Merge pull request #461 from alibaba-damo-academy/dev_sx

Update contextual finetune and bias_grad_times

egs_modelscope/asr/paraformer/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404/finetune.py

@@ -14,6 +14,7 @@ def modelscope_finetune(params):
     ds_dict = MsDataset.load(params.data_path)
     kwargs = dict(
         model=params.model,
+        model_revision="v1.0.2",
         data_dir=ds_dict,
         dataset_type=params.dataset_type,
         work_dir=params.output_dir,
@@ -31,6 +32,6 @@ if __name__ == '__main__':
     params.dataset_type = "large"                   # finetune contextual paraformer模型只能使用large dataset
     params.batch_bins = 200000                      # 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 = 0.00005                             # 设置学习率
+    params.lr = 0.0002                              # 设置学习率
 
     modelscope_finetune(params)

funasr/tasks/abs_task.py

@@ -548,6 +548,12 @@ class AbsTask(ABC):
             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",
             type=str2bool,

funasr/train/trainer.py

@@ -95,6 +95,7 @@ class TrainerOptions:
     use_pai: bool
     oss_bucket: Union[oss2.Bucket, None]
     batch_interval: int
+    bias_grad_times: float
 
 class Trainer:
     """Trainer having a optimizer.
@@ -546,8 +547,11 @@ class Trainer:
         no_forward_run = options.no_forward_run
         ngpu = options.ngpu
         use_wandb = options.use_wandb
+        bias_grad_times = options.bias_grad_times
         distributed = distributed_option.distributed
 
+        if bias_grad_times != 1.0:
+            logging.warning("Using bias_grad_times: {} for gradient scaling".format(bias_grad_times))
         if log_interval is None:
             try:
                 log_interval = max(len(iterator) // 20, 10)
@@ -690,6 +694,16 @@ class Trainer:
                         scale_factor=0.55,
                     )
 
+                # for contextual training
+                if bias_grad_times != 1.0:
+                    # contextual related parameter names
+                    cr_pnames = ["bias_encoder", "bias_embed", "decoder.bias_decoder", "decoder.bias_output"]
+                    for name, param in model.named_parameters():
+                        for cr_pname in cr_pnames:
+                            if cr_pname in name:
+                                param.grad *= bias_grad_times
+                                continue
+
                 # compute the gradient norm to check if it is normal or not
                 grad_norm = torch.nn.utils.clip_grad_norm_(
                     model.parameters(),