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9a8086bdf5c1
diff --git a/funasr/bin/train.py b/funasr/bin/train.py
index 97516eb8b..d20915c57 100644
--- a/funasr/bin/train.py
+++ b/funasr/bin/train.py
@@ -205,7 +205,6 @@ def main(**kwargs):
dataloader_tr, dataloader_val = dataloader.build_iter(
epoch, data_split_i=data_split_i, start_step=trainer.start_step
)
- trainer.start_step = 0
trainer.train_epoch(
model=model,
@@ -218,7 +217,9 @@ def main(**kwargs):
writer=writer,
data_split_i=data_split_i,
data_split_num=dataloader.data_split_num,
+ start_step=trainer.start_step,
)
+ trainer.start_step = 0
torch.cuda.empty_cache()
diff --git a/funasr/datasets/audio_datasets/update_jsonl.py b/funasr/datasets/audio_datasets/update_jsonl.py
index ad47e1285..05870fe1f 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) > 100 or (source_len - source_len_old) > 100:
- logging.info(f"old: {source_len_old}, new: {source_len}, wav: {wav_path}")
+ # if (source_len_old - source_len) > 100 or (source_len - source_len_old) > 100:
+ # logging.info(f"old: {source_len_old}, new: {source_len}, wav: {wav_path}")
data["source_len"] = source_len
data["source"] = wav_path
jsonl_line = json.dumps(data, ensure_ascii=False)
diff --git a/funasr/train_utils/trainer.py b/funasr/train_utils/trainer.py
index e86420cdf..a28ca5183 100644
--- a/funasr/train_utils/trainer.py
+++ b/funasr/train_utils/trainer.py
@@ -456,7 +456,7 @@ class Trainer:
batch_num_epoch = len(dataloader_train)
self.log(
epoch,
- batch_idx,
+ batch_idx + kwargs.get("start_step", 0),
step_in_epoch=self.step_in_epoch,
batch_num_epoch=batch_num_epoch,
lr=lr,