From 7984a37f8ce2d0db89913b22bad521e6a64aed27 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E4=BB=81=E8=BF=B7?= Date: Wed, 1 Mar 2023 17:03:45 +0800 Subject: [PATCH] update large dataset for sampling rate --- funasr/datasets/large_datasets/build_dataloader.py | 5 +++-- funasr/datasets/large_datasets/dataset.py | 11 +++++++++-- funasr/tasks/abs_task.py | 4 +++- 3 files changed, 15 insertions(+), 5 deletions(-) diff --git a/funasr/datasets/large_datasets/build_dataloader.py b/funasr/datasets/large_datasets/build_dataloader.py index 093ad608e..0ad1889ca 100644 --- a/funasr/datasets/large_datasets/build_dataloader.py +++ b/funasr/datasets/large_datasets/build_dataloader.py @@ -34,7 +34,7 @@ def load_seg_dict(seg_dict_file): return seg_dict class ArkDataLoader(AbsIterFactory): - def __init__(self, data_list, dict_file, dataset_conf, seg_dict_file=None, punc_dict_file=None, mode="train"): + def __init__(self, data_list, dict_file, dataset_conf, frontend_conf=None, seg_dict_file=None, punc_dict_file=None, mode="train"): symbol_table = read_symbol_table(dict_file) if dict_file is not None else None if seg_dict_file is not None: seg_dict = load_seg_dict(seg_dict_file) @@ -45,10 +45,11 @@ class ArkDataLoader(AbsIterFactory): else: punc_dict = None self.dataset_conf = dataset_conf + self.frontend_conf = frontend_conf logging.info("dataloader config: {}".format(self.dataset_conf)) batch_mode = self.dataset_conf.get("batch_mode", "padding") self.dataset = Dataset(data_list, symbol_table, seg_dict, punc_dict, - self.dataset_conf, mode=mode, batch_mode=batch_mode) + self.dataset_conf, self.frontend_conf, mode=mode, batch_mode=batch_mode) def build_iter(self, epoch, shuffle=True): self.dataset.set_epoch(epoch) diff --git a/funasr/datasets/large_datasets/dataset.py b/funasr/datasets/large_datasets/dataset.py index 61231d24c..1942371e5 100644 --- a/funasr/datasets/large_datasets/dataset.py +++ b/funasr/datasets/large_datasets/dataset.py @@ -28,10 +28,11 @@ def read_lists(list_file): class AudioDataset(IterableDataset): - def __init__(self, scp_lists, data_names, data_types, shuffle=True, mode="train"): + def __init__(self, scp_lists, data_names, data_types, frontend_conf=None, shuffle=True, 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 @@ -119,6 +120,11 @@ class AudioDataset(IterableDataset): elif data_type == "sound": key, path = item.strip().split() waveform, sampling_rate = torchaudio.load(path) + 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] sample_dict[data_name] = mat @@ -153,13 +159,14 @@ def Dataset(data_list_file, seg_dict, punc_dict, conf, + frontend_conf, 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") - dataset = AudioDataset(scp_lists, data_names, data_types, shuffle=shuffle, mode=mode) + dataset = AudioDataset(scp_lists, data_names, data_types, frontend_conf=frontend_conf, shuffle=shuffle, mode=mode) filter_conf = conf.get('filter_conf', {}) filter_fn = partial(filter, **filter_conf) diff --git a/funasr/tasks/abs_task.py b/funasr/tasks/abs_task.py index d2a00b2fc..a643acb78 100644 --- a/funasr/tasks/abs_task.py +++ b/funasr/tasks/abs_task.py @@ -1348,11 +1348,13 @@ class AbsTask(ABC): if args.dataset_type == "large": from funasr.datasets.large_datasets.build_dataloader import ArkDataLoader train_iter_factory = ArkDataLoader(args.train_data_file, args.token_list, args.dataset_conf, + frontend_conf=args.frontend_conf if hasattr(args, "frontend_conf") else None, seg_dict_file=args.seg_dict_file if hasattr(args, "seg_dict_file") else None, punc_dict_file=args.punc_list if hasattr(args, "punc_list") else None, mode="train") - valid_iter_factory = ArkDataLoader(args.valid_data_file, args.token_list, args.dataset_conf, + valid_iter_factory = ArkDataLoader(args.valid_data_file, args.token_list, args.dataset_conf, + frontend_conf=args.frontend_conf if hasattr(args, "frontend_conf") else None, seg_dict_file=args.seg_dict_file if hasattr(args, "seg_dict_file") else None, punc_dict_file=args.punc_list if hasattr(args, "punc_list") else None,