Codex/create training and inference scripts

examples/industrial_data_pretraining/paraformer/README_LoRA_zh.md

@@ -0,0 +1,112 @@
+# Paraformer LoRA 微调说明
+
+本文档说明如何在 FunASR 中使用 LoRA 微调 Paraformer，并提供训练、推理与 CER 评测的完整示例。
+
+## 1. 前置准备
+
+1. 已准备好符合 FunASR 要求的 `train.jsonl` 与 `val.jsonl`。
+2. 进入仓库根目录（示例路径）：
+
+```bash
+cd /workspace/FunASR
+```
+
+## 2. 训练配置
+
+LoRA 配置文件：
+
+```
+examples/industrial_data_pretraining/paraformer/conf/paraformer_lora.yaml
+```
+
+关键字段说明：
+- `model`: 基座模型名称或本地模型路径。
+- `lora_only`: 是否只训练 LoRA 参数。
+- `lora_bias`: LoRA 偏置训练策略（`none`/`all`/`lora_only`）。
+- `encoder_conf.lora_*` / `decoder_conf.lora_*`: LoRA 参数（rank/alpha/dropout）。
+- `train_data_set_list`/`valid_data_set_list`: 训练/验证集 jsonl。
+
+如需覆盖配置，请通过命令行 `++key=value` 传参。
+
+## 3. 训练脚本
+
+脚本：
+
+```
+examples/industrial_data_pretraining/paraformer/lora_finetune.sh
+```
+
+你只需要确认脚本中的数据路径：
+
+```bash
+data_dir="${workspace}/data/list"
+train_data="${data_dir}/train.jsonl"
+val_data="${data_dir}/val.jsonl"
+```
+
+运行：
+
+```bash
+bash examples/industrial_data_pretraining/paraformer/lora_finetune.sh
+```
+
+训练日志与模型输出将保存在：
+
+```
+examples/industrial_data_pretraining/paraformer/outputs_lora
+```
+
+## 4. 推理脚本
+
+推理脚本会读取 jsonl 输入并生成 `text.hyp` / `text.ref`：
+
+- Python 脚本：`examples/industrial_data_pretraining/paraformer/lora_infer.py`
+- Shell 封装：`examples/industrial_data_pretraining/paraformer/lora_infer.sh`
+
+修改 `lora_infer.sh` 中路径后运行：
+
+```bash
+bash examples/industrial_data_pretraining/paraformer/lora_infer.sh
+```
+
+输出目录默认：
+
+```
+examples/industrial_data_pretraining/paraformer/outputs_lora/infer
+```
+
+## 5. CER 评测
+
+评测脚本：
+
+```
+examples/industrial_data_pretraining/paraformer/lora_cer.sh
+```
+
+运行：
+
+```bash
+bash examples/industrial_data_pretraining/paraformer/lora_cer.sh
+```
+
+结果会输出 CER 统计到：
+
+```
+examples/industrial_data_pretraining/paraformer/outputs_lora/infer/text.cer
+```
+
+## 6. 常见问题
+
+1. **训练不收敛或效果差**
+   - 尝试调整 `lora_rank`、`lora_alpha`、`lora_dropout`。
+   - 调整 `optim_conf.lr` 与 `train_conf.max_epoch`。
+
+2. **推理报错找不到配置**
+   - 确保训练输出目录中存在 `config.yaml`，并在推理脚本中设置正确的 `config_path` 和 `config_name`。
+
+3. **多卡训练**
+   - 设置 `CUDA_VISIBLE_DEVICES`，脚本会自动计算 `gpu_num`。
+
+---
+
+如需进一步定制，可直接在 `paraformer_lora.yaml` 中修改配置或在命令行传参覆盖。

examples/industrial_data_pretraining/paraformer/conf/paraformer_lora.yaml

@@ -0,0 +1,51 @@
+# LoRA finetune config for Paraformer
+# You can override data paths and hyper-parameters by command-line ++key=value.
+
+# model hub name or local model dir
+model: iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch
+model_revision: master
+
+# LoRA settings
+lora_only: true
+lora_bias: none
+
+encoder_conf:
+  lora_list: ["q", "k", "v", "o"]
+  lora_rank: 8
+  lora_alpha: 16
+  lora_dropout: 0.05
+
+decoder_conf:
+  lora_list: ["q", "k", "v", "o"]
+  lora_rank: 8
+  lora_alpha: 16
+  lora_dropout: 0.05
+
+# dataset
+train_data_set_list: data/list/train.jsonl
+valid_data_set_list: data/list/val.jsonl
+
+dataset: AudioDataset
+dataset_conf:
+  index_ds: IndexDSJsonl
+  data_split_num: 1
+  batch_sampler: BatchSampler
+  batch_size: 6000
+  sort_size: 1024
+  batch_type: token
+  num_workers: 4
+
+# training
+train_conf:
+  max_epoch: 30
+  log_interval: 10
+  resume: true
+  validate_interval: 2000
+  save_checkpoint_interval: 2000
+  keep_nbest_models: 10
+  avg_nbest_model: 5
+  use_deepspeed: false
+
+optim: adam
+optim_conf:
+  lr: 0.0001

examples/industrial_data_pretraining/paraformer/lora_cer.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
+#  MIT License  (https://opensource.org/licenses/MIT)
+
+set -euo pipefail
+
+workspace=$(pwd)
+
+infer_dir="${workspace}/examples/industrial_data_pretraining/paraformer/outputs_lora/infer"
+ref_file="${infer_dir}/text.ref"
+hyp_file="${infer_dir}/text.hyp"
+cer_file="${infer_dir}/text.cer"
+
+python -m funasr.metrics.wer \
+  ++ref_file="${ref_file}" \
+  ++hyp_file="${hyp_file}" \
+  ++cer_file="${cer_file}" \
+  ++cn_postprocess=false
+
+# Show final CER summary
+if [ -f "${cer_file}" ]; then
+  tail -n 3 "${cer_file}"
+fi

examples/industrial_data_pretraining/paraformer/lora_finetune.sh

@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
+#  MIT License  (https://opensource.org/licenses/MIT)
+
+set -euo pipefail
+
+workspace=$(pwd)
+
+# which gpu to train or finetune
+export CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-"0"}
+gpu_num=$(echo ${CUDA_VISIBLE_DEVICES} | awk -F "," '{print NF}')
+
+# data dir, which contains train.jsonl/val.jsonl
+# NOTE: update these paths to your dataset jsonl files.
+data_dir="${workspace}/data/list"
+train_data="${data_dir}/train.jsonl"
+val_data="${data_dir}/val.jsonl"
+
+# config
+config_path="${workspace}/examples/industrial_data_pretraining/paraformer/conf"
+config_name="paraformer_lora.yaml"
+
+# exp output dir
+output_dir="${workspace}/examples/industrial_data_pretraining/paraformer/outputs_lora"
+log_file="${output_dir}/log.txt"
+
+mkdir -p "${output_dir}"
+
+DISTRIBUTED_ARGS="
+    --nnodes ${WORLD_SIZE:-1} \
+    --nproc_per_node ${gpu_num} \
+    --node_rank ${RANK:-0} \
+    --master_addr ${MASTER_ADDR:-127.0.0.1} \
+    --master_port ${MASTER_PORT:-26669}
+"
+
+echo "log_file: ${log_file}"
+
+torchrun ${DISTRIBUTED_ARGS} \
+  funasr/bin/train_ds.py \
+  --config-path "${config_path}" \
+  --config-name "${config_name}" \
+  ++train_data_set_list="${train_data}" \
+  ++valid_data_set_list="${val_data}" \
+  ++output_dir="${output_dir}" \
+  &> "${log_file}"

examples/industrial_data_pretraining/paraformer/lora_infer.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+# -*- encoding: utf-8 -*-
+
+import argparse
+import json
+import os
+from typing import List, Tuple
+
+from omegaconf import OmegaConf
+
+from funasr import AutoModel
+
+
+def load_jsonl(jsonl_path: str) -> Tuple[List[str], List[str]]:
+    keys = []
+    targets = []
+    with open(jsonl_path, "r", encoding="utf-8") as f:
+        for line in f:
+            if not line.strip():
+                continue
+            record = json.loads(line)
+            key = record.get("key")
+            if key is None and isinstance(record.get("source"), dict):
+                key = record["source"].get("key")
+            keys.append(key or "")
+            targets.append(record.get("target", ""))
+    return keys, targets
+
+
+def build_model(args: argparse.Namespace):
+    kwargs = {}
+    if args.config_path and args.config_name:
+        cfg_path = os.path.join(args.config_path, args.config_name)
+        cfg = OmegaConf.load(cfg_path)
+        kwargs.update(OmegaConf.to_container(cfg, resolve=True))
+    if args.model:
+        kwargs["model"] = args.model
+    if args.init_param:
+        kwargs["init_param"] = args.init_param
+    kwargs["device"] = args.device
+    if args.batch_size:
+        kwargs["batch_size"] = args.batch_size
+    return AutoModel(**kwargs)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", type=str, default=None, help="model name or model dir")
+    parser.add_argument("--config-path", type=str, default=None, help="config directory")
+    parser.add_argument("--config-name", type=str, default=None, help="config filename")
+    parser.add_argument("--init-param", type=str, default=None, help="model checkpoint path")
+    parser.add_argument("--input-jsonl", type=str, required=True, help="input jsonl with source/target")
+    parser.add_argument("--output-dir", type=str, required=True, help="output directory")
+    parser.add_argument("--device", type=str, default="cuda:0", help="cuda:0 or cpu")
+    parser.add_argument("--batch-size", type=int, default=1, help="batch size for inference")
+    args = parser.parse_args()
+
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    keys, targets = load_jsonl(args.input_jsonl)
+
+    model = build_model(args)
+    results = model.generate(input=args.input_jsonl, batch_size=args.batch_size)
+
+    hyp_path = os.path.join(args.output_dir, "text.hyp")
+    ref_path = os.path.join(args.output_dir, "text.ref")
+
+    with open(hyp_path, "w", encoding="utf-8") as hyp_f, open(
+        ref_path, "w", encoding="utf-8"
+    ) as ref_f:
+        for idx, result in enumerate(results):
+            key = keys[idx] if idx < len(keys) else result.get("key", f"utt_{idx}")
+            hyp = result.get("text", "")
+            ref = targets[idx] if idx < len(targets) else ""
+            hyp_f.write(f"{key} {hyp}\n")
+            ref_f.write(f"{key} {ref}\n")
+
+    print(f"hyp saved to: {hyp_path}")
+    print(f"ref saved to: {ref_path}")
+
+
+if __name__ == "__main__":
+    main()

examples/industrial_data_pretraining/paraformer/lora_infer.sh

@@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
+#  MIT License  (https://opensource.org/licenses/MIT)
+
+set -euo pipefail
+
+workspace=$(pwd)
+
+# model path and config (from training output)
+model_dir="${workspace}/examples/industrial_data_pretraining/paraformer/outputs_lora"
+init_param="${model_dir}/model.pt"
+config_path="${model_dir}"
+config_name="config.yaml"
+
+# input jsonl (must contain source/target)
+input_jsonl="${workspace}/data/list/val.jsonl"
+
+# output directory
+output_dir="${model_dir}/infer"
+
+# device
+device="cuda:0"
+
+python ${workspace}/examples/industrial_data_pretraining/paraformer/lora_infer.py \
+  --model "${model_dir}" \
+  --config-path "${config_path}" \
+  --config-name "${config_name}" \
+  --init-param "${init_param}" \
+  --input-jsonl "${input_jsonl}" \
+  --output-dir "${output_dir}" \
+  --device "${device}" \
+  --batch-size 1

funasr/bin/train.py

@@ -108,6 +108,11 @@ def main(**kwargs):
                 if k.startswith(t + ".") or k == t:
                     logging.info(f"Setting {k}.requires_grad = False")
                     p.requires_grad = False
+    lora_only = kwargs.get("lora_only", False)
+    if lora_only:
+        lora_bias = kwargs.get("lora_bias", "none")
+        logging.info("Enable LoRA-only training with bias=%s", lora_bias)
+        mark_only_lora_as_trainable(model, bias=lora_bias)
     if local_rank == 0:
         logging.info(f"{model_summary(model)}")
 

funasr/bin/train_ds.py

@@ -119,6 +119,11 @@ def main(**kwargs):
                 if k.startswith(t + ".") or k == t:
                     logging.info(f"Setting {k}.requires_grad = False")
                     p.requires_grad = False
+    lora_only = kwargs.get("lora_only", False)
+    if lora_only:
+        lora_bias = kwargs.get("lora_bias", "none")
+        logging.info("Enable LoRA-only training with bias=%s", lora_bias)
+        mark_only_lora_as_trainable(model, bias=lora_bias)
     if local_rank == 0:
         logging.info(f"{model_summary(model)}")