auto frontend

游雁

2024-06-07 c350b6becb887d578126e363306e7398782ef675

auto frontend

 funasr/datasets/audio_datasets/samplers.py

@@ -365,6 +365,7 @@
        self.max_token_length = kwargs.get("max_token_length", 2048)
        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
        self.start_step = kwargs.get("start_step", 2048)
        self.batch_size_sample_max = kwargs.get("batch_size_sample_max", 200)

        super().__init__(
            dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle, drop_last=drop_last
@@ -388,19 +389,25 @@
            )
            batch = []
            max_len_in_batch = 0
            count = 0
            for idx in buffer:
                original_sample_length = self.dataset.get_source_len(idx)
                if original_sample_length > self.max_token_length:
                    continue
                sample_length = 1 if self.batch_type == "example" else original_sample_length
                potential_batch_length = max(max_len_in_batch, sample_length) * (len(batch) + 1)
                if potential_batch_length <= self.batch_size:
                if (
                    potential_batch_length <= self.batch_size
                    and count <= self.batch_size_sample_max
                ):
                    batch.append(idx)
                    max_len_in_batch = max(max_len_in_batch, sample_length)
                    count += 1
                else:
                    buffer_batches.append(batch)
                    batch = [idx]
                    max_len_in_batch = sample_length
                    count = 0
            if batch:
                buffer_batches.append(batch)


 funasr/datasets/openai_datasets/datasets.py

@@ -154,13 +154,13 @@
                fbank_mask += fbank_mask_i
                fbank_beg.append(fbank_beg_i)

            # if len(input_ids) > self.max_token_length:
            #     badcase_flag = True
            if len(input_ids) > self.max_token_length:
                badcase_flag = True
            if badcase_flag:
                continue
            input_ids = torch.tensor(input_ids, dtype=torch.int64)[: self.max_token_length]
            input_ids = torch.tensor(input_ids, dtype=torch.int64)  # [: self.max_token_length]
            attention_mask = torch.tensor([len(input_ids)], dtype=torch.int32)
            labels = torch.tensor(labels, dtype=torch.int64)[: self.max_token_length]
            labels = torch.tensor(labels, dtype=torch.int64)  # [: self.max_token_length]

            fbank = speech[0, :, :]
            fbank_lens = speech_lengths