resume from step

游雁

2024-04-26 ab6391da4f226627b2f827c102d1f28aa21a3cb8

resume from step

 funasr/datasets/sense_voice_datasets/datasets.py

@@ -159,7 +159,7 @@

    def _filter_badcase(self, outputs, i=0):
        b, t, _ = outputs["speech"].shape
        

        if b * t > self.batch_size * 1.25:
            beg = torch.randint(0, 2, ()).item()
            if b < 2:
@@ -170,7 +170,6 @@
            for key, data_list in outputs.items():
                outputs[key] = outputs[key][beg : beg + b : 2]

                
            speech_lengths_max = outputs["speech_lengths"].max().item()
            outputs["speech"] = outputs["speech"][:, :speech_lengths_max, :]
            text_lengths_max = outputs["text_lengths"].max().item()