From 4ace5a95b052d338947fc88809a440ccd55cf6b4 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 16 十一月 2023 16:39:52 +0800
Subject: [PATCH] funasr pages
---
funasr/tasks/abs_task.py | 84 +++++++++++++++++++----------------------
1 files changed, 39 insertions(+), 45 deletions(-)
diff --git a/funasr/tasks/abs_task.py b/funasr/tasks/abs_task.py
index 55a5d79..f7f13d2 100644
--- a/funasr/tasks/abs_task.py
+++ b/funasr/tasks/abs_task.py
@@ -30,9 +30,8 @@
import torch.nn
import torch.optim
import yaml
+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
from funasr import __version__
from funasr.datasets.dataset import AbsDataset
@@ -44,19 +43,18 @@
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.optimizers.sgd import SGD
from funasr.samplers.build_batch_sampler import BATCH_TYPES
from funasr.samplers.build_batch_sampler import build_batch_sampler
from funasr.samplers.unsorted_batch_sampler import UnsortedBatchSampler
from funasr.schedulers.noam_lr import NoamLR
-from funasr.schedulers.warmup_lr import WarmupLR
from funasr.schedulers.tri_stage_scheduler import TriStageLR
+from funasr.schedulers.warmup_lr import WarmupLR
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.train.abs_espnet_model import AbsESPnetModel
from funasr.train.class_choices import ClassChoices
from funasr.train.distributed_utils import DistributedOption
from funasr.train.trainer import Trainer
@@ -73,6 +71,7 @@
from funasr.utils.types import str_or_none
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
+from funasr.modules.lora.utils import mark_only_lora_as_trainable
try:
import wandb
@@ -230,8 +229,8 @@
>>> 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 +250,8 @@
>>> 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
@@ -263,12 +262,12 @@
@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:
- assert check_argument_types()
class ArgumentDefaultsRawTextHelpFormatter(
argparse.RawTextHelpFormatter,
@@ -954,11 +953,22 @@
default=None,
help="oss bucket.",
)
+ group.add_argument(
+ "--enable_lora",
+ type=str2bool,
+ default=False,
+ help="Apply lora for finetuning.",
+ )
+ group.add_argument(
+ "--lora_bias",
+ type=str,
+ default="none",
+ help="lora bias.",
+ )
cls.trainer.add_arguments(parser)
cls.add_task_arguments(parser)
- assert check_return_type(parser)
return parser
@classmethod
@@ -1006,7 +1016,6 @@
return _cls
# This method is used only for --print_config
- assert check_argument_types()
parser = cls.get_parser()
args, _ = parser.parse_known_args()
config = vars(args)
@@ -1046,7 +1055,6 @@
@classmethod
def check_required_command_args(cls, args: argparse.Namespace):
- assert check_argument_types()
if hasattr(args, "required"):
for k in vars(args):
if "-" in k:
@@ -1076,7 +1084,6 @@
inference: bool = False,
) -> None:
"""Check if the dataset satisfy the requirement of current Task"""
- assert check_argument_types()
mes = (
f"If you intend to use an additional input, modify "
f'"{cls.__name__}.required_data_names()" or '
@@ -1103,14 +1110,12 @@
@classmethod
def print_config(cls, file=sys.stdout) -> None:
- assert check_argument_types()
# Shows the config: e.g. python train.py asr --print_config
config = cls.get_default_config()
file.write(yaml_no_alias_safe_dump(config, indent=4, sort_keys=False))
@classmethod
def main(cls, args: argparse.Namespace = None, cmd: Sequence[str] = None):
- assert check_argument_types()
print(get_commandline_args(), file=sys.stderr)
if args is None:
parser = cls.get_parser()
@@ -1147,7 +1152,6 @@
@classmethod
def main_worker(cls, args: argparse.Namespace):
- assert check_argument_types()
# 0. Init distributed process
distributed_option = build_dataclass(DistributedOption, args)
@@ -1172,7 +1176,8 @@
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)
@@ -1181,8 +1186,10 @@
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")]
@@ -1244,14 +1251,16 @@
# 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),
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:
@@ -1372,19 +1381,10 @@
# 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,
@@ -1561,7 +1561,6 @@
- 4 epoch with "--num_iters_per_epoch" == 4
"""
- assert check_argument_types()
iter_options = cls.build_iter_options(args, distributed_option, mode)
# Overwrite iter_options if any kwargs is given
@@ -1594,7 +1593,6 @@
def build_sequence_iter_factory(
cls, args: argparse.Namespace, iter_options: IteratorOptions, mode: str
) -> AbsIterFactory:
- assert check_argument_types()
if hasattr(args, "frontend_conf"):
if args.frontend_conf is not None and "fs" in args.frontend_conf:
@@ -1688,7 +1686,6 @@
iter_options: IteratorOptions,
mode: str,
) -> AbsIterFactory:
- assert check_argument_types()
dataset = ESPnetDataset(
iter_options.data_path_and_name_and_type,
@@ -1793,7 +1790,6 @@
def build_multiple_iter_factory(
cls, args: argparse.Namespace, distributed_option: DistributedOption, mode: str
):
- assert check_argument_types()
iter_options = cls.build_iter_options(args, distributed_option, mode)
assert len(iter_options.data_path_and_name_and_type) > 0, len(
iter_options.data_path_and_name_and_type
@@ -1890,7 +1886,6 @@
inference: bool = False,
) -> DataLoader:
"""Build DataLoader using iterable dataset"""
- assert check_argument_types()
# For backward compatibility for pytorch DataLoader
if collate_fn is not None:
kwargs = dict(collate_fn=collate_fn)
@@ -1929,7 +1924,7 @@
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.
@@ -1940,7 +1935,6 @@
device: Device type, "cpu", "cuda", or "cuda:N".
"""
- assert check_argument_types()
if config_file is None:
assert model_file is not None, (
"The argument 'model_file' must be provided "
@@ -1956,9 +1950,9 @@
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