From 723488d97b256a2682af3bf8eb8a8da2c1a6990d Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 23 十一月 2023 16:16:20 +0800
Subject: [PATCH] funasr v2
---
funasr/datasets/data_sampler.py | 2
funasr/datasets/dataloader_fn.py | 49 +++++++++++++++---------
funasr/datasets/dataset_jsonl.py | 24 ++++++++----
3 files changed, 48 insertions(+), 27 deletions(-)
diff --git a/funasr/datasets/data_sampler.py b/funasr/datasets/data_sampler.py
index 6b3407c..c8e7b0d 100644
--- a/funasr/datasets/data_sampler.py
+++ b/funasr/datasets/data_sampler.py
@@ -4,7 +4,7 @@
class BatchSampler(torch.utils.data.BatchSampler):
- def __init__(self, dataset, batch_size_type: str="example", batch_size: int=14, sort_size: int=30, drop_last: bool=False, shuffle: bool=True, **kwargs):
+ def __init__(self, dataset, batch_size_type: str="example", batch_size: int=100, sort_size: int=30, drop_last: bool=False, shuffle: bool=True, **kwargs):
self.drop_last = drop_last
self.pre_idx = -1
diff --git a/funasr/datasets/dataloader_fn.py b/funasr/datasets/dataloader_fn.py
index 8e8e423..3393a33 100644
--- a/funasr/datasets/dataloader_fn.py
+++ b/funasr/datasets/dataloader_fn.py
@@ -1,4 +1,4 @@
-
+import time
import torch
from funasr.datasets.dataset_jsonl import AudioDataset
from funasr.datasets.data_sampler import BatchSampler
@@ -8,7 +8,7 @@
collate_fn = None
# collate_fn = collate_fn,
-jsonl = "/Users/zhifu/funasr_github/test_local/all_task_debug_len.jsonl"
+jsonl = "/Users/zhifu/funasr_github/test_local/aishell2_dev_ios/asr_task_debug_len.jsonl"
frontend = WavFrontend()
token_type = 'char'
@@ -26,7 +26,7 @@
non_linguistic_symbols=non_linguistic_symbols,
g2p_type=g2p_type,
)
-token_list = ""
+token_list = "/Users/zhifu/.cache/modelscope/hub/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/tokens.txt"
unk_symbol = "<unk>"
token_id_converter = TokenIDConverter(
@@ -34,20 +34,33 @@
unk_symbol=unk_symbol,
)
-dataset = AudioDataset(jsonl, frontend=frontend, tokenizer=tokenizer)
+dataset = AudioDataset(jsonl, frontend=frontend, tokenizer=tokenizer, token_id_converter=token_id_converter)
batch_sampler = BatchSampler(dataset)
-dataloader_tr = torch.utils.data.DataLoader(dataset,
- collate_fn=dataset.collator,
- batch_sampler=batch_sampler,
- shuffle=False,
- num_workers=0,
- pin_memory=True)
-print(len(dataset))
-for i in range(3):
- print(i)
- for data in dataloader_tr:
- print(len(data), data)
-# data_iter = iter(dataloader_tr)
-# data = next(data_iter)
-pass
+
+def collator(samples: list = None):
+ return samples
+
+if __name__ == "__main__":
+
+ dataloader_tr = torch.utils.data.DataLoader(dataset,
+ collate_fn=dataset.collator,
+ batch_sampler=batch_sampler,
+ shuffle=False,
+ num_workers=8,
+ pin_memory=True)
+
+ print(len(dataset))
+ for i in range(3):
+ print(i)
+ beg = time.time()
+ for j, data in enumerate(dataloader_tr):
+ end = time.time()
+ time_cost = end - beg
+ beg = end
+ print(j, time_cost)
+ # data_iter = iter(dataloader_tr)
+ # data = next(data_iter)
+ pass
+
+
\ No newline at end of file
diff --git a/funasr/datasets/dataset_jsonl.py b/funasr/datasets/dataset_jsonl.py
index 72d9a99..9e4ee6f 100644
--- a/funasr/datasets/dataset_jsonl.py
+++ b/funasr/datasets/dataset_jsonl.py
@@ -22,7 +22,10 @@
def extract_features(data, date_type: str="sound", frontend=None):
if date_type == "sound":
- feat, feats_lens = frontend(data, len(data))
+ 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)
@@ -74,13 +77,14 @@
class AudioDataset(torch.utils.data.Dataset):
- def __init__(self, path, frontend=None, tokenizer=None):
+ def __init__(self, path, frontend=None, tokenizer=None, token_id_converter=None):
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
@@ -92,13 +96,15 @@
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)
target = item["target"]
- text = self.tokenizer.encode(target)
- text_lengths = len(text)
- text, text_lengths = torch.tensor(text, dtype=torch.int64), torch.tensor([text_lengths], dtype=torch.int32)
+ text = self.tokenizer.text2tokens(target)
+ ids = self.token_id_converter.tokens2ids(text)
+ ids_lengths = len(ids)
+ text, text_lengths = torch.tensor(ids, dtype=torch.int64), torch.tensor([ids_lengths], dtype=torch.int32)
return {"speech": speech,
"speech_lengths": speech_lengths,
"text": text,
@@ -108,17 +114,19 @@
def collator(self, samples: list=None):
+ # return samples
+
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 data_list[0].dtype.kind == "i":
+ if data_list[0].dtype == torch.int64:
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 samples
\ No newline at end of file
+ return outputs
\ No newline at end of file
--
Gitblit v1.9.1