ parser.add_argument(
+ "--use_preprocessor",
+ type=str2bool,
+ default=True,
+ help="Apply preprocessing to data or not",
+ )
+
+ # optimization related
+ parser.add_argument(
+ "--optim",
+ type=lambda x: x.lower(),
+ default="adam",
+ help="The optimizer type",
+ )
+ parser.add_argument(
+ "--optim_conf",
+ action=NestedDictAction,
+ default=dict(),
+ help="The keyword arguments for optimizer",
+ )
+ parser.add_argument(
+ "--scheduler",
+ type=lambda x: str_or_none(x.lower()),
+ default=None,
+ help="The lr scheduler type",
+ )
+ parser.add_argument(
+ "--scheduler_conf",
+ action=NestedDictAction,
+ default=dict(),
+ help="The keyword arguments for lr scheduler",
+ )
+
+ # most task related
+ parser.add_argument(
+ "--init",
+ type=lambda x: str_or_none(x.lower()),
+ default=None,
+ help="The initialization method",
+ choices=[
+ "chainer",
+ "xavier_uniform",
+ "xavier_normal",
+ "kaiming_uniform",
+ "kaiming_normal",
+ None,
+ ],
+ )
+ parser.add_argument(
+ "--token_list",
+ type=str_or_none,
+ default=None,
+ help="A text mapping int-id to token",
+ )
+ parser.add_argument(
+ "--token_type",
+ type=str,
+ default="bpe",
+ choices=["bpe", "char", "word"],
+ help="",
+ )
+ parser.add_argument(
+ "--bpemodel",
+ type=str_or_none,
+ default=None,
+ help="The model file fo sentencepiece",
+ )
+ parser.add_argument(
+ "--cleaner",
+ type=str_or_none,
+ choices=[None, "tacotron", "jaconv", "vietnamese"],
+ default=None,
+ help="Apply text cleaning",
+ )
+ parser.add_argument(
+ "--g2p",
+ type=str_or_none,
+ choices=g2p_choices,
+ default=None,
+ help="Specify g2p method if --token_type=phn",
+ )
+
+ # pai related
+ parser.add_argument(
+ "--use_pai",
+ type=str2bool,
+ default=False,
+ help="flag to indicate whether training on PAI",
+ )
+ parser.add_argument(
+ "--simple_ddp",
+ type=str2bool,
+ default=False,
+ )
+ parser.add_argument(
+ "--num_worker_count",
+ type=int,
+ default=1,
+ help="The number of machines on PAI.",
+ )
+ parser.add_argument(
+ "--access_key_id",
+ type=str,
+ default=None,
+ help="The username for oss.",
+ )
+ parser.add_argument(
+ "--access_key_secret",
+ type=str,
+ default=None,
+ help="The password for oss.",
+ )
+ parser.add_argument(
+ "--endpoint",
+ type=str,
+ default=None,
+ help="The endpoint for oss.",
+ )
+ parser.add_argument(
+ "--bucket_name",
+ type=str,
+ default=None,
+ help="The bucket name for oss.",
+ )
+ parser.add_argument(
+ "--oss_bucket",
+ default=None,
+ help="oss bucket.",
+ )
+ parser.add_argument(
+ "--enable_lora",
+ type=str2bool,
+ default=False,
+ help="Apply lora for finetuning.",
+ )
+ parser.add_argument(
+ "--lora_bias",
+ type=str,
+ default="none",
+ help="lora bias.",
+ )
+
+ return parser
def build_trainer(modelscope_dict,
@@ -56,9 +527,10 @@
specaug_conf=None,
mate_params=None,
**kwargs):
- mode = modelscope_dict['mode']
- args, ASRTask = parse_args(mode=mode)
- # ddp related
+ parser = get_parser()
+ args, extra_task_params = parser.parse_known_args()
+ args = build_args(args, parser, extra_task_params)
+
if args.local_rank is not None:
distributed = True
else:
@@ -97,21 +569,9 @@
setattr(args, key, value)
if mate_params is not None and "lora_params" in mate_params:
lora_params = mate_params['lora_params']
- configs['encoder_conf'].update(lora_params)
- configs['decoder_conf'].update(lora_params)
-
- # prepare data
+ configs['encoder_conf'].update(lora_params)
+ configs['decoder_conf'].update(lora_params)
args.dataset_type = dataset_type
- if args.dataset_type == "small":
- args.train_data_path_and_name_and_type = [["{}/{}/wav.scp".format(data_dir, train_set), "speech", "sound"],
- ["{}/{}/text".format(data_dir, train_set), "text", "text"]]
- args.valid_data_path_and_name_and_type = [["{}/{}/wav.scp".format(data_dir, dev_set), "speech", "sound"],
- ["{}/{}/text".format(data_dir, dev_set), "text", "text"]]
- elif args.dataset_type == "large":
- args.train_data_file = None
- args.valid_data_file = None
- else:
- raise ValueError(f"Not supported dataset_type={args.dataset_type}")
args.init_param = [init_param]
if mate_params is not None and "init_param" in mate_params:
if len(mate_params["init_param"]) != 0:
@@ -127,6 +587,16 @@
args.output_dir = output_dir
args.gpu_id = args.local_rank
args.config = finetune_config
+ args.use_pai = False
+ args.batch_type = "length"
+ args.oss_bucket = None
+ args.input_size = None
+ if distributed:
+ args.distributed = True
+ args.simple_ddp = True
+ else:
+ args.distributed = False
+ args.ngpu = 1
if optim is not None:
args.optim = optim
if lr is not None:
@@ -144,6 +614,7 @@
if batch_bins is not None:
if args.dataset_type == "small":
args.batch_bins = batch_bins
+ args.dataset_conf["batch_conf"]["batch_size"] = batch_bins
elif args.dataset_type == "large":
args.dataset_conf["batch_conf"]["batch_size"] = batch_bins
else:
@@ -153,7 +624,94 @@
if args.patience in ["null", "none", "None"]:
args.patience = None
args.local_rank = local_rank
- args.distributed = distributed
- ASRTask.finetune_args = args
- return ASRTask
+ # set random seed
+ set_all_random_seed(args.seed)
+ torch.backends.cudnn.enabled = args.cudnn_enabled
+ torch.backends.cudnn.benchmark = args.cudnn_benchmark
+ torch.backends.cudnn.deterministic = args.cudnn_deterministic
+
+ # ddp init
+ distributed_option = build_distributed(args)
+
+ # for logging
+ if not distributed_option.distributed or distributed_option.dist_rank == 0:
+ logging.basicConfig(
+ level="INFO",
+ format=f"[{os.uname()[1].split('.')[0]}]"
+ f" %(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
+ )
+ else:
+ logging.basicConfig(
+ level="ERROR",
+ format=f"[{os.uname()[1].split('.')[0]}]"
+ f" %(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
+ )
+
+ # prepare files for dataloader
+ prepare_data(args, distributed_option)
+
+ model = build_model(args)
+ model = model.to(
+ dtype=getattr(torch, args.train_dtype),
+ device="cuda" if args.ngpu > 0 else "cpu",
+ )
+ if args.enable_lora:
+ mark_only_lora_as_trainable(model, args.lora_bias)
+ for t in args.freeze_param:
+ for k, p in model.named_parameters():
+ if k.startswith(t + ".") or k == t:
+ logging.info(f"Setting {k}.requires_grad = False")
+ p.requires_grad = False
+
+ optimizers = build_optimizer(args, model=model)
+ schedulers = build_scheduler(args, optimizers)
+
+ logging.info("world size: {}, rank: {}, local_rank: {}".format(distributed_option.dist_world_size,
+ distributed_option.dist_rank,
+ distributed_option.local_rank))
+ logging.info(pytorch_cudnn_version())
+ logging.info("Args: {}".format(args))
+ logging.info(model_summary(model))
+ logging.info("Optimizer: {}".format(optimizers))
+ logging.info("Scheduler: {}".format(schedulers))
+
+ # dump args to config.yaml
+ if not distributed_option.distributed or distributed_option.dist_rank == 0:
+ os.makedirs(args.output_dir, exist_ok=True)
+ with open(os.path.join(args.output_dir, "config.yaml"), "w") as f:
+ logging.info("Saving the configuration in {}/{}".format(args.output_dir, "config.yaml"))
+ if args.use_pai:
+ buffer = BytesIO()
+ torch.save({"config": vars(args)}, buffer)
+ args.oss_bucket.put_object(os.path.join(args.output_dir, "config.dict"), buffer.getvalue())
+ else:
+ yaml_no_alias_safe_dump(vars(args), f, indent=4, sort_keys=False)
+
+ for p in args.init_param:
+ logging.info(f"Loading pretrained params from {p}")
+ load_pretrained_model(
+ model=model,
+ init_param=p,
+ ignore_init_mismatch=args.ignore_init_mismatch,
+ map_location=f"cuda:{torch.cuda.current_device()}"
+ if args.ngpu > 0
+ else "cpu",
+ oss_bucket=args.oss_bucket,
+ )
+
+ # dataloader for training/validation
+ train_dataloader, valid_dataloader = build_dataloader(args)
+
+ # Trainer, including model, optimizers, etc.
+ trainer = build_trainer_modelscope(
+ args=args,
+ model=model,
+ optimizers=optimizers,
+ schedulers=schedulers,
+ train_dataloader=train_dataloader,
+ valid_dataloader=valid_dataloader,
+ distributed_option=distributed_option
+ )
+
+ return trainer
diff --git a/funasr/datasets/small_datasets/sequence_iter_factory.py b/funasr/datasets/small_datasets/sequence_iter_factory.py
index e748c3d..8a7279a 100644
--- a/funasr/datasets/small_datasets/sequence_iter_factory.py
+++ b/funasr/datasets/small_datasets/sequence_iter_factory.py
@@ -66,8 +66,9 @@
batch_bins=dataset_conf["batch_conf"]["batch_size"] * args.ngpu,
shape_files=shape_files,
sort_in_batch=dataset_conf["sort_in_batch"] if hasattr(dataset_conf, "sort_in_batch") else "descending",
- sort_batch=dataset_conf["sort_batch"] if hasattr(dataset_conf, "sort_batch") else "ascending",
+ sort_batch=dataset_conf["sort_batch"] if hasattr(dataset_conf, "sort_batch") else "descending",
drop_last=False,
+ min_batch_size=torch.distributed.get_world_size(),
padding=True,
)
diff --git a/funasr/utils/prepare_data.py b/funasr/utils/prepare_data.py
index 8d82a2f..c9615e7 100644
--- a/funasr/utils/prepare_data.py
+++ b/funasr/utils/prepare_data.py
@@ -195,11 +195,35 @@
def prepare_data(args, distributed_option):
- distributed = distributed_option.distributed
data_names = args.dataset_conf.get("data_names", "speech,text").split(",")
data_types = args.dataset_conf.get("data_types", "sound,text").split(",")
file_names = args.data_file_names.split(",")
batch_type = args.dataset_conf["batch_conf"]["batch_type"]
+ print("data_names: {}, data_types: {}, file_names: {}".format(data_names, data_types, file_names))
+ assert len(data_names) == len(data_types) == len(file_names)
+ if args.dataset_type == "small":
+ args.train_shape_file = [os.path.join(args.data_dir, args.train_set, "{}_shape".format(data_names[0]))]
+ args.valid_shape_file = [os.path.join(args.data_dir, args.valid_set, "{}_shape".format(data_names[0]))]
+ args.train_data_path_and_name_and_type, args.valid_data_path_and_name_and_type = [], []
+ for file_name, data_name, data_type in zip(file_names, data_names, data_types):
+ args.train_data_path_and_name_and_type.append(
+ ["{}/{}/{}".format(args.data_dir, args.train_set, file_name), data_name, data_type])
+ args.valid_data_path_and_name_and_type.append(
+ ["{}/{}/{}".format(args.data_dir, args.valid_set, file_name), data_name, data_type])
+ if os.path.exists(args.train_shape_file[0]):
+ assert os.path.exists(args.valid_shape_file[0])
+ print('shape file for small dataset already exists.')
+ return
+ else:
+ concat_data_name = "_".join(data_names)
+ args.train_data_file = os.path.join(args.data_dir, args.train_set, "{}_data.list".format(concat_data_name))
+ args.valid_data_file = os.path.join(args.data_dir, args.valid_set, "{}_data.list".format(concat_data_name))
+ if os.path.exists(args.train_data_file):
+ assert os.path.exists(args.valid_data_file)
+ print('data list for large dataset already exists.')
+ return
+
+ distributed = distributed_option.distributed
if not distributed or distributed_option.dist_rank == 0:
if hasattr(args, "filter_input") and args.filter_input:
filter_wav_text(args.data_dir, args.train_set)
@@ -213,20 +237,5 @@
generate_data_list(args, args.data_dir, args.train_set)
generate_data_list(args, args.data_dir, args.valid_set)
- print("data_names: {}, data_types: {}, file_names: {}".format(data_names, data_types, file_names))
- assert len(data_names) == len(data_types) == len(file_names)
- if args.dataset_type == "small":
- args.train_shape_file = [os.path.join(args.data_dir, args.train_set, "{}_shape".format(data_names[0]))]
- args.valid_shape_file = [os.path.join(args.data_dir, args.valid_set, "{}_shape".format(data_names[0]))]
- args.train_data_path_and_name_and_type, args.valid_data_path_and_name_and_type = [], []
- for file_name, data_name, data_type in zip(file_names, data_names, data_types):
- args.train_data_path_and_name_and_type.append(
- ["{}/{}/{}".format(args.data_dir, args.train_set, file_name), data_name, data_type])
- args.valid_data_path_and_name_and_type.append(
- ["{}/{}/{}".format(args.data_dir, args.valid_set, file_name), data_name, data_type])
- else:
- concat_data_name = "_".join(data_names)
- args.train_data_file = os.path.join(args.data_dir, args.train_set, "{}_data.list".format(concat_data_name))
- args.valid_data_file = os.path.join(args.data_dir, args.valid_set, "{}_data.list".format(concat_data_name))
if distributed:
dist.barrier()
--
Gitblit v1.9.1