From 54931dd4e1a099d7d6f144c4e12e5453deb3aa26 Mon Sep 17 00:00:00 2001
From: 雾聪 <wucong.lyb@alibaba-inc.com>
Date: 星期三, 28 六月 2023 10:41:57 +0800
Subject: [PATCH] Merge branch 'main' of https://github.com/alibaba-damo-academy/FunASR into main
---
funasr/tasks/abs_task.py | 55 +++++++++++++++++++++++++++++--------------------------
1 files changed, 29 insertions(+), 26 deletions(-)
diff --git a/funasr/tasks/abs_task.py b/funasr/tasks/abs_task.py
index 5f9e8fc..0fb77a9 100644
--- a/funasr/tasks/abs_task.py
+++ b/funasr/tasks/abs_task.py
@@ -266,6 +266,7 @@
def build_model(cls, args: argparse.Namespace) -> FunASRModel:
raise NotImplementedError
+
@classmethod
def get_parser(cls) -> config_argparse.ArgumentParser:
assert check_argument_types()
@@ -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",
)
@@ -1153,10 +1166,10 @@
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
@@ -1319,6 +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,
@@ -1330,6 +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,
@@ -1361,25 +1376,10 @@
# 7. Build iterator factories
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,
- 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,
- bpemodel_file=args.bpemodel if hasattr(args, "bpemodel") else None,
- mode="eval")
+ from funasr.datasets.large_datasets.build_dataloader import LargeDataLoader
+ train_iter_factory = LargeDataLoader(args, mode="train")
+ valid_iter_factory = LargeDataLoader(args, mode="eval")
+
elif args.dataset_type == "small":
train_iter_factory = cls.build_iter_factory(
args=args,
@@ -1591,8 +1591,11 @@
) -> 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
--
Gitblit v1.9.1