From 3d9f094e9652d4b84894c6fd4eae39a4a753b0f0 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期二, 16 五月 2023 23:48:00 +0800
Subject: [PATCH] train
---
funasr/tasks/abs_task.py | 83 +++++++++++++++++++++++++++--------------
1 files changed, 55 insertions(+), 28 deletions(-)
diff --git a/funasr/tasks/abs_task.py b/funasr/tasks/abs_task.py
index 86957d9..fd4e190 100644
--- a/funasr/tasks/abs_task.py
+++ b/funasr/tasks/abs_task.py
@@ -30,7 +30,7 @@
import torch.nn
import torch.optim
import yaml
-from funasr.train.abs_espnet_model import AbsESPnetModel
+from funasr.models.base_model import FunASRModel
from torch.utils.data import DataLoader
from typeguard import check_argument_types
from typeguard import check_return_type
@@ -230,8 +230,8 @@
>>> cls.check_task_requirements()
If your model is defined as following,
- >>> from funasr.train.abs_espnet_model import AbsESPnetModel
- >>> class Model(AbsESPnetModel):
+ >>> from funasr.models.base_model import FunASRModel
+ >>> class Model(FunASRModel):
... def forward(self, input, output, opt=None): pass
then "required_data_names" should be as
@@ -251,8 +251,8 @@
>>> cls.check_task_requirements()
If your model is defined as follows,
- >>> from funasr.train.abs_espnet_model import AbsESPnetModel
- >>> class Model(AbsESPnetModel):
+ >>> from funasr.models.base_model import FunASRModel
+ >>> class Model(FunASRModel):
... def forward(self, input, output, opt=None): pass
then "optional_data_names" should be as
@@ -263,8 +263,9 @@
@classmethod
@abstractmethod
- def build_model(cls, args: argparse.Namespace) -> AbsESPnetModel:
+ def build_model(cls, args: argparse.Namespace) -> FunASRModel:
raise NotImplementedError
+
@classmethod
def get_parser(cls) -> config_argparse.ArgumentParser:
@@ -445,6 +446,12 @@
help='Perform on "collect stats" mode',
)
group.add_argument(
+ "--mc",
+ type=bool,
+ default=False,
+ help="MultiChannel input",
+ )
+ group.add_argument(
"--write_collected_feats",
type=str2bool,
default=False,
@@ -467,7 +474,7 @@
parser.add_argument(
"--batch_interval",
type=int,
- default=10000,
+ default=-1,
help="The batch interval for saving model.",
)
group.add_argument(
@@ -547,6 +554,12 @@
type=int,
default=1,
help="The number of gradient accumulation",
+ )
+ group.add_argument(
+ "--bias_grad_times",
+ type=float,
+ default=1.0,
+ help="To scale the gradient of contextual related params",
)
group.add_argument(
"--no_forward_run",
@@ -635,8 +648,8 @@
group.add_argument(
"--init_param",
type=str,
+ action="append",
default=[],
- nargs="*",
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, "
@@ -662,7 +675,7 @@
"--freeze_param",
type=str,
default=[],
- nargs="*",
+ action="append",
help="Freeze parameters",
)
@@ -1153,14 +1166,15 @@
elif args.distributed and args.simple_ddp:
distributed_option.init_torch_distributed_pai(args)
args.ngpu = dist.get_world_size()
- if args.dataset_type == "small":
+ if args.dataset_type == "small" and args.ngpu > 0:
if args.batch_size is not None:
args.batch_size = args.batch_size * args.ngpu
- if args.batch_bins is not None:
+ if args.batch_bins is not None and args.ngpu > 0:
args.batch_bins = args.batch_bins * args.ngpu
# filter samples if wav.scp and text are mismatch
- if (args.train_shape_file is None and args.dataset_type == "small") or args.train_data_file is None and args.dataset_type == "large":
+ if (
+ args.train_shape_file is None and args.dataset_type == "small") or args.train_data_file is None and args.dataset_type == "large":
if not args.simple_ddp or distributed_option.dist_rank == 0:
filter_wav_text(args.data_dir, args.train_set)
filter_wav_text(args.data_dir, args.dev_set)
@@ -1169,8 +1183,10 @@
if args.train_shape_file is None and args.dataset_type == "small":
if not args.simple_ddp or distributed_option.dist_rank == 0:
- calc_shape(args.data_dir, args.train_set, args.frontend_conf, args.speech_length_min, args.speech_length_max)
- calc_shape(args.data_dir, args.dev_set, args.frontend_conf, args.speech_length_min, args.speech_length_max)
+ calc_shape(args.data_dir, args.train_set, args.frontend_conf, args.speech_length_min,
+ args.speech_length_max)
+ calc_shape(args.data_dir, args.dev_set, args.frontend_conf, args.speech_length_min,
+ args.speech_length_max)
if args.simple_ddp:
dist.barrier()
args.train_shape_file = [os.path.join(args.data_dir, args.train_set, "speech_shape")]
@@ -1232,9 +1248,9 @@
# 2. Build model
model = cls.build_model(args=args)
- if not isinstance(model, AbsESPnetModel):
+ if not isinstance(model, FunASRModel):
raise RuntimeError(
- f"model must inherit {AbsESPnetModel.__name__}, but got {type(model)}"
+ f"model must inherit {FunASRModel.__name__}, but got {type(model)}"
)
model = model.to(
dtype=getattr(torch, args.train_dtype),
@@ -1316,6 +1332,7 @@
data_path_and_name_and_type=args.train_data_path_and_name_and_type,
key_file=train_key_file,
batch_size=args.batch_size,
+ mc=args.mc,
dtype=args.train_dtype,
num_workers=args.num_workers,
allow_variable_data_keys=args.allow_variable_data_keys,
@@ -1327,6 +1344,7 @@
data_path_and_name_and_type=args.valid_data_path_and_name_and_type,
key_file=valid_key_file,
batch_size=args.valid_batch_size,
+ mc=args.mc,
dtype=args.train_dtype,
num_workers=args.num_workers,
allow_variable_data_keys=args.allow_variable_data_keys,
@@ -1360,15 +1378,21 @@
if args.dataset_type == "large":
from funasr.datasets.large_datasets.build_dataloader import ArkDataLoader
train_iter_factory = ArkDataLoader(args.train_data_file, args.token_list, args.dataset_conf,
- frontend_conf=args.frontend_conf if hasattr(args, "frontend_conf") else None,
- seg_dict_file=args.seg_dict_file if hasattr(args, "seg_dict_file") else None,
- punc_dict_file=args.punc_list if hasattr(args, "punc_list") else None,
+ frontend_conf=args.frontend_conf if hasattr(args,
+ "frontend_conf") else None,
+ seg_dict_file=args.seg_dict_file if hasattr(args,
+ "seg_dict_file") else None,
+ punc_dict_file=args.punc_list if hasattr(args,
+ "punc_list") else None,
bpemodel_file=args.bpemodel if hasattr(args, "bpemodel") else None,
mode="train")
- valid_iter_factory = ArkDataLoader(args.valid_data_file, args.token_list, args.dataset_conf,
- frontend_conf=args.frontend_conf if hasattr(args, "frontend_conf") else None,
- seg_dict_file=args.seg_dict_file if hasattr(args, "seg_dict_file") else None,
- punc_dict_file=args.punc_list if hasattr(args, "punc_list") else None,
+ valid_iter_factory = ArkDataLoader(args.valid_data_file, args.token_list, args.dataset_conf,
+ frontend_conf=args.frontend_conf if hasattr(args,
+ "frontend_conf") else None,
+ seg_dict_file=args.seg_dict_file if hasattr(args,
+ "seg_dict_file") else None,
+ punc_dict_file=args.punc_list if hasattr(args,
+ "punc_list") else None,
bpemodel_file=args.bpemodel if hasattr(args, "bpemodel") else None,
mode="eval")
elif args.dataset_type == "small":
@@ -1582,8 +1606,11 @@
) -> AbsIterFactory:
assert check_argument_types()
- if args.frontend_conf is not None and "fs" in args.frontend_conf:
- dest_sample_rate = args.frontend_conf["fs"]
+ if hasattr(args, "frontend_conf"):
+ if args.frontend_conf is not None and "fs" in args.frontend_conf:
+ dest_sample_rate = args.frontend_conf["fs"]
+ else:
+ dest_sample_rate = 16000
else:
dest_sample_rate = 16000
@@ -1912,7 +1939,7 @@
model_file: Union[Path, str] = None,
cmvn_file: Union[Path, str] = None,
device: str = "cpu",
- ) -> Tuple[AbsESPnetModel, argparse.Namespace]:
+ ) -> Tuple[FunASRModel, argparse.Namespace]:
"""Build model from the files.
This method is used for inference or fine-tuning.
@@ -1939,9 +1966,9 @@
args["cmvn_file"] = cmvn_file
args = argparse.Namespace(**args)
model = cls.build_model(args)
- if not isinstance(model, AbsESPnetModel):
+ if not isinstance(model, FunASRModel):
raise RuntimeError(
- f"model must inherit {AbsESPnetModel.__name__}, but got {type(model)}"
+ f"model must inherit {FunASRModel.__name__}, but got {type(model)}"
)
model.to(device)
if model_file is not None:
--
Gitblit v1.9.1