From b15db52e4e67da8a133a67e8ffa415386de48b40 Mon Sep 17 00:00:00 2001
From: zhuyunfeng <10596244@qq.com>
Date: 星期二, 09 五月 2023 23:03:15 +0800
Subject: [PATCH] Add contributor
---
funasr/tasks/abs_task.py | 124 ++++++++++++++++++++++++++++++++++++----
1 files changed, 110 insertions(+), 14 deletions(-)
diff --git a/funasr/tasks/abs_task.py b/funasr/tasks/abs_task.py
index 7899400..55a5d79 100644
--- a/funasr/tasks/abs_task.py
+++ b/funasr/tasks/abs_task.py
@@ -43,6 +43,7 @@
from funasr.iterators.chunk_iter_factory import ChunkIterFactory
from funasr.iterators.multiple_iter_factory import MultipleIterFactory
from funasr.iterators.sequence_iter_factory import SequenceIterFactory
+from funasr.main_funcs.collect_stats import collect_stats
from funasr.optimizers.sgd import SGD
from funasr.optimizers.fairseq_adam import FairseqAdam
from funasr.samplers.build_batch_sampler import BATCH_TYPES
@@ -70,7 +71,7 @@
from funasr.utils.types import str2triple_str
from funasr.utils.types import str_or_int
from funasr.utils.types import str_or_none
-from funasr.utils.wav_utils import calc_shape, generate_data_list
+from funasr.utils.wav_utils import calc_shape, generate_data_list, filter_wav_text
from funasr.utils.yaml_no_alias_safe_dump import yaml_no_alias_safe_dump
try:
@@ -444,6 +445,12 @@
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,
@@ -462,6 +469,12 @@
type=int,
default=sys.maxsize,
help="The maximum number update step to train",
+ )
+ parser.add_argument(
+ "--batch_interval",
+ type=int,
+ default=-1,
+ help="The batch interval for saving model.",
)
group.add_argument(
"--patience",
@@ -540,6 +553,12 @@
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",
@@ -628,8 +647,8 @@
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, "
@@ -638,12 +657,12 @@
"and exclude_keys excludes keys of model states for the initialization."
"e.g.\n"
" # Load all parameters"
- " --init_param some/where/model.pth\n"
+ " --init_param some/where/model.pb\n"
" # Load only decoder parameters"
- " --init_param some/where/model.pth:decoder:decoder\n"
+ " --init_param some/where/model.pb:decoder:decoder\n"
" # Load only decoder parameters excluding decoder.embed"
- " --init_param some/where/model.pth:decoder:decoder:decoder.embed\n"
- " --init_param some/where/model.pth:decoder:decoder:decoder.embed\n",
+ " --init_param some/where/model.pb:decoder:decoder:decoder.embed\n"
+ " --init_param some/where/model.pb:decoder:decoder:decoder.embed\n",
)
group.add_argument(
"--ignore_init_mismatch",
@@ -655,7 +674,7 @@
"--freeze_param",
type=str,
default=[],
- nargs="*",
+ action="append",
help="Freeze parameters",
)
@@ -1146,11 +1165,19 @@
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 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)
+ if args.simple_ddp:
+ dist.barrier()
if args.train_shape_file is None and args.dataset_type == "small":
if not args.simple_ddp or distributed_option.dist_rank == 0:
@@ -1184,12 +1211,18 @@
# logging.basicConfig() is invoked in main_worker() instead of main()
# because it can be invoked only once in a process.
# FIXME(kamo): Should we use logging.getLogger()?
+ # BUGFIX: Remove previous handlers and reset log level
+ for handler in logging.root.handlers[:]:
+ logging.root.removeHandler(handler)
logging.basicConfig(
level=args.log_level,
format=f"[{os.uname()[1].split('.')[0]}]"
f" %(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
else:
+ # BUGFIX: Remove previous handlers and reset log level
+ for handler in logging.root.handlers[:]:
+ logging.root.removeHandler(handler)
# Suppress logging if RANK != 0
logging.basicConfig(
level="ERROR",
@@ -1272,6 +1305,54 @@
if args.dry_run:
pass
+ elif args.collect_stats:
+ # Perform on collect_stats mode. This mode has two roles
+ # - Derive the length and dimension of all input data
+ # - Accumulate feats, square values, and the length for whitening
+
+ if args.valid_batch_size is None:
+ args.valid_batch_size = args.batch_size
+
+ if len(args.train_shape_file) != 0:
+ train_key_file = args.train_shape_file[0]
+ else:
+ train_key_file = None
+ if len(args.valid_shape_file) != 0:
+ valid_key_file = args.valid_shape_file[0]
+ else:
+ valid_key_file = None
+
+ collect_stats(
+ model=model,
+ train_iter=cls.build_streaming_iterator(
+ 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,
+ ngpu=args.ngpu,
+ preprocess_fn=cls.build_preprocess_fn(args, train=False),
+ collate_fn=cls.build_collate_fn(args, train=False),
+ ),
+ valid_iter=cls.build_streaming_iterator(
+ 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,
+ ngpu=args.ngpu,
+ preprocess_fn=cls.build_preprocess_fn(args, train=False),
+ collate_fn=cls.build_collate_fn(args, train=False),
+ ),
+ output_dir=output_dir,
+ ngpu=args.ngpu,
+ log_interval=args.log_interval,
+ write_collected_feats=args.write_collected_feats,
+ )
else:
logging.info("Training args: {}".format(args))
# 6. Loads pre-trained model
@@ -1293,12 +1374,16 @@
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,
- seg_dict_file=args.seg_dict_file if hasattr(args,
- "seg_dict_file") 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,
- seg_dict_file=args.seg_dict_file if hasattr(args,
- "seg_dict_file") 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":
train_iter_factory = cls.build_iter_factory(
@@ -1511,12 +1596,21 @@
) -> AbsIterFactory:
assert check_argument_types()
+ 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
+
dataset = ESPnetDataset(
iter_options.data_path_and_name_and_type,
float_dtype=args.train_dtype,
preprocess=iter_options.preprocess_fn,
max_cache_size=iter_options.max_cache_size,
max_cache_fd=iter_options.max_cache_fd,
+ dest_sample_rate=dest_sample_rate,
)
cls.check_task_requirements(
dataset, args.allow_variable_data_keys, train=iter_options.train
@@ -1788,6 +1882,7 @@
key_file: str = None,
batch_size: int = 1,
fs: dict = None,
+ mc: bool = False,
dtype: str = np.float32,
num_workers: int = 1,
allow_variable_data_keys: bool = False,
@@ -1806,6 +1901,7 @@
data_path_and_name_and_type,
float_dtype=dtype,
fs=fs,
+ mc=mc,
preprocess=preprocess_fn,
key_file=key_file,
)
--
Gitblit v1.9.1