{"id":26464,"date":"2025-02-23T12:46:56","date_gmt":"2025-02-23T04:46:56","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=26464"},"modified":"2025-02-23T12:46:56","modified_gmt":"2025-02-23T04:46:56","slug":"vlm-r1","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/vlm-r1\/","title":{"rendered":"VLM-R1\uff1a\u901a\u8fc7\u81ea\u7136\u8bed\u8a00\u5b9a\u4f4d\u56fe\u50cf\u76ee\u6807\u7684\u89c6\u89c9\u8bed\u8a00\u6a21\u578b"},"content":{"rendered":"

VLM-R1 \u662f\u7531 Om AI Lab \u5f00\u53d1\u7684\u4e00\u4e2a\u5f00\u6e90\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u9879\u76ee\uff0c\u6258\u7ba1\u5728 GitHub \u4e0a\u3002\u8be5\u9879\u76ee\u57fa\u4e8e DeepSeek<\/a> \u7684 R1 \u65b9\u6cd5\uff0c\u7ed3\u5408 Qwen2.5-VL<\/a> \u6a21\u578b\uff0c\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08R1\uff09\u548c\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u6280\u672f\uff0c\u663e\u8457\u63d0\u5347\u4e86\u6a21\u578b\u5728\u89c6\u89c9\u7406\u89e3\u4efb\u52a1\u4e2d\u7684\u7a33\u5b9a\u6027\u4e0e\u6cdb\u5316\u80fd\u529b\u3002VLM-R1 \u7279\u522b\u64c5\u957f\u5904\u7406\u6307\u4ee3\u8868\u8fbe\u7406\u89e3\uff08REC\uff09\uff0c\u4f8b\u5982\u56de\u7b54\u201c\u56fe\u4e2d\u7ea2\u8272\u7684\u676f\u5b50\u5728\u54ea\u91cc\u201d\u8fd9\u7c7b\u95ee\u9898\uff0c\u5e76\u80fd\u5728\u56fe\u50cf\u4e2d\u7cbe\u786e\u5b9a\u4f4d\u76ee\u6807\u3002\u9879\u76ee\u63d0\u4f9b\u8be6\u7ec6\u7684\u5b89\u88c5\u811a\u672c\u3001\u6570\u636e\u96c6\u652f\u6301\u4ee5\u53ca\u8bad\u7ec3\u4ee3\u7801\uff0c\u9002\u5408\u5f00\u53d1\u8005\u3001\u7814\u7a76\u4eba\u5458\u7528\u4e8e\u89c6\u89c9\u8bed\u8a00\u4efb\u52a1\u7684\u63a2\u7d22\u4e0e\u5f00\u53d1\u3002\u622a\u81f3 2025 \u5e74 2 \u6708\uff0c\u8be5\u9879\u76ee\u5728 GitHub \u4e0a\u5df2\u83b7\u5f97\u8fd1 2000 \u661f\u6807\uff0c\u663e\u793a\u51fa\u5176\u5728\u591a\u6a21\u6001 AI \u9886\u57df\u7684\u5e7f\u6cdb\u5173\u6ce8\u3002<\/p>\n

\"VLM-R1\uff1a\u64c5\u957f\u89c6\u89c9\u7406\u89e3\u4efb\u52a1\u7684\u5f3a\u5316\u5b66\u4e60\u89c6\u89c9\u8bed\u8a00\u6a21\u578b-1\"

\u6f14\u793a\u5730\u5740\uff1ahttps:\/\/huggingface.co\/spaces\/omlab\/VLM-R1-Referral-Expression<\/p><\/div>\n

 <\/p>\n

\u529f\u80fd\u5217\u8868<\/h2>\n
    \n
  • \u6307\u4ee3\u8868\u8fbe\u7406\u89e3\uff08REC\uff09<\/strong>\uff1a\u80fd\u591f\u89e3\u6790\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\uff0c\u5b9a\u4f4d\u56fe\u50cf\u4e2d\u7684\u7279\u5b9a\u76ee\u6807\u3002<\/li>\n
  • \u56fe\u50cf\u4e0e\u6587\u672c\u8054\u5408\u5904\u7406<\/strong>\uff1a\u652f\u6301\u540c\u65f6\u8f93\u5165\u56fe\u50cf\u548c\u6587\u5b57\uff0c\u751f\u6210\u51c6\u786e\u7684\u5206\u6790\u7ed3\u679c\u3002<\/li>\n
  • \u5f3a\u5316\u5b66\u4e60\u4f18\u5316<\/strong>\uff1a\u901a\u8fc7 R1 \u65b9\u6cd5\u8bad\u7ec3\uff0c\u63d0\u5347\u6a21\u578b\u5728\u590d\u6742\u89c6\u89c9\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002<\/li>\n
  • \u5f00\u6e90\u8bad\u7ec3\u4ee3\u7801<\/strong>\uff1a\u63d0\u4f9b\u5b8c\u6574\u7684\u8bad\u7ec3\u811a\u672c\u548c\u914d\u7f6e\u6587\u4ef6\uff0c\u65b9\u4fbf\u7528\u6237\u81ea\u5b9a\u4e49\u6a21\u578b\u3002<\/li>\n
  • \u6570\u636e\u96c6\u652f\u6301<\/strong>\uff1a\u5185\u7f6e COCO \u548c RefCOCO \u6570\u636e\u96c6\u4e0b\u8f7d\u4e0e\u5904\u7406\u529f\u80fd\uff0c\u7b80\u5316\u5f00\u53d1\u6d41\u7a0b\u3002<\/li>\n
  • \u9ad8\u6027\u80fd\u63a8\u7406\u652f\u6301<\/strong>\uff1a\u517c\u5bb9 Flash Attention \u7b49\u6280\u672f\uff0c\u63d0\u5347\u8ba1\u7b97\u6548\u7387\u3002<\/li>\n<\/ul>\n

     <\/p>\n

    \u4f7f\u7528\u5e2e\u52a9<\/h2>\n

    \u5b89\u88c5\u6d41\u7a0b<\/h3>\n

    VLM-R1 \u662f\u4e00\u4e2a\u57fa\u4e8e Python \u7684\u9879\u76ee\uff0c\u9700\u8981\u4e00\u5b9a\u7684\u73af\u5883\u914d\u7f6e\u624d\u80fd\u8fd0\u884c\u3002\u4ee5\u4e0b\u662f\u8be6\u7ec6\u7684\u5b89\u88c5\u4e0e\u4f7f\u7528\u6b65\u9aa4\uff0c\u5e2e\u52a9\u7528\u6237\u5feb\u901f\u4e0a\u624b\u3002<\/p>\n

    1. \u73af\u5883\u51c6\u5907<\/h4>\n
      \n
    • \u5b89\u88c5 Anaconda<\/strong>\uff1a\u5efa\u8bae\u4f7f\u7528 Anaconda \u7ba1\u7406 Python \u73af\u5883\uff0c\u786e\u4fdd\u7cfb\u7edf\u517c\u5bb9\u6027\u3002\u4e0b\u8f7d\u5730\u5740\uff1aAnaconda \u5b98\u7f51\u3002\u5b89\u88c5\u5b8c\u6210\u540e\uff0c\u6253\u5f00\u7ec8\u7aef\u3002<\/li>\n
    • \u521b\u5efa\u865a\u62df\u73af\u5883<\/strong>\uff1a\u5728\u7ec8\u7aef\u8f93\u5165\u4ee5\u4e0b\u547d\u4ee4\uff0c\u521b\u5efa\u4e00\u4e2a\u540d\u4e3a\u00a0vlm-r1<\/code>\u00a0\u7684 Python 3.10 \u73af\u5883\uff1a\n
      conda create -n vlm-r1 python=3.10<\/code><\/pre>\n<\/li>\n<\/ul>\n
        \n
      • \u6fc0\u6d3b\u73af\u5883<\/strong>\uff1a\u6fc0\u6d3b\u521a\u521a\u521b\u5efa\u7684\u73af\u5883\uff1a\n
        conda activate vlm-r1\r\n<\/code><\/pre>\n<\/li>\n<\/ul>\n
        2. \u5b89\u88c5\u9879\u76ee\u4f9d\u8d56<\/h4>\n
          \n
        • \u514b\u9686\u9879\u76ee<\/strong>\uff1a\u5c06 VLM-R1 \u7684\u4ee3\u7801\u4ed3\u5e93\u4e0b\u8f7d\u5230\u672c\u5730\u3002\u6253\u5f00\u7ec8\u7aef\uff0c\u8f93\u5165\uff1a\n
          git clone https:\/\/github.com\/om-ai-lab\/VLM-R1.git  \r\ncd VLM-R1\r\n<\/code><\/pre>\n<\/li>\n
        • \u8fd0\u884c\u5b89\u88c5\u811a\u672c<\/strong>\uff1a\u9879\u76ee\u63d0\u4f9b\u4e86\u4e00\u4e2a\u00a0setup.sh<\/code>\u00a0\u811a\u672c\uff0c\u7528\u4e8e\u81ea\u52a8\u5b89\u88c5\u4f9d\u8d56\u3002\u5728\u7ec8\u7aef\u4e2d\u8fd0\u884c\uff1a\n
          bash setup.sh\r\n<\/code><\/pre>\n
          \u8be5\u811a\u672c\u4f1a\u5b89\u88c5 PyTorch\u3001Transformers \u7b49\u6838\u5fc3\u5e93\uff0c\u786e\u4fdd\u73af\u5883\u5c31\u7eea\u3002<\/li>\n<\/ul>\n
          3. \u6570\u636e\u51c6\u5907<\/h4>\n
            \n
          • \u4e0b\u8f7d COCO \u6570\u636e\u96c6<\/strong>\uff1aVLM-R1 \u4f7f\u7528 COCO Train2014 \u56fe\u50cf\u6570\u636e\u96c6\u8fdb\u884c\u8bad\u7ec3\u3002\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\u4e0b\u8f7d\u5e76\u89e3\u538b\uff1a\n
            wget http:\/\/images.cocodataset.org\/train2014\/train2014.zip  \r\nunzip train2014.zip -d <your_image_root>\r\n<\/code><\/pre>\n
            \u8bb0\u4e0b\u89e3\u538b\u8def\u5f84\u00a0<your_image_root><\/code>\uff0c\u540e\u7eed\u914d\u7f6e\u4e2d\u9700\u8981\u7528\u5230\u3002<\/li>\n
          • \u4e0b\u8f7d RefCOCO \u6807\u6ce8\u6587\u4ef6<\/strong>\uff1aRefCOCO \u6570\u636e\u96c6\u7528\u4e8e\u6307\u4ee3\u8868\u8fbe\u4efb\u52a1\u3002\u4e0b\u8f7d\u94fe\u63a5\u53ef\u5728\u9879\u76ee\u6587\u6863\u4e2d\u627e\u5230\uff0c\u89e3\u538b\u540e\u653e\u7f6e\u5728\u5408\u9002\u76ee\u5f55\u3002<\/li>\n<\/ul>\n
            4. \u8bad\u7ec3\u6a21\u578b<\/h4>\n
              \n
            • \u914d\u7f6e\u8bad\u7ec3\u53c2\u6570<\/strong>\uff1a\u8fdb\u5165\u00a0src\/open-r1-multimodal<\/code>\u00a0\u76ee\u5f55\uff0c\u7f16\u8f91\u8bad\u7ec3\u811a\u672c\u4e2d\u7684\u53c2\u6570\u3002\u4f8b\u5982\uff1a\n
              cd src\/open-r1-multimodal\r\n<\/code><\/pre>\n
              \u4fee\u6539\u00a0grpo_rec.py<\/code>\u00a0\u6216\u8fd0\u884c\u547d\u4ee4\u65f6\u6307\u5b9a\u53c2\u6570\u3002\u4ee5\u4e0b\u662f\u4e00\u4e2a\u793a\u4f8b\u547d\u4ee4\uff1a<\/p>\n
              torchrun --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr=\"127.0.0.1\" --master_port=\"12346\" \\  \r\nsrc\/open_r1\/grpo_rec.py \\  \r\n--deepspeed local_scripts\/zero3.json \\  \r\n--output_dir output\/my_model \\  \r\n--model_name_or_path Qwen\/Qwen2.5-VL-3B-Instruct \\  \r\n--dataset_name data_config\/rec.yaml \\  \r\n--image_root <your_image_root> \\  \r\n--max_prompt_length 1024 \\  \r\n--num_generations 8 \\  \r\n--per_device_train_batch_size 1 \\  \r\n--gradient_accumulation_steps 2 \\  \r\n--logging_steps 1 \\  \r\n--bf16 \\  \r\n--torch_dtype bfloat16 \\  \r\n--num_train_epochs 2 \\  \r\n--save_steps 100\r\n<\/code><\/pre>\n
                \n
              • \u53c2\u6570\u8bf4\u660e\uff1a\n
                  \n
                • --nproc_per_node<\/code>\uff1aGPU \u6570\u91cf\uff0c\u9700\u6839\u636e\u4f60\u7684\u786c\u4ef6\u8c03\u6574\u3002<\/li>\n
                • --image_root<\/code>\uff1a\u66ff\u6362\u4e3a\u4f60\u7684 COCO \u6570\u636e\u96c6\u8def\u5f84\u3002<\/li>\n
                • --output_dir<\/code>\uff1a\u6a21\u578b\u4fdd\u5b58\u8def\u5f84\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
                  5. \u529f\u80fd\u64cd\u4f5c\u6d41\u7a0b<\/h4>\n
                  \u6307\u4ee3\u8868\u8fbe\u7406\u89e3\uff08REC\uff09<\/strong><\/h5>\n
                    \n
                  • \u8fd0\u884c\u6d4b\u8bd5\u811a\u672c<\/strong>\uff1a\u8bad\u7ec3\u5b8c\u6210\u540e\uff0c\u4f7f\u7528\u63d0\u4f9b\u7684\u6d4b\u8bd5\u811a\u672c\u9a8c\u8bc1\u6a21\u578b\u6548\u679c\u3002\u8fdb\u5165\u00a0src\/eval<\/code>\u00a0\u76ee\u5f55\uff1a\n
                    cd src\/eval  \r\npython test_rec_r1.py --model_path <your_trained_model> --image_root <your_image_root> --annotation_path <refcoco_annotation>\r\n<\/code><\/pre>\n<\/li>\n
                  • \u8f93\u5165\u793a\u4f8b<\/strong>\uff1a\u4e0a\u4f20\u4e00\u5f20\u56fe\u7247\u5e76\u8f93\u5165\u95ee\u9898\uff0c\u4f8b\u5982\u201c\u56fe\u4e2d\u7684\u84dd\u8272\u6c7d\u8f66\u5728\u54ea\u91cc\uff1f\u201d\u3002\u6a21\u578b\u4f1a\u8fd4\u56de\u76ee\u6807\u4f4d\u7f6e\u7684\u5750\u6807\u6216\u63cf\u8ff0\u3002<\/li>\n<\/ul>\n
                    \u56fe\u50cf\u4e0e\u6587\u672c\u5206\u6790<\/strong><\/h5>\n
                      \n
                    • \u51c6\u5907\u8f93\u5165<\/strong>\uff1a\u5c06\u56fe\u50cf\u6587\u4ef6\u548c\u95ee\u9898\u6587\u672c\u653e\u5165\u6307\u5b9a\u76ee\u5f55\uff0c\u6216\u76f4\u63a5\u5728\u811a\u672c\u4e2d\u6307\u5b9a\u8def\u5f84\u3002<\/li>\n
                    • \u8fd0\u884c\u63a8\u7406<\/strong>\uff1a\u4f7f\u7528\u4e0a\u8ff0\u6d4b\u8bd5\u811a\u672c\uff0c\u6a21\u578b\u4f1a\u8f93\u51fa\u5bf9\u56fe\u50cf\u5185\u5bb9\u7684\u5206\u6790\u7ed3\u679c\uff0c\u4f8b\u5982\u7269\u4f53\u7c7b\u522b\u3001\u4f4d\u7f6e\u7b49\u3002<\/li>\n<\/ul>\n
                      \u81ea\u5b9a\u4e49\u8bad\u7ec3<\/strong><\/h5>\n
                        \n
                      • \u4fee\u6539\u6570\u636e\u96c6<\/strong>\uff1a\u82e5\u9700\u4f7f\u7528\u81ea\u5df1\u7684\u6570\u636e\u96c6\uff0c\u7f16\u8f91\u00a0data_config\/rec.yaml<\/code>\uff0c\u6dfb\u52a0\u56fe\u7247\u8def\u5f84\u548c\u6807\u6ce8\u6587\u4ef6\u3002<\/li>\n
                      • \u8c03\u6574\u8d85\u53c2\u6570<\/strong>\uff1a\u6839\u636e\u4efb\u52a1\u9700\u6c42\u4fee\u6539\u00a0grpo_rec.py<\/code>\u00a0\u4e2d\u7684\u5b66\u4e60\u7387\u3001\u6279\u6b21\u5927\u5c0f\u7b49\u53c2\u6570\u3002<\/li>\n<\/ul>\n
                        6. \u6ce8\u610f\u4e8b\u9879<\/h4>\n
                          \n
                        • \u786c\u4ef6\u8981\u6c42<\/strong>\uff1a\u5efa\u8bae\u4f7f\u7528\u81f3\u5c11 8GB \u663e\u5b58\u7684 GPU\uff0c\u82e5\u8d44\u6e90\u6709\u9650\uff0c\u53ef\u51cf\u5c11\u00a0num_generations<\/code>\u00a0\u53c2\u6570\u4ee5\u964d\u4f4e\u5185\u5b58\u5360\u7528\u3002<\/li>\n
                        • \u8c03\u8bd5\u6a21\u5f0f<\/strong>\uff1a\u5728\u8bad\u7ec3\u65f6\u53ef\u8bbe\u7f6e\u00a0export DEBUG_MODE=\"true\"<\/code>\uff0c\u67e5\u770b\u8be6\u7ec6\u65e5\u5fd7\u3002<\/li>\n
                        • \u793e\u533a\u652f\u6301<\/strong>\uff1a\u5982\u9047\u5230\u95ee\u9898\uff0c\u53ef\u5728 GitHub Issues \u9875\u9762\u63d0\u95ee\uff0cOm AI Lab \u56e2\u961f\u548c\u793e\u533a\u4f1a\u63d0\u4f9b\u5e2e\u52a9\u3002<\/li>\n<\/ul>\n
                          \u901a\u8fc7\u4ee5\u4e0a\u6b65\u9aa4\uff0c\u7528\u6237\u53ef\u4ee5\u5b8c\u6574\u5b89\u88c5\u5e76\u4f7f\u7528 VLM-R1\uff0c\u65e0\u8bba\u662f\u8fdb\u884c\u89c6\u89c9\u4efb\u52a1\u7814\u7a76\u8fd8\u662f\u5f00\u53d1\u5b9e\u9645\u5e94\u7528\uff0c\u90fd\u80fd\u5feb\u901f\u4e0a\u624b\u5e76\u53d1\u6325\u5176\u5f3a\u5927\u529f\u80fd\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
                          VLM-R1 \u662f\u7531 Om AI Lab \u5f00\u53d1\u7684\u4e00\u4e2a\u5f00\u6e90\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u9879\u76ee\uff0c\u6258\u7ba1\u5728 GitHub \u4e0a\u3002\u8be5\u9879\u76ee\u57fa\u4e8e DeepSeek \u7684 R1 \u65b9\u6cd5\uff0c\u7ed3\u5408 Qwen2.5-VL \u6a21\u578b\uff0c\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08R1\uff09\u548c\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u6280\u672f\uff0c\u663e\u8457\u63d0\u5347\u4e86\u6a21\u578b\u5728…<\/p>\n","protected":false},"author":1,"featured_media":61895,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[230,378],"class_list":["post-26464","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","tag-aikaiyuanxiangmu","tag-shijuemubiaojiance"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/26464","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=26464"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/26464\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media\/61895"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=26464"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=26464"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=26464"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}