version

egs_modelscope/asr/paraformer/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/finetune2.py

@@ -0,0 +1,37 @@
+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)
+    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 = 50                           # 最大训练轮数
+    params.lr = 0.0005                              # 设置学习率
+    params.scheduler_conf = {"warmup_steps": 30000}
+    
+    modelscope_finetune(params)

funasr/version.txt

@@ -1 +1 @@
-0.2.3
+0.3.0