first commit for takway.ai

funasr/datasets/audio_datasets/datasets.py

@@ -0,0 +1,260 @@
+import torch
+import random
+
+from funasr.register import tables
+from funasr.utils.load_utils import extract_fbank, load_audio_text_image_video
+
+
+@tables.register("dataset_classes", "AudioDataset")
+class AudioDataset(torch.utils.data.Dataset):
+    """
+    AudioDataset
+    """
+
+    def __init__(
+        self,
+        path,
+        index_ds: str = None,
+        frontend=None,
+        tokenizer=None,
+        int_pad_value: int = -1,
+        float_pad_value: float = 0.0,
+        **kwargs,
+    ):
+        super().__init__()
+        index_ds_class = tables.index_ds_classes.get(index_ds)
+        self.index_ds = index_ds_class(path, **kwargs)
+        preprocessor_speech = kwargs.get("preprocessor_speech", None)
+        if preprocessor_speech:
+            preprocessor_speech_class = tables.preprocessor_classes.get(preprocessor_speech)
+            preprocessor_speech = preprocessor_speech_class(
+                **kwargs.get("preprocessor_speech_conf")
+            )
+        self.preprocessor_speech = preprocessor_speech
+        preprocessor_text = kwargs.get("preprocessor_text", None)
+        if preprocessor_text:
+            preprocessor_text_class = tables.preprocessor_classes.get(preprocessor_text)
+            preprocessor_text = preprocessor_text_class(**kwargs.get("preprocessor_text_conf"))
+        self.preprocessor_text = preprocessor_text
+
+        self.frontend = frontend
+        self.fs = 16000 if frontend is None else frontend.fs
+        self.data_type = "sound"
+        self.tokenizer = tokenizer
+
+        self.int_pad_value = int_pad_value
+        self.float_pad_value = float_pad_value
+
+    def get_source_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_source_len(item)
+
+    def get_target_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_target_len(item)
+
+    def __len__(self):
+        return len(self.index_ds)
+
+    def __getitem__(self, index):
+        item = self.index_ds[index]
+        # import pdb;
+        # pdb.set_trace()
+        source = item["source"]
+        data_src = load_audio_text_image_video(source, fs=self.fs)
+        if self.preprocessor_speech:
+            data_src = self.preprocessor_speech(data_src, fs=self.fs)
+        speech, speech_lengths = extract_fbank(
+            data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
+        )  # speech: [b, T, d]
+
+        target = item["target"]
+        if self.preprocessor_text:
+            target = self.preprocessor_text(target)
+        if self.tokenizer:
+            ids = self.tokenizer.encode(target)
+            text = torch.tensor(ids, dtype=torch.int64)
+        else:
+            ids = target
+            text = ids
+        ids_lengths = len(ids)
+        text_lengths = torch.tensor([ids_lengths], dtype=torch.int32)
+
+        return {
+            "speech": speech[0, :, :],
+            "speech_lengths": speech_lengths,
+            "text": text,
+            "text_lengths": text_lengths,
+        }
+
+    def collator(self, samples: list = None):
+        outputs = {}
+        for sample in samples:
+            for key in sample.keys():
+                if key not in outputs:
+                    outputs[key] = []
+                outputs[key].append(sample[key])
+
+        for key, data_list in outputs.items():
+            if isinstance(data_list[0], torch.Tensor):
+                if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:
+
+                    pad_value = self.int_pad_value
+                else:
+                    pad_value = self.float_pad_value
+
+                outputs[key] = torch.nn.utils.rnn.pad_sequence(
+                    data_list, batch_first=True, padding_value=pad_value
+                )
+        return outputs
+
+
+@tables.register("dataset_classes", "AudioDatasetHotword")
+class AudioDatasetHotword(AudioDataset):
+    # for finetuning contextual_paraformer and seaco_paraformer
+    def __init__(
+        self,
+        *args,
+        seaco_id: bool = 0,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        self.seaco_id = seaco_id
+
+    def __getitem__(self, index):
+        item = self.index_ds[index]
+        # import pdb;
+        # pdb.set_trace()
+        source = item["source"]
+        data_src = load_audio_text_image_video(source, fs=self.fs)
+        if self.preprocessor_speech:
+            data_src = self.preprocessor_speech(data_src, fs=self.fs)
+        speech, speech_lengths = extract_fbank(
+            data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
+        )  # speech: [b, T, d]
+
+        target = item["target"]
+        if self.preprocessor_text:
+            target = self.preprocessor_text(target)
+        if self.tokenizer:
+            ids = self.tokenizer.encode(target)
+            text = torch.tensor(ids, dtype=torch.int64)
+        else:
+            ids = target
+            text = ids
+        ids_lengths = len(ids)
+        text_lengths = torch.tensor([ids_lengths], dtype=torch.int32)
+
+        def generate_index(
+            length,
+            hotword_min_length=2,
+            hotword_max_length=8,
+            sample_rate=0.75,
+            double_rate=0.1,
+            pre_prob=0.0,
+            pre_index=None,
+            pre_hwlist=None,
+        ):
+            if length < hotword_min_length:
+                return [-1]
+            if random.random() < sample_rate:
+                if pre_prob > 0 and random.random() < pre_prob and pre_index is not None:
+                    return pre_index
+                if length == hotword_min_length:
+                    return [0, length - 1]
+                elif (
+                    random.random() < double_rate
+                    and length > hotword_max_length + hotword_min_length + 2
+                ):
+                    # sample two hotwords in a sentence
+                    _max_hw_length = min(hotword_max_length, length // 2)
+                    # first hotword
+                    start1 = random.randint(0, length // 3)
+                    end1 = random.randint(
+                        start1 + hotword_min_length - 1, start1 + _max_hw_length - 1
+                    )
+                    # second hotword
+                    start2 = random.randint(end1 + 1, length - hotword_min_length)
+                    end2 = random.randint(
+                        min(length - 1, start2 + hotword_min_length - 1),
+                        min(length - 1, start2 + hotword_max_length - 1),
+                    )
+                    return [start1, end1, start2, end2]
+                else:  # single hotword
+                    start = random.randint(0, length - hotword_min_length)
+                    end = random.randint(
+                        min(length - 1, start + hotword_min_length - 1),
+                        min(length - 1, start + hotword_max_length - 1),
+                    )
+                    return [start, end]
+            else:
+                return [-1]
+
+        hotword_indx = generate_index(text_lengths[0])
+        return {
+            "speech": speech[0, :, :],
+            "speech_lengths": speech_lengths,
+            "text": text,
+            "text_lengths": text_lengths,
+            "hotword_indx": hotword_indx,
+            "seaco_id": self.seaco_id,
+        }
+
+    def collator(self, samples: list = None):
+        outputs = {}
+        hotword_indxs = []
+        seaco_id = samples[0]["seaco_id"]
+        for sample in samples:
+            for key in sample.keys():
+                if key == "seaco_id":
+                    continue
+                elif key == "hotword_indx":
+                    hotword_indxs.append(sample[key])
+                else:
+                    if key not in outputs:
+                        outputs[key] = []
+                    outputs[key].append(sample[key])
+
+        for key, data_list in outputs.items():
+            if isinstance(data_list[0], torch.Tensor):
+                if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:
+                    pad_value = self.int_pad_value
+                else:
+                    pad_value = self.float_pad_value
+                outputs[key] = torch.nn.utils.rnn.pad_sequence(
+                    data_list, batch_first=True, padding_value=pad_value
+                )
+
+        hotword_list, hotword_lengths = [], []
+        text = outputs["text"]
+        seaco_label_pad = torch.ones_like(text) * -1 if seaco_id else None
+        for b, (hotword_indx, one_text, length) in enumerate(
+            zip(hotword_indxs, text, outputs["text_lengths"])
+        ):
+            length = length[0]
+            if seaco_label_pad is not None:
+                seaco_label_pad[b][:length] = seaco_id
+            if hotword_indx[0] != -1:
+                start, end = int(hotword_indx[0]), int(hotword_indx[1])
+                hotword = one_text[start : end + 1]
+                hotword_list.append(hotword)
+                hotword_lengths.append(end - start + 1)
+                if seaco_label_pad is not None:
+                    seaco_label_pad[b][start : end + 1] = one_text[start : end + 1]
+                if len(hotword_indx) == 4 and hotword_indx[2] != -1:
+                    # the second hotword if exist
+                    start, end = int(hotword_indx[2]), int(hotword_indx[3])
+                    hotword_list.append(one_text[start : end + 1])
+                    hotword_lengths.append(end - start + 1)
+                    if seaco_label_pad is not None:
+                        seaco_label_pad[b][start : end + 1] = one_text[start : end + 1]
+        hotword_list.append(torch.tensor([1]))
+        hotword_lengths.append(1)
+        hotword_pad = torch.nn.utils.rnn.pad_sequence(
+            hotword_list, batch_first=True, padding_value=0
+        )
+        outputs["hotword_pad"] = hotword_pad
+        outputs["hotword_lengths"] = torch.tensor(hotword_lengths, dtype=torch.int32)
+        if seaco_label_pad is not None:
+            outputs["seaco_label_pad"] = seaco_label_pad
+        return outputs

funasr/datasets/audio_datasets/espnet_samplers.py

@@ -0,0 +1,161 @@
+import torch
+import numpy as np
+import logging
+import math
+import torch.distributed as dist
+from torch.utils.data import DistributedSampler
+from torch.utils.data import BatchSampler, Sampler
+import torch.distributed as dist
+import random
+from funasr.register import tables
+
+
+@tables.register("batch_sampler_classes", "EspnetStyleBatchSampler")
+def EspnetStyleBatchSampler_fn(dataset, **kwargs):
+    dataloader_args = {}
+
+    batch_sampler = EspnetStyleBatchSampler(dataset, **kwargs)
+    dataloader_args["batch_sampler"] = batch_sampler
+    dataloader_args["num_workers"] = kwargs.get("num_workers", 4)
+    dataloader_args["pin_memory"] = kwargs.get("pin_memory", True)
+
+    return dataloader_args
+
+
+import torch
+from torch.utils.data import Dataset, DistributedSampler
+import math
+import random
+
+
+class EspnetStyleBatchSampler(DistributedSampler):
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        batch_type="token",
+        rank=None,
+        num_replicas=None,
+        rank_split=False,
+        shuffle=True,
+        drop_last=False,
+        is_training: bool = True,
+        sort_size: int = 1024,
+        start_step: int = 0,
+        **kwargs,
+    ):
+
+        try:
+            rank = dist.get_rank()
+            num_replicas = dist.get_world_size()
+        except:
+            rank = 0
+            num_replicas = 1
+        # if rank_split:
+        #     logging.info(f"Warning, rank_split: {rank_split}, batch and shuffle data in local rank")
+        #     rank = 0
+        #     num_replicas = 1
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.batch_type = batch_type
+        self.is_training = is_training
+        self.shuffle = shuffle and is_training
+        self.drop_last = drop_last
+
+        self.total_size = len(self.dataset)
+        self.num_samples = int(math.ceil(self.total_size / self.num_replicas))
+        self.epoch = 0
+        self.sort_size = sort_size * num_replicas
+        self.max_token_length = kwargs.get("max_token_length", 2048)
+        self.min_token_length = kwargs.get("min_token_length", 0)
+        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
+        self.start_step = start_step
+        if self.start_step > 0:
+            logging.info(f"Warning, start_step > 0, dataloader start from step: {self.start_step}")
+        # super().__init__(dataset, num_replicas=num_replicas, rank=rank,
+        #                  shuffle=shuffle, drop_last=drop_last)
+
+    def __iter__(self):
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            random.seed(self.epoch)
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = list(range(len(self.dataset)))
+
+        # Sort indices by sample length
+        sorted_indices = sorted(indices, key=lambda idx: self.dataset.get_source_len(idx))
+
+        # Organize batches based on 'length' or 'example'
+        buffer_batches = []
+        batch = []
+        max_len_in_batch = 0  # Tracks the max sample length within the current batch
+
+        for idx in sorted_indices:
+
+            # original_sample_length = self.dataset.get_source_len(idx)
+            # if (
+            #     original_sample_length < self.min_token_length
+            #     or original_sample_length > self.max_token_length
+            # ):  # Skip samples that exceed the max length
+            #     continue
+
+            # sample_length = 1 if self.batch_type == "example" else original_sample_length
+
+            # Set sample_length based on the batch type
+            if self.batch_type == "example":
+                sample_length = 1
+            elif self.batch_type == "token":
+                sample_length = self.dataset.get_source_len(idx) + int(
+                    self.dataset.get_target_len(idx) * 1.2
+                )
+            else:
+                sample_length = self.dataset.get_source_len(idx)
+            # Calculate potential batch size with the new sample
+            potential_batch_length = max(max_len_in_batch, sample_length) * (len(batch) + 1)
+            # Add index to batch if it doesn't exceed batch size limit
+            if potential_batch_length <= self.batch_size:
+                batch.append(idx)
+                max_len_in_batch = max(max_len_in_batch, sample_length)
+            else:
+                # Save the current batch and start a new one
+                buffer_batches.append(batch)
+                batch = [idx]
+                max_len_in_batch = sample_length
+
+        # Add the last batch if it shouldn't be dropped
+        if batch and (not self.drop_last or len(batch) * max_len_in_batch == self.batch_size):
+            buffer_batches.append(batch)
+
+        # Shuffle the list of batches
+        if self.shuffle:
+            random.seed(self.epoch)
+            random.shuffle(buffer_batches)
+
+        # Ensure each rank gets the same number of batches
+        batches_per_rank = int(math.ceil(len(buffer_batches) / self.num_replicas))
+        total_batches_needed = batches_per_rank * self.num_replicas
+        extra_batches = total_batches_needed - len(buffer_batches)
+        # Add extra batches by random selection, if needed
+        buffer_batches += random.choices(buffer_batches, k=extra_batches)
+
+        # Allocate the batches to the current rank
+        start_idx = self.rank * batches_per_rank
+        end_idx = start_idx + batches_per_rank
+        rank_batches = buffer_batches[start_idx + self.start_step : end_idx]
+        logging.info(
+            f"rank: {self.rank}, dataloader start from step: {self.start_step}, batch_num: {end_idx-start_idx}, batch_num_after_step: {len(rank_batches)}"
+        )
+        # Return an iterator over the batches for the current rank
+        return iter(rank_batches)
+
+    def __len__(self):
+        # Calculate the number of batches per epoch for the current rank
+        return 1
+
+    def set_epoch(self, epoch):
+        # Set the epoch for shuffling
+        self.epoch = epoch

funasr/datasets/audio_datasets/index_ds.py

@@ -0,0 +1,144 @@
+import os
+import json
+import torch
+import logging
+
+import librosa
+import random
+import torch.distributed as dist
+
+from funasr.register import tables
+
+
+@tables.register("index_ds_classes", "IndexDSJsonl")
+@tables.register("index_ds_classes", "IndexDSJsonlRankFull")
+@tables.register("index_ds_classes", "IndexDSJsonlRankSplit")
+class IndexDSJsonlRankFull(torch.utils.data.Dataset):
+
+    def __init__(self, path: str, **kwargs):
+        super().__init__()
+        self.max_source_length = kwargs.get("max_source_length", 2048)
+        self.min_source_length = kwargs.get("min_source_length", 0)
+        self.max_target_length = kwargs.get("max_target_length", 2048)
+        self.min_target_length = kwargs.get("min_target_length", 0)
+        self.max_token_length = kwargs.get("max_token_length", 2200)
+
+        is_training = kwargs.get("is_training", True)
+        if not (path.endswith(".jsonl") or path.endswith(".json")):
+            # jsonl list file
+            data_split_num = kwargs.get("data_split_num", 1)
+            data_split_i = kwargs.get("data_split_i", 0)
+
+            if not is_training:
+                data_split_num = 1
+                data_split_i = 0
+            with open(path, encoding="utf-8") as fin:
+                file_list_all = fin.readlines()
+
+                num_per_slice = (len(file_list_all) - 1) // data_split_num + 1  # 16
+                file_list = file_list_all[
+                    data_split_i * num_per_slice : (data_split_i + 1) * num_per_slice
+                ]
+                logging.info(
+                    f"is_training: {is_training}, data_split_num: {data_split_num}, data_split_i: {data_split_i}, \nfile_list: {file_list}, \nfile_list_all: {file_list_all}"
+                )
+
+        else:
+            file_list = [path]
+
+        # total_num = len(file_list)
+        # try:
+        #     rank = dist.get_rank()
+        #     world_size = dist.get_world_size()
+        # except:
+        #     rank = 0
+        #     world_size = 1
+        #     logging.info("distributed is not initialized, only single shard")
+        #
+        # if not kwargs.get("rank_split", False):
+        #     logging.info(f"Warning, rank_split disenabled, batch and shuffle data in global")
+        #     rank = 0
+        #     world_size = 1
+        #
+        # num_per_rank = total_num // world_size
+        # if num_per_rank * world_size < total_num:
+        #     logging.info(f"Warning, jsonl file:{total_num} could not be divided by world_size: {world_size}, {path}")
+        #     total_num_needed = num_per_rank * world_size
+        #
+        #     extra_num = total_num_needed - total_num
+        #     file_list_tmp = random.choices(file_list, k=extra_num)
+        #     file_list += file_list_tmp
+        #     logging.info(f"Warning, after random choices: {file_list}")
+        #
+        # file_list_rank = file_list[rank * num_per_rank:(rank + 1) * num_per_rank]
+        #
+        # logging.info(
+        #     f"is_training: {is_training}, file_list_rank: {file_list_rank}")
+
+        # contents = []
+        # for file_json in file_list_rank:
+        contents = []
+        for file_json in file_list:
+            with open(file_json.strip(), encoding="utf-8") as fin:
+                for line in fin:
+                    data = json.loads(line.strip())
+                    if "text" in data:  # for sft
+                        contents.append(data["text"])
+                    if "source" in data:  # for speech lab pretrain
+                        prompt = data.get("prompt", "<ASR>")
+                        source = data["source"].replace(
+                            "/cpfs01", "/cpfs_speech/data"
+                        )  # only use in alibaba gpu group: .replace("/cpfs01", "/cpfs_speech/data")
+                        target = data["target"]
+                        source_len = data.get("source_len", 1)
+                        target_len = data.get("target_len", 0)
+                        if "aishell" in source:
+                            target = target.replace(" ", "")
+                        if (
+                            source_len < self.min_source_length
+                            or source_len > self.max_source_length
+                        ):
+                            continue
+                        if (
+                            target_len < self.min_target_length
+                            or target_len > self.max_target_length
+                        ):
+                            continue
+
+                        if (source_len + target_len) > self.max_token_length:
+                            continue
+
+                        contents_i = {
+                            "source": source,
+                            "prompt": prompt,
+                            "target": target,
+                            "source_len": source_len,
+                            "target_len": target_len,
+                        }
+                        text_language = data.get("text_language", None)
+                        if text_language is not None:
+                            contents_i["text_language"] = text_language
+                        # audio_language = data.get("audio_language", None)
+                        # if audio_language is not None:
+                        #     contents_i["audio_language"] = audio_language
+                        contents.append(contents_i)
+
+        self.contents = contents
+
+        logging.info("total_num of samplers: {}, {}".format(len(self.contents), path))
+
+    def __len__(self):
+        return len(self.contents)
+
+    def __getitem__(self, index):
+
+        data = self.contents[index]
+
+        return data
+
+    def get_source_len(self, data_dict):
+        return data_dict.get("source_len", 1)
+
+    def get_target_len(self, data_dict):
+
+        return data_dict.get("target_len", 0)

funasr/datasets/audio_datasets/jsonl2scp.py

@@ -0,0 +1,64 @@
+import os
+import json
+import torch
+import logging
+import hydra
+from omegaconf import DictConfig, OmegaConf
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+
+
+def gen_scp_from_jsonl(jsonl_file, data_type_list, wav_scp_file, text_file):
+
+    wav_f = open(wav_scp_file, "w")
+    text_f = open(text_file, "w")
+    with open(jsonl_file, encoding="utf-8") as fin:
+        for line in fin:
+            data = json.loads(line.strip())
+
+            prompt = data.get("prompt", "<ASR>")
+            source = data[data_type_list[0]]
+            target = data[data_type_list[1]]
+            source_len = data.get("source_len", 1)
+            target_len = data.get("target_len", 0)
+            if "aishell" in source:
+                target = target.replace(" ", "")
+            key = data["key"]
+            wav_f.write(f"{key}\t{source}\n")
+            wav_f.flush()
+            text_f.write(f"{key}\t{target}\n")
+            text_f.flush()
+
+    wav_f.close()
+    text_f.close()
+
+
+@hydra.main(config_name=None, version_base=None)
+def main_hydra(cfg: DictConfig):
+
+    kwargs = OmegaConf.to_container(cfg, resolve=True)
+    print(kwargs)
+
+    scp_file_list = kwargs.get(
+        "scp_file_list",
+        ("/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"),
+    )
+    if isinstance(scp_file_list, str):
+        scp_file_list = eval(scp_file_list)
+    data_type_list = kwargs.get("data_type_list", ("source", "target"))
+    jsonl_file = kwargs.get(
+        "jsonl_file_in", "/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl"
+    )
+    gen_scp_from_jsonl(jsonl_file, data_type_list, *scp_file_list)
+
+
+"""
+python -m funasr.datasets.audio_datasets.json2scp \
+++scp_file_list='["/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"]' \
+++data_type_list='["source", "target"]' \
+++jsonl_file_in=/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl
+"""
+
+if __name__ == "__main__":
+    main_hydra()

funasr/datasets/audio_datasets/preprocessor.py

@@ -0,0 +1,55 @@
+import os
+import json
+import torch
+import logging
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+from typing import Collection
+import torch
+import torchaudio
+from torch import nn
+import random
+import re
+from funasr.tokenizer.cleaner import TextCleaner
+from funasr.register import tables
+
+
+@tables.register("preprocessor_classes", "SpeechPreprocessSpeedPerturb")
+class SpeechPreprocessSpeedPerturb(nn.Module):
+    def __init__(self, speed_perturb: list = None, **kwargs):
+        super().__init__()
+        self.speed_perturb = speed_perturb
+
+    def forward(self, waveform, fs, **kwargs):
+        if self.speed_perturb is None:
+            return waveform
+        speed = random.choice(self.speed_perturb)
+        if speed != 1.0:
+            if not isinstance(waveform, torch.Tensor):
+                waveform = torch.tensor(waveform)
+            waveform, _ = torchaudio.sox_effects.apply_effects_tensor(
+                waveform.view(1, -1), fs, [["speed", str(speed)], ["rate", str(fs)]]
+            )
+            waveform = waveform.view(-1)
+
+        return waveform
+
+
+@tables.register("preprocessor_classes", "TextPreprocessSegDict")
+class TextPreprocessSegDict(nn.Module):
+    def __init__(
+        self,
+        seg_dict: str = None,
+        text_cleaner: Collection[str] = None,
+        split_with_space: bool = False,
+        **kwargs
+    ):
+        super().__init__()
+
+        self.text_cleaner = TextCleaner(text_cleaner)
+
+    def forward(self, text, **kwargs):
+        text = self.text_cleaner(text)
+
+        return text

funasr/datasets/audio_datasets/samplers.py

@@ -0,0 +1,467 @@
+import torch
+import numpy as np
+import logging
+import math
+import random
+import torch.distributed as dist
+from torch.utils.data import DistributedSampler
+from torch.utils.data import BatchSampler, Sampler
+import torch.distributed as dist
+
+from funasr.register import tables
+
+
+@tables.register("batch_sampler_classes", "BatchSampler")
+@tables.register("batch_sampler_classes", "CustomDistributedBatchSampler")
+@tables.register("batch_sampler_classes", "CustomDistributedDynamicBatchSampler")
+@tables.register("batch_sampler_classes", "DynamicBatchLocalShuffleSampler")
+@tables.register("batch_sampler_classes", "RankFullLocalShuffleBatchSampler")
+@tables.register("batch_sampler_classes", "RankFullLocalShuffleDynamicBatchSampler")
+def CustomDistributedBatchSampler_fn(dataset, **kwargs):
+    dataloader_args = {}
+    batch_type = kwargs.get("batch_type", "example")
+    if batch_type == "example":
+        batch_sampler = CustomDistributedBatchSampler(dataset, **kwargs)
+
+    else:
+        if kwargs.get("sort_size", -1) > 0:
+            batch_sampler = CustomDistributedBufferDynamicBatchSampler(dataset, **kwargs)
+        else:
+            batch_sampler = CustomDistributedDynamicBatchSampler(dataset, **kwargs)
+        # batch_sampler = CustomDistributedDynamicBatchSampler(dataset, **kwargs)
+
+    dataloader_args["batch_sampler"] = batch_sampler
+    dataloader_args["num_workers"] = kwargs.get("num_workers", 4)
+    dataloader_args["pin_memory"] = kwargs.get("pin_memory", True)
+
+    return dataloader_args
+
+
+class CustomDistributedBatchSampler(Sampler):
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        num_replicas=None,
+        rank=None,
+        shuffle=True,
+        drop_last=False,
+        is_training: bool = True,
+        **kwargs,
+    ):
+
+        try:
+            rank = dist.get_rank()
+            num_replicas = dist.get_world_size()
+        except:
+            rank = 0
+            num_replicas = 1
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.is_training = is_training
+        self.shuffle = shuffle and is_training
+        self.drop_last = drop_last
+        # self.total_size = len(dataset)
+        if self.drop_last:
+            self.total_size = (len(self.dataset) // (batch_size * num_replicas)) * (
+                batch_size * num_replicas
+            )
+        else:
+            self.total_size = math.ceil(len(self.dataset) / (batch_size * num_replicas)) * (
+                batch_size * num_replicas
+            )
+        self.num_samples = int(self.total_size // self.num_replicas)
+        self.epoch = 0
+        self.max_token_length = kwargs.get("max_token_length", None)
+        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
+
+    def __iter__(self):
+        # Generate a list of indices
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = list(range(len(self.dataset)))
+
+        # Add extra samples to make it evenly divisible
+        padding_size = self.total_size - len(indices)
+        if padding_size <= len(indices):
+            indices += indices[:padding_size]
+        else:
+            indices += (
+                indices * (padding_size // len(indices)) + indices[: padding_size % len(indices)]
+            )
+
+        assert len(indices) == self.total_size
+
+        # Subsample
+        indices = indices[self.rank : self.total_size : self.num_replicas]
+        assert len(indices) == self.num_samples
+
+        # Filter out indices with length greater than the max length, if provided
+        if self.max_token_length is not None:
+            filtered_indices = []
+            for idx in indices:
+                source_len = self.dataset.get_source_len(idx) / self.length_scale_source
+                if source_len <= self.max_token_length:
+                    filtered_indices.append(idx)
+            indices = filtered_indices
+
+        # Now that we have only the indices for this replica, chunk them into batches
+        batches = [
+            indices[i : i + self.batch_size] for i in range(0, len(indices), self.batch_size)
+        ]
+
+        # Drop the last batch if it's not full and drop_last is True
+        if self.drop_last and len(batches[-1]) != self.batch_size:
+            batches = batches[:-1]
+
+        return iter(batches)
+
+    def __len__(self):
+
+        return self.num_samples // self.batch_size
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
+
+
+class CustomDistributedBufferBatchSampler(Sampler):
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        num_replicas=None,
+        rank=None,
+        shuffle=True,
+        drop_last=False,
+        is_training: bool = True,
+        sort_size: int = 1024,
+        **kwargs,
+    ):
+
+        try:
+            rank = dist.get_rank()
+            num_replicas = dist.get_world_size()
+        except:
+            rank = 0
+            num_replicas = 1
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.is_training = is_training
+        self.shuffle = shuffle and is_training
+        self.drop_last = drop_last
+        # self.total_size = len(dataset)
+        if self.drop_last:
+            self.total_size = (len(self.dataset) // (batch_size * num_replicas)) * (
+                batch_size * num_replicas
+            )
+        else:
+            self.total_size = math.ceil(len(self.dataset) / (batch_size * num_replicas)) * (
+                batch_size * num_replicas
+            )
+        self.num_samples = int(self.total_size // self.num_replicas)
+        self.epoch = 0
+        self.max_token_length = kwargs.get("max_token_length", None)
+        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
+        self.sort_size = sort_size
+
+    def __iter__(self):
+        # Generate a list of indices
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = list(range(len(self.dataset)))
+
+        # Add extra samples to make it evenly divisible
+        padding_size = self.total_size - len(indices)
+        if padding_size <= len(indices):
+            indices += indices[:padding_size]
+        else:
+            indices += (
+                indices * (padding_size // len(indices)) + indices[: padding_size % len(indices)]
+            )
+
+        assert len(indices) == self.total_size
+
+        # Subsample
+        indices = indices[self.rank : self.total_size : self.num_replicas]
+        assert len(indices) == self.num_samples
+
+        # Filter out indices with length greater than the max length, if provided
+        if self.max_token_length is not None:
+            filtered_indices = []
+            for idx in indices:
+                source_len = self.dataset.get_source_len(idx) / self.length_scale_source
+                if source_len <= self.max_token_length:
+                    filtered_indices.append(idx)
+            indices = filtered_indices
+
+        # Buffer sorting logic
+        sorted_batches = []
+        buffer = []
+
+        for idx in indices:
+            buffer.append(idx)
+            if len(buffer) >= self.sort_size:
+                # Sort the buffer based on some criteria, e.g., dataset sample length
+                buffer.sort(key=lambda x: self.dataset.get_source_len(x))
+                sorted_batches.extend(self._create_batches_from_buffer(buffer))
+                buffer = []
+
+        # Handle the remaining items in the buffer
+        if buffer:
+            buffer.sort(key=lambda x: self.dataset.get_source_len(x))
+            sorted_batches.extend(self._create_batches_from_buffer(buffer))
+
+        return iter(sorted_batches)
+
+    def _create_batches_from_buffer(self, buffer):
+        # Function to convert the sorted buffer into batches
+        batched_buffer = [
+            buffer[i : i + self.batch_size] for i in range(0, len(buffer), self.batch_size)
+        ]
+        if self.drop_last and len(batched_buffer[-1]) != self.batch_size:
+            batched_buffer = batched_buffer[:-1]
+        return batched_buffer
+
+    def __len__(self):
+
+        return self.num_samples // self.batch_size
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
+
+
+class CustomDistributedDynamicBatchSampler(DistributedSampler):
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        num_replicas=None,
+        rank=None,
+        shuffle=True,
+        drop_last=False,
+        is_training: bool = True,
+        **kwargs,
+    ):
+
+        try:
+            rank = dist.get_rank()
+            num_replicas = dist.get_world_size()
+        except:
+            rank = 0
+            num_replicas = 1
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.is_training = is_training
+        self.shuffle = shuffle and is_training
+        self.drop_last = drop_last
+
+        self.total_size = len(self.dataset)
+        # self.num_samples = int(math.ceil(self.total_size / self.num_replicas))
+        self.epoch = 0
+        self.max_token_length = kwargs.get("max_token_length", 2048)
+        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
+
+    def __iter__(self):
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = list(range(len(self.dataset)))
+
+        indices = indices[self.rank : self.total_size : self.num_replicas]
+
+        batches = []
+        batch = []
+        max_len_in_batch = 0
+        current_batch_length = 0
+
+        for idx in indices:
+            sample_length = self.dataset.get_source_len(idx)
+            if sample_length > self.max_token_length:
+                continue
+            potential_batch_length = (
+                max_len_in_batch if sample_length < max_len_in_batch else sample_length
+            ) * (len(batch) + 1)
+
+            if potential_batch_length <= self.batch_size:
+                batch.append(idx)
+                if sample_length > max_len_in_batch:
+                    max_len_in_batch = sample_length
+                    # current_batch_length = max_len_in_batch * len(batch)
+            else:
+                batches.append(batch)
+                batch = [idx]
+                max_len_in_batch = sample_length
+                # current_batch_length = max_len_in_batch
+
+        # Add the last batch if it's not empty and we're not dropping it
+        if batch and (not self.drop_last or len(batch) * max_len_in_batch == self.batch_size):
+            batches.append(batch)
+
+        return iter(batches)
+
+    def __len__(self):
+
+        return 1
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
+
+
+class CustomDistributedBufferDynamicBatchSampler(DistributedSampler):
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        batch_type="token",
+        num_replicas=None,
+        rank=None,
+        rank_split=False,
+        shuffle=True,
+        drop_last=False,
+        is_training: bool = True,
+        sort_size: int = 1024,
+        **kwargs,
+    ):
+
+        try:
+            rank = dist.get_rank()
+            num_replicas = dist.get_world_size()
+        except:
+            rank = 0
+            num_replicas = 1
+
+        # if rank_split:
+        #     logging.info(f"Warning, rank_split: {rank_split}, batch and shuffle data in local rank")
+        #     rank = 0
+        #     num_replicas = 1
+
+        self.rank = rank
+        self.num_replicas = num_replicas
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.batch_type = batch_type
+        self.is_training = is_training
+        self.shuffle = shuffle and is_training
+        self.drop_last = drop_last
+
+        self.total_size = len(self.dataset)
+        self.num_samples = int(math.ceil(self.total_size / self.num_replicas))
+        self.epoch = 0
+        self.sort_size = sort_size * num_replicas
+        self.max_token_length = kwargs.get("max_token_length", 2048)
+        self.length_scale_source = kwargs.get("length_scale_source", 1.0)
+        super().__init__(
+            dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle, drop_last=drop_last
+        )
+
+    def __iter__(self):
+        if self.shuffle:
+            g = torch.Generator()
+            g.manual_seed(self.epoch)
+            random.seed(self.epoch)
+
+            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+        else:
+            indices = list(range(len(self.dataset)))
+
+        # Create sorted buffers and form batches
+        buffer_batches = []
+        for i in range(0, len(indices), self.sort_size):
+            buffer = sorted(
+                indices[i : i + self.sort_size], key=lambda idx: self.dataset.get_source_len(idx)
+            )
+            batch = []
+            max_len_in_batch = 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:
+                    batch.append(idx)
+                    max_len_in_batch = max(max_len_in_batch, sample_length)
+                else:
+                    buffer_batches.append(batch)
+                    batch = [idx]
+                    max_len_in_batch = sample_length
+            if batch:
+                buffer_batches.append(batch)
+
+        # Ensure each rank gets the same number of batches, duplicate data if needed
+        batches_per_rank = math.ceil(len(buffer_batches) / self.num_replicas)
+        total_batches_needed = batches_per_rank * self.num_replicas
+
+        extra_batches = total_batches_needed - len(buffer_batches)
+        buffer_batches += random.choices(buffer_batches, k=extra_batches)
+
+        # Evenly distribute batches from buffer_batches to each rank
+        rank_batches = [[] for _ in range(self.num_replicas)]
+        for i, batch in enumerate(buffer_batches):
+            rank_batches[i % self.num_replicas].append(batch)
+
+        # Assign all batches for the current rank directly
+        final_batches = rank_batches[self.rank]
+
+        return iter(final_batches)
+
+    def __len__(self):
+
+        return 1
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
+
+
+class DistributedSamplerWarp(BatchSampler):
+    def __init__(
+        self, dataset, batch_size, num_replicas=None, rank=None, shuffle=True, drop_last=False
+    ):
+        if num_replicas is None:
+            if not torch.distributed.is_available():
+                raise RuntimeError("Requires distributed package to be available")
+            num_replicas = torch.distributed.get_world_size()
+        if rank is None:
+            if not torch.distributed.is_available():
+                raise RuntimeError("Requires distributed package to be available")
+            rank = torch.distributed.get_rank()
+
+        self.dataset = dataset
+        self.batch_size = batch_size
+        self.num_replicas = num_replicas
+        self.rank = rank
+        self.shuffle = shuffle
+        self.drop_last = drop_last
+
+        # Create an instance of the DistributedSampler
+        self.sampler = DistributedSampler(
+            self.dataset, num_replicas=self.num_replicas, rank=self.rank, shuffle=self.shuffle
+        )
+
+        # Call BatchSampler's constructor
+        super().__init__(self.sampler, batch_size, drop_last)
+
+    def __iter__(self):
+        # If we shuffle, we need to call the set_epoch method
+        if self.shuffle:
+            self.sampler.set_epoch(self.epoch)
+
+        # Generate batch indices using the parent class
+        return super().__iter__()
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch

funasr/datasets/audio_datasets/scp2jsonl.py

@@ -0,0 +1,123 @@
+import os
+import json
+import torch
+import logging
+import hydra
+from omegaconf import DictConfig, OmegaConf
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+from tqdm import tqdm
+
+
+def gen_jsonl_from_wav_text_list(
+    path, data_type_list=("source", "target"), jsonl_file_out: str = None, **kwargs
+):
+    try:
+        rank = dist.get_rank()
+        world_size = dist.get_world_size()
+    except:
+        rank = 0
+        world_size = 1
+
+    cpu_cores = os.cpu_count() or 1
+    print(f"convert wav.scp text to jsonl, ncpu: {cpu_cores}")
+    if rank == 0:
+        json_dict = {}
+        for data_type, data_file in zip(data_type_list, path):
+            json_dict[data_type] = {}
+            with open(data_file, "r") as f:
+
+                data_file_lists = f.readlines()
+                lines_for_each_th = (len(data_file_lists) - 1) // cpu_cores + 1
+                task_num = cpu_cores if len(data_file_lists) > cpu_cores else 1
+                # import pdb;pdb.set_trace()
+                if task_num > 1:
+                    with concurrent.futures.ThreadPoolExecutor(max_workers=cpu_cores) as executor:
+
+                        futures = [
+                            executor.submit(
+                                parse_context_length,
+                                data_file_lists[
+                                    i * lines_for_each_th : (i + 1) * lines_for_each_th
+                                ],
+                                data_type,
+                                i,
+                            )
+                            for i in range(task_num)
+                        ]
+
+                        for future in concurrent.futures.as_completed(futures):
+
+                            json_dict[data_type].update(future.result())
+                else:
+                    res = parse_context_length(data_file_lists, data_type)
+                    json_dict[data_type].update(res)
+
+        with open(jsonl_file_out, "w") as f:
+            for key in json_dict[data_type_list[0]].keys():
+                jsonl_line = {"key": key}
+                for data_file in data_type_list:
+                    jsonl_line.update(json_dict[data_file][key])
+                jsonl_line = json.dumps(jsonl_line, ensure_ascii=False)
+                f.write(jsonl_line + "\n")
+                f.flush()
+        print(f"processed {len(json_dict[data_type_list[0]])} samples")
+
+    else:
+        pass
+
+    if world_size > 1:
+        dist.barrier()
+
+
+def parse_context_length(data_list: list, data_type: str, id=0):
+    pbar = tqdm(total=len(data_list), dynamic_ncols=True)
+    res = {}
+    for i, line in enumerate(data_list):
+        pbar.update(1)
+        pbar.set_description(f"cpu: {id}")
+        lines = line.strip().split(maxsplit=1)
+        key = lines[0]
+        line = lines[1] if len(lines) > 1 else ""
+        line = line.strip()
+        if os.path.exists(line):
+            waveform, _ = librosa.load(line, sr=16000)
+            sample_num = len(waveform)
+            context_len = int(sample_num / 16000 * 1000 / 10)
+        else:
+            context_len = len(line.split()) if " " in line else len(line)
+        res[key] = {data_type: line, f"{data_type}_len": context_len}
+    return res
+
+
+@hydra.main(config_name=None, version_base=None)
+def main_hydra(cfg: DictConfig):
+
+    kwargs = OmegaConf.to_container(cfg, resolve=True)
+    print(kwargs)
+
+    scp_file_list = kwargs.get(
+        "scp_file_list",
+        ("/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"),
+    )
+    if isinstance(scp_file_list, str):
+        scp_file_list = eval(scp_file_list)
+    data_type_list = kwargs.get("data_type_list", ("source", "target"))
+    jsonl_file_out = kwargs.get(
+        "jsonl_file_out", "/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl"
+    )
+    gen_jsonl_from_wav_text_list(
+        scp_file_list, data_type_list=data_type_list, jsonl_file_out=jsonl_file_out
+    )
+
+
+"""
+python -m funasr.datasets.audio_datasets.scp2jsonl \
+++scp_file_list='["/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"]' \
+++data_type_list='["source", "target"]' \
+++jsonl_file_out=/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl
+"""
+
+if __name__ == "__main__":
+    main_hydra()

funasr/datasets/audio_datasets/scp2len.py

@@ -0,0 +1,121 @@
+import os
+import json
+import torch
+import logging
+import hydra
+from omegaconf import DictConfig, OmegaConf
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+from tqdm import tqdm
+
+
+def gen_jsonl_from_wav_text_list(
+    path, data_type_list=("source",), jsonl_file_out: str = None, **kwargs
+):
+    try:
+        rank = dist.get_rank()
+        world_size = dist.get_world_size()
+    except:
+        rank = 0
+        world_size = 1
+
+    cpu_cores = os.cpu_count() or 1
+    print(f"convert wav.scp text to jsonl, ncpu: {cpu_cores}")
+    if rank == 0:
+        json_dict = {}
+        # for data_type, data_file in zip(data_type_list, path):
+        data_type = data_type_list[0]
+        data_file = path
+        json_dict[data_type] = {}
+        with open(data_file, "r") as f:
+
+            data_file_lists = f.readlines()
+            print("")
+            lines_for_each_th = (len(data_file_lists) - 1) // cpu_cores + 1
+            task_num = cpu_cores if len(data_file_lists) > cpu_cores else 1
+            # import pdb;pdb.set_trace()
+            if task_num > 1:
+                with concurrent.futures.ThreadPoolExecutor(max_workers=cpu_cores) as executor:
+
+                    futures = [
+                        executor.submit(
+                            parse_context_length,
+                            data_file_lists[i * lines_for_each_th : (i + 1) * lines_for_each_th],
+                            data_type,
+                            i,
+                        )
+                        for i in range(task_num)
+                    ]
+
+                    for future in concurrent.futures.as_completed(futures):
+
+                        json_dict[data_type].update(future.result())
+            else:
+                res = parse_context_length(data_file_lists, data_type)
+                json_dict[data_type].update(res)
+
+        with open(jsonl_file_out, "w") as f:
+            for key in json_dict[data_type_list[0]].keys():
+                jsonl_line = {"key": key}
+                for data_file in data_type_list:
+                    jsonl_line.update(json_dict[data_file][key])
+                # jsonl_line = json.dumps(jsonl_line, ensure_ascii=False)
+                source_len = jsonl_line["source_len"]
+                jsonl_line = f"{key} {source_len}"
+                f.write(jsonl_line + "\n")
+                f.flush()
+        print(f"processed {len(json_dict[data_type_list[0]])} samples")
+
+    else:
+        pass
+
+    if world_size > 1:
+        dist.barrier()
+
+
+def parse_context_length(data_list: list, data_type: str, id=0):
+    pbar = tqdm(total=len(data_list), dynamic_ncols=True)
+    res = {}
+    for i, line in enumerate(data_list):
+        pbar.update(1)
+        pbar.set_description(f"cpu: {id}")
+        lines = line.strip().split(maxsplit=1)
+        key = lines[0]
+        line = lines[1] if len(lines) > 1 else ""
+        line = line.strip()
+        if os.path.exists(line):
+            waveform, _ = librosa.load(line, sr=16000)
+            sample_num = len(waveform)
+            context_len = int(sample_num / 16000 * 1000 / 10)
+        else:
+            context_len = len(line.split()) if " " in line else len(line)
+        res[key] = {data_type: line, f"{data_type}_len": context_len}
+    return res
+
+
+@hydra.main(config_name=None, version_base=None)
+def main_hydra(cfg: DictConfig):
+
+    kwargs = OmegaConf.to_container(cfg, resolve=True)
+    print(kwargs)
+
+    scp_file_list = kwargs.get("scp_file_list", "/Users/zhifu/funasr1.0/data/list/train_wav.scp")
+    # if isinstance(scp_file_list, str):
+    #     scp_file_list = eval(scp_file_list)
+    data_type_list = kwargs.get("data_type_list", ("source",))
+    jsonl_file_out = kwargs.get("jsonl_file_out", "/Users/zhifu/funasr1.0/data/list/wav_len.txt")
+    gen_jsonl_from_wav_text_list(
+        scp_file_list, data_type_list=data_type_list, jsonl_file_out=jsonl_file_out
+    )
+
+
+"""
+python -m funasr.datasets.audio_datasets.scp2jsonl \
+++scp_file_list='["/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"]' \
+++data_type_list='["source", "target"]' \
+++jsonl_file_out=/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl
+"""
+
+if __name__ == "__main__":
+    main_hydra()

funasr/datasets/audio_datasets/update_jsonl.py

@@ -0,0 +1,99 @@
+import os
+import json
+import torch
+import logging
+import hydra
+from omegaconf import DictConfig, OmegaConf
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+import threading
+from tqdm import tqdm
+from concurrent.futures import ThreadPoolExecutor
+
+
+def gen_scp_from_jsonl(jsonl_file, jsonl_file_out, ncpu):
+    jsonl_file_out_f = open(jsonl_file_out, "w")
+    with open(jsonl_file, encoding="utf-8") as fin:
+        lines = fin.readlines()
+
+        num_total = len(lines)
+        if ncpu > 1:
+            # 使用ThreadPoolExecutor限制并发线程数
+            with ThreadPoolExecutor(max_workers=ncpu) as executor:
+                # 提交任务到线程池
+                futures = {executor.submit(update_data, lines, i) for i in tqdm(range(num_total))}
+
+                # 等待所有任务完成，这会阻塞直到所有提交的任务完成
+                for future in concurrent.futures.as_completed(futures):
+                    # 这里可以添加额外的逻辑来处理完成的任务，但在这个例子中我们只是等待
+                    pass
+        else:
+            for i in range(num_total):
+                update_data(lines, i)
+        logging.info("All audio durations have been processed.")
+
+        for line in lines:
+
+            jsonl_file_out_f.write(line + "\n")
+            jsonl_file_out_f.flush()
+
+    jsonl_file_out_f.close()
+
+
+def update_data(lines, i):
+    line = lines[i]
+    data = json.loads(line.strip())
+
+    wav_path = data["source"].replace("/cpfs01", "/cpfs_speech/data")
+    if os.path.exists(wav_path):
+        waveform, _ = librosa.load(wav_path, sr=16000)
+        sample_num = len(waveform)
+        source_len = int(sample_num / 16000 * 1000 / 10)
+        source_len_old = data["source_len"]
+        # if (source_len_old - source_len) > 100 or (source_len - source_len_old) > 100:
+        #     logging.info(f"old: {source_len_old}, new: {source_len}, wav: {wav_path}")
+        data["source_len"] = source_len
+        data["source"] = wav_path
+        jsonl_line = json.dumps(data, ensure_ascii=False)
+        lines[i] = jsonl_line
+
+
+def update_wav_len(jsonl_file_list_in, jsonl_file_out_dir, ncpu=1):
+
+    os.makedirs(jsonl_file_out_dir, exist_ok=True)
+    with open(jsonl_file_list_in, "r") as f:
+        data_file_lists = f.readlines()
+
+        for i, jsonl in enumerate(data_file_lists):
+            filename_with_extension = os.path.basename(jsonl.strip())
+            jsonl_file_out = os.path.join(jsonl_file_out_dir, filename_with_extension)
+            logging.info(f"{i}/{len(data_file_lists)}, jsonl: {jsonl}, {jsonl_file_out}")
+
+            gen_scp_from_jsonl(jsonl.strip(), jsonl_file_out, ncpu)
+
+
+@hydra.main(config_name=None, version_base=None)
+def main_hydra(cfg: DictConfig):
+
+    kwargs = OmegaConf.to_container(cfg, resolve=True)
+    logging.info(kwargs)
+
+    jsonl_file_list_in = kwargs.get(
+        "jsonl_file_list_in", "/Users/zhifu/funasr1.0/data/list/data_jsonl.list"
+    )
+    jsonl_file_out_dir = kwargs.get("jsonl_file_out_dir", "/Users/zhifu/funasr1.0/data_tmp")
+    ncpu = kwargs.get("ncpu", 1)
+    update_wav_len(jsonl_file_list_in, jsonl_file_out_dir, ncpu)
+    # gen_scp_from_jsonl(jsonl_file_list_in, jsonl_file_out_dir)
+
+
+"""
+python -m funasr.datasets.audio_datasets.json2scp \
+++scp_file_list='["/Users/zhifu/funasr1.0/test_local/wav.scp", "/Users/zhifu/funasr1.0/test_local/text.txt"]' \
+++data_type_list='["source", "target"]' \
+++jsonl_file_in=/Users/zhifu/funasr1.0/test_local/audio_datasets.jsonl
+"""
+
+if __name__ == "__main__":
+    main_hydra()

funasr/datasets/dataloader_entry.py

@@ -0,0 +1,136 @@
+import logging
+import torch
+
+from funasr.register import tables
+
+
+# @tables.register("dataloader_classes", "DataloaderMapStyle")
+def DataloaderMapStyle(frontend=None, tokenizer=None, **kwargs):
+    # dataset
+    logging.info("Build dataloader")
+    dataset_class = tables.dataset_classes.get(kwargs.get("dataset", "AudioDataset"))
+    dataset_tr = dataset_class(
+        kwargs.get("train_data_set_list"),
+        frontend=frontend,
+        tokenizer=tokenizer,
+        is_training=True,
+        **kwargs.get("dataset_conf"),
+    )
+    dataset_val = dataset_class(
+        kwargs.get("valid_data_set_list"),
+        frontend=frontend,
+        tokenizer=tokenizer,
+        is_training=False,
+        **kwargs.get("dataset_conf"),
+    )
+
+    # dataloader
+    batch_sampler = kwargs["dataset_conf"].get("batch_sampler", "BatchSampler")
+    batch_sampler_val = None
+    if batch_sampler is not None:
+        batch_sampler_class = tables.batch_sampler_classes.get(batch_sampler)
+        batch_sampler = batch_sampler_class(dataset_tr, **kwargs.get("dataset_conf"))
+        batch_sampler_val = batch_sampler_class(
+            dataset_val, is_training=False, **kwargs.get("dataset_conf")
+        )
+
+    dataloader_tr = torch.utils.data.DataLoader(
+        dataset_tr, collate_fn=dataset_tr.collator, **batch_sampler
+    )
+    dataloader_val = torch.utils.data.DataLoader(
+        dataset_val, collate_fn=dataset_val.collator, **batch_sampler_val
+    )
+
+    return dataloader_tr, dataloader_val
+
+
+@tables.register("dataloader_classes", "DataloaderMapStyle")
+class DataloaderMapStyle:
+    def __init__(self, frontend=None, tokenizer=None, **kwargs):
+        # dataset
+        logging.info("Build dataloader")
+        dataset_class = tables.dataset_classes.get(kwargs.get("dataset", "AudioDataset"))
+        dataset_tr = dataset_class(
+            kwargs.get("train_data_set_list"),
+            frontend=frontend,
+            tokenizer=tokenizer,
+            is_training=True,
+            **kwargs.get("dataset_conf"),
+        )
+        dataset_val = dataset_class(
+            kwargs.get("valid_data_set_list"),
+            frontend=frontend,
+            tokenizer=tokenizer,
+            is_training=False,
+            **kwargs.get("dataset_conf"),
+        )
+
+        self.dataset_tr = dataset_tr
+        self.dataset_val = dataset_val
+        self.kwargs = kwargs
+
+        # split dataset
+        self.data_split_num = kwargs["dataset_conf"].get("data_split_num", 1)
+        self.dataset_class = dataset_class
+        self.frontend = frontend
+        self.tokenizer = tokenizer
+        self.kwargs = kwargs
+
+    def build_iter(self, epoch=0, data_split_i=0, start_step=0, **kwargs):
+
+        # reload dataset slice
+        if self.data_split_num > 1:
+            del self.dataset_tr
+            self.dataset_tr = self.dataset_class(
+                self.kwargs.get("train_data_set_list"),
+                frontend=self.frontend,
+                tokenizer=self.tokenizer,
+                is_training=True,
+                **self.kwargs.get("dataset_conf"),
+                data_split_i=data_split_i,
+            )
+
+        # dataloader
+        batch_sampler = self.kwargs["dataset_conf"].get("batch_sampler", "BatchSampler")
+        batch_sampler_val = None
+        if batch_sampler is not None:
+            batch_sampler_class = tables.batch_sampler_classes.get(batch_sampler)
+            batch_sampler = batch_sampler_class(
+                self.dataset_tr, start_step=start_step, **self.kwargs.get("dataset_conf")
+            )
+            batch_sampler_val = batch_sampler_class(
+                self.dataset_val, is_training=False, **self.kwargs.get("dataset_conf")
+            )
+
+        batch_sampler["batch_sampler"].set_epoch(epoch)
+        batch_sampler_val["batch_sampler"].set_epoch(epoch)
+        dataloader_tr = torch.utils.data.DataLoader(
+            self.dataset_tr, collate_fn=self.dataset_tr.collator, **batch_sampler
+        )
+        dataloader_val = torch.utils.data.DataLoader(
+            self.dataset_val, collate_fn=self.dataset_val.collator, **batch_sampler_val
+        )
+
+        return dataloader_tr, dataloader_val
+
+
+@tables.register("dataloader_classes", "DataloaderIterable")
+def DataloaderIterable(frontend=None, tokenizer=None, **kwargs):
+    logging.info("Build dataloader")
+    dataset_class = tables.dataset_classes.get(kwargs.get("dataset", "LargeDataset"))
+    dataset_tr = dataset_class(
+        kwargs.get("train_data_set_list"),
+        frontend=frontend,
+        tokenizer=tokenizer,
+        is_training=True,
+        **kwargs.get("dataset_conf"),
+    )
+    dataset_val = dataset_class(
+        kwargs.get("valid_data_set_list"),
+        frontend=frontend,
+        tokenizer=tokenizer,
+        is_training=False,
+        **kwargs.get("dataset_conf"),
+    )
+
+    return dataset_tr, dataset_val

funasr/datasets/large_datasets/abs_iter_factory.py

@@ -0,0 +1,9 @@
+from abc import ABC
+from abc import abstractmethod
+from typing import Iterator
+
+
+class AbsIterFactory(ABC):
+    @abstractmethod
+    def build_iter(self, epoch: int, shuffle: bool = None) -> Iterator:
+        raise NotImplementedError

funasr/datasets/large_datasets/build_dataloader.py

@@ -0,0 +1,109 @@
+import logging
+from pathlib import Path
+from typing import Iterable
+from typing import List
+from typing import Union
+
+import sentencepiece as spm
+from torch.utils.data import DataLoader
+
+from funasr.datasets.large_datasets.dataset import Dataset
+from funasr.datasets.large_datasets.abs_iter_factory import AbsIterFactory
+from funasr.tokenizer.abs_tokenizer import AbsTokenizer
+
+from funasr.register import tables
+
+
+def read_symbol_table(symbol_table_file):
+    if isinstance(symbol_table_file, str):
+        symbol_table = {}
+        with open(symbol_table_file, "r", encoding="utf8") as fin:
+            for i, line in enumerate(fin):
+                char = line.strip()
+                symbol_table[char] = i
+    else:
+        assert isinstance(symbol_table_file, list)
+        symbol_table = {}
+        for i, char in enumerate(symbol_table_file):
+            symbol_table[char] = i
+    return symbol_table
+
+
+def load_seg_dict(seg_dict_file):
+    seg_dict = {}
+    assert isinstance(seg_dict_file, str)
+    with open(seg_dict_file, "r", encoding="utf8") as f:
+        lines = f.readlines()
+        for line in lines:
+            s = line.strip().split()
+            key = s[0]
+            value = s[1:]
+            seg_dict[key] = " ".join(value)
+    return seg_dict
+
+
+class SentencepiecesTokenizer(AbsTokenizer):
+    def __init__(self, model: Union[Path, str]):
+        self.model = str(model)
+        self.sp = None
+
+    def __repr__(self):
+        return f'{self.__class__.__name__}(model="{self.model}")'
+
+    def _build_sentence_piece_processor(self):
+        if self.sp is None:
+            self.sp = spm.SentencePieceProcessor()
+            self.sp.load(self.model)
+
+    def text2tokens(self, line: str) -> List[str]:
+        self._build_sentence_piece_processor()
+        return self.sp.EncodeAsPieces(line)
+
+    def tokens2text(self, tokens: Iterable[str]) -> str:
+        self._build_sentence_piece_processor()
+        return self.sp.DecodePieces(list(tokens))
+
+
+@tables.register("dataset_classes", "LargeDataset")
+class LargeDataLoader(AbsIterFactory):
+    def __init__(self, args, mode="train"):
+        symbol_table, seg_dict, punc_dict, bpe_tokenizer = None, None, None, None
+        if hasattr(args, "token_list") and args.token_list is not None:
+            symbol_table = read_symbol_table(args.token_list)
+        if hasattr(args, "seg_dict_file") and args.seg_dict_file is not None:
+            seg_dict = load_seg_dict(args.seg_dict_file)
+        if hasattr(args, "punc_list") and args.punc_list is not None:
+            punc_dict = read_symbol_table(args.punc_list)
+        if hasattr(args, "bpemodel") and args.bpemodel is not None:
+            bpe_tokenizer = SentencepiecesTokenizer(args.bpemodel)
+        self.dataset_conf = args.dataset_conf
+        if "frontend_conf" not in args:
+            self.frontend_conf = None
+        else:
+            self.frontend_conf = args.frontend_conf
+        self.speed_perturb = args.speed_perturb if hasattr(args, "speed_perturb") else None
+        logging.info("dataloader config: {}".format(self.dataset_conf))
+        batch_mode = self.dataset_conf.get("batch_mode", "padding")
+        data_list = args.train_data_file if mode == "train" else args.valid_data_file
+        self.dataset = Dataset(
+            data_list,
+            symbol_table,
+            seg_dict,
+            punc_dict,
+            bpe_tokenizer,
+            self.dataset_conf,
+            self.frontend_conf,
+            speed_perturb=self.speed_perturb if mode == "train" else None,
+            mode=mode,
+            batch_mode=batch_mode,
+        )
+
+    def build_iter(self, epoch, shuffle=True):
+        self.dataset.set_epoch(epoch)
+        data_loader = DataLoader(
+            self.dataset,
+            batch_size=None,
+            pin_memory=True,
+            num_workers=self.dataset_conf.get("num_workers", 8),
+        )
+        return data_loader

funasr/datasets/large_datasets/collate_fn.py

@@ -0,0 +1,194 @@
+from typing import Collection
+from typing import Dict
+from typing import List
+from typing import Tuple
+from typing import Union
+
+import numpy as np
+import torch
+from funasr.models.transformer.utils.nets_utils import pad_list, pad_list_all_dim
+
+
+class CommonCollateFn:
+    """Functor class of common_collate_fn()"""
+
+    def __init__(
+        self,
+        float_pad_value: Union[float, int] = 0.0,
+        int_pad_value: int = -32768,
+        not_sequence: Collection[str] = (),
+        max_sample_size=None,
+    ):
+        self.float_pad_value = float_pad_value
+        self.int_pad_value = int_pad_value
+        self.not_sequence = set(not_sequence)
+        self.max_sample_size = max_sample_size
+
+    def __repr__(self):
+        return (
+            f"{self.__class__}(float_pad_value={self.float_pad_value}, "
+            f"int_pad_value={self.float_pad_value})"
+        )
+
+    def __call__(
+        self, data: Collection[Tuple[str, Dict[str, np.ndarray]]]
+    ) -> Tuple[List[str], Dict[str, torch.Tensor]]:
+        return common_collate_fn(
+            data,
+            float_pad_value=self.float_pad_value,
+            int_pad_value=self.int_pad_value,
+            not_sequence=self.not_sequence,
+        )
+
+
+def common_collate_fn(
+    data: Collection[Tuple[str, Dict[str, np.ndarray]]],
+    float_pad_value: Union[float, int] = 0.0,
+    int_pad_value: int = -32768,
+    not_sequence: Collection[str] = (),
+) -> Tuple[List[str], Dict[str, torch.Tensor]]:
+    """Concatenate ndarray-list to an array and convert to torch.Tensor."""
+    uttids = [u for u, _ in data]
+    data = [d for _, d in data]
+
+    assert all(set(data[0]) == set(d) for d in data), "dict-keys mismatching"
+    assert all(
+        not k.endswith("_lengths") for k in data[0]
+    ), f"*_lengths is reserved: {list(data[0])}"
+
+    output = {}
+    for key in data[0]:
+        if data[0][key].dtype.kind == "i":
+            pad_value = int_pad_value
+        else:
+            pad_value = float_pad_value
+
+        array_list = [d[key] for d in data]
+        tensor_list = [torch.from_numpy(a) for a in array_list]
+        tensor = pad_list(tensor_list, pad_value)
+        output[key] = tensor
+
+        if key not in not_sequence:
+            lens = torch.tensor([d[key].shape[0] for d in data], dtype=torch.long)
+            output[key + "_lengths"] = lens
+
+    output = (uttids, output)
+    return output
+
+
+class DiarCollateFn:
+    """Functor class of common_collate_fn()"""
+
+    def __init__(
+        self,
+        float_pad_value: Union[float, int] = 0.0,
+        int_pad_value: int = -32768,
+        not_sequence: Collection[str] = (),
+        max_sample_size=None,
+    ):
+        self.float_pad_value = float_pad_value
+        self.int_pad_value = int_pad_value
+        self.not_sequence = set(not_sequence)
+        self.max_sample_size = max_sample_size
+
+    def __repr__(self):
+        return (
+            f"{self.__class__}(float_pad_value={self.float_pad_value}, "
+            f"int_pad_value={self.float_pad_value})"
+        )
+
+    def __call__(
+        self, data: Collection[Tuple[str, Dict[str, np.ndarray]]]
+    ) -> Tuple[List[str], Dict[str, torch.Tensor]]:
+        return diar_collate_fn(
+            data,
+            float_pad_value=self.float_pad_value,
+            int_pad_value=self.int_pad_value,
+            not_sequence=self.not_sequence,
+        )
+
+
+def diar_collate_fn(
+    data: Collection[Tuple[str, Dict[str, np.ndarray]]],
+    float_pad_value: Union[float, int] = 0.0,
+    int_pad_value: int = -32768,
+    not_sequence: Collection[str] = (),
+) -> Tuple[List[str], Dict[str, torch.Tensor]]:
+    """Concatenate ndarray-list to an array and convert to torch.Tensor."""
+    uttids = [u for u, _ in data]
+    data = [d for _, d in data]
+
+    assert all(set(data[0]) == set(d) for d in data), "dict-keys mismatching"
+    assert all(
+        not k.endswith("_lengths") for k in data[0]
+    ), f"*_lengths is reserved: {list(data[0])}"
+
+    output = {}
+    for key in data[0]:
+        if data[0][key].dtype.kind == "i":
+            pad_value = int_pad_value
+        else:
+            pad_value = float_pad_value
+
+        array_list = [d[key] for d in data]
+        tensor_list = [torch.from_numpy(a) for a in array_list]
+        tensor = pad_list_all_dim(tensor_list, pad_value)
+        output[key] = tensor
+
+        if key not in not_sequence:
+            lens = torch.tensor([d[key].shape[0] for d in data], dtype=torch.long)
+            output[key + "_lengths"] = lens
+
+    output = (uttids, output)
+    return output
+
+
+def crop_to_max_size(feature, target_size):
+    size = len(feature)
+    diff = size - target_size
+    if diff <= 0:
+        return feature
+
+    start = np.random.randint(0, diff + 1)
+    end = size - diff + start
+    return feature[start:end]
+
+
+def clipping_collate_fn(
+    data: Collection[Tuple[str, Dict[str, np.ndarray]]],
+    max_sample_size=None,
+    not_sequence: Collection[str] = (),
+) -> Tuple[List[str], Dict[str, torch.Tensor]]:
+    # mainly for pre-training
+    uttids = [u for u, _ in data]
+    data = [d for _, d in data]
+
+    assert all(set(data[0]) == set(d) for d in data), "dict-keys mismatching"
+    assert all(
+        not k.endswith("_lengths") for k in data[0]
+    ), f"*_lengths is reserved: {list(data[0])}"
+
+    output = {}
+    for key in data[0]:
+        array_list = [d[key] for d in data]
+        tensor_list = [torch.from_numpy(a) for a in array_list]
+        sizes = [len(s) for s in tensor_list]
+        if max_sample_size is None:
+            target_size = min(sizes)
+        else:
+            target_size = min(min(sizes), max_sample_size)
+        tensor = tensor_list[0].new_zeros(len(tensor_list), target_size, tensor_list[0].shape[1])
+        for i, (source, size) in enumerate(zip(tensor_list, sizes)):
+            diff = size - target_size
+            if diff == 0:
+                tensor[i] = source
+            else:
+                tensor[i] = crop_to_max_size(source, target_size)
+        output[key] = tensor
+
+        if key not in not_sequence:
+            lens = torch.tensor([source.shape[0] for source in tensor], dtype=torch.long)
+            output[key + "_lengths"] = lens
+
+    output = (uttids, output)
+    return output

funasr/datasets/large_datasets/datapipes/batch.py

@@ -0,0 +1,213 @@
+import random
+
+from itertools import count
+from functools import partial
+from torch.utils.data import IterableDataset
+from funasr.datasets.large_datasets.datapipes.map import MapperIterDataPipe
+
+tiebreaker = count()
+
+
+def _default_len_fn(token):
+    return len(token), next(tiebreaker)
+
+
+def _token_len_fn(token, len_fn):
+    return len_fn(token), next(tiebreaker), token
+
+
+class MaxTokenBucketizerIterDataPipe(IterableDataset):
+
+    def __init__(
+        self,
+        datapipe,
+        batch_size=8000,
+        len_fn=_default_len_fn,
+        buffer_size=10240,
+        sort_size=500,
+        batch_mode="padding",
+    ):
+        assert batch_size > 0, "Batch size is required to be larger than 0!"
+        assert buffer_size >= -1, "Buffer size is required to be larger than -1!"
+        assert sort_size > 0, "Sort size is required to be larger than 0!"
+
+        datapipe = MapperIterDataPipe(datapipe, fn=partial(_token_len_fn, len_fn=len_fn))
+        self.datapipe = datapipe
+        self.batch_size = batch_size
+        self.buffer_size = buffer_size
+        self.sort_size = sort_size
+        self.batch_mode = batch_mode
+
+    def set_epoch(self, epoch):
+        self.datapipe.set_epoch(epoch)
+
+    def __iter__(self):
+        buffer = []
+        batch = []
+        bucket = []
+        max_lengths = 0
+        min_lengths = 999999
+        batch_lengths = 0
+
+        if self.batch_mode == "clipping":
+            assert self.buffer_size > 0, "for clipping batch_mode, buffer_size must be > 1"
+            for d in self.datapipe:
+                if d[0] > self.batch_size:
+                    continue
+                buffer.append(d)
+                if len(buffer) == self.buffer_size:
+                    random.shuffle(buffer)
+                    for sample in buffer:
+                        bucket.append(sample)
+                        if len(bucket) == self.sort_size:
+                            bucket.sort()
+                            for x in bucket:
+                                length, _, token = x
+                                if length < min_lengths:
+                                    min_lengths = length
+                                batch_lengths = min_lengths * (len(batch) + 1)
+                                if batch_lengths > self.batch_size:
+                                    yield batch
+                                    batch = []
+                                    min_lengths = length
+                                batch.append(token)
+                            bucket = []
+                    buffer = []
+
+            if buffer:
+                random.shuffle(buffer)
+                for sample in buffer:
+                    bucket.append(sample)
+                    if len(bucket) == self.sort_size:
+                        bucket.sort()
+                        for x in bucket:
+                            length, _, token = x
+                            if length < min_lengths:
+                                min_lengths = length
+                            batch_lengths = min_lengths * (len(batch) + 1)
+                            if batch_lengths > self.batch_size:
+                                yield batch
+                                batch = []
+                                min_lengths = length
+                            batch.append(token)
+                        bucket = []
+                buffer = []
+
+            if bucket:
+                bucket.sort()
+                for x in bucket:
+                    length, _, token = x
+                    if length < min_lengths:
+                        min_lengths = length
+                    batch_lengths = min_lengths * (len(batch) + 1)
+                    if batch_lengths > self.batch_size:
+                        yield batch
+                        batch = []
+                        min_lengths = length
+                    batch.append(token)
+                bucket = []
+
+            if batch:
+                yield batch
+
+        else:
+            if self.buffer_size == -1:
+                for d in self.datapipe:
+                    if d[0] > self.batch_size:
+                        continue
+                    buffer.append(d)
+                buffer.sort()
+                for sample in buffer:
+                    length, _, token = sample
+                    if length > max_lengths:
+                        max_lengths = length
+                    batch_lengths = max_lengths * (len(batch) + 1)
+                    if batch_lengths > self.batch_size:
+                        bucket.append(batch)
+                        batch = []
+                        max_lengths = length
+                    batch.append(token)
+                random.shuffle(bucket)
+                if bucket:
+                    for batch_sample in bucket:
+                        yield batch_sample
+                if batch:
+                    yield batch
+
+            elif self.buffer_size == 0:
+                for d in self.datapipe:
+                    if d[0] > self.batch_size:
+                        continue
+                    length, _, token = d
+                    if length > self.batch_size:
+                        continue
+                    if length > max_lengths:
+                        max_lengths = length
+                    batch_lengths = max_lengths * (len(batch) + 1)
+                    if batch_lengths > self.batch_size:
+                        yield batch
+                        batch = []
+                        max_lengths = length
+                    batch.append(token)
+                if batch:
+                    yield batch
+
+            else:
+                for d in self.datapipe:
+                    if d[0] > self.batch_size:
+                        continue
+                    buffer.append(d)
+                    if len(buffer) == self.buffer_size:
+                        random.shuffle(buffer)
+                        for sample in buffer:
+                            bucket.append(sample)
+                            if len(bucket) == self.sort_size:
+                                bucket.sort()
+                                for x in bucket:
+                                    length, _, token = x
+                                    if length > max_lengths:
+                                        max_lengths = length
+                                    batch_lengths = max_lengths * (len(batch) + 1)
+                                    if batch_lengths > self.batch_size:
+                                        yield batch
+                                        batch = []
+                                        max_lengths = length
+                                    batch.append(token)
+                                bucket = []
+                        buffer = []
+
+                if buffer:
+                    random.shuffle(buffer)
+                    for sample in buffer:
+                        bucket.append(sample)
+                        if len(bucket) == self.sort_size:
+                            bucket.sort()
+                            for x in bucket:
+                                length, _, token = x
+                                if length > max_lengths:
+                                    max_lengths = length
+                                batch_lengths = max_lengths * (len(batch) + 1)
+                                if batch_lengths > self.batch_size:
+                                    yield batch
+                                    batch = []
+                                    max_lengths = length
+                                batch.append(token)
+                            bucket = []
+                    buffer = []
+
+                if bucket:
+                    bucket.sort()
+                    for x in bucket:
+                        length, _, token = x
+                        if length > max_lengths:
+                            max_lengths = length
+                        batch_lengths = max_lengths * (len(batch) + 1)
+                        if batch_lengths > self.batch_size:
+                            yield batch
+                            batch = []
+                            max_lengths = length
+                        batch.append(token)
+                    bucket = []
+
+                if batch:
+                    yield batch

funasr/datasets/large_datasets/datapipes/filter.py

@@ -0,0 +1,23 @@
+from torch.utils.data import IterableDataset
+
+
+def default_fn(data):
+    return data
+
+
+class FilterIterDataPipe(IterableDataset):
+
+    def __init__(self, datapipe, fn=default_fn):
+        self.datapipe = datapipe
+        self.fn = fn
+
+    def set_epoch(self, epoch):
+        self.datapipe.set_epoch(epoch)
+
+    def __iter__(self):
+        assert callable(self.fn)
+        for data in self.datapipe:
+            if self.fn(data):
+                yield data
+            else:
+                continue

funasr/datasets/large_datasets/datapipes/map.py

@@ -0,0 +1,20 @@
+from torch.utils.data import IterableDataset
+
+
+def default_fn(data):
+    return data
+
+
+class MapperIterDataPipe(IterableDataset):
+
+    def __init__(self, datapipe, fn=default_fn):
+        self.datapipe = datapipe
+        self.fn = fn
+
+    def set_epoch(self, epoch):
+        self.datapipe.set_epoch(epoch)
+
+    def __iter__(self):
+        assert callable(self.fn)
+        for data in self.datapipe:
+            yield self.fn(data)

funasr/datasets/large_datasets/dataset.py

@@ -0,0 +1,299 @@
+import logging
+import os
+import random
+from functools import partial
+
+import torch
+import torch.distributed as dist
+import torchaudio
+import numpy as np
+
+# import librosa
+import librosa
+from kaldiio import ReadHelper
+from torch.utils.data import IterableDataset
+
+from funasr.datasets.large_datasets.datapipes.batch import MaxTokenBucketizerIterDataPipe
+from funasr.datasets.large_datasets.datapipes.filter import FilterIterDataPipe
+from funasr.datasets.large_datasets.datapipes.map import MapperIterDataPipe
+from funasr.datasets.large_datasets.utils.clipping import clipping
+from funasr.datasets.large_datasets.utils.filter import filter
+from funasr.datasets.large_datasets.utils.padding import padding
+from funasr.datasets.large_datasets.utils.tokenize import tokenize
+
+
+def read_lists(list_file):
+    lists = []
+    with open(list_file, "r", encoding="utf8") as fin:
+        for line in fin:
+            parts = line.strip()
+            lists.append(parts)
+    return lists
+
+
+class AudioDataset(IterableDataset):
+    def __init__(
+        self,
+        scp_lists,
+        data_names,
+        data_types,
+        frontend_conf=None,
+        shuffle=True,
+        speed_perturb=None,
+        mode="train",
+    ):
+        self.scp_lists = scp_lists
+        self.data_names = data_names
+        self.data_types = data_types
+        self.frontend_conf = frontend_conf
+        self.shuffle = shuffle
+        self.mode = mode
+        self.epoch = -1
+        self.rank = 0
+        self.world_size = 1
+        self.worker_id = 0
+        self.num_workers = 1
+        self.speed_perturb = speed_perturb
+        if self.speed_perturb is not None:
+            logging.info("Using speed_perturb: {}".format(speed_perturb))
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
+
+    def get_rank_data_list(self, data_index):
+        assert dist.is_available()
+        if dist.is_initialized():
+            self.rank = dist.get_rank()
+            self.world_size = dist.get_world_size()
+        else:
+            self.rank = 0
+            self.world_size = 1
+
+        if self.mode == "train":
+            if self.shuffle:
+                random.seed(self.epoch)
+                random.shuffle(data_index)
+            return data_index[self.rank :: self.world_size]
+
+        return data_index
+
+    def get_worker_data_list(self, rank_data_index):
+        worker_info = torch.utils.data.get_worker_info()
+        if worker_info is None:
+            self.worker_id = 0
+            self.num_workers = 1
+        else:
+            self.worker_id = worker_info.id
+            self.num_workers = worker_info.num_workers
+
+        return rank_data_index[self.worker_id :: self.num_workers]
+
+    def close_reader(self, reader_list):
+        for reader in reader_list:
+            reader.close()
+
+    def __iter__(self):
+        data_index = list(range(len(self.scp_lists)))
+        rank_data_index = self.get_rank_data_list(data_index)
+        worker_data_index = self.get_worker_data_list(rank_data_index)
+
+        for index in worker_data_index:
+            data = dict(scp=self.scp_lists[index])
+
+            assert "scp" in data
+            scp = data["scp"]
+            data_file_list = scp.strip().split()
+            data_name_list = self.data_names.split(",")
+            data_type_list = self.data_types.split(",")
+
+            for file in data_file_list:
+                assert os.path.exists(file), "{} not exists".format(file)
+
+            assert (
+                len(data_file_list) == len(data_name_list) == len(data_type_list)
+            ), "The item number of data, data_names, data_types must be the same "
+
+            reader_list = []
+            for data_file, data_type in zip(data_file_list, data_type_list):
+                if data_type == "kaldi_ark":
+                    ark_reader = ReadHelper("ark:{}".format(data_file))
+                    reader_list.append(ark_reader)
+                elif data_type == "text" or data_type == "sound" or data_type == "text_hotword":
+                    text_reader = open(data_file, "r", encoding="utf-8")
+                    reader_list.append(text_reader)
+                elif data_type == "none":
+                    continue
+                else:
+                    raise TypeError("Data type {} is not supported".format(data_type))
+
+            for items in zip(*reader_list):
+                sample_dict = {}
+                for item, (data_name, data_type) in zip(items, zip(data_name_list, data_type_list)):
+                    if data_type == "kaldi_ark":
+                        key, mat = item
+                        sample_dict[data_name] = mat
+                        if data_name == "speech":
+                            sample_dict["key"] = key
+                    elif data_type == "sound":
+                        key, path = item.strip().split()
+                        try:
+                            waveform, sampling_rate = torchaudio.load(path)
+                        except:
+                            # waveform, sampling_rate = librosa.load(path, dtype='float32')
+                            waveform, sampling_rate = librosa.load(path, dtype="float32")
+                            if waveform.ndim == 2:
+                                waveform = waveform[:, 0]
+                            waveform = np.expand_dims(waveform, axis=0)
+                            waveform = torch.tensor(waveform)
+                        if self.frontend_conf is not None:
+                            if sampling_rate != self.frontend_conf["fs"]:
+                                waveform = torchaudio.transforms.Resample(
+                                    orig_freq=sampling_rate, new_freq=self.frontend_conf["fs"]
+                                )(waveform)
+                                sampling_rate = self.frontend_conf["fs"]
+                        waveform = waveform.numpy()
+                        mat = waveform[0]
+                        if self.speed_perturb is not None:
+                            speed = random.choice(self.speed_perturb)
+                            if speed != 1.0:
+                                mat, _ = torchaudio.sox_effects.apply_effects_tensor(
+                                    torch.tensor(mat).view(1, -1),
+                                    sampling_rate,
+                                    [["speed", str(speed)], ["rate", str(sampling_rate)]],
+                                )
+                                mat = mat.view(-1).numpy()
+                        sample_dict[data_name] = mat
+                        sample_dict["sampling_rate"] = sampling_rate
+                        if data_name == "speech":
+                            sample_dict["key"] = key
+                    elif data_type == "text_hotword":
+                        text = item
+                        segs = text.strip().split()
+                        sample_dict[data_name] = segs[1:]
+                        if "key" not in sample_dict:
+                            sample_dict["key"] = segs[0]
+                        sample_dict["hw_tag"] = 1
+                    elif data_type == "text_nospace":
+                        text = item
+                        segs = text.strip().split(maxsplit=1)
+                        sample_dict[data_name] = [x for x in segs[1]]
+                        if "key" not in sample_dict:
+                            sample_dict["key"] = segs[0]
+                    else:
+                        text = item
+                        segs = text.strip().split()
+                        sample_dict[data_name] = segs[1:]
+                        if "key" not in sample_dict:
+                            sample_dict["key"] = segs[0]
+                yield sample_dict
+
+            self.close_reader(reader_list)
+
+
+def len_fn_example(data):
+    return 1
+
+
+def len_fn_token(data):
+    assert "speech" in data
+    if "sampling_rate" in data:
+        return (data["speech"].shape[0] / data["sampling_rate"]) * 1000.0
+    else:
+        return data["speech"].shape[0]
+
+
+def Dataset(
+    data_list_file,
+    dict,
+    seg_dict,
+    punc_dict,
+    bpe_tokenizer,
+    conf,
+    frontend_conf,
+    speed_perturb=None,
+    mode="train",
+    batch_mode="padding",
+):
+    scp_lists = read_lists(data_list_file)
+    shuffle = conf.get("shuffle", True)
+    data_names = conf.get("data_names", "speech,text")
+    data_types = conf.get("data_types", "kaldi_ark,text")
+
+    pre_hwfile = conf.get("pre_hwlist", None)
+    # pre_prob = conf.get("pre_prob", 0)  # unused yet
+    if pre_hwfile is not None:
+        pre_hwlist = []
+        with open(pre_hwfile, "r", encoding="utf-8") as fin:
+            for line in fin.readlines():
+                pre_hwlist.append(line.strip())
+    else:
+        pre_hwlist = None
+
+    hw_config = {
+        "sample_rate": conf.get("sample_rate", 0.6),
+        "double_rate": conf.get("double_rate", 0.1),
+        "hotword_min_length": conf.get("hotword_min_length", 2),
+        "hotword_max_length": conf.get("hotword_max_length", 8),
+        "pre_prob": conf.get("pre_prob", 0.0),
+        "pre_hwlist": pre_hwlist,
+    }
+
+    dataset = AudioDataset(
+        scp_lists,
+        data_names,
+        data_types,
+        frontend_conf=frontend_conf,
+        shuffle=shuffle,
+        speed_perturb=speed_perturb,
+        mode=mode,
+    )
+
+    if "text" in data_names:
+        vocab = {
+            "vocab": dict,
+            "seg_dict": seg_dict,
+            "punc_dict": punc_dict,
+            "bpe_tokenizer": bpe_tokenizer,
+            "hw_config": hw_config,
+        }
+        tokenize_fn = partial(tokenize, **vocab)
+        dataset = MapperIterDataPipe(dataset, fn=tokenize_fn)
+
+    filter_conf = conf.get("filter_conf", {})
+    filter_fn = partial(filter, **filter_conf)
+    dataset = FilterIterDataPipe(dataset, fn=filter_fn)
+
+    if shuffle:
+        buffer_conf = conf.get("shuffle_conf", {})
+        buffer_size = buffer_conf["shuffle_size"]
+        sort_size = buffer_conf["sort_size"]
+    else:
+        buffer_size = 0
+        sort_size = 1
+
+    batch_conf = conf.get("batch_conf", {})
+    batch_size = batch_conf["batch_size"]
+    batch_type = batch_conf["batch_type"]
+
+    assert batch_type in ["example", "token"]
+    if batch_type == "example":
+        len_fn = len_fn_example
+    else:
+        len_fn = len_fn_token
+
+    dataset = MaxTokenBucketizerIterDataPipe(
+        dataset,
+        batch_size=batch_size,
+        len_fn=len_fn,
+        buffer_size=buffer_size,
+        sort_size=sort_size,
+        batch_mode=batch_mode,
+    )
+
+    int_pad_value = conf.get("int_pad_value", -1)
+    float_pad_value = conf.get("float_pad_value", 0.0)
+    padding_conf = {"int_pad_value": int_pad_value, "float_pad_value": float_pad_value}
+    padding_fn = partial(padding, **padding_conf)
+    dataset = MapperIterDataPipe(dataset, fn=padding_fn if batch_mode == "padding" else clipping)
+
+    return dataset

funasr/datasets/large_datasets/utils/clipping.py

@@ -0,0 +1,44 @@
+import numpy as np
+import torch
+
+from funasr.datasets.large_datasets.collate_fn import crop_to_max_size
+
+
+def clipping(data):
+    assert isinstance(data, list)
+    assert "key" in data[0]
+
+    keys = [x["key"] for x in data]
+
+    batch = {}
+    data_names = data[0].keys()
+    for data_name in data_names:
+        if data_name == "key":
+            continue
+        else:
+            if data[0][data_name].dtype.kind == "i":
+                tensor_type = torch.int64
+            else:
+                tensor_type = torch.float32
+
+            tensor_list = [torch.tensor(np.copy(d[data_name]), dtype=tensor_type) for d in data]
+            tensor_lengths = torch.tensor([len(d[data_name]) for d in data], dtype=torch.int32)
+
+            length_clip = min(tensor_lengths)
+            tensor_clip = tensor_list[0].new_zeros(
+                len(tensor_list), length_clip, tensor_list[0].shape[1]
+            )
+            for i, (tensor, length) in enumerate(zip(tensor_list, tensor_lengths)):
+                diff = length - length_clip
+                assert diff >= 0
+                if diff == 0:
+                    tensor_clip[i] = tensor
+                else:
+                    tensor_clip[i] = crop_to_max_size(tensor, length_clip)
+
+            batch[data_name] = tensor_clip
+            batch[data_name + "_lengths"] = torch.tensor(
+                [tensor.shape[0] for tensor in tensor_clip], dtype=torch.long
+            )
+
+    return keys, batch

funasr/datasets/large_datasets/utils/filter.py

@@ -0,0 +1,27 @@
+#!/usr/bin/env python
+
+
+def filter(
+    data, speech_length_min=100, speech_length_max=15000, token_length_min=0, token_length_max=200
+):
+    assert "speech" in data or "text" in data
+
+    if "speech" in data and "text" in data:
+        if "sampling_rate" in data:
+            speech_length = (data["speech"].shape[0] / data["sampling_rate"]) * 1000.0
+        else:
+            speech_length = data["speech"].shape[0]
+        num_tokens = len(data["text"])
+        return (
+            speech_length_min < speech_length < speech_length_max
+            and token_length_min < num_tokens < token_length_max
+        )
+    elif "speech" in data:
+        if "sampling_rate" in data:
+            speech_length = (data["speech"].shape[0] / data["sampling_rate"]) * 1000.0
+        else:
+            speech_length = data["speech"].shape[0]
+        return speech_length_min < speech_length < speech_length_max
+    else:
+        num_tokens = len(data["text"])
+        return token_length_min < num_tokens < token_length_max

funasr/datasets/large_datasets/utils/hotword_utils.py

@@ -0,0 +1,42 @@
+import random
+
+
+def sample_hotword(
+    length,
+    hotword_min_length,
+    hotword_max_length,
+    sample_rate,
+    double_rate,
+    pre_prob,
+    pre_index=None,
+    pre_hwlist=None,
+):
+    if length < hotword_min_length:
+        return [-1]
+    if random.random() < sample_rate:
+        if pre_prob > 0 and random.random() < pre_prob and pre_index is not None:
+            return pre_index
+        if length == hotword_min_length:
+            return [0, length - 1]
+        elif random.random() < double_rate and length > hotword_max_length + hotword_min_length + 2:
+            # sample two hotwords in a sentence
+            _max_hw_length = min(hotword_max_length, length // 2)
+            # first hotword
+            start1 = random.randint(0, length // 3)
+            end1 = random.randint(start1 + hotword_min_length - 1, start1 + _max_hw_length - 1)
+            # second hotword
+            start2 = random.randint(end1 + 1, length - hotword_min_length)
+            end2 = random.randint(
+                min(length - 1, start2 + hotword_min_length - 1),
+                min(length - 1, start2 + hotword_max_length - 1),
+            )
+            return [start1, end1, start2, end2]
+        else:  # single hotword
+            start = random.randint(0, length - hotword_min_length)
+            end = random.randint(
+                min(length - 1, start + hotword_min_length - 1),
+                min(length - 1, start + hotword_max_length - 1),
+            )
+            return [start, end]
+    else:
+        return [-1]

funasr/datasets/large_datasets/utils/low_frame_rate.py

@@ -0,0 +1,30 @@
+import numpy as np
+
+
+def build_LFR_features(data, m, n):
+    """
+    Actually, this implements stacking frames and skipping frames.
+    if m = 1 and n = 1, just return the origin features.
+    if m = 1 and n > 1, it works like skipping.
+    if m > 1 and n = 1, it works like stacking but only support right frames.
+    if m > 1 and n > 1, it works like LFR.
+
+    Args:
+        inputs_batch: inputs is T x D np.ndarray
+        m: number of frames to stack
+        n: number of frames to skip
+    """
+
+    LFR_inputs = []
+    T = data.shape[0]
+    T_lfr = int(np.ceil(T / n))
+    for i in range(T_lfr):
+        if m <= T - i * n:
+            LFR_inputs.append(np.hstack(data[i * n : i * n + m]))
+        else:
+            num_padding = m - (T - i * n)
+            frame = np.hstack(data[i * n :])
+            for _ in range(num_padding):
+                frame = np.hstack((frame, data[-1]))
+            LFR_inputs.append(frame)
+    return np.vstack(LFR_inputs)

funasr/datasets/large_datasets/utils/padding.py

@@ -0,0 +1,72 @@
+import numpy as np
+import torch
+from torch.nn.utils.rnn import pad_sequence
+
+
+def padding(data, float_pad_value=0.0, int_pad_value=-1):
+    assert isinstance(data, list)
+    assert "key" in data[0]
+    assert "speech" in data[0] or "text" in data[0]
+
+    keys = [x["key"] for x in data]
+
+    batch = {}
+    data_names = data[0].keys()
+    for data_name in data_names:
+        if data_name == "key" or data_name == "sampling_rate":
+            continue
+        else:
+            if data_name != "hotword_indxs":
+                if data[0][data_name].dtype.kind == "i":
+                    pad_value = int_pad_value
+                    tensor_type = torch.int64
+                else:
+                    pad_value = float_pad_value
+                    tensor_type = torch.float32
+
+            tensor_list = [torch.tensor(np.copy(d[data_name]), dtype=tensor_type) for d in data]
+            tensor_lengths = torch.tensor([len(d[data_name]) for d in data], dtype=torch.int32)
+            tensor_pad = pad_sequence(tensor_list, batch_first=True, padding_value=pad_value)
+            batch[data_name] = tensor_pad
+            batch[data_name + "_lengths"] = tensor_lengths
+
+    # SAC LABEL INCLUDE
+    if "hotword_indxs" in batch:
+        # if hotword indxs in batch
+        # use it to slice hotwords out
+        hotword_list = []
+        hotword_lengths = []
+        text = batch["text"]
+        text_lengths = batch["text_lengths"]
+        hotword_indxs = batch["hotword_indxs"]
+        dha_pad = torch.ones_like(text) * -1
+        _, t1 = text.shape
+        t1 += 1  # TODO: as parameter which is same as predictor_bias
+        nth_hw = 0
+        for b, (hotword_indx, one_text, length) in enumerate(
+            zip(hotword_indxs, text, text_lengths)
+        ):
+            dha_pad[b][:length] = 8405
+            if hotword_indx[0] != -1:
+                start, end = int(hotword_indx[0]), int(hotword_indx[1])
+                hotword = one_text[start : end + 1]
+                hotword_list.append(hotword)
+                hotword_lengths.append(end - start + 1)
+                dha_pad[b][start : end + 1] = one_text[start : end + 1]
+                nth_hw += 1
+                if len(hotword_indx) == 4 and hotword_indx[2] != -1:
+                    # the second hotword if exist
+                    start, end = int(hotword_indx[2]), int(hotword_indx[3])
+                    hotword_list.append(one_text[start : end + 1])
+                    hotword_lengths.append(end - start + 1)
+                    dha_pad[b][start : end + 1] = one_text[start : end + 1]
+                    nth_hw += 1
+        hotword_list.append(torch.tensor([1]))
+        hotword_lengths.append(1)
+        hotword_pad = pad_sequence(hotword_list, batch_first=True, padding_value=0)
+        batch["hotword_pad"] = hotword_pad
+        batch["hotword_lengths"] = torch.tensor(hotword_lengths, dtype=torch.int32)
+        batch["dha_pad"] = dha_pad
+        del batch["hotword_indxs"]
+        del batch["hotword_indxs_lengths"]
+    return keys, batch

funasr/datasets/large_datasets/utils/tokenize.py

@@ -0,0 +1,93 @@
+#!/usr/bin/env python
+import re
+import numpy as np
+from funasr.datasets.large_datasets.utils.hotword_utils import sample_hotword
+
+
+def forward_segment(text, seg_dict):
+    word_list = []
+    i = 0
+    while i < len(text):
+        longest_word = text[i]
+        for j in range(i + 1, len(text) + 1):
+            word = text[i:j]
+            if word in seg_dict:
+                if len(word) > len(longest_word):
+                    longest_word = word
+        word_list.append(longest_word)
+        i += len(longest_word)
+    return word_list
+
+
+def seg_tokenize(txt, seg_dict):
+    pattern = re.compile(r"^[\u4E00-\u9FA50-9]+$")
+    out_txt = ""
+    for word in txt:
+        word = word.lower()
+        if word in seg_dict:
+            out_txt += seg_dict[word] + " "
+        else:
+            if pattern.match(word):
+                for char in word:
+                    if char in seg_dict:
+                        out_txt += seg_dict[char] + " "
+                    else:
+                        out_txt += "<unk>" + " "
+            else:
+                out_txt += "<unk>" + " "
+    return out_txt.strip().split()
+
+
+def tokenize(data, vocab=None, seg_dict=None, punc_dict=None, bpe_tokenizer=None, hw_config=None):
+    assert "text" in data
+    assert isinstance(vocab, dict)
+    text = data["text"]
+    token = []
+    vad = -2
+    if bpe_tokenizer is not None:
+        text = bpe_tokenizer.text2tokens(" ".join(text))
+    if seg_dict is not None:
+        assert isinstance(seg_dict, dict)
+        text = seg_tokenize(text, seg_dict)
+
+    length = len(text)
+    if "hw_tag" in data:
+        pre_index = None
+        if hw_config["pre_hwlist"] is not None and hw_config["pre_prob"] > 0:
+            # enable preset hotword detect in sampling
+            for hw in hw_config["pre_hwlist"]:
+                hw = " ".join(seg_tokenize(hw, seg_dict))
+                _find = " ".join(text).find(hw)
+                if _find != -1:
+                    # _find = text[:_find].count(" ")  # bpe sometimes
+                    pre_index = [_find, _find + max(hw.count(" "), 1)]
+                    break
+        hotword_indxs = sample_hotword(length, **hw_config, pre_index=pre_index)
+        data["hotword_indxs"] = hotword_indxs
+        del data["hw_tag"]
+    for i in range(length):
+        x = text[i]
+        if i == length - 1 and "punc" in data and x.startswith("vad:"):
+            vad = x[4:]
+            if len(vad) == 0:
+                vad = -1
+            else:
+                vad = int(vad)
+        elif x in vocab:
+            token.append(vocab[x])
+        else:
+            token.append(vocab["<unk>"])
+
+    if "punc" in data and punc_dict is not None:
+        punc_token = []
+        for punc in data["punc"]:
+            if punc in punc_dict:
+                punc_token.append(punc_dict[punc])
+            else:
+                punc_token.append(punc_dict["_"])
+        data["punc"] = np.array(punc_token)
+
+    data["text"] = np.array(token)
+    if vad is not -2:
+        data["vad_indexes"] = np.array([vad], dtype=np.int64)
+    return data

funasr/datasets/llm_datasets/datasets.py

@@ -0,0 +1,436 @@
+import torch
+import copy
+
+from funasr.register import tables
+from funasr.utils.load_utils import extract_fbank, load_audio_text_image_video
+
+
+@tables.register("dataset_classes", "AudioLLMNARDataset")
+class AudioLLMNARDataset(torch.utils.data.Dataset):
+    """
+    AudioLLMDataset
+    """
+
+    def __init__(
+        self,
+        path,
+        index_ds: str = None,
+        frontend=None,
+        tokenizer=None,
+        int_pad_value: int = -1,
+        float_pad_value: float = 0.0,
+        **kwargs
+    ):
+        super().__init__()
+        index_ds_class = tables.index_ds_classes.get(index_ds)
+        self.index_ds = index_ds_class(path, **kwargs)
+        preprocessor_speech = kwargs.get("preprocessor_speech", None)
+        if preprocessor_speech:
+            preprocessor_speech_class = tables.preprocessor_classes.get(preprocessor_speech)
+            preprocessor_speech = preprocessor_speech_class(
+                **kwargs.get("preprocessor_speech_conf", {})
+            )
+        self.preprocessor_speech = preprocessor_speech
+        preprocessor_text = kwargs.get("preprocessor_text", None)
+        if preprocessor_text:
+            preprocessor_text_class = tables.preprocessor_classes.get(preprocessor_text)
+            preprocessor_text = preprocessor_text_class(**kwargs.get("preprocessor_text_conf", {}))
+        self.preprocessor_text = preprocessor_text
+
+        self.frontend = frontend
+        self.fs = 16000 if frontend is None else frontend.fs
+        self.data_type = "sound"
+        self.tokenizer = tokenizer
+
+        self.float_pad_value = float_pad_value
+        self.prompt = kwargs.get("prompt", "Please copy the following text.")
+        self.prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(
+            self.prompt
+        )  # "USER: \nINSTRUCTION: {}\nINPUT: {}\nASSISTANT: "
+        self.prompt_af = ""
+        self.IGNORE_INDEX = kwargs.get("IGNORE_INDEX", -100)
+        self.int_pad_value = self.IGNORE_INDEX
+
+    def get_source_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_source_len(item)
+
+    def get_target_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_target_len(item)
+
+    def __len__(self):
+        return len(self.index_ds)
+
+    def __getitem__(self, index):
+        item = self.index_ds[index]
+        # import pdb;
+        # pdb.set_trace()
+        source = item["source"]
+        data_src = load_audio_text_image_video(source, fs=self.fs)
+        if self.preprocessor_speech:
+            data_src = self.preprocessor_speech(data_src, fs=self.fs)
+        speech, speech_lengths = extract_fbank(
+            data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
+        )  # speech: [b, T, d]
+        speech = speech.squeeze(0)
+
+        target = item["target"]
+        if self.preprocessor_text:
+            target = self.preprocessor_text(target)
+
+        prompt_ids_pre = self.tokenizer.encode(self.prompt_pre)  # [bos,prompt]
+        prompt_ids_length = len(prompt_ids_pre)
+
+        # bos prompt audio bos target
+        # prompt_input = "{}{}".format(self.prompt_pre, target)
+        # prompt_input_ids = self.tokenizer.encode(prompt_input) #[bos, prompt, input]
+        # audio_length = len(prompt_input_ids) - prompt_ids_length
+        target_ids = self.tokenizer.encode(target)
+        if target_ids[0] == self.tokenizer.bos_token_id:
+            target_ids = target_ids[1:]
+        target_ids_length = len(target_ids)
+        audio_length = target_ids_length
+        input_ids = (
+            prompt_ids_pre + target_ids + [self.tokenizer.pad_token_id] + target_ids
+        )  # [bos, prompt, input, pad, target]
+        input_ids = torch.tensor(
+            copy.deepcopy(input_ids), dtype=torch.int64
+        )  # [bos, prompt, input, pad, target]
+        input_ids[prompt_ids_length : prompt_ids_length + audio_length] = (
+            -1
+        )  # [bos, prompt,-1, pad, target] # it is no need, only for check
+        attention_mask = input_ids.ge(-1)  # [true, true, true, true, true], length mask
+
+        # bos prompt audio target eos
+        # prompt_answer = "{}{}".format(self.prompt_pre, target)
+        # prompt_answer_ids = self.tokenizer.encode(prompt_answer) #[bos, prompt, input]
+        # answer_length = len(prompt_answer_ids) - prompt_ids_length
+        target_ids = self.tokenizer.encode(target)
+        if target_ids[0] == self.tokenizer.bos_token_id:
+            target_ids = target_ids[1:]
+        # target_ids_length = len(target_ids)
+        labels_ids = (
+            prompt_ids_pre + target_ids + target_ids + [self.tokenizer.eos_token_id]
+        )  # [bos, prompt, input, target, eos]
+        labels_ids = torch.tensor(
+            copy.deepcopy(labels_ids), dtype=torch.int64
+        )  # [bos, prompt, input, target, eos]
+        labels_ids[:prompt_ids_length] = -1  # [-1, -1, input, target, eos]
+        label_mask = labels_ids.ge(0)  # [false, false, true, true, true], length mask
+        labels_ids[~label_mask] = self.IGNORE_INDEX  # [-1, -1, input, target, eos]
+
+        audio_mask = (
+            [0] * prompt_ids_length + [1] * audio_length + [0] * target_ids_length + [0]
+        )  # [0, 0, 1, 0, 0]
+        audio_mask = torch.tensor(audio_mask, dtype=torch.float32)
+
+        ids = target_ids  # self.tokenizer.encode(target) # token ids is different from labels_ids
+        text = torch.tensor(ids, dtype=torch.int64)
+        text_lengths = torch.tensor([len(ids)], dtype=torch.int32)
+
+        prompt_bos_length = torch.tensor([len(prompt_ids_pre)], dtype=torch.int32)
+
+        return {
+            "speech": speech,
+            "speech_lengths": speech_lengths,
+            "text": text,
+            "text_lengths": text_lengths,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "labels_ids": labels_ids,
+            "label_mask": label_mask,
+            "audio_mask": audio_mask,
+            "prompt_bos_length": prompt_bos_length,
+        }
+
+    def collator(self, samples: list = None):
+        outputs = {}
+        for sample in samples:
+            for key in sample.keys():
+                if key not in outputs:
+                    outputs[key] = []
+                outputs[key].append(sample[key])
+
+        for key, data_list in outputs.items():
+            if isinstance(data_list[0], torch.Tensor):
+                if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:
+
+                    pad_value = self.int_pad_value
+                else:
+                    pad_value = self.float_pad_value
+
+                outputs[key] = torch.nn.utils.rnn.pad_sequence(
+                    data_list, batch_first=True, padding_value=pad_value
+                )
+        return outputs
+
+
+@tables.register("dataset_classes", "AudioLLMDataset")
+class AudioLLMDataset(torch.utils.data.Dataset):
+    """
+    AudioLLMDataset
+    """
+
+    def __init__(
+        self,
+        path,
+        index_ds: str = None,
+        frontend=None,
+        tokenizer=None,
+        int_pad_value: int = -1,
+        float_pad_value: float = 0.0,
+        **kwargs
+    ):
+        super().__init__()
+        index_ds_class = tables.index_ds_classes.get(index_ds)
+        self.index_ds = index_ds_class(path, **kwargs)
+        preprocessor_speech = kwargs.get("preprocessor_speech", None)
+        if preprocessor_speech:
+            preprocessor_speech_class = tables.preprocessor_classes.get(preprocessor_speech)
+            preprocessor_speech = preprocessor_speech_class(
+                **kwargs.get("preprocessor_speech_conf", {})
+            )
+        self.preprocessor_speech = preprocessor_speech
+        preprocessor_text = kwargs.get("preprocessor_text", None)
+        if preprocessor_text:
+            preprocessor_text_class = tables.preprocessor_classes.get(preprocessor_text)
+            preprocessor_text = preprocessor_text_class(**kwargs.get("preprocessor_text_conf", {}))
+        self.preprocessor_text = preprocessor_text
+
+        self.frontend = frontend
+        self.fs = 16000 if frontend is None else frontend.fs
+        self.data_type = "sound"
+        self.tokenizer = tokenizer
+
+        self.float_pad_value = float_pad_value
+        self.prompt = kwargs.get("prompt", "Transcribe speech to text.")
+        self.prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(
+            self.prompt
+        )  # "USER: \nINSTRUCTION: {}\nnINPUT: {}\nASSISTANT: "
+        self.prompt_af = ""
+        self.IGNORE_INDEX = kwargs.get("IGNORE_INDEX", -100)
+        self.int_pad_value = self.IGNORE_INDEX
+
+    def get_source_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_source_len(item)
+
+    def get_target_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_target_len(item)
+
+    def __len__(self):
+        return len(self.index_ds)
+
+    def __getitem__(self, index):
+        item = self.index_ds[index]
+        # import pdb;
+        # pdb.set_trace()
+        source = item["source"]
+        data_src = load_audio_text_image_video(source, fs=self.fs)
+        if self.preprocessor_speech:
+            data_src = self.preprocessor_speech(data_src, fs=self.fs)
+        speech, speech_lengths = extract_fbank(
+            data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
+        )  # speech: [b, T, d]
+        speech = speech.squeeze(0)
+
+        target = item["target"]
+        if self.preprocessor_text:
+            target = self.preprocessor_text(target)
+
+        prompt_ids_pre = self.tokenizer.encode(self.prompt_pre)  # [bos,prompt]
+        prompt_ids_length = len(prompt_ids_pre)
+
+        prompt_input = "{}{}".format(self.prompt_pre, target)
+        prompt_input_ids = self.tokenizer.encode(prompt_input)
+        audio_length = len(prompt_input_ids) - prompt_ids_length
+        input_ids = prompt_input_ids + [self.tokenizer.pad_token_id]
+        input_ids = torch.tensor(input_ids, dtype=torch.int64)  # [bos, prompt, input, pad]
+        input_ids[prompt_ids_length:] = -1  # [bos, prompt,-1,-1]
+        attention_mask = input_ids.ge(-1)  # [true, true, true, true], length mask
+
+        prompt_answer = "{}{}".format(self.prompt_pre, target)
+        prompt_answer_ids = self.tokenizer.encode(prompt_answer)
+        answer_length = len(prompt_answer_ids) - prompt_ids_length
+        labels_ids = copy.deepcopy(prompt_input_ids) + [self.tokenizer.eos_token_id]
+        labels_ids = torch.tensor(labels_ids, dtype=torch.int64)  # [bos, prompt, input, eos]
+        labels_ids[:prompt_ids_length] = -1  # [-1, -1, input, eos]
+        label_mask = labels_ids.ge(0)  # [False,False,True,True]
+        labels_ids[~label_mask] = self.IGNORE_INDEX  # [-100,-100,input,eos]
+
+        audio_mask = [0] * prompt_ids_length + [1] * audio_length + [0]
+        audio_mask = torch.tensor(audio_mask, dtype=torch.float32)
+
+        ids = self.tokenizer.encode(target)  # token ids is different from labels_ids
+        text = torch.tensor(ids, dtype=torch.int64)
+        text_lengths = torch.tensor([len(ids)], dtype=torch.int32)
+
+        return {
+            "speech": speech,
+            "speech_lengths": speech_lengths,
+            "text": text,
+            "text_lengths": text_lengths,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "labels_ids": labels_ids,
+            "label_mask": label_mask,
+            "audio_mask": audio_mask,
+        }
+
+    def collator(self, samples: list = None):
+        outputs = {}
+        for sample in samples:
+            for key in sample.keys():
+                if key not in outputs:
+                    outputs[key] = []
+                outputs[key].append(sample[key])
+
+        for key, data_list in outputs.items():
+            if isinstance(data_list[0], torch.Tensor):
+                if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:
+
+                    pad_value = self.int_pad_value
+                else:
+                    pad_value = self.float_pad_value
+
+                outputs[key] = torch.nn.utils.rnn.pad_sequence(
+                    data_list, batch_first=True, padding_value=pad_value
+                )
+        return outputs
+
+
+@tables.register("dataset_classes", "AudioLLMARDataset")
+class AudioLLMARDataset(torch.utils.data.Dataset):
+    """
+    AudioLLMDataset
+    """
+
+    def __init__(
+        self,
+        path,
+        index_ds: str = None,
+        frontend=None,
+        tokenizer=None,
+        int_pad_value: int = -1,
+        float_pad_value: float = 0.0,
+        **kwargs
+    ):
+        super().__init__()
+        index_ds_class = tables.index_ds_classes.get(index_ds)
+        self.index_ds = index_ds_class(path, **kwargs)
+        preprocessor_speech = kwargs.get("preprocessor_speech", None)
+        if preprocessor_speech:
+            preprocessor_speech_class = tables.preprocessor_classes.get(preprocessor_speech)
+            preprocessor_speech = preprocessor_speech_class(
+                **kwargs.get("preprocessor_speech_conf", {})
+            )
+        self.preprocessor_speech = preprocessor_speech
+        preprocessor_text = kwargs.get("preprocessor_text", None)
+        if preprocessor_text:
+            preprocessor_text_class = tables.preprocessor_classes.get(preprocessor_text)
+            preprocessor_text = preprocessor_text_class(**kwargs.get("preprocessor_text_conf", {}))
+        self.preprocessor_text = preprocessor_text
+
+        self.frontend = frontend
+        self.fs = 16000 if frontend is None else frontend.fs
+        self.data_type = "sound"
+        self.tokenizer = tokenizer
+
+        self.float_pad_value = float_pad_value
+        self.prompt = kwargs.get("prompt", "Transcribe speech to text.")
+        self.prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(
+            self.prompt
+        )  # "USER: \nINSTRUCTION: {}\nnINPUT: {}\nASSISTANT: "
+        self.prompt_af = ""
+        self.IGNORE_INDEX = kwargs.get("IGNORE_INDEX", -100)
+        self.int_pad_value = self.IGNORE_INDEX
+
+    def get_source_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_source_len(item)
+
+    def get_target_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_target_len(item)
+
+    def __len__(self):
+        return len(self.index_ds)
+
+    def __getitem__(self, index):
+        item = self.index_ds[index]
+        # import pdb;
+        # pdb.set_trace()
+        source = item["source"]
+        data_src = load_audio_text_image_video(source, fs=self.fs)
+        if self.preprocessor_speech:
+            data_src = self.preprocessor_speech(data_src, fs=self.fs)
+        speech, speech_lengths = extract_fbank(
+            data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
+        )  # speech: [b, T, d]
+        speech = speech.squeeze(0)
+
+        target = item["target"]
+        if self.preprocessor_text:
+            target = self.preprocessor_text(target)
+
+        prompt_ids_pre = self.tokenizer.encode(self.prompt_pre)  # [bos,prompt]
+        prompt_ids_length = len(prompt_ids_pre)
+
+        prompt_input = "{}{}".format(self.prompt_pre, target)
+        prompt_input_ids = self.tokenizer.encode(prompt_input)
+        audio_length = len(prompt_input_ids) - prompt_ids_length
+        input_ids = prompt_input_ids + [self.tokenizer.pad_token_id]
+        input_ids = torch.tensor(input_ids, dtype=torch.int64)  # [bos, prompt, input, pad]
+        input_ids[prompt_ids_length:] = -1  # [bos, prompt,-1,-1]
+        attention_mask = input_ids.ge(-1)  # [true, true, true, true], length mask
+
+        prompt_answer = "{}{}".format(self.prompt_pre, target)
+        prompt_answer_ids = self.tokenizer.encode(prompt_answer)
+        answer_length = len(prompt_answer_ids) - prompt_ids_length
+        labels_ids = copy.deepcopy(prompt_input_ids) + [self.tokenizer.eos_token_id]
+        labels_ids = torch.tensor(labels_ids, dtype=torch.int64)  # [bos, prompt, input, eos]
+        labels_ids[:prompt_ids_length] = -1  # [-1, -1, input, eos]
+        label_mask = labels_ids.ge(0)  # [False,False,True,True]
+        labels_ids[~label_mask] = self.IGNORE_INDEX  # [-100,-100,input,eos]
+
+        audio_mask = [0] * prompt_ids_length + [1] * audio_length + [0]
+        audio_mask = torch.tensor(audio_mask, dtype=torch.float32)
+
+        ids = self.tokenizer.encode(target)  # token ids is different from labels_ids
+        text = torch.tensor(ids, dtype=torch.int64)
+        text_lengths = torch.tensor([len(ids)], dtype=torch.int32)
+
+        return {
+            "speech": speech,
+            "speech_lengths": speech_lengths,
+            "text": text,
+            "text_lengths": text_lengths,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "labels_ids": labels_ids,
+            "label_mask": label_mask,
+            "audio_mask": audio_mask,
+        }
+
+    def collator(self, samples: list = None):
+        outputs = {}
+        for sample in samples:
+            for key in sample.keys():
+                if key not in outputs:
+                    outputs[key] = []
+                outputs[key].append(sample[key])
+
+        for key, data_list in outputs.items():
+            if isinstance(data_list[0], torch.Tensor):
+                if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:
+
+                    pad_value = self.int_pad_value
+                else:
+                    pad_value = self.float_pad_value
+
+                outputs[key] = torch.nn.utils.rnn.pad_sequence(
+                    data_list, batch_first=True, padding_value=pad_value
+                )
+        return outputs

funasr/datasets/llm_datasets/preprocessor.py

@@ -0,0 +1,34 @@
+import os
+import json
+import torch
+import logging
+import concurrent.futures
+import librosa
+import torch.distributed as dist
+from typing import Collection
+import torch
+import torchaudio
+from torch import nn
+import random
+import re
+import string
+from funasr.tokenizer.cleaner import TextCleaner
+from funasr.register import tables
+
+
+@tables.register("preprocessor_classes", "TextPreprocessRemovePunctuation")
+class TextPreprocessRemovePunctuation(nn.Module):
+    def __init__(self, **kwargs):
+        super().__init__()
+
+    def forward(self, text, **kwargs):
+        # 定义英文标点符号
+        en_punct = string.punctuation
+        # 定义中文标点符号（部分常用的）
+        cn_punct = "。？！，、；：“”‘’（）《》【】…—～·"
+        # 合并英文和中文标点符号
+        all_punct = en_punct + cn_punct
+        # 创建正则表达式模式，匹配任何在all_punct中的字符
+        punct_pattern = re.compile("[{}]".format(re.escape(all_punct)))
+        # 使用正则表达式的sub方法替换掉这些字符
+        return punct_pattern.sub("", text)

funasr/datasets/llm_datasets_qwenaudio/datasets.py

@@ -0,0 +1,161 @@
+import torch
+import copy
+
+from funasr.register import tables
+from funasr.utils.load_utils import extract_fbank, load_audio_text_image_video
+
+
+@tables.register("dataset_classes", "AudioLLMQwenAudioDataset")
+class AudioLLMQwenAudioDataset(torch.utils.data.Dataset):
+    """
+    AudioLLMDataset
+    """
+
+    def __init__(
+        self,
+        path,
+        index_ds: str = None,
+        frontend=None,
+        tokenizer=None,
+        int_pad_value: int = -1,
+        float_pad_value: float = 0.0,
+        **kwargs
+    ):
+        super().__init__()
+        index_ds_class = tables.index_ds_classes.get(index_ds)
+        self.index_ds = index_ds_class(path, **kwargs)
+        preprocessor_speech = kwargs.get("preprocessor_speech", None)
+        if preprocessor_speech:
+            preprocessor_speech_class = tables.preprocessor_classes.get(preprocessor_speech)
+            preprocessor_speech = preprocessor_speech_class(
+                **kwargs.get("preprocessor_speech_conf", {})
+            )
+        self.preprocessor_speech = preprocessor_speech
+        preprocessor_text = kwargs.get("preprocessor_text", None)
+        if preprocessor_text:
+            preprocessor_text_class = tables.preprocessor_classes.get(preprocessor_text)
+            preprocessor_text = preprocessor_text_class(**kwargs.get("preprocessor_text_conf", {}))
+        self.preprocessor_text = preprocessor_text
+
+        self.frontend = frontend
+        self.fs = 16000 if frontend is None else frontend.fs
+        self.data_type = "sound"
+        self.tokenizer = tokenizer
+
+        self.float_pad_value = float_pad_value
+        self.prompt = kwargs.get("prompt", "Transcribe speech to text.")
+        # self.prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(self.prompt)  # "USER: \nINSTRUCTION: {}\nnINPUT: {}\nASSISTANT: "
+        self.prompt_af = ""
+        self.IGNORE_INDEX = kwargs.get("IGNORE_INDEX", -100)
+        self.int_pad_value = self.IGNORE_INDEX
+        self.audio_adaptor_downsample_rate = kwargs.get("audio_adaptor_downsample_rate", 5)
+        self.audio_encoder_downsample_rate = kwargs.get("audio_encoder_downsample_rate", 2)
+        self.prompt_template = "{}"
+        self.answer_template = "{}"
+
+    def get_source_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_source_len(item)
+
+    def get_target_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_target_len(item)
+
+    def __len__(self):
+        return len(self.index_ds)
+
+    def __getitem__(self, index):
+        item = self.index_ds[index]
+        # import pdb;
+        # pdb.set_trace()
+        source = item["source"]
+        data_src = load_audio_text_image_video(source, fs=self.fs)
+        if self.preprocessor_speech:
+            data_src = self.preprocessor_speech(data_src, fs=self.fs)
+        speech, speech_lengths = extract_fbank(
+            data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
+        )  # speech: [b, T, d]
+        speech = speech.squeeze(0)
+
+        audio_pseudo_length = (
+            (speech.shape[0] + 1)
+            // self.audio_adaptor_downsample_rate
+            // self.audio_encoder_downsample_rate
+        )
+        audio_pseudo = torch.full((audio_pseudo_length,), -1)  # placeholder
+
+        target = item["target"]
+        if self.preprocessor_text:
+            target = self.preprocessor_text(target)
+
+        self.prompt_pre = self.prompt_template.format(self.prompt)
+        prompt_ids_pre = self.tokenizer.encode(self.prompt_pre)  # [bos,prompt]
+        prompt_pre_length = len(prompt_ids_pre)
+
+        # input
+        input = self.answer_template.format(target.lower())
+        prompt_input = "{}{}".format(self.prompt_pre, input)
+        prompt_input_ids = self.tokenizer.encode(prompt_input)  # [bos, prompt, input]
+        # audio_length = len(prompt_input_ids) - prompt_pre_length
+        input_ids = prompt_input_ids + [self.tokenizer.pad_token_id]  # [bos, prompt, input, pad]
+        input_ids_length = len(input_ids)
+        input_ids = torch.tensor(input_ids, dtype=torch.int64)  # [bos, prompt, input, pad]
+        input_ids = torch.cat((audio_pseudo, input_ids))  # [audio, bos, prompt, input, pad]
+        # input_ids[:audio_pseudo_length] = -1 # [-1, bos, prompt, input, pad]
+        attention_mask = input_ids.ge(-1)  # [true, true, true, true, true], length mask
+        # input_ids[prompt_pre_length:] = -1  # [bos, prompt,-1,-1]
+        # attention_mask = input_ids.ge(-1)  # [true, true, true, true], length mask
+
+        # label
+        answer = self.answer_template.format(target.lower())
+        prompt_answer = "{}{}".format(self.prompt_pre, answer)
+        prompt_answer_ids = self.tokenizer.encode(prompt_answer)
+        # answer_length = len(prompt_answer_ids) - prompt_pre_length
+        labels_ids = copy.deepcopy(prompt_answer_ids) + [self.tokenizer.eos_token_id]
+        labels_ids = torch.tensor(labels_ids, dtype=torch.int64)  # [bos, prompt, answer, eos]
+        labels_ids = torch.cat((audio_pseudo, labels_ids))  # [audio, bos, prompt, answer, eos]
+        labels_ids[: audio_pseudo_length + prompt_pre_length] = -1  # [-1, -1, -1, answer, eos]
+        # labels_ids[:prompt_pre_length] = -1  # [-1, -1, input, eos]
+        label_mask = labels_ids.ge(0)  # [false, false, false, true, true]
+        labels_ids[~label_mask] = self.IGNORE_INDEX  # [-100, -100, -100, answer, eos]
+
+        # audio_mask for input_ids
+        audio_mask = [1] * audio_pseudo_length + [0] * input_ids_length
+        audio_mask = torch.tensor(audio_mask, dtype=torch.float32)
+
+        ids = self.tokenizer.encode(target)  # token ids is different from labels_ids
+        text = torch.tensor(ids, dtype=torch.int64)
+        text_lengths = torch.tensor([len(ids)], dtype=torch.int32)
+
+        return {
+            "speech": speech,
+            "speech_lengths": speech_lengths,
+            "text": text,
+            "text_lengths": text_lengths,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "labels_ids": labels_ids,
+            "label_mask": label_mask,
+            "audio_mask": audio_mask,
+        }
+
+    def collator(self, samples: list = None):
+        outputs = {}
+        for sample in samples:
+            for key in sample.keys():
+                if key not in outputs:
+                    outputs[key] = []
+                outputs[key].append(sample[key])
+
+        for key, data_list in outputs.items():
+            if isinstance(data_list[0], torch.Tensor):
+                if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:
+
+                    pad_value = self.int_pad_value
+                else:
+                    pad_value = self.float_pad_value
+
+                outputs[key] = torch.nn.utils.rnn.pad_sequence(
+                    data_list, batch_first=True, padding_value=pad_value
+                )
+        return outputs

funasr/datasets/llm_datasets_vicuna/datasets.py

@@ -0,0 +1,161 @@
+import torch
+import copy
+
+from funasr.register import tables
+from funasr.utils.load_utils import extract_fbank, load_audio_text_image_video
+
+
+@tables.register("dataset_classes", "AudioLLMVicunaDataset")
+class AudioLLMVicunaDataset(torch.utils.data.Dataset):
+    """
+    AudioLLMDataset
+    """
+
+    def __init__(
+        self,
+        path,
+        index_ds: str = None,
+        frontend=None,
+        tokenizer=None,
+        int_pad_value: int = -1,
+        float_pad_value: float = 0.0,
+        **kwargs
+    ):
+        super().__init__()
+        index_ds_class = tables.index_ds_classes.get(index_ds)
+        self.index_ds = index_ds_class(path, **kwargs)
+        preprocessor_speech = kwargs.get("preprocessor_speech", None)
+        if preprocessor_speech:
+            preprocessor_speech_class = tables.preprocessor_classes.get(preprocessor_speech)
+            preprocessor_speech = preprocessor_speech_class(
+                **kwargs.get("preprocessor_speech_conf", {})
+            )
+        self.preprocessor_speech = preprocessor_speech
+        preprocessor_text = kwargs.get("preprocessor_text", None)
+        if preprocessor_text:
+            preprocessor_text_class = tables.preprocessor_classes.get(preprocessor_text)
+            preprocessor_text = preprocessor_text_class(**kwargs.get("preprocessor_text_conf", {}))
+        self.preprocessor_text = preprocessor_text
+
+        self.frontend = frontend
+        self.fs = 16000 if frontend is None else frontend.fs
+        self.data_type = "sound"
+        self.tokenizer = tokenizer
+
+        self.float_pad_value = float_pad_value
+        self.prompt = kwargs.get("prompt", "Transcribe speech to text.")
+        # self.prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(self.prompt)  # "USER: \nINSTRUCTION: {}\nnINPUT: {}\nASSISTANT: "
+        self.prompt_af = ""
+        self.IGNORE_INDEX = kwargs.get("IGNORE_INDEX", -100)
+        self.int_pad_value = self.IGNORE_INDEX
+        self.audio_adaptor_downsample_rate = kwargs.get("audio_adaptor_downsample_rate", 5)
+        self.audio_encoder_downsample_rate = kwargs.get("audio_encoder_downsample_rate", 2)
+        self.prompt_template = "USER: {}\n ASSISTANT:"
+        self.answer_template = "{}"
+
+    def get_source_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_source_len(item)
+
+    def get_target_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_target_len(item)
+
+    def __len__(self):
+        return len(self.index_ds)
+
+    def __getitem__(self, index):
+        item = self.index_ds[index]
+        # import pdb;
+        # pdb.set_trace()
+        source = item["source"]
+        data_src = load_audio_text_image_video(source, fs=self.fs)
+        if self.preprocessor_speech:
+            data_src = self.preprocessor_speech(data_src, fs=self.fs)
+        speech, speech_lengths = extract_fbank(
+            data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
+        )  # speech: [b, T, d]
+        speech = speech.squeeze(0)
+
+        audio_pseudo_length = (
+            (speech.shape[0] + 1)
+            // self.audio_adaptor_downsample_rate
+            // self.audio_encoder_downsample_rate
+        )
+        audio_pseudo = torch.full((audio_pseudo_length,), -1)  # placeholder
+
+        target = item["target"]
+        if self.preprocessor_text:
+            target = self.preprocessor_text(target)
+
+        self.prompt_pre = self.prompt_template.format(self.prompt)
+        prompt_ids_pre = self.tokenizer.encode(self.prompt_pre)  # [bos,prompt]
+        prompt_pre_length = len(prompt_ids_pre)
+
+        # input
+        input = self.answer_template.format(target.lower())
+        prompt_input = "{}{}".format(self.prompt_pre, input)
+        prompt_input_ids = self.tokenizer.encode(prompt_input)  # [bos, prompt, input]
+        # audio_length = len(prompt_input_ids) - prompt_pre_length
+        input_ids = prompt_input_ids + [self.tokenizer.pad_token_id]  # [bos, prompt, input, pad]
+        input_ids_length = len(input_ids)
+        input_ids = torch.tensor(input_ids, dtype=torch.int64)  # [bos, prompt, input, pad]
+        input_ids = torch.cat((audio_pseudo, input_ids))  # [audio, bos, prompt, input, pad]
+        # input_ids[:audio_pseudo_length] = -1 # [-1, bos, prompt, input, pad]
+        attention_mask = input_ids.ge(-1)  # [true, true, true, true, true], length mask
+        # input_ids[prompt_pre_length:] = -1  # [bos, prompt,-1,-1]
+        # attention_mask = input_ids.ge(-1)  # [true, true, true, true], length mask
+
+        # label
+        answer = self.answer_template.format(target.lower())
+        prompt_answer = "{}{}".format(self.prompt_pre, answer)
+        prompt_answer_ids = self.tokenizer.encode(prompt_answer)
+        # answer_length = len(prompt_answer_ids) - prompt_pre_length
+        labels_ids = copy.deepcopy(prompt_answer_ids) + [self.tokenizer.eos_token_id]
+        labels_ids = torch.tensor(labels_ids, dtype=torch.int64)  # [bos, prompt, answer, eos]
+        labels_ids = torch.cat((audio_pseudo, labels_ids))  # [audio, bos, prompt, answer, eos]
+        labels_ids[: audio_pseudo_length + prompt_pre_length] = -1  # [-1, -1, -1, answer, eos]
+        # labels_ids[:prompt_pre_length] = -1  # [-1, -1, input, eos]
+        label_mask = labels_ids.ge(0)  # [false, false, false, true, true]
+        labels_ids[~label_mask] = self.IGNORE_INDEX  # [-100, -100, -100, answer, eos]
+
+        # audio_mask for input_ids
+        audio_mask = [1] * audio_pseudo_length + [0] * input_ids_length
+        audio_mask = torch.tensor(audio_mask, dtype=torch.float32)
+
+        ids = self.tokenizer.encode(target)  # token ids is different from labels_ids
+        text = torch.tensor(ids, dtype=torch.int64)
+        text_lengths = torch.tensor([len(ids)], dtype=torch.int32)
+
+        return {
+            "speech": speech,
+            "speech_lengths": speech_lengths,
+            "text": text,
+            "text_lengths": text_lengths,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "labels_ids": labels_ids,
+            "label_mask": label_mask,
+            "audio_mask": audio_mask,
+        }
+
+    def collator(self, samples: list = None):
+        outputs = {}
+        for sample in samples:
+            for key in sample.keys():
+                if key not in outputs:
+                    outputs[key] = []
+                outputs[key].append(sample[key])
+
+        for key, data_list in outputs.items():
+            if isinstance(data_list[0], torch.Tensor):
+                if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:
+
+                    pad_value = self.int_pad_value
+                else:
+                    pad_value = self.float_pad_value
+
+                outputs[key] = torch.nn.utils.rnn.pad_sequence(
+                    data_list, batch_first=True, padding_value=pad_value
+                )
+        return outputs

funasr/datasets/sense_voice_datasets/datasets.py

@@ -0,0 +1,232 @@
+import logging
+
+import torch
+import random
+import traceback
+from funasr.register import tables
+from funasr.utils.load_utils import extract_fbank, load_audio_text_image_video
+
+
+@tables.register("dataset_classes", "SenseVoiceDataset")
+class SenseVoiceDataset(torch.utils.data.Dataset):
+    """
+    SenseVoiceDataset
+    """
+
+    def __init__(
+        self,
+        path,
+        index_ds: str = None,
+        frontend=None,
+        tokenizer=None,
+        int_pad_value: int = -1,
+        float_pad_value: float = 0.0,
+        **kwargs,
+    ):
+        super().__init__()
+        index_ds_class = tables.index_ds_classes.get(index_ds)
+        self.index_ds = index_ds_class(path, **kwargs)
+        preprocessor_speech = kwargs.get("preprocessor_speech", None)
+        if preprocessor_speech:
+            preprocessor_speech_class = tables.preprocessor_classes.get(preprocessor_speech)
+            preprocessor_speech = preprocessor_speech_class(
+                **kwargs.get("preprocessor_speech_conf")
+            )
+        self.preprocessor_speech = preprocessor_speech
+        preprocessor_text = kwargs.get("preprocessor_text", None)
+        if preprocessor_text:
+            preprocessor_text_class = tables.preprocessor_classes.get(preprocessor_text)
+            preprocessor_text = preprocessor_text_class(**kwargs.get("preprocessor_text_conf"))
+        self.preprocessor_text = preprocessor_text
+
+        self.frontend = frontend
+        self.fs = 16000 if frontend is None else frontend.fs
+        self.data_type = "sound"
+        self.tokenizer = tokenizer
+
+        self.int_pad_value = int_pad_value
+        self.float_pad_value = float_pad_value
+        self.sos = kwargs.get("sos", "<|startoftranscript|>")
+        self.eos = kwargs.get("eos", "<|endoftext|>")
+        self.batch_size = kwargs.get("batch_size")
+        self.batch_type = kwargs.get("batch_type")
+        self.prompt_ids_len = 0
+        self.retry = kwargs.get("retry", 5)
+
+        self.permute = False
+        from funasr.frontends.whisper_frontend import WhisperFrontend
+
+        if isinstance(self.frontend, WhisperFrontend):
+            self.permute = True
+
+    def get_source_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_source_len(item)
+
+    def get_target_len(self, index):
+        item = self.index_ds[index]
+        return self.index_ds.get_target_len(item)
+
+    def __len__(self):
+        return len(self.index_ds)
+
+    def __getitem__(self, index):
+        # import pdb;
+        # pdb.set_trace()
+
+        output = None
+        for idx in range(self.retry):
+            if idx == 0:
+                index_cur = index
+            else:
+                index_cur = torch.randint(0, len(self.index_ds), ()).item()
+
+            item = self.index_ds[index_cur]
+
+            source = item["source"]
+            try:
+                data_src = load_audio_text_image_video(source, fs=self.fs)
+            except Exception as e:
+                logging.error(f"Loading wav failed! {str(e)}, {traceback.format_exc()}")
+                continue
+
+            if self.preprocessor_speech:
+                data_src = self.preprocessor_speech(data_src, fs=self.fs)
+            speech, speech_lengths = extract_fbank(
+                data_src, data_type=self.data_type, frontend=self.frontend, is_final=True
+            )  # speech: [b, T, d]
+
+            if speech_lengths > self.batch_size:
+                continue
+            if self.permute:
+                speech = speech.permute(0, 2, 1)
+            target = item["target"]
+            if self.preprocessor_text:
+                target = self.preprocessor_text(target)
+
+            task = item.get("prompt", "<|ASR|>")
+            text_language = item.get("text_language", "<|zh|>")
+
+            if isinstance(self.sos, str):
+                prompt = f"{self.sos}{task}{text_language}"
+                prompt_ids = self.tokenizer.encode(prompt, allowed_special="all")
+            else:
+                prompt = f"{task}{text_language}"
+                prompt_ids = self.tokenizer.encode(prompt, allowed_special="all")
+                prompt_ids = [self.sos] + prompt_ids
+
+            prompt_ids_len = len(prompt_ids) - 1  # [sos, task]
+            self.prompt_ids_len = prompt_ids_len
+
+            target_ids = self.tokenizer.encode(target, allowed_special="all")
+            target_ids_len = len(target_ids) + 1  # [lid, text]
+            if target_ids_len > 200:
+                continue
+
+            if isinstance(self.eos, str):
+                eos = self.tokenizer.encode(self.eos, allowed_special="all")  # [eos]
+            else:
+                eos = [self.eos]
+
+            ids = prompt_ids + target_ids + eos  # [sos, task, lid, text, eos]
+            ids_lengths = len(ids)
+
+            text = torch.tensor(ids, dtype=torch.int64)
+            text_lengths = torch.tensor([ids_lengths], dtype=torch.int32)
+
+            target_mask = (
+                [0] * (prompt_ids_len) + [1] * (target_ids_len) + [1]
+            )  # [sos, task, lid, text, eos]: [0, 0, 1, 1, 1]
+            target_mask_lengths = len(target_mask)
+            target_mask = torch.tensor(target_mask, dtype=torch.float32)
+            target_mask_lengths = torch.tensor([target_mask_lengths], dtype=torch.int32)
+
+            output = {
+                "speech": speech[0, :, :],
+                "speech_lengths": speech_lengths,
+                "text": text,
+                "text_lengths": text_lengths,
+                "target_mask": target_mask,
+                "target_mask_lengths": target_mask_lengths,
+            }
+            break
+
+        return output
+
+    def collator(self, samples: list = None):
+        outputs = {}
+        for sample in samples:
+            if sample is None:
+                continue
+            for key in sample.keys():
+                if key not in outputs:
+                    outputs[key] = []
+                outputs[key].append(sample[key])
+
+        if len(outputs) < 1:
+            logging.error(f"ERROR: data is empty!")
+            outputs = {
+                "speech": torch.rand((10, 128), dtype=torch.float32)[None, :, :],
+                "speech_lengths": torch.tensor(
+                    [
+                        10,
+                    ],
+                    dtype=torch.int32,
+                )[:, None],
+                "text": torch.tensor(
+                    [
+                        58836,
+                    ],
+                    dtype=torch.int32,
+                )[None, :],
+                "text_lengths": torch.tensor(
+                    [
+                        1,
+                    ],
+                    dtype=torch.int32,
+                )[:, None],
+                "target_mask": torch.tensor([[0] * (self.prompt_ids_len) + [1] * (1) + [1]])[
+                    None, :
+                ],
+            }
+            return outputs
+
+        for key, data_list in outputs.items():
+            if isinstance(data_list[0], torch.Tensor):
+                if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:
+
+                    pad_value = self.int_pad_value
+                else:
+                    pad_value = self.float_pad_value
+
+                outputs[key] = torch.nn.utils.rnn.pad_sequence(
+                    data_list, batch_first=True, padding_value=pad_value
+                )
+
+        if self.batch_type != "example":
+            for i in range(10):
+                outputs = self._filter_badcase(outputs, i=i)
+
+        return outputs
+
+    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:
+                beg = 0
+            logging.info(
+                f"Warning, b * t: {b * t} > {self.batch_size}, drop half data {i}th, beg:{beg}"
+            )
+            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()
+            outputs["text"] = outputs["text"][:, :text_lengths_max]
+            target_mask_lengths_max = outputs["target_mask_lengths"].max().item()
+            outputs["target_mask"] = outputs["target_mask"][:, :target_mask_lengths_max]
+
+        return outputs