From 806a03609df033d61f824f1ab8527eb88fe837ad Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期三, 13 十二月 2023 19:43:13 +0800
Subject: [PATCH] funasr2 paraformer biciparaformer contextuaparaformer
---
funasr/datasets/dataset_jsonl.py | 71 ++++++++++++++---------------------
1 files changed, 28 insertions(+), 43 deletions(-)
diff --git a/funasr/datasets/dataset_jsonl.py b/funasr/datasets/dataset_jsonl.py
index 543b60e..21df89e 100644
--- a/funasr/datasets/dataset_jsonl.py
+++ b/funasr/datasets/dataset_jsonl.py
@@ -6,34 +6,9 @@
import librosa
import torchaudio
import time
+import logging
-def load_audio(audio_path: str, fs: int=16000):
- audio = None
- if audio_path.startswith("oss:"):
- pass
- elif audio_path.startswith("odps:"):
- pass
- else:
- if ".ark:" in audio_path:
- audio = kaldiio.load_mat(audio_path)
- else:
- # audio, fs = librosa.load(audio_path, sr=fs)
- audio, fs = torchaudio.load(audio_path)
- audio = audio[0, :]
- return audio
-
-def extract_features(data, date_type: str="sound", frontend=None):
- if date_type == "sound":
-
- if isinstance(data, np.ndarray):
- data = torch.from_numpy(data).to(torch.float32)
- data_len = torch.tensor([data.shape[0]]).to(torch.int32)
- feat, feats_lens = frontend(data[None, :], data_len)
-
- feat = feat[0, :, :]
- else:
- feat, feats_lens = torch.from_numpy(data).to(torch.float32), torch.tensor([data.shape[0]]).to(torch.int32)
- return feat, feats_lens
+from funasr.datasets.fun_datasets.load_audio_extract_fbank import load_audio, extract_fbank
@@ -41,8 +16,7 @@
def __init__(self, path):
super().__init__()
- # data_parallel_size = dist.get_world_size()
- data_parallel_size = 1
+
contents = []
with open(path, encoding='utf-8') as fin:
for line in fin:
@@ -66,33 +40,46 @@
self.contents = []
total_num = len(contents)
- num_per_rank = total_num // data_parallel_size
- # rank = dist.get_rank()
- rank = 0
+ try:
+ rank = dist.get_rank()
+ world_size = dist.get_world_size()
+ except:
+ rank = 0
+ world_size = 1
+ logging.warning("distributed is not initialized, only single shard")
+ num_per_rank = total_num // world_size
+
+ # rank = 0
# import ipdb; ipdb.set_trace()
self.contents = contents[rank * num_per_rank:(rank + 1) * num_per_rank]
-
+
+ logging.info("in rank: {}, num of samplers: {}, total_num of samplers across ranks: {}".format(rank, len(self.contents), len(contents)))
def __len__(self):
return len(self.contents)
def __getitem__(self, index):
return self.contents[index]
+
+ def get_source_len(self, data_dict):
+ return data_dict["source_len"]
+
+ def get_target_len(self, data_dict):
+
+ return data_dict["target_len"] if "target_len" in data_dict else 0
class AudioDataset(torch.utils.data.Dataset):
- def __init__(self, path, frontend=None, tokenizer=None, token_id_converter=None):
-
+ def __init__(self, path, frontend=None, tokenizer=None, int_pad_value: int = -1, float_pad_value: float = 0.0, **kwargs):
super().__init__()
self.indexed_dataset = IndexedDatasetJsonl(path)
self.frontend = frontend.forward
self.fs = 16000 if frontend is None else frontend.fs
self.data_type = "sound"
self.tokenizer = tokenizer
- self.token_id_converter = token_id_converter
- self.int_pad_value = -1
- self.float_pad_value = 0.0
+ self.int_pad_value = int_pad_value
+ self.float_pad_value = float_pad_value
@@ -102,18 +89,16 @@
def __getitem__(self, index):
item = self.indexed_dataset[index]
- # return item
source = item["source"]
data_src = load_audio(source, fs=self.fs)
- speech, speech_lengths = extract_features(data_src, self.data_type, self.frontend)
+ speech, speech_lengths = extract_fbank(data_src, self.data_type, self.frontend) # speech: [b, T, d]
target = item["target"]
- text = self.tokenizer.text2tokens(target)
- ids = self.token_id_converter.tokens2ids(text)
+ ids = self.tokenizer.encode(target)
ids_lengths = len(ids)
text, text_lengths = torch.tensor(ids, dtype=torch.int64), torch.tensor([ids_lengths], dtype=torch.int32)
- return {"speech": speech,
+ return {"speech": speech[0, :, :],
"speech_lengths": speech_lengths,
"text": text,
"text_lengths": text_lengths,
--
Gitblit v1.9.1