{"id":65,"date":"2024-07-21T17:15:10","date_gmt":"2024-07-21T09:15:10","guid":{"rendered":"https:\/\/www.onetts.com\/ai\/?p=65"},"modified":"2024-07-22T09:29:37","modified_gmt":"2024-07-22T01:29:37","slug":"qwen-vl","status":"publish","type":"post","link":"https:\/\/www.onetts.com\/ai\/models\/qwen-vl\/","title":{"rendered":"Qwen-VL"},"content":{"rendered":"<p data-spm-anchor-id=\"a2c6h.13066512.0.i4.763536afeW9dPt\">Qwen-VL\uff0c\u5168\u79f0Qwen Large Vision Language Model\uff0c\u662f\u7531\u963f\u91cc\u4e91\u7814\u53d1\u7684\u5927\u89c4\u6a21\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u3002\u5b83\u7ed3\u5408\u4e86\u6df1\u5ea6\u5b66\u4e60\u548c\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u6700\u65b0\u8fdb\u5c55\uff0c\u65e8\u5728\u5904\u7406\u548c\u7406\u89e3\u56fe\u50cf\u4e0e\u6587\u672c\u4e4b\u95f4\u7684\u590d\u6742\u5173\u7cfb\u3002Qwen-VL\u6a21\u578b\u7684\u521b\u65b0\u4e4b\u5904\u5728\u4e8e\u5176\u80fd\u591f\u63a5\u53d7\u56fe\u50cf\u3001\u6587\u672c\u548c\u8fb9\u754c\u6846\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u751f\u6210\u6587\u672c\u548c\u8fb9\u754c\u6846\u4f5c\u4e3a\u8f93\u51fa\uff0c\u8fd9\u4f7f\u5f97\u5b83\u5728\u591a\u6a21\u6001\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272\u3002<\/p>\n<h3 id=\"-\">\u6838\u5fc3\u7279\u6027<\/h3>\n<ul>\n<li><strong>\u591a\u8bed\u8a00\u652f\u6301<\/strong>\uff1aQwen-VL\u4e0d\u4ec5\u652f\u6301\u4e2d\u6587\uff0c\u8fd8\u80fd\u591f\u5904\u7406\u82f1\u6587\u7b49\u591a\u79cd\u8bed\u8a00\uff0c\u4f7f\u5176\u6210\u4e3a\u8de8\u6587\u5316\u4ea4\u6d41\u7684\u5f3a\u5927\u5de5\u5177\u3002<\/li>\n<li><strong>\u7ec6\u7c92\u5ea6\u8bc6\u522b<\/strong>\uff1a\u6a21\u578b\u91c7\u7528448\u5206\u8fa8\u7387\u7684\u8f93\u5165\uff0c\u76f8\u8f83\u4e8e\u5176\u4ed6\u6a21\u578b\u7684224\u5206\u8fa8\u7387\uff0c\u80fd\u591f\u8fdb\u884c\u66f4\u7cbe\u7ec6\u7684\u56fe\u50cf\u8bc6\u522b\u548c\u7406\u89e3\u3002<\/li>\n<li><strong>\u591a\u56fe\u8f93\u5165<\/strong>\uff1aQwen-VL\u80fd\u591f\u5904\u7406\u591a\u5f20\u56fe\u7247\u7684\u8f93\u5165\uff0c\u652f\u6301\u56fe\u7247\u95f4\u7684\u6bd4\u8f83\u548c\u591a\u56fe\u6587\u5b66\u521b\u4f5c\u3002<\/li>\n<li><strong>\u7aef\u5230\u7aef\u8bc6\u522b<\/strong>\uff1a\u6a21\u578b\u80fd\u591f\u76f4\u63a5\u4ece\u56fe\u50cf\u4e2d\u8bc6\u522b\u548c\u7406\u89e3\u6587\u672c\uff0c\u65e0\u9700\u989d\u5916\u7684\u56fe\u50cf\u5904\u7406\u6b65\u9aa4\u3002<\/li>\n<\/ul>\n<h2 id=\"-\">\u6a21\u578b\u8bc4\u6d4b<\/h2>\n<p>Qwen-VL\u5728\u591a\u4e2a\u6807\u51c6\u57fa\u51c6\u4e0a\u8fdb\u884c\u4e86\u5168\u9762\u7684\u8bc4\u6d4b\uff0c\u4ee5\u9a8c\u8bc1\u5176\u6027\u80fd\u548c\u6548\u679c\u3002<\/p>\n<h3 id=\"-\">\u82f1\u6587\u6807\u51c6\u57fa\u51c6<\/h3>\n<ul>\n<li><strong>Zero-shot Captioning<\/strong>\uff1a\u5728Flickr30K\u6570\u636e\u96c6\u4e0a\uff0cQwen-VL\u53d6\u5f97\u4e86SOTA\uff08State of the Art\uff09\u7684\u7ed3\u679c\uff0c\u663e\u793a\u51fa\u5176\u5728\u672a\u89c1\u6570\u636e\u96c6\u4e0a\u7684\u56fe\u7247\u63cf\u8ff0\u80fd\u529b\u3002<\/li>\n<li><strong>General VQA<\/strong>\uff1a\u5728VQAv2\u548cOK-VQA\u6570\u636e\u96c6\u4e0a\uff0cQwen-VL\u540c\u6837\u5c55\u73b0\u4e86\u5176\u5728\u901a\u7528\u95ee\u7b54\u4efb\u52a1\u4e0a\u7684\u5f3a\u5927\u80fd\u529b\u3002<\/li>\n<\/ul>\n<h3 id=\"-\">\u4e2d\u6587\u5b9a\u4f4d\u4efb\u52a1<\/h3>\n<ul>\n<li><strong>RefCOCO<\/strong>\uff1a\u5728RefCOCO\u6570\u636e\u96c6\u4e0a\uff0cQwen-VL\u5728\u591a\u4e2a\u5b50\u4efb\u52a1\u4e0a\u5747\u53d6\u5f97\u4e86SOTA\uff0c\u8bc1\u660e\u4e86\u5176\u5728\u4e2d\u6587\u5b9a\u4f4d\u4efb\u52a1\u4e0a\u7684\u4f18\u52bf\u3002<\/li>\n<\/ul>\n<h3 id=\"touchstone-\">TouchStone\u8bc4\u6d4b<\/h3>\n<ul>\n<li><strong>TouchStone<\/strong>\uff1a\u8fd9\u662f\u4e00\u4e2a\u57fa\u4e8eGPT4\u6253\u5206\u7684\u8bc4\u6d4b\u57fa\u51c6\uff0cQwen-VL\u5728\u4e2d\u82f1\u6587\u8bc4\u6d4b\u4e2d\u5747\u53d6\u5f97\u4e86\u4f18\u5f02\u7684\u6210\u7ee9\uff0c\u663e\u793a\u51fa\u5176\u5728\u56fe\u6587\u5bf9\u8bdd\u80fd\u529b\u548c\u4e0e\u4eba\u7c7b\u5bf9\u9f50\u6c34\u5e73\u4e0a\u7684\u4f18\u52bf\u3002<\/li>\n<\/ul>\n<h2 id=\"-\">\u90e8\u7f72\u4f7f\u7528<\/h2>\n<p>\u90e8\u7f72Qwen-VL\u6a21\u578b\u9700\u8981\u9075\u5faa\u4e00\u7cfb\u5217\u6b65\u9aa4\uff0c\u4ee5\u786e\u4fdd\u6a21\u578b\u80fd\u591f\u6b63\u786e\u8fd0\u884c\u5e76\u53d1\u6325\u5176\u529f\u80fd\u3002<\/p>\n<h3 id=\"-\">\u73af\u5883\u51c6\u5907<\/h3>\n<ul>\n<li>\u786e\u4fdd\u7cfb\u7edf\u5b89\u88c5\u4e86Python 3.8\u6216\u4ee5\u4e0a\u7248\u672c\u3002<\/li>\n<li>\u5b89\u88c5PyTorch 1.12\u6216\u4ee5\u4e0a\u7248\u672c\uff0c\u63a8\u8350\u4f7f\u75282.0\u6216\u4ee5\u4e0a\u7248\u672c\u3002<\/li>\n<li>\u5982\u679c\u4f7f\u7528GPU\uff0c\u5efa\u8bae\u4f7f\u7528CUDA 11.4\u6216\u4ee5\u4e0a\u7248\u672c\u3002<\/li>\n<\/ul>\n<h3 id=\"-\">\u5b89\u88c5\u4f9d\u8d56<\/h3>\n<p>\u901a\u8fc7pip\u5b89\u88c5\u6240\u9700\u7684\u4f9d\u8d56\u5e93\uff1a<\/p>\n<pre><code class=\"lang-bash\">pip <span class=\"token function\">install<\/span> modelscope -U\r\npip <span class=\"token function\">install<\/span> transformers accelerate tiktoken -U\r\npip <span class=\"token function\">install<\/span> einops transformers_stream_generator -U\r\npip <span class=\"token function\">install<\/span> <span class=\"token string\">\"pillow==9.*\"<\/span> -U\r\npip <span class=\"token function\">install<\/span> torchvision\r\npip <span class=\"token function\">install<\/span> matplotlib -U\r\n<\/code><\/pre>\n<h3 id=\"-\">\u6a21\u578b\u4e0b\u8f7d\u4e0e\u521d\u59cb\u5316<\/h3>\n<p>\u4f7f\u7528ModelScope\u63d0\u4f9b\u7684API\u4e0b\u8f7d\u6a21\u578b\u5e76\u521d\u59cb\u5316\uff1a<\/p>\n<pre><code class=\"lang-python\"><span class=\"token keyword\">from<\/span> modelscope <span class=\"token keyword\">import<\/span> snapshot_download<span class=\"token punctuation\">,<\/span> AutoModelForCausalLM<span class=\"token punctuation\">,<\/span> AutoTokenizer<span class=\"token punctuation\">,<\/span> GenerationConfig\r\n\r\nmodel_id <span class=\"token operator\">=<\/span> <span class=\"token string\">'qwen\/Qwen-VL'<\/span>\r\nrevision <span class=\"token operator\">=<\/span> <span class=\"token string\">'v1.0.3'<\/span>\r\nmodel_dir <span class=\"token operator\">=<\/span> snapshot_download<span class=\"token punctuation\">(<\/span>model_id<span class=\"token punctuation\">,<\/span> revision<span class=\"token operator\">=<\/span>revision<span class=\"token punctuation\">)<\/span>\r\n\r\ntokenizer <span class=\"token operator\">=<\/span> AutoTokenizer<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span>model_dir<span class=\"token punctuation\">,<\/span> trust_remote_code<span class=\"token operator\">=<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span>\r\nmodel <span class=\"token operator\">=<\/span> AutoModelForCausalLM<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span>model_dir<span class=\"token punctuation\">,<\/span> device_map<span class=\"token operator\">=<\/span><span class=\"token string\">\"auto\"<\/span><span class=\"token punctuation\">,<\/span> trust_remote_code<span class=\"token operator\">=<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span> fp16<span class=\"token operator\">=<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>eval<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span>\r\n<\/code><\/pre>\n<h3 id=\"-\">\u6a21\u578b\u4f7f\u7528<\/h3>\n<p>\u7f16\u5199\u4ee3\u7801\u4ee5\u4f7f\u7528\u6a21\u578b\u8fdb\u884c\u56fe\u50cf\u548c\u6587\u672c\u7684\u5904\u7406\uff1a<\/p>\n<pre><code class=\"lang-python\">query <span class=\"token operator\">=<\/span> tokenizer<span class=\"token punctuation\">.<\/span>from_list_format<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">[<\/span>\r\n    <span class=\"token punctuation\">{<\/span><span class=\"token string\">'image'<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">'https:\/\/qianwen-res.oss-cn-beijing.aliyuncs.com\/Qwen-VL\/assets\/demo.jpeg'<\/span><span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">,<\/span>\r\n    <span class=\"token punctuation\">{<\/span><span class=\"token string\">'text'<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">'Generate the caption in English with grounding:'<\/span><span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">,<\/span>\r\n<span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span>\r\ninputs <span class=\"token operator\">=<\/span> tokenizer<span class=\"token punctuation\">(<\/span>query<span class=\"token punctuation\">,<\/span> return_tensors<span class=\"token operator\">=<\/span><span class=\"token string\">'pt'<\/span><span class=\"token punctuation\">)<\/span>\r\ninputs <span class=\"token operator\">=<\/span> inputs<span class=\"token punctuation\">.<\/span>to<span class=\"token punctuation\">(<\/span>model<span class=\"token punctuation\">.<\/span>device<span class=\"token punctuation\">)<\/span>\r\npred <span class=\"token operator\">=<\/span> model<span class=\"token punctuation\">.<\/span>generate<span class=\"token punctuation\">(<\/span><span class=\"token operator\">**<\/span>inputs<span class=\"token punctuation\">)<\/span>\r\nresponse <span class=\"token operator\">=<\/span> tokenizer<span class=\"token punctuation\">.<\/span>decode<span class=\"token punctuation\">(<\/span>pred<span class=\"token punctuation\">.<\/span>cpu<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> skip_special_tokens<span class=\"token operator\">=<\/span><span class=\"token boolean\">False<\/span><span class=\"token punctuation\">)<\/span>\r\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span>response<span class=\"token punctuation\">)<\/span>\r\n<\/code><\/pre>\n<h2 id=\"-\">\u5e38\u89c1\u95ee\u9898<\/h2>\n<p>\u5728\u4f7f\u7528Qwen-VL\u6a21\u578b\u65f6\uff0c\u7528\u6237\u53ef\u80fd\u4f1a\u9047\u5230\u4e00\u4e9b\u5e38\u89c1\u95ee\u9898\uff0c\u4ee5\u4e0b\u662f\u4e00\u4e9b\u5e38\u89c1\u95ee\u9898\u53ca\u5176\u89e3\u7b54\u3002<\/p>\n<h3 id=\"q-\">Q: \u5982\u4f55\u89e3\u51b3\u6a21\u578b\u5728\u7279\u5b9a\u786c\u4ef6\u4e0a\u7684\u6027\u80fd\u95ee\u9898\uff1f<\/h3>\n<p>A: \u786e\u4fdd\u60a8\u7684\u786c\u4ef6\u6ee1\u8db3\u6a21\u578b\u7684\u6700\u4f4e\u8981\u6c42\uff0c\u5e76\u6839\u636e\u786c\u4ef6\u914d\u7f6e\u8c03\u6574\u6a21\u578b\u7684\u8fd0\u884c\u53c2\u6570\uff0c\u4f8b\u5982\u4f7f\u7528\u9002\u5f53\u7684\u7cbe\u5ea6\u6a21\u5f0f\uff08\u5982fp16\u6216bf16\uff09\u3002<\/p>\n<h3 id=\"q-\">Q: \u5982\u4f55\u5904\u7406\u6a21\u578b\u8f93\u51fa\u7684\u8fb9\u754c\u6846\u4e0d\u51c6\u786e\u7684\u60c5\u51b5\uff1f<\/h3>\n<p>A: \u8fb9\u754c\u6846\u7684\u51c6\u786e\u6027\u53ef\u80fd\u53d7\u5230\u591a\u79cd\u56e0\u7d20\u7684\u5f71\u54cd\uff0c\u5305\u62ec\u8f93\u5165\u56fe\u50cf\u7684\u5206\u8fa8\u7387\u548c\u8d28\u91cf\u3002\u5c1d\u8bd5\u8c03\u6574\u8f93\u5165\u56fe\u50cf\u7684\u5206\u8fa8\u7387\u6216\u4f7f\u7528\u66f4\u9ad8\u5206\u8fa8\u7387\u7684\u56fe\u50cf\u3002<\/p>\n<h3 id=\"q-qwen-vl-\">Q: \u5982\u4f55\u5728\u4e0d\u540c\u7684\u64cd\u4f5c\u7cfb\u7edf\u4e0a\u90e8\u7f72Qwen-VL\u6a21\u578b\uff1f<\/h3>\n<p>A: \u90e8\u7f72\u6b65\u9aa4\u5728\u4e0d\u540c\u7684\u64cd\u4f5c\u7cfb\u7edf\u4e0a\u5927\u81f4\u76f8\u540c\uff0c\u4f46\u53ef\u80fd\u9700\u8981\u6839\u636e\u64cd\u4f5c\u7cfb\u7edf\u8c03\u6574\u73af\u5883\u914d\u7f6e\u548c\u4f9d\u8d56\u5e93\u7684\u5b89\u88c5\u65b9\u5f0f\u3002<\/p>\n<h2 id=\"-\">\u76f8\u5173\u8d44\u6e90<\/h2>\n<p>\u4ee5\u4e0b\u662f\u4e00\u4e9b\u53ef\u4ee5\u8bbf\u95ee\u7684\u76f8\u5173\u8d44\u6e90\u94fe\u63a5\uff0c\u4ee5\u83b7\u53d6\u66f4\u591a\u5173\u4e8eQwen-VL\u7684\u4fe1\u606f\u548c\u652f\u6301\u3002<\/p>\n<ul>\n<li>Qwen-VL\u6a21\u578b\u5e93\u4e3b\u9875\uff1a<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https:\/\/modelscope.cn\/models\/qwen\/Qwen-VL\">ModelScope Qwen-VL<\/a><\/li>\n<li>\u6280\u672f\u5907\u5fd8\u5f55\uff1a<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https:\/\/modelscope.cn\/models\/qwen\/Qwen-VL\/technical_memo\">Qwen-VL Technical Memo<\/a><\/li>\n<li>\u5fae\u8c03(SFT)\u4ee3\u7801\u793a\u4f8b\uff1a<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https:\/\/github.com\/modelscope\/swift\/tree\/main\/examples\/pytorch\/llm\">Qwen-VL SFT Example<\/a><\/li>\n<li>\u8bc4\u6d4b\u811a\u672c\uff1a<a target=\"_blank\" rel=\"noopener noreferrer nofollow\" href=\"https:\/\/github.com\/modelscope\/swift\/tree\/main\/eval\">Qwen-VL Evaluation Scripts<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Qwen-VL\uff0c\u5168\u79f0Qwen Large Vision Language Model\uff0c\u662f\u7531\u963f\u91cc\u4e91\u7814\u53d1\u7684\u5927\u89c4\u6a21\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u3002\u5b83\u7ed3\u5408\u4e86\u6df1\u5ea6\u5b66\u4e60\u548c\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u6700\u65b0\u8fdb\u5c55\uff0c\u65e8\u5728\u5904\u7406\u548c\u7406\u89e3\u56fe\u50cf\u4e0e\u6587\u672c\u4e4b\u95f4\u7684\u590d\u6742\u5173\u7cfb\u3002Qwen-VL\u6a21\u578b\u7684\u521b\u65b0\u4e4b\u5904\u5728\u4e8e\u5176\u80fd\u591f\u63a5\u53d7\u56fe\u50cf\u3001\u6587\u672c\u548c\u8fb9\u754c\u6846\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u751f\u6210\u6587\u672c\u548c\u8fb9\u754c\u6846\u4f5c\u4e3a\u8f93\u51fa\uff0c\u8fd9\u4f7f\u5f97\u5b83\u5728\u591a\u6a21\u6001\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272\u3002 \u6838\u5fc3\u7279\u6027 \u591a\u8bed\u8a00\u652f\u6301\uff1aQwen-VL\u4e0d\u4ec5\u652f\u6301\u4e2d\u6587\uff0c\u8fd8\u80fd\u591f\u5904\u7406\u82f1\u6587\u7b49\u591a\u79cd\u8bed\u8a00\uff0c\u4f7f\u5176\u6210\u4e3a\u8de8\u6587\u5316\u4ea4\u6d41\u7684\u5f3a\u5927\u5de5\u5177\u3002 \u7ec6\u7c92\u5ea6\u8bc6\u522b\uff1a\u6a21\u578b\u91c7\u7528448\u5206\u8fa8\u7387\u7684\u8f93\u5165\uff0c\u76f8\u8f83\u4e8e\u5176\u4ed6\u6a21\u578b\u7684224\u5206\u8fa8\u7387\uff0c\u80fd\u591f\u8fdb<\/p>\n","protected":false},"author":1,"featured_media":66,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[12],"tags":[],"collection":[48],"company":[6],"rank":[],"class_list":["post-65","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-multimodal","collection-qwen-vl","company-alibaba"],"_links":{"self":[{"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts\/65","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/comments?post=65"}],"version-history":[{"count":1,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts\/65\/revisions"}],"predecessor-version":[{"id":67,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/posts\/65\/revisions\/67"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/media\/66"}],"wp:attachment":[{"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/media?parent=65"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/categories?post=65"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/tags?post=65"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/collection?post=65"},{"taxonomy":"company","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/company?post=65"},{"taxonomy":"rank","embeddable":true,"href":"https:\/\/www.onetts.com\/ai\/wp-json\/wp\/v2\/rank?post=65"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}