feat✨: 更新训练脚本以支持MOELORA，调整梯度累积步数，优化配置文件

.vscode/settings.json

@@ -13,7 +13,7 @@
     // ],
     "python.analysis.languageServerMode": "default",
     "python.analysis.typeCheckingMode": "basic",
-    "python.analysis.userFileIndexingLimit": -1,
+    "python.analysis.userFileIndexingLimit": 10000,
     "python.analysis.usePullDiagnostics": false,
     "python.analysis.importFormat": "relative"
 }

src/configs/accelerate_configs/deepspeed_zero1.yaml

@@ -2,7 +2,7 @@ compute_environment: LOCAL_MACHINE
 debug: false
 deepspeed_config:
   deepspeed_multinode_launcher: standard
-  gradient_accumulation_steps: 4
+  gradient_accumulation_steps: 2
   zero3_init_flag: false
   zero_stage: 1
 distributed_type: DEEPSPEED

src/scripts/train_omni.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+
+accelerate launch --config_file configs/accelerate_configs/deepspeed_zero1.yaml train.py \
+    --dataset_name textvqa \
+    --use_peft \
+    --peft_type MOELORA \
+    --model_name_or_path Qwen/Qwen2.5-Omni-3B \
+    --lora_target_modules .*model\.layers.*proj \
+    --lora_r 8 \
+    --lora_alpha 32 \
+    --per_device_train_batch_size 3 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 2 \
+    --num_train_epochs 1 \
+    --output_dir checkpoint/qwen2_alllinear/ \
+    --learning_rate 2e-4 \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type cosine \
+    --bf16 \
+    --torch_dtype bfloat16 \
+    --logging_steps 10 \
+    --gradient_checkpointing \
+    --weight_decay 0.1 \
+    --resume_from_checkpoint /root/autodl-tmp/zhouyunyao/projects/CL-LMM/src/checkpoint/qwen2_alllinear/checkpoint-1000

src/todo.md

@@ -27,4 +27,16 @@
 
 [2025.05.16]
 
-- [ ] 处理不同的持续学习框架，使得整体框架能够兼容
+- [ ] 处理不同的持续学习框架，使得整体框架能够兼容
+
+[2025.05.28]
+
+- [x] MoeLora
+- [ ] Coin Benchmark
+- [x] 确定保存什么，便于后期测试
+- [ ] Olora
+- [ ] Hide-Llava
+
+[2025.05.30]
+
+- [ ] 评价指标

src/train.py

@@ -65,6 +65,13 @@ if __name__ == "__main__":
 
         model.add_adapter(peft_config)
 
+    elif model_args.peft_type == "MOELORA":
+        from peft.tuners import MOELoraConfig
+
+        peft_config = MOELoraConfig(target_modules=model_args.lora_target_modules)
+
+        model.add_adapter(peft_config)
+
     elif model_args.peft_type == "LORA":
         from peft.tuners.lora import LoraConfig
 

src/train.sh

@@ -3,14 +3,14 @@
 accelerate launch --config_file configs/accelerate_configs/deepspeed_zero1.yaml train.py \
     --dataset_name textvqa \
     --use_peft \
-    --peft_type LORA \
+    --peft_type MOELORA \
     --model_name_or_path Qwen/Qwen2.5-Omni-3B \
     --lora_target_modules .\*proj.\*\|.\*fc.\*\|.\*mlp\.0\|.\*mlp\.2 \
     --lora_r 8 \
     --lora_alpha 32 \
     --per_device_train_batch_size 3 \
-    --per_device_eval_batch_size 1 \
     --gradient_accumulation_steps 4 \
+    --per_device_eval_batch_size 1 \
     --num_train_epochs 1 \
     --output_dir checkpoint/qwen2_alllinear/ \
     --learning_rate 5e-5 \