import os
|
|
from modelscope.metainfo import Trainers
|
from modelscope.trainers import build_trainer
|
|
from funasr.datasets.ms_dataset import MsDataset
|
from funasr.utils.modelscope_param import modelscope_args
|
|
|
def modelscope_finetune(params):
|
if not os.path.exists(params.output_dir):
|
os.makedirs(params.output_dir, exist_ok=True)
|
# dataset split ["train", "validation"]
|
ds_dict = MsDataset.load(params.data_path)
|
kwargs = dict(
|
model=params.model,
|
data_dir=ds_dict,
|
dataset_type=params.dataset_type,
|
work_dir=params.output_dir,
|
batch_bins=params.batch_bins,
|
max_epoch=params.max_epoch,
|
lr=params.lr,
|
mate_params=params.param_dict)
|
trainer = build_trainer(Trainers.speech_asr_trainer, default_args=kwargs)
|
trainer.train()
|
|
|
if __name__ == '__main__':
|
params = modelscope_args(model="damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch", data_path="./data")
|
params.output_dir = "./checkpoint" # m模型保存路径
|
params.data_path = "./example_data/" # 数据路径
|
params.dataset_type = "small" # 小数据量设置small，若数据量大于1000小时，请使用large
|
params.batch_bins = 2000 # batch size，如果dataset_type="small"，batch_bins单位为fbank特征帧数，如果dataset_type="large"，batch_bins单位为毫秒，
|
params.max_epoch = 20 # 最大训练轮数
|
params.lr = 0.0002 # 设置学习率
|
init_param = [] # 初始模型路径，默认加载modelscope模型初始化，例如: ["checkpoint/20epoch.pb"]
|
freeze_param = [] # 模型参数freeze, 例如: ["encoder"]
|
ignore_init_mismatch = True # 是否忽略模型参数初始化不匹配
|
use_lora = False # 是否使用lora进行模型微调
|
params.param_dict = {"init_param":init_param, "freeze_param": freeze_param, "ignore_init_mismatch": ignore_init_mismatch}
|
if use_lora:
|
enable_lora = True
|
lora_bias = "all"
|
lora_params = {"lora_list":['q','v'], "lora_rank":8, "lora_alpha":16, "lora_dropout":0.1}
|
lora_config = {"enable_lora": enable_lora, "lora_bias": lora_bias, "lora_params": lora_params}
|
params.param_dict.update(lora_config)
|
|
modelscope_finetune(params)
|
