train update

游雁

2024-03-27 9b4e9cc8a0311e5243d69b73ed073e7ea441982e

 funasr/models/paraformer/model.py

@@ -13,13 +13,14 @@
from funasr.models.ctc.ctc import CTC
from funasr.utils import postprocess_utils
from funasr.metrics.compute_acc import th_accuracy
from funasr.train_utils.device_funcs import to_device
from funasr.utils.datadir_writer import DatadirWriter
from funasr.models.paraformer.search import Hypothesis
from funasr.models.paraformer.cif_predictor import mae_loss
from funasr.train_utils.device_funcs import force_gatherable
from funasr.losses.label_smoothing_loss import LabelSmoothingLoss
from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
from funasr.models.transformer.utils.nets_utils import make_pad_mask, pad_list
from funasr.models.transformer.utils.nets_utils import make_pad_mask
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank


@@ -154,8 +155,8 @@
        self.predictor_bias = predictor_bias
        self.sampling_ratio = sampling_ratio
        self.criterion_pre = mae_loss(normalize_length=length_normalized_loss)
        # self.step_cur = 0
        #


        self.share_embedding = share_embedding
        if self.share_embedding:
            self.decoder.embed = None
@@ -230,6 +231,7 @@
        stats["loss_pre"] = loss_pre.detach().cpu() if loss_pre is not None else None
        
        stats["loss"] = torch.clone(loss.detach())
        stats["batch_size"] = batch_size
        
        # force_gatherable: to-device and to-tensor if scalar for DataParallel
        if self.length_normalized_loss:
@@ -549,3 +551,10 @@
                
        return results, meta_data

    def export(self, **kwargs):
        from .export_meta import export_rebuild_model
        if 'max_seq_len' not in kwargs:
            kwargs['max_seq_len'] = 512
        models = export_rebuild_model(model=self, **kwargs)
        return models