Merge branch 'main' of github.com:alibaba-damo-academy/FunASR add

游雁

2023-02-13 fcc9c89eaba9a4e36c54958aeedeec7ab3756cd7

 funasr/datasets/large_datasets/dataset.py

@@ -1,8 +1,10 @@
import os
import random
import numpy
from functools import partial

import torch
import torchaudio
import torch.distributed as dist
from kaldiio import ReadHelper
from torch.utils.data import IterableDataset
@@ -12,6 +14,7 @@
from funasr.datasets.large_datasets.datapipes.map import MapperIterDataPipe
from funasr.datasets.large_datasets.utils.filter import filter
from funasr.datasets.large_datasets.utils.padding import padding
from funasr.datasets.large_datasets.utils.clipping import clipping
from funasr.datasets.large_datasets.utils.tokenize import tokenize


@@ -97,9 +100,11 @@
                if data_type == "kaldi_ark":
                    ark_reader = ReadHelper('ark:{}'.format(data_file))
                    reader_list.append(ark_reader)
                elif data_type == "text":
                elif data_type == "text" or data_type == "sound":
                    text_reader = open(data_file, "r")
                    reader_list.append(text_reader)
                elif data_type == "none":
                    continue
                else:
                    raise TypeError("Data type {} is not supported".format(data_type))

@@ -109,6 +114,15 @@
                    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()
                        waveform, sampling_rate = torchaudio.load(path)
                        waveform = waveform.numpy()
                        mat = waveform[0]
                        sample_dict[data_name] = mat
                        sample_dict["sampling_rate"] = sampling_rate
                        if data_name == "speech":
                            sample_dict["key"] = key
                    else:
@@ -125,13 +139,18 @@

def len_fn_token(data):
    assert "speech" in data
    return data["speech"].shape[0]
    if "sampling_rate" in data:
        return (data["speech"].shape[0] / data["sampling_rate"]) * 1000.
    else:
        return data["speech"].shape[0]


def Dataset(data_list_file,
            dict,
            seg_dict,
            conf,
            mode="train"):
            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")
@@ -142,9 +161,10 @@
    filter_fn = partial(filter, **filter_conf)
    dataset = FilterIterDataPipe(dataset, fn=filter_fn)

    vocab = {'vocab': dict}
    tokenize_fn = partial(tokenize, **vocab)
    dataset = MapperIterDataPipe(dataset, fn=tokenize_fn)
    if "text" in data_names:
        vocab = {'vocab': dict, 'seg_dict': seg_dict}
        tokenize_fn = partial(tokenize, **vocab)
        dataset = MapperIterDataPipe(dataset, fn=tokenize_fn)

    if shuffle:
        buffer_conf = conf.get('shuffle_conf', {})
@@ -168,8 +188,9 @@
                                             batch_size=batch_size,
                                             len_fn=len_fn,
                                             buffer_size=buffer_size,
                                             sort_size=sort_size)
                                             sort_size=sort_size,
                                             batch_mode=batch_mode)

    dataset = MapperIterDataPipe(dataset, fn=padding)
    dataset = MapperIterDataPipe(dataset, fn=padding if batch_mode == "padding" else clipping)

    return dataset