Skip to content

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

batch

游雁a883f2342aae
diff --git a/funasr/datasets/audio_datasets/espnet_samplers.py b/funasr/datasets/audio_datasets/espnet_samplers.py
index cb30a28e2..e155cd7fc 100644
--- a/funasr/datasets/audio_datasets/espnet_samplers.py
+++ b/funasr/datasets/audio_datasets/espnet_samplers.py
@@ -71,7 +71,7 @@ class EspnetStyleBatchSampler(DistributedSampler):
self.max_token_length = kwargs.get("max_token_length", 2048)
self.min_token_length = kwargs.get("min_token_length", 0)
self.length_scale_source = kwargs.get("length_scale_source", 1.0)
- self.start_step = 0
+ self.start_step = start_step
if self.start_step > 0:
logging.info(f"Warning, start_step > 0, dataloader start from step: {self.start_step}")
# super().__init__(dataset, num_replicas=num_replicas, rank=rank,
@@ -146,7 +146,10 @@ class EspnetStyleBatchSampler(DistributedSampler):
start_idx = self.rank * batches_per_rank
end_idx = start_idx + batches_per_rank
rank_batches = buffer_batches[start_idx + self.start_step : end_idx]
-
+ if self.start_step > 0:
+ logging.info(
+ f"Warning, rank: {self.rank}, dataloader start from step: {self.start_step}, batch_num_before: {end_idx-start_idx}, now: {len(rank_batches)}"
+ )
# Return an iterator over the batches for the current rank
return iter(rank_batches)
diff --git a/funasr/datasets/audio_datasets/index_ds.py b/funasr/datasets/audio_datasets/index_ds.py
index 70581e825..385218ae3 100644
--- a/funasr/datasets/audio_datasets/index_ds.py
+++ b/funasr/datasets/audio_datasets/index_ds.py
@@ -35,7 +35,7 @@ class IndexDSJsonlRankFull(torch.utils.data.Dataset):
with open(path, encoding="utf-8") as fin:
file_list_all = fin.readlines()
- num_per_slice = (len(file_list_all) - 1) // data_split_num + 1
+ num_per_slice = (len(file_list_all) - 1) // data_split_num + 1 # 16
file_list = file_list_all[
data_split_i * num_per_slice : (data_split_i + 1) * num_per_slice
]
diff --git a/funasr/datasets/audio_datasets/update_jsonl.py b/funasr/datasets/audio_datasets/update_jsonl.py
index 6fe377cf5..561811e8e 100644
--- a/funasr/datasets/audio_datasets/update_jsonl.py
+++ b/funasr/datasets/audio_datasets/update_jsonl.py
@@ -50,8 +50,8 @@ def update_data(lines, i):
sample_num = len(waveform)
source_len = int(sample_num / 16000 * 1000 / 10)
source_len_old = data["source_len"]
- if source_len_old != source_len:
- print(f"wav: {wav_path}, old: {source_len_old}, new: {source_len}")
+ if (source_len_old - source_len) > 100 or (source_len - source_len_old) > 100:
+ print(f"old: {source_len_old}, new: {source_len}, wav: {wav_path}")
data["source_len"] = source_len
jsonl_line = json.dumps(data, ensure_ascii=False)
lines[i] = jsonl_line