Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

auto frontend

游雁c350b6becb88
diff --git a/funasr/datasets/audio_datasets/samplers.py b/funasr/datasets/audio_datasets/samplers.py
index 27fbdac25..2e271af6b 100644
--- a/funasr/datasets/audio_datasets/samplers.py
+++ b/funasr/datasets/audio_datasets/samplers.py
@@ -365,6 +365,7 @@ class CustomDistributedBufferDynamicBatchSampler(DistributedSampler):
self.max_token_length = kwargs.get("max_token_length", 2048)
self.length_scale_source = kwargs.get("length_scale_source", 1.0)
self.start_step = kwargs.get("start_step", 2048)
+ self.batch_size_sample_max = kwargs.get("batch_size_sample_max", 200)
super().__init__(
dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle, drop_last=drop_last
@@ -388,19 +389,25 @@ class CustomDistributedBufferDynamicBatchSampler(DistributedSampler):
)
batch = []
max_len_in_batch = 0
+ count = 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:
+ if (
+ potential_batch_length <= self.batch_size
+ and count <= self.batch_size_sample_max
+ ):
batch.append(idx)
max_len_in_batch = max(max_len_in_batch, sample_length)
+ count += 1
else:
buffer_batches.append(batch)
batch = [idx]
max_len_in_batch = sample_length
+ count = 0
if batch:
buffer_batches.append(batch)
diff --git a/funasr/datasets/openai_datasets/datasets.py b/funasr/datasets/openai_datasets/datasets.py
index c51ace997..46735ba2c 100644
--- a/funasr/datasets/openai_datasets/datasets.py
+++ b/funasr/datasets/openai_datasets/datasets.py
@@ -154,13 +154,13 @@ class OpenAIDataset(torch.utils.data.Dataset):
fbank_mask += fbank_mask_i
fbank_beg.append(fbank_beg_i)
- # if len(input_ids) > self.max_token_length:
- # badcase_flag = True
+ if len(input_ids) > self.max_token_length:
+ badcase_flag = True
if badcase_flag:
continue
- input_ids = torch.tensor(input_ids, dtype=torch.int64)[: self.max_token_length]
+ input_ids = torch.tensor(input_ids, dtype=torch.int64) # [: self.max_token_length]
attention_mask = torch.tensor([len(input_ids)], dtype=torch.int32)
- labels = torch.tensor(labels, dtype=torch.int64)[: self.max_token_length]
+ labels = torch.tensor(labels, dtype=torch.int64) # [: self.max_token_length]
fbank = speech[0, :, :]
fbank_lens = speech_lengths