train

游雁

2023-05-16 3d9f094e9652d4b84894c6fd4eae39a4a753b0f0

 funasr/train/trainer.py

@@ -39,7 +39,7 @@
from funasr.torch_utils.device_funcs import to_device
from funasr.torch_utils.recursive_op import recursive_average
from funasr.torch_utils.set_all_random_seed import set_all_random_seed
from funasr.train.abs_espnet_model import AbsESPnetModel
from funasr.models.base_model import FunASRModel
from funasr.train.distributed_utils import DistributedOption
from funasr.train.reporter import Reporter
from funasr.train.reporter import SubReporter
@@ -95,6 +95,7 @@
    use_pai: bool
    oss_bucket: Union[oss2.Bucket, None]
    batch_interval: int
    bias_grad_times: float

class Trainer:
    """Trainer having a optimizer.
@@ -165,7 +166,7 @@
    @classmethod
    def run(
        cls,
        model: AbsESPnetModel,
        model: FunASRModel,
        optimizers: Sequence[torch.optim.Optimizer],
        schedulers: Sequence[Optional[AbsScheduler]],
        train_iter_factory: AbsIterFactory,
@@ -546,8 +547,11 @@
        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 @@
                        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(),