Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
Dev gzf new (#1553)
zhifu gao2297d7515afb
diff --git a/docs/tutorial/README_zh.md b/docs/tutorial/README_zh.mdindex 78acb5825..4d5e31079 100644--- a/docs/tutorial/README_zh.md+++ b/docs/tutorial/README_zh.md@@ -268,7 +268,7 @@ torchrun --nnodes 1 --nproc_per_node ${gpu_num} \export CUDA_VISIBLE_DEVICES="0,1"gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')-torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr=192.168.1.1 --master_port=12345 \+torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr 192.168.1.1 --master_port 12345 \../../../funasr/bin/train.py ${train_args}```在从节点上(假设IP为192.168.1.2),你需要确保MASTER_ADDR和MASTER_PORT环境变量与主节点设置的一致,并运行同样的命令:@@ -276,7 +276,7 @@ torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr=192.export CUDA_VISIBLE_DEVICES="0,1"gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')-torchrun --nnodes 2 --node_rank 1 --nproc_per_node ${gpu_num} --master_addr=192.168.1.1 --master_port=12345 \+torchrun --nnodes 2 --node_rank 1 --nproc_per_node ${gpu_num} --master_addr 192.168.1.1 --master_port 12345 \../../../funasr/bin/train.py ${train_args}```diff --git a/examples/README_zh.md b/examples/README_zh.mdindex 78acb5825..4d5e31079 100644--- a/examples/README_zh.md+++ b/examples/README_zh.md@@ -268,7 +268,7 @@ torchrun --nnodes 1 --nproc_per_node ${gpu_num} \export CUDA_VISIBLE_DEVICES="0,1"gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')-torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr=192.168.1.1 --master_port=12345 \+torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr 192.168.1.1 --master_port 12345 \../../../funasr/bin/train.py ${train_args}```在从节点上(假设IP为192.168.1.2),你需要确保MASTER_ADDR和MASTER_PORT环境变量与主节点设置的一致,并运行同样的命令:@@ -276,7 +276,7 @@ torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr=192.export CUDA_VISIBLE_DEVICES="0,1"gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')-torchrun --nnodes 2 --node_rank 1 --nproc_per_node ${gpu_num} --master_addr=192.168.1.1 --master_port=12345 \+torchrun --nnodes 2 --node_rank 1 --nproc_per_node ${gpu_num} --master_addr 192.168.1.1 --master_port 12345 \../../../funasr/bin/train.py ${train_args}```diff --git a/examples/industrial_data_pretraining/paraformer-zh-spk/README_zh.md b/examples/industrial_data_pretraining/paraformer-zh-spk/README_zh.mdindex 78acb5825..4d5e31079 100644--- a/examples/industrial_data_pretraining/paraformer-zh-spk/README_zh.md+++ b/examples/industrial_data_pretraining/paraformer-zh-spk/README_zh.md@@ -268,7 +268,7 @@ torchrun --nnodes 1 --nproc_per_node ${gpu_num} \export CUDA_VISIBLE_DEVICES="0,1"gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')-torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr=192.168.1.1 --master_port=12345 \+torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr 192.168.1.1 --master_port 12345 \../../../funasr/bin/train.py ${train_args}```在从节点上(假设IP为192.168.1.2),你需要确保MASTER_ADDR和MASTER_PORT环境变量与主节点设置的一致,并运行同样的命令:@@ -276,7 +276,7 @@ torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr=192.export CUDA_VISIBLE_DEVICES="0,1"gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')-torchrun --nnodes 2 --node_rank 1 --nproc_per_node ${gpu_num} --master_addr=192.168.1.1 --master_port=12345 \+torchrun --nnodes 2 --node_rank 1 --nproc_per_node ${gpu_num} --master_addr 192.168.1.1 --master_port 12345 \../../../funasr/bin/train.py ${train_args}```diff --git a/examples/industrial_data_pretraining/paraformer/README_zh.md b/examples/industrial_data_pretraining/paraformer/README_zh.mdindex 78acb5825..4d5e31079 100644--- a/examples/industrial_data_pretraining/paraformer/README_zh.md+++ b/examples/industrial_data_pretraining/paraformer/README_zh.md@@ -268,7 +268,7 @@ torchrun --nnodes 1 --nproc_per_node ${gpu_num} \export CUDA_VISIBLE_DEVICES="0,1"gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')-torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr=192.168.1.1 --master_port=12345 \+torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr 192.168.1.1 --master_port 12345 \../../../funasr/bin/train.py ${train_args}```在从节点上(假设IP为192.168.1.2),你需要确保MASTER_ADDR和MASTER_PORT环境变量与主节点设置的一致,并运行同样的命令:@@ -276,7 +276,7 @@ torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr=192.export CUDA_VISIBLE_DEVICES="0,1"gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')-torchrun --nnodes 2 --node_rank 1 --nproc_per_node ${gpu_num} --master_addr=192.168.1.1 --master_port=12345 \+torchrun --nnodes 2 --node_rank 1 --nproc_per_node ${gpu_num} --master_addr 192.168.1.1 --master_port 12345 \../../../funasr/bin/train.py ${train_args}```diff --git a/examples/industrial_data_pretraining/paraformer_streaming/README_zh.md b/examples/industrial_data_pretraining/paraformer_streaming/README_zh.mdindex 78acb5825..4d5e31079 100644--- a/examples/industrial_data_pretraining/paraformer_streaming/README_zh.md+++ b/examples/industrial_data_pretraining/paraformer_streaming/README_zh.md@@ -268,7 +268,7 @@ torchrun --nnodes 1 --nproc_per_node ${gpu_num} \export CUDA_VISIBLE_DEVICES="0,1"gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')-torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr=192.168.1.1 --master_port=12345 \+torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr 192.168.1.1 --master_port 12345 \../../../funasr/bin/train.py ${train_args}```在从节点上(假设IP为192.168.1.2),你需要确保MASTER_ADDR和MASTER_PORT环境变量与主节点设置的一致,并运行同样的命令:@@ -276,7 +276,7 @@ torchrun --nnodes 2 --node_rank 0 --nproc_per_node ${gpu_num} --master_addr=192.export CUDA_VISIBLE_DEVICES="0,1"gpu_num=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')-torchrun --nnodes 2 --node_rank 1 --nproc_per_node ${gpu_num} --master_addr=192.168.1.1 --master_port=12345 \+torchrun --nnodes 2 --node_rank 1 --nproc_per_node ${gpu_num} --master_addr 192.168.1.1 --master_port 12345 \../../../funasr/bin/train.py ${train_args}```diff --git a/funasr/datasets/audio_datasets/samplers.py b/funasr/datasets/audio_datasets/samplers.pyindex a56a980f7..01f5e6a3e 100644--- a/funasr/datasets/audio_datasets/samplers.py+++ b/funasr/datasets/audio_datasets/samplers.py@@ -23,11 +23,11 @@ def CustomDistributedBatchSampler_fn(dataset, **kwargs):batch_sampler = CustomDistributedBatchSampler(dataset, **kwargs)else:- # if kwargs.get("sort_size", -1) > 0:- # batch_sampler = CustomDistributedBufferDynamicBatchSampler(dataset, **kwargs)- # else:- # batch_sampler = CustomDistributedDynamicBatchSampler(dataset, **kwargs)- batch_sampler = CustomDistributedDynamicBatchSampler(dataset, **kwargs)+ if kwargs.get("sort_size", -1) > 0:+ batch_sampler = CustomDistributedBufferDynamicBatchSampler(dataset, **kwargs)+ else:+ batch_sampler = CustomDistributedDynamicBatchSampler(dataset, **kwargs)+ # batch_sampler = CustomDistributedDynamicBatchSampler(dataset, **kwargs)dataloader_args["batch_sampler"] = batch_samplerdataloader_args["num_workers"] = kwargs.get("num_workers", 4)@@ -244,6 +244,8 @@ class CustomDistributedDynamicBatchSampler(DistributedSampler):self.total_size = len(self.dataset)# self.num_samples = int(math.ceil(self.total_size / self.num_replicas))self.epoch = 0+ self.max_token_length = kwargs.get("max_token_length", 2048)+ self.length_scale_source = kwargs.get("length_scale_source", 1.0)def __iter__(self):if self.shuffle:@@ -262,6 +264,8 @@ class CustomDistributedDynamicBatchSampler(DistributedSampler):for idx in indices:sample_length = self.dataset.get_source_len(idx)+ if sample_length > self.max_token_length:+ continuepotential_batch_length = (max_len_in_batch if sample_length < max_len_in_batch else sample_length) * (len(batch) + 1)@@ -269,12 +273,12 @@ class CustomDistributedDynamicBatchSampler(DistributedSampler):batch.append(idx)if sample_length > max_len_in_batch:max_len_in_batch = sample_length- current_batch_length = max_len_in_batch * len(batch)+ # current_batch_length = max_len_in_batch * len(batch)else:batches.append(batch)batch = [idx]max_len_in_batch = sample_length- current_batch_length = max_len_in_batch+ # current_batch_length = max_len_in_batch# Add the last batch if it's not empty and we're not dropping itif batch and (not self.drop_last or len(batch) * max_len_in_batch == self.batch_size):@@ -293,6 +297,7 @@ class CustomDistributedDynamicBatchSampler(DistributedSampler):class CustomDistributedBufferDynamicBatchSampler(DistributedSampler):def __init__(self, dataset,batch_size,+ batch_type="token",num_replicas=None,rank=None,shuffle=True,@@ -312,6 +317,7 @@ class CustomDistributedBufferDynamicBatchSampler(DistributedSampler):self.num_replicas = num_replicasself.dataset = datasetself.batch_size = batch_size+ self.batch_type = batch_typeself.is_training = is_trainingself.shuffle = shuffle and is_trainingself.drop_last = drop_last@@ -319,42 +325,54 @@ class CustomDistributedBufferDynamicBatchSampler(DistributedSampler):self.total_size = len(self.dataset)# self.num_samples = int(math.ceil(self.total_size / self.num_replicas))self.epoch = 0- self.sort_size = sort_size-+ self.sort_size = sort_size * num_replicas+ self.max_token_length = kwargs.get("max_token_length", 2048)+ self.length_scale_source = kwargs.get("length_scale_source", 1.0)+def __iter__(self):if self.shuffle:g = torch.Generator()g.manual_seed(self.epoch)- indices = torch.randperm(self.total_size, generator=g).tolist()+ indices = torch.randperm(len(self.dataset), generator=g).tolist()else:- indices = list(range(self.total_size))-- # Distribute indices among replicas- indices = indices[self.rank:self.total_size:self.num_replicas]-- # Sort indices into buffers- sorted_buffers = [sorted(indices[i:i + self.sort_size], key=lambda idx: self.dataset.get_source_len(idx)) for i in range(0, len(indices), self.sort_size)]+ indices = list(range(len(self.dataset)))- batches = []- for buffer in sorted_buffers:+ # Create sorted buffers and form batches+ buffer_batches = []+ for i in range(0, len(indices), self.sort_size):+ buffer = sorted(indices[i:i + self.sort_size], key=lambda idx: self.dataset.get_source_len(idx))batch = []max_len_in_batch = 0for idx in buffer:- sample_length = self.dataset.get_source_len(idx)+ original_sample_length = self.dataset.get_source_len(idx)+ if original_sample_length > self.max_sample_length:+ continue+ sample_length = 1 if self.batch_type == "example" else original_sample_lengthpotential_batch_length = max(max_len_in_batch, sample_length) * (len(batch) + 1)if potential_batch_length <= self.batch_size:batch.append(idx)max_len_in_batch = max(max_len_in_batch, sample_length)else:- batches.append(batch)+ buffer_batches.append(batch)batch = [idx]max_len_in_batch = sample_length-- # Add the last batch if it's not empty and we're not dropping it- if batch and (not self.drop_last or len(batch) * max_len_in_batch == self.batch_size):- batches.append(batch)+ if batch:+ buffer_batches.append(batch)- return iter(batches)+ # Ensure each rank gets the same number of batches, duplicate data if needed+ batches_per_rank = math.ceil(len(buffer_batches) / self.num_replicas)+ total_batches_needed = batches_per_rank * self.num_replicas+ buffer_batches.extend(buffer_batches[:total_batches_needed - len(buffer_batches)])++ # Evenly distribute batches from buffer_batches to each rank+ rank_batches = [[] for _ in range(self.num_replicas)]+ for i, batch in enumerate(buffer_batches):+ rank_batches[i % self.num_replicas].append(batch)++ # Assign all batches for the current rank directly+ final_batches = rank_batches[self.rank]++ return iter(final_batches)def __len__(self):diff --git a/funasr/train_utils/trainer.py b/funasr/train_utils/trainer.pyindex 55dce9996..35a266fbf 100644--- a/funasr/train_utils/trainer.py+++ b/funasr/train_utils/trainer.py@@ -161,17 +161,17 @@ class Trainer:self.best_step_or_epoch = ckpt_namebest_ckpt = Path(os.path.join(self.output_dir, f'model.pt.best'))torch.save(state, best_ckpt)- logging.info(f"Update best acc: {self.val_acc_step_or_eoch[self.best_step_or_epoch]}, {best_ckpt}")+ logging.info(f"Update best acc: {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}, {best_ckpt}")else:- logging.info(f"No improvement in acc: {self.val_acc_step_or_eoch[ckpt_name]} < {self.val_acc_step_or_eoch[self.best_step_or_epoch]}")+ logging.info(f"No improvement in acc: {self.val_acc_step_or_eoch[ckpt_name]:.4f} < {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}")elif self.avg_keep_nbest_models_type == "loss":if self.val_loss_step_or_eoch[ckpt_name] <= self.val_loss_step_or_eoch[self.best_step_or_epoch]:self.best_step_or_epoch = ckpt_namebest_ckpt = Path(os.path.join(self.output_dir, f'model.pt.best'))torch.save(state, best_ckpt)- logging.info(f"Update best loss: {self.val_loss_step_or_eoch[self.best_step_or_epoch]}, {best_ckpt}")+ logging.info(f"Update best loss: {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}, {best_ckpt}")else:- logging.info(f"No improvement in loss: {self.val_loss_step_or_eoch[ckpt_name]} > {self.val_loss_step_or_eoch[self.best_step_or_epoch]}")+ logging.info(f"No improvement in loss: {self.val_loss_step_or_eoch[ckpt_name]:.4f} > {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}")else:print("Undo")self.saved_ckpts[ckpt_name] = getattr(self, f"val_{self.avg_keep_nbest_models_type}_step_or_eoch")[ckpt_name]