{"id":91890,"date":"2026-08-08T17:24:50","date_gmt":"2026-08-08T09:24:50","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/91890.html"},"modified":"2026-08-08T17:24:50","modified_gmt":"2026-08-08T09:24:50","slug":"%e4%b8%ba%e4%bb%80%e4%b9%88-pt-%e8%83%bd%e8%bd%ac-onnx%ef%bc%9fonnx-%e5%88%b0%e5%ba%95%e6%98%af%e4%bb%80%e4%b9%88%ef%bc%9f%e4%b8%ba%e4%bb%80%e4%b9%88%e9%83%a8%e7%bd%b2%e9%83%bd%e8%a6%81%e5%85%88","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/91890.html","title":{"rendered":"\u4e3a\u4ec0\u4e48 .pt \u80fd\u8f6c onnx\uff1fonnx \u5230\u5e95\u662f\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u90e8\u7f72\u90fd\u8981\u5148\u8fc7\u4e00\u9053 ONNX\uff1f\u2014\u2014 \u4eb2\u624b\u8dd1\u4e00\u904d\uff0c\u5feb 5 \u500d\u7684\u539f\u56e0\u5168\u5728\u8fd9"},"content":{"rendered":"

\u52a8\u624b\u8dd1\u8fc7\u624d\u6562\u5199\u3002\u672c\u6587\u6240\u6709\u7ed3\u8bba\u90fd\u6765\u81ea\u6211\u5728\u81ea\u5df1\u73af\u5883\u91cc\u4eb2\u624b\u8dd1\u51fa\u6765\u7684\u771f\u5b9e\u8f93\u51fa，\u4e00\u4e2a\u63a8\u6f14\u3001\u4e00\u4e2a\u62cd\u8111\u888b\u7684\u7ed3\u8bba\u90fd\u6ca1\u6709\u3002\u73af\u5883：torch 2.10.0+cu128 \/ onnx 1.21.0 \/ onnxruntime 1.26.0\u3002<\/p>\n

\u5f15\u8a00：\u4e09\u4e2a\u6bcf\u5929\u90fd\u60f3\u95ee\u7684\u95ee\u9898<\/h3>\n

\u505a\u63a8\u7406\u90e8\u7f72\u7684\u4eba，\u51e0\u4e4e\u5929\u5929\u548c .pt\u3001.onnx \u6253\u4ea4\u9053\u3002\u4f46\u5927\u591a\u6570\u4eba\u662f"\u4f1a\u590d\u5236\u547d\u4ee4，\u4e0d\u61c2\u539f\u7406"：<\/p>\n

  • \u4e3a\u4ec0\u4e48 torch.onnx.export \u4e00\u4e0b，.pt \u5c31\u80fd\u53d8\u6210 .onnx？\u5b83\u4e0d\u662f\u4e24\u79cd\u5b8c\u5168\u4e0d\u540c\u7684\u6587\u4ef6\u5417？<\/li>\n
  • ONNX \u5230\u5e95\u662f\u4ec0\u4e48\u4e1c\u897f？\u4e00\u4e2a\u6587\u4ef6\u683c\u5f0f？<\/li>\n
  • \u4e3a\u4ec0\u4e48\u5e02\u9762\u4e0a\u7684\u6a21\u578b\u90fd\u8981\u5148\u8f6c\u6210 ONNX \u8fd9\u4e2a"\u4e2d\u95f4\u6001"，\u518d\u8f6c\u6210 TensorRT \/ OpenVINO \/ RKNN？<\/li>\n
  • \u4e3a\u4ec0\u4e48\u90fd\u8bf4 ONNX \u63a8\u7406\u6bd4 .pt \u5feb？\u5b83\u54ea\u6765\u7684\u5e95\u6c14？<\/li>\n

    \u8fd9\u7bc7\u6211\u7528\u4e00\u4e2a 8.7 \u4e07\u53c2\u6570\u7684\u5c0f CNN \u5f53\u5b9e\u9a8c\u5bf9\u8c61，\u628a .pt \u548c .onnx \u90fd\u62c6\u5f00\u7ed9\u4f60\u770b，\u518d\u5b9e\u6d4b\u63a8\u7406\u901f\u5ea6，\u7528\u6570\u636e\u56de\u7b54\u8fd9\u56db\u4e2a\u95ee\u9898\u3002
    \n\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    \n

    \u7b2c\u4e00\u7ae0\u3001\u6495\u5f00\u5916\u58f3：.pt \u548c .onnx \u5206\u522b\u662f\u4ec0\u4e48\u683c\u5f0f？<\/h3>\n

    \u548c\u4e0a\u4e00\u7bc7\u62c6 .pt \u4e00\u6837，\u5148\u770b\u6587\u4ef6\u5934（\u9b54\u6570）\u3002\u4fdd\u5b58\u4e00\u4e2a state_dict \u518d\u7528\u4e8c\u8fdb\u5236\u8bfb\u524d 20 \u5b57\u8282：<\/p>\n

    with<\/span> open<\/span>(<\/span>'pt_onnx_blog\/tinycnn.pt'<\/span>,<\/span> 'rb'<\/span>)<\/span> as<\/span> f:<\/span>
    \n print<\/span>(<\/span>repr<\/span>(<\/span>f.<\/span>read(<\/span>20<\/span>)<\/span>)<\/span>)<\/span><\/p>\n

    \u6211\u73af\u5883\u91cc\u771f\u5b9e\u8f93\u51fa：<\/p>\n

    b'PK\\\\x03\\\\x04\\\\x00\\\\x00\\\\x08\\\\x08\\\\x00\\\\x00\\\\x00\\\\x00\\\\x00\\\\x00\\\\x00\\\\x00\\\\x00\\\\x00\\\\x00\\\\x00'<\/p>\n

    PK\\\\x03\\\\x04 \u662f ZIP \u7684\u9b54\u6570\u3002\u6240\u4ee5 .pt \u672c\u8d28\u662f\u4e2a ZIP \u538b\u7f29\u5305\u3002\u7528 zipfile \u770b\u91cc\u9762\u88c5\u4e86\u4ec0\u4e48：<\/p>\n

    tinycnn\/data.pkl
    \ntinycnn\/.format_version
    \ntinycnn\/.storage_alignment
    \ntinycnn\/byteorder
    \ntinycnn\/data\/0
    \ntinycnn\/data\/1
    \ntinycnn\/data\/2
    \ntinycnn\/data\/3<\/p>\n

    data.pkl \u662f Pickle \u5e8f\u5217\u5316\u7684\u5bf9\u8c61\u7ed3\u6784，data\/0\u3001data\/1\u2026 \u662f\u6bcf\u4e2a\u5f20\u91cf\u7684\u4e8c\u8fdb\u5236\u6570\u636e\u3002\u4e5f\u5c31\u662f\u8bf4：.pt = ZIP \u5916\u58f3 + Pickle \u5185\u6838 + \u72ec\u7acb\u5b58\u50a8\u7684\u5f20\u91cf\u3002\u5b83\u5b58\u7684\u662f\u4e00\u5806\u5e26\u540d\u5b57\u7684 numpy \u5f20\u91cf，\u5916\u52a0"\u600e\u4e48\u628a\u5b83\u4eec\u88c5\u56de\u7c7b"\u7684\u8bf4\u660e\u3002<\/p>\n

    \u90a3 .onnx \u5462？\u540c\u6837\u8bfb\u6587\u4ef6\u5934：<\/p>\n

    b'\\\\x08\\\\n\\\\x12\\\\x07pyto…'<\/p>\n

    \u8fd9\u662f Protobuf（Protocol Buffers） \u7684\u7f16\u7801\u3002\\\\x08\\\\n \u8868\u793a ir_version=10，\\\\x12\\\\x07pyto… \u662f producer \u5b57\u6bb5，\u5185\u5bb9\u662f pytorch_export\u2026\u3002\u4e5f\u5c31\u662f\u8bf4 ONNX \u662f\u4e00\u4e2a\u7528 Protobuf \u5e8f\u5217\u5316\u7684\u3001\u63cf\u8ff0"\u8ba1\u7b97\u56fe"\u7684\u6807\u51c6\u6587\u4ef6\u3002<\/p>\n

    \u7528 onnx \u5e93\u628a\u5b83\u89e3\u6790\u6210\u4eba\u7c7b\u80fd\u770b\u61c2\u7684\u6837\u5b50：<\/p>\n

    %input[FLOAT, 1x3x112x112] # \u8f93\u5165
    \ninitializers( # \u6743\u91cd（initializer \u76f4\u63a5\u6302\u5728\u56fe\u91cc）
    \n %conv1.weight[FLOAT, 16x3x3x3]
    \n %fc.weight[FLOAT, 10×25088]
    \n …
    \n)
    \n%getitem = Conv(…)(%input, %conv1.weight, %conv1.bias)
    \n%relu = Relu(%getitem)
    \n%max_pool2d = MaxPool(%relu)
    \n…<\/p>\n

    \u5230\u8fd9\u91cc\u4e24\u4e2a\u6587\u4ef6\u7684\u672c\u8d28\u5c31\u6e05\u695a\u4e86：<\/p>\n\n.pt.onnx<\/tr>\n\n\n\n\n\n
    \u672c\u8d28<\/td>\nZIP + Pickle<\/td>\nProtobuf \u5e8f\u5217\u5316\u7684\u8ba1\u7b97\u56fe<\/td>\n<\/tr>\n
    \u5b58\u4ec0\u4e48<\/td>\n\u5f20\u91cf + \u201c\u5982\u4f55\u88c5\u56de\u7c7b\u201d<\/td>\n\u7b97\u5b50\u8282\u70b9 + \u6743\u91cd + \u6570\u636e\u6d41（\u6709\u5411\u56fe）<\/td>\n<\/tr>\n
    \u4f9d\u8d56\u4ec0\u4e48<\/td>\n\u5fc5\u987b\u6709\u5bf9\u5e94\u7684 Python \u7c7b\u5b9a\u4e49<\/td>\n\u4e0d\u4f9d\u8d56\u4efb\u4f55\u6846\u67b6，\u7eaf\u6570\u636e\u63cf\u8ff0<\/td>\n<\/tr>\n
    \u8ba1\u91cf\u5355\u4f4d<\/td>\n\u5f20\u91cf<\/td>\n\u56fe\u7684\u8282\u70b9（\u7b97\u5b50）<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u4e00\u4e2a\u5b58\u7684\u662f"\u53c2\u6570\u548c\u7c7b"，\u4e00\u4e2a\u5b58\u7684\u662f"\u4e00\u5f20\u7b97\u5b50\u7ec4\u6210\u7684\u6570\u636e\u6d41\u56fe + \u53c2\u6570"\u3002\u8fd9\u5c31\u662f\u4e24\u8005\u6700\u6839\u672c\u7684\u533a\u522b，\u4e5f\u662f\u540e\u9762\u6240\u6709\u95ee\u9898\u7684\u94a5\u5319\u3002<\/p>\n

    \n

    \u7b2c\u4e8c\u7ae0\u3001\u4e3a\u4ec0\u4e48 .pt \u80fd\u8f6c\u6210 onnx？<\/h3>\n

    \u56e0\u4e3a\u4efb\u4f55\u4e00\u4e2a PyTorch \u6a21\u578b，\u672c\u8d28\u4e0a\u90fd"\u957f\u7740\u4e00\u5f20\u8ba1\u7b97\u56fe"，\u53ea\u662f\u5e73\u65f6\u88ab\u85cf\u8d77\u6765\u4e86\u3002<\/p>\n

    \u8bad\u6a21\u578b\u65f6\u4f60\u5199\u7684\u662f forward()，\u91cc\u9762\u662f\u4e00\u4e32\u7b97\u5b50：Conv\u3001ReLU\u3001BatchNorm\u3001MaxPool\u3001Linear\u2026\u2026PyTorch \u5e95\u5c42（ATen）\u672c\u6765\u5c31\u628a\u5b83\u4eec\u7f16\u6392\u6210\u4e86\u4e00\u5f20 \u52a8\u6001\u8ba1\u7b97\u56fe\u3002\u8f6c ONNX \u8981\u505a\u7684，\u5c31\u662f\u628a\u8fd9\u56e2"\u85cf\u7740\u56fe"\u7684 Python \u5bf9\u8c61，\u7528\u4e00\u6279\u771f\u5b9e\u8f93\u5165\u5582\u5b83\u8dd1\u4e00\u904d，\u628a\u8fd9\u6761\u8def\u5f84\u4e0a\u7684\u7b97\u5b50\u3001\u5f62\u72b6\u3001\u6743\u91cd\u5168\u90e8\u56fa\u5316\u4e0b\u6765，\u518d\u7ffb\u8bd1\u6210 ONNX \u7684\u56fe\u683c\u5f0f\u3002<\/p>\n

    \u4e00\u53e5\u8bdd：torch.onnx.export \u505a\u7684\u4e8b = \u7528\u771f\u5b9e\u8f93\u5165\u8dd1\u4e00\u904d\u524d\u5411（trace）+ \u628a\u7b97\u5b50\u7ffb\u8bd1\u6210 ONNX \u6807\u51c6\u7b97\u5b50 + \u628a\u6743\u91cd\u56fa\u5316\u6210 initializer\u3002<\/p>\n

    \u6211\u5bfc\u51fa\u7684\u56fe\u4e0a，torch.onnx \u5728\u7ffb\u8bd1\u65f6\u987a\u624b\u505a\u4e86\u4e00\u4ef6\u4e8b：\u628a BatchNorm \u76f4\u63a5\u7194\u8fdb\u4e86 Conv\u3002\u770b\u7b97\u5b50\u5206\u5e03\u5c31\u9732\u9985\u4e86：<\/p>\n

    \u7b97\u5b50\u8282\u70b9\u603b\u6570: 8
    \n\u7b97\u5b50\u7c7b\u578b\u5206\u5e03: {'Conv': 2, 'Relu': 2, 'MaxPool': 2, 'Reshape': 1, 'Gemm': 1}<\/p>\n

    \u6211\u7684\u6a21\u578b\u660e\u660e\u6709 2 \u4e2a BatchNorm \u5c42，\u4f46\u5bfc\u51fa\u7684 ONNX \u56fe\u91cc\u4e00\u4e2a BatchNorm \u8282\u70b9\u90fd\u6ca1\u6709\u3002\u56e0\u4e3a BN \u5728\u63a8\u7406\u65f6\u53ea\u662f\u4e00\u7ec4 scale\/bias\/shift，\u53ef\u4ee5\u6298\u53e0\u8fdb\u524d\u9762\u7684 Conv \u6743\u91cd\u91cc\u3002\u8fd9\u5c31\u662f\u56fe\u4f18\u5316（\u7b97\u5b50\u878d\u5408）\u7684\u7b2c\u4e00\u6b65，\u4e5f\u662f ONNX \u80fd\u53d8\u5feb\u7684\u7b2c\u4e00\u4e2a\u539f\u56e0，\u540e\u9762\u8fd8\u4f1a\u5c55\u5f00\u3002<\/p>\n

    \u6240\u4ee5"\u4e3a\u4ec0\u4e48\u80fd\u8f6c"：\u540c\u4e00\u4e2a\u6a21\u578b，PyTorch \u773c\u91cc\u662f"\u7c7b + \u53c2\u6570"，ONNX \u773c\u91cc\u662f"\u7b97\u5b50\u56fe + \u53c2\u6570"\u3002\u5bfc\u51fa\u53ea\u662f\u505a\u4e86\u4e2a\u7ffb\u8bd1，\u628a\u524d\u8005\u7ffb\u8bd1\u6210\u540e\u8005，\u4fe1\u606f\u4e0d\u4e22，\u8fd8\u987a\u5e26\u4f18\u5316\u4e86\u4e00\u4e0b\u3002<\/p>\n

    \u5b8c\u6574\u5bfc\u51fa\u4ee3\u7801（\u53ef\u590d\u73b0）<\/h4>\n

    import<\/span> torch,<\/span> torch.<\/span>nn as<\/span> nn<\/p>\n

    class<\/span> TinyCNN<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
    \n def<\/span> __init__<\/span>(<\/span>self)<\/span>:<\/span>
    \n super<\/span>(<\/span>)<\/span>.<\/span>__init__(<\/span>)<\/span>
    \n self.<\/span>conv1 =<\/span> nn.<\/span>Conv2d(<\/span>3<\/span>,<\/span> 16<\/span>,<\/span> 3<\/span>,<\/span> padding=<\/span>1<\/span>)<\/span>
    \n self.<\/span>bn1 =<\/span> nn.<\/span>BatchNorm2d(<\/span>16<\/span>)<\/span>
    \n self.<\/span>relu =<\/span> nn.<\/span>ReLU(<\/span>)<\/span>
    \n self.<\/span>pool =<\/span> nn.<\/span>MaxPool2d(<\/span>2<\/span>)<\/span>
    \n self.<\/span>conv2 =<\/span> nn.<\/span>Conv2d(<\/span>16<\/span>,<\/span> 32<\/span>,<\/span> 3<\/span>,<\/span> padding=<\/span>1<\/span>)<\/span>
    \n self.<\/span>bn2 =<\/span> nn.<\/span>BatchNorm2d(<\/span>32<\/span>)<\/span>
    \n self.<\/span>fc =<\/span> nn.<\/span>Linear(<\/span>32<\/span> *<\/span> 28<\/span> *<\/span> 28<\/span>,<\/span> 10<\/span>)<\/span>
    \n def<\/span> forward<\/span>(<\/span>self,<\/span> x)<\/span>:<\/span>
    \n x =<\/span> self.<\/span>pool(<\/span>self.<\/span>relu(<\/span>self.<\/span>bn1(<\/span>self.<\/span>conv1(<\/span>x)<\/span>)<\/span>)<\/span>)<\/span>
    \n x =<\/span> self.<\/span>pool(<\/span>self.<\/span>relu(<\/span>self.<\/span>bn2(<\/span>self.<\/span>conv2(<\/span>x)<\/span>)<\/span>)<\/span>)<\/span>
    \n x =<\/span> x.<\/span>view(<\/span>x.<\/span>size(<\/span>0<\/span>)<\/span>,<\/span> –<\/span>1<\/span>)<\/span>
    \n return<\/span> self.<\/span>fc(<\/span>x)<\/span><\/p>\n

    model =<\/span> TinyCNN(<\/span>)<\/span>.<\/span>eval<\/span>(<\/span>)<\/span>
    \nx =<\/span> torch.<\/span>randn(<\/span>1<\/span>,<\/span> 3<\/span>,<\/span> 112<\/span>,<\/span> 112<\/span>)<\/span><\/p>\n

    torch.<\/span>save(<\/span>model.<\/span>state_dict(<\/span>)<\/span>,<\/span> "tinycnn.pt"<\/span>)<\/span> # \u5b58\u6743\u91cd<\/span>
    \ntorch.<\/span>onnx.<\/span>export(<\/span>
    \n model,<\/span> x,<\/span> "tinycnn.onnx"<\/span>,<\/span>
    \n input_names=<\/span>[<\/span>"input"<\/span>]<\/span>,<\/span> output_names=<\/span>[<\/span>"output"<\/span>]<\/span>,<\/span>
    \n opset_version=<\/span>17<\/span>,<\/span>
    \n)<\/span>
    \nprint<\/span>(<\/span>"\u53c2\u6570\u91cf:"<\/span>,<\/span> sum<\/span>(<\/span>p.<\/span>numel(<\/span>)<\/span> for<\/span> p in<\/span> model.<\/span>parameters(<\/span>)<\/span>)<\/span>)<\/span> # 87114<\/span><\/p>\n

    \u6ce8\u610f：torch.onnx.export \u9700\u8981\u4f60\u4f20\u5165\u4e00\u4e2a\u771f\u5b9e\u8f93\u5165 x，\u8fd9\u6b63\u662f"trace"\u7684\u8981\u6c42\u2014\u2014\u5b83\u5f97\u8dd1\u4e00\u904d\u624d\u77e5\u9053\u56fe\u957f\u4ec0\u4e48\u6837\u3002<\/p>\n

    \n

    \u7b2c\u4e09\u7ae0\u3001\u4e3a\u4ec0\u4e48\u5e02\u9762\u4e0a\u7684\u6a21\u578b\u90fd\u8981\u4e00\u4e2a ONNX \u4e2d\u95f4\u6001？<\/h3>\n

    \u56e0\u4e3a ONNX \u4e0d\u662f"\u67d0\u79cd\u63a8\u7406\u5f15\u64ce\u7684\u79c1\u6709\u683c\u5f0f"，\u800c\u662f\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684"\u901a\u7528\u8bed\u8a00 \/ \u4e2d\u95f4\u8868\u793a（IR）"\u3002<\/p>\n

    \u73b0\u5b9e\u91cc，\u8bad\u5b8c\u4e00\u4e2a\u6a21\u578b，\u4f60\u60f3\u90e8\u7f72\u5230：<\/p>\n

      \n
    • NVIDIA \u663e\u5361 \u2192 \u7528 TensorRT（TRT \u5f15\u64ce）<\/li>\n
    • Intel CPU \/ \u96c6\u663e \u2192 \u7528 OpenVINO（IR）<\/li>\n
    • \u745e\u82af\u5fae Rockchip \u677f\u5b50 \u2192 \u7528 RKNN<\/li>\n
    • \u624b\u673a\/\u6811\u8393\u6d3e \u2192 \u7528 TFLite \/ NCNN \/ MNN\u2026\u2026<\/li>\n<\/ul>\n

      \u4f60\u4e0d\u53ef\u80fd\u6bcf\u4e2a\u786c\u4ef6\u90fd\u8ba9 PyTorch \u539f\u751f\u652f\u6301\u3002\u4e8e\u662f ONNX \u5c31\u6210\u4e86"\u4e07\u80fd\u4e2d\u8f6c\u7ad9"：<\/p>\n

      PyTorch(.pt) \u2500\u2500\u5bfc\u51fa\u2500\u2500\u25b6 ONNX \u2500\u2500\u8f6c\u6362\u2500\u2500\u25b6 TensorRT \/ OpenVINO \/ RKNN \/ NCNN …
      \nTensorFlow(.pb) \u2500\u2500\u25b6 ONNX \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u25b6 \u540c\u4e00\u4e2a ONNX，\u5582\u7ed9\u4efb\u610f\u5f15\u64ce<\/p>\n

      \u4e3a\u4ec0\u4e48\u9700\u8981\u4e00\u4e2a"\u4e2d\u95f4\u6001"\u800c\u4e0d\u662f"\u6bcf\u4e2a\u5f15\u64ce\u5404\u81ea\u8bfb .pt"？<\/p>\n

    • \u8de8\u6846\u67b6：ONNX \u662f\u5f00\u653e\u6807\u51c6，PyTorch\u3001TensorFlow\u3001Paddle \u90fd\u80fd\u5bfc\u51fa ONNX，\u5404\u5bb6\u7684\u6a21\u578b\u90fd\u80fd\u6c47\u5230\u540c\u4e00\u6761\u90e8\u7f72\u6d41\u6c34\u7ebf\u4e0a\u3002<\/li>\n
    • \u8de8\u786c\u4ef6：\u6bcf\u4e2a\u786c\u4ef6\u5382\u5546\u53ea\u8981\u5b9e\u73b0"\u8bfb ONNX \u2192 \u8f6c\u6210\u81ea\u5df1\u7684\u683c\u5f0f"，\u5c31\u4e00\u6b21\u6027\u652f\u6301\u4e86\u6240\u6709\u6846\u67b6\u7684\u6a21\u578b，\u4e0d\u7528\u4e3a\u6bcf\u4e2a\u6846\u67b6\u5404\u5199\u4e00\u904d\u8f6c\u6362\u5668\u3002<\/li>\n
    • \u89e3\u8026\u8bad\u7ec3\u4e0e\u90e8\u7f72：ONNX \u4e0d\u4f9d\u8d56\u4efb\u4f55 Python \u7c7b\u3002\u8bad\u7ec3\u7aef\u5347\u7ea7\u6846\u67b6，\u53ea\u8981\u5bfc\u51fa\u7684 ONNX \u6807\u51c6\u4e0d\u53d8，\u90e8\u7f72\u7aef\u5b8c\u5168\u65e0\u611f\u3002<\/li>\n

      \u6240\u4ee5"ONNX \u4e2d\u95f4\u6001"\u7684\u4ef7\u503c，\u7c7b\u6bd4\u4e00\u4e0b\u5c31\u662f：\u5927\u5bb6\u90fd\u5728\u7528\u901a\u7528\u683c\u5f0f（\u6bd4\u5982 PDF），\u800c\u4e0d\u662f\u6bcf\u5bb6\u6d4f\u89c8\u5668\u5404\u51fa\u4e2a\u79c1\u6709\u683c\u5f0f\u3002ONNX \u5c31\u662f\u6a21\u578b\u754c\u7684 PDF\u3002<\/p>\n

      \u6ce8\u610f\u4e00\u4e2a\u5bb9\u6613\u6df7\u6dc6\u7684\u70b9：ONNX \u672c\u8eab\u4e0d\u662f"\u6700\u7ec8\u90e8\u7f72\u683c\u5f0f"，\u800c\u662f"\u5206\u53d1\/\u4ea4\u6362\u683c\u5f0f"\u3002\u771f\u6b63\u4e0a\u7ebf\u8dd1\u5f97\u98de\u5feb\u7684，\u662f\u5404\u5f15\u64ce\u628a ONNX \u518d\u8f6c\u51fa\u6765\u7684\u539f\u751f\u5f15\u64ce\u6587\u4ef6（TRT \u5f15\u64ce\u3001RKNN \u6a21\u578b\u7b49）\u3002ONNX \u7ad9\u5728"\u6e90\u5934\u548c\u7ec8\u70b9\u4e4b\u95f4"\u3002<\/p>\n

      \n

      \u7b2c\u56db\u7ae0\u3001\u4e3a\u4ec0\u4e48 ONNX \u63a8\u7406\u6bd4 .pt \u5feb？\u2014\u2014 \u5b9e\u6d4b 5.11 \u500d<\/h3>\n

      \u5148\u8bf4\u7834\u4e00\u4e2a\u8bef\u533a：\u5feb\u7684\u4e0d\u662f"ONNX \u8fd9\u4e2a\u6587\u4ef6"，\u800c\u662f\u8fd0\u884c\u5b83\u7684\u63a8\u7406\u5f15\u64ce ONNX Runtime（ORT）\u3002.onnx \u6587\u4ef6\u672c\u8eab\u53ea\u662f\u4e00\u5f20\u56fe，\u662f\u4e00\u6bb5\u63cf\u8ff0，\u4e0d\u662f\u7f16\u8bd1\u597d\u7684\u4e8c\u8fdb\u5236\u3002\u901f\u5ea6\u6765\u81ea"\u7528\u8c01\u53bb\u6267\u884c\u8fd9\u5f20\u56fe"\u3002<\/p>\n

      \u6211\u62ff\u540c\u4e00\u4e2a\u6a21\u578b，\u540c\u4e00\u5f20 1\u00d73\u00d7112\u00d7112 \u8f93\u5165，\u5728 CPU \u4e0a\u8dd1 200 \u6b21\u53d6\u5e73\u5747（\u5148 warmup）：<\/p>\n

      PyTorch eager \u5e73\u5747: 0.877 ms
      \nONNX Runtime \u5e73\u5747: 0.172 ms
      \n\u52a0\u901f\u6bd4: 5.11x<\/p>\n

      \u8f93\u51fa\u4e00\u81f4\u6027\u4e5f\u9a8c\u8bc1\u4e86（\u6d6e\u70b9\u8bef\u5dee\u53ef\u5ffd\u7565）：<\/p>\n

      \u6700\u5927\u7edd\u5bf9\u8bef\u5dee: 1.45e-07<\/p>\n

      \u4e00\u4e2a eager \u6a21\u5f0f，\u4e00\u4e2a ORT \u5f15\u64ce，\u4e3a\u4ec0\u4e48\u5dee 5 \u500d？\u4e09\u4e2a\u539f\u56e0，\u5176\u4e2d\u7b2c\u4e8c\u70b9\u662f\u6838\u5fc3\u3002<\/p>\n

      \u539f\u56e0\u4e00：\u6ca1\u6709 Python \u89e3\u91ca\u5668\u5f00\u9500<\/h4>\n

      PyTorch eager \u6a21\u5f0f，\u6bcf\u6267\u884c\u4e00\u4e2a\u7b97\u5b50\u90fd\u8981\u8d70\u4e00\u904d Python \u89e3\u91ca\u5668 \u2192 \u6d3e\u53d1\u5230 C++\/CUDA kernel \u7684\u94fe\u8def，\u4e00\u5c42\u5c42\u51fd\u6570\u8c03\u7528\u3001\u5bf9\u8c61\u521b\u5efa\u3001GIL\u3001\u52a8\u6001\u5f62\u72b6\u68c0\u67e5\u3002ORT \u662f\u7eaf C++ \u8fd0\u884c\u65f6，\u628a\u6574\u5f20\u56fe load \u8fdb\u53bb\u540e，\u4e00\u6b21\u8c03\u7528 sess.run() \u5c31\u80fd\u628a\u6574\u6761\u56fe\u8dd1\u5b8c，\u7b97\u5b50\u4e4b\u95f4\u7684\u6570\u636e\u5728\u5185\u5b58\u91cc\u76f4\u63a5\u4f20\u9012，\u4e0d\u9700\u8981\u6765\u56de\u8d8a\u8fc7 Python \u8fb9\u754c\u3002<\/p>\n

      \u5bf9\u5c0f\u7b97\u5b50\u3001\u5c0f\u6a21\u578b，Python \u7684\u8c03\u5ea6\u5f00\u9500\u5360\u6bd4\u5c24\u5176\u5927，\u6240\u4ee5\u52a0\u901f\u6bd4\u8d8a\u660e\u663e\u3002\u6a21\u578b\u8d85\u5927\u3001\u7b97\u5b50\u8d85\u5927\u65f6，Python \u8c03\u5ea6\u5360\u6bd4\u4e0b\u964d，\u52a0\u901f\u6bd4\u4f1a\u7f29\u6c34\u2014\u2014\u4f46"\u7701\u6389 Python \u5f00\u9500"\u8fd9\u70b9\u59cb\u7ec8\u6210\u7acb\u3002<\/p>\n

      \u539f\u56e0\u4e8c：\u56fe\u7ea7\u4f18\u5316（\u7b97\u5b50\u878d\u5408 + \u5e38\u91cf\u6298\u53e0）\u2014\u2014 \u5b9e\u6d4b 17.9 \u500d<\/h4>\n

      \u8fd9\u662f\u6700\u786c\u6838\u7684\u4e00\u70b9\u3002ORT \u5728\u52a0\u8f7d ONNX \u65f6\u4f1a\u505a\u4e00\u7cfb\u5217\u56fe\u53d8\u6362：<\/p>\n

        \n
      • \u7b97\u5b50\u878d\u5408：\u628a Conv + BatchNorm + ReLU \u7194\u6210\u4e00\u4e2a ConvRelu\u3001Conv+Add+Relu \u7194\u6210\u4e00\u4e2a FusedConv \u7b49，\u51cf\u5c11\u5185\u5b58\u8bfb\u5199\u6b21\u6570\u548c kernel \u542f\u52a8\u6b21\u6570\u3002<\/li>\n
      • \u5e38\u91cf\u6298\u53e0：\u56fe\u91cc\u80fd\u63d0\u524d\u7b97\u7684\u5e38\u91cf（\u5982\u67d0\u4e9b shape \u8ba1\u7b97）\u5728\u52a0\u8f7d\u65f6\u5c31\u7b97\u5b8c，\u63a8\u7406\u65f6\u4e0d\u518d\u7b97\u3002<\/li>\n
      • \u6b7b\u4ee3\u7801\u6d88\u9664 \/ \u5197\u4f59\u6d88\u9664：\u5220\u6389\u4e0d\u5f71\u54cd\u8f93\u51fa\u7684\u8282\u70b9\u3002<\/li>\n<\/ul>\n

        \u6211\u7528\u540c\u4e00\u4e2a ONNX \u6587\u4ef6，\u4e00\u53ea\u5f00\u56fe\u4f18\u5316\u3001\u4e00\u53ea\u5173\u6389，\u5b9e\u6d4b：<\/p>\n

        \u56fe\u4f18\u5316\u5f00\u542f (\u9ed8\u8ba4 ALL): 0.111 ms
        \n\u56fe\u4f18\u5316\u5173\u95ed (DISABLE) : 1.993 ms
        \n\u4f18\u5316\u5e26\u6765\u7684\u52a0\u901f : 17.91x<\/p>\n

        \u5149"\u56fe\u4f18\u5316"\u4e00\u9879\u5c31\u5e26\u6765 18 \u500d\u5dee\u8ddd\u3002 \u8fd9\u5145\u5206\u8bf4\u660e：ONNX\/ORT \u7684\u52a0\u901f\u5927\u5934，\u662f\u628a"\u4e00\u5f20\u6734\u7d20\u7b97\u5b50\u56fe"\u548c"\u878d\u5408\u540e\u7684\u9ad8\u6548\u7b97\u5b50\u56fe"\u7684\u533a\u522b，\u800c\u4e0d\u662f\u6587\u4ef6\u683c\u5f0f\u672c\u8eab\u6709\u591a\u7384\u3002<\/p>\n

        \u539f\u56e0\u4e09：\u9759\u6001\u56fe + \u56fa\u5b9a\u5f62\u72b6 \u2192 \u5185\u5b58\u89c4\u5212\u4e0e kernel \u9009\u62e9<\/h4>\n

        ONNX \u56fe\u7ed3\u6784\u56fa\u5b9a\u3001\u5f62\u72b6\u56fa\u5b9a（\u5bfc\u51fa\u65f6\u5b9a\u4e86 1\u00d73\u00d7112\u00d7112），ORT \u53ef\u4ee5\u5728\u4f1a\u8bdd\u521d\u59cb\u5316\u9636\u6bb5\u5c31\u5b8c\u6210\u5185\u5b58\u6c60\u89c4\u5212\u3001kernel \u9009\u62e9\u3001\u591a\u7ebf\u7a0b\u8c03\u5ea6\u7f16\u6392，\u63a8\u7406\u65f6\u96f6\u52a8\u6001\u51b3\u7b56\u3002\u800c PyTorch eager \u6bcf\u6b21\u524d\u5411\u90fd\u8981\u73b0\u573a\u51b3\u5b9a\u5f62\u72b6\u3001\u5206\u914d\u5185\u5b58\u3002<\/p>\n

        \u5c0f\u7ed3：\u4e3a\u4ec0\u4e48 onnx \u5feb<\/h4>\n\n\u52a0\u901f\u6765\u6e90\u539f\u7406\u5b9e\u6d4b\u5f71\u54cd<\/tr>\n\n\n\n\n
        \u65e0 Python \u89e3\u91ca\u5668\u5f00\u9500<\/td>\n\u4e00\u6b21 sess.run() \u8dd1\u5b8c\u6574\u56fe<\/td>\n\u660e\u663e<\/td>\n<\/tr>\n
        \u56fe\u7ea7\u4f18\u5316（\u7b97\u5b50\u878d\u5408\u7b49）<\/td>\nConv+BN+ReLU \u878d\u5408\u3001\u5e38\u91cf\u6298\u53e0\u3001\u6b7b\u7801\u6d88\u9664<\/td>\n\u5355\u72ec 17.9\u00d7<\/td>\n<\/tr>\n
        \u9759\u6001\u56fe + \u56fa\u5b9a\u5f62\u72b6<\/td>\n\u63d0\u524d\u89c4\u5212\u5185\u5b58\u4e0e kernel<\/td>\n\u660e\u663e<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

        \u516c\u5e73\u8d77\u89c1\u8865\u5145\u4e00\u53e5：PyTorch \u4e5f\u6709\u4e00\u5957 torch.compile \/ TorchScript \u9759\u6001\u5316\u65b9\u6848\u6765\u7f29\u5c0f\u5dee\u8ddd，\u4f46 ONNX \u751f\u6001\u7684\u8de8\u786c\u4ef6\u901a\u7528\u6027 + \u6210\u719f\u7684\u56fe\u4f18\u5316，\u8ba9\u5b83\u6210\u4e3a\u90e8\u7f72\u4e8b\u5b9e\u6807\u51c6\u3002<\/p>\n

        \n

        \u7b2c\u4e94\u7ae0\u3001\u5b8c\u6574\u53ef\u590d\u73b0\u811a\u672c<\/h3>\n

        conda activate ysj310-gpu
        \npip install<\/span> onnx onnxruntime onnxscript<\/p>\n

        python exp1_export.py # \u5bfc\u51fa .pt \u548c .onnx<\/span>
        \npython exp2_inspect_speed.py # \u89e3\u5256\u56fe + \u63a8\u7406\u901f\u5ea6\u5bf9\u6bd4<\/span>
        \npython exp3_opt.py # \u56fe\u4f18\u5316\u5f00\/\u5173\u5bf9\u6bd4<\/span>
        \npython exp4_magic.py # \u770b\u4e24\u4e2a\u6587\u4ef6\u7684\u9b54\u6570<\/span><\/p>\n

        exp2_inspect_speed.py \u4ee3\u7801：<\/p>\n

        import<\/span> onnx
        \nimport<\/span> time
        \nimport<\/span> torch
        \nimport<\/span> torch.<\/span>nn as<\/span> nn
        \nimport<\/span> numpy as<\/span> np
        \nimport<\/span> onnxruntime as<\/span> ort<\/p>\n

        # ———- 1. \u89e3\u5256 ONNX \u56fe ———-<\/span>
        \nm =<\/span> onnx.<\/span>load(<\/span>"pt_onnx_blog\/tinycnn.onnx"<\/span>)<\/span>
        \nprint<\/span>(<\/span>"=== ONNX \u56fe\u7ed3\u6784 ==="<\/span>)<\/span>
        \nprint<\/span>(<\/span>"\u8f93\u5165:"<\/span>,<\/span> [<\/span>(<\/span>i.<\/span>name,<\/span> [<\/span>d.<\/span>dim_value for<\/span> d in<\/span> i.<\/span>type<\/span>.<\/span>tensor_type.<\/span>shape.<\/span>dim]<\/span>)<\/span> for<\/span> i in<\/span> m.<\/span>graph.<\/span>input<\/span>]<\/span>)<\/span>
        \nprint<\/span>(<\/span>"\u8f93\u51fa:"<\/span>,<\/span> [<\/span>(<\/span>o.<\/span>name,<\/span> [<\/span>d.<\/span>dim_value for<\/span> d in<\/span> o.<\/span>type<\/span>.<\/span>tensor_type.<\/span>shape.<\/span>dim]<\/span>)<\/span> for<\/span> o in<\/span> m.<\/span>graph.<\/span>output]<\/span>)<\/span>
        \nprint<\/span>(<\/span>"\u7b97\u5b50\u8282\u70b9\u603b\u6570:"<\/span>,<\/span> len<\/span>(<\/span>m.<\/span>graph.<\/span>node)<\/span>)<\/span>
        \nfrom<\/span> collections import<\/span> Counter
        \nops =<\/span> Counter(<\/span>n.<\/span>op_type for<\/span> n in<\/span> m.<\/span>graph.<\/span>node)<\/span>
        \nprint<\/span>(<\/span>"\u7b97\u5b50\u7c7b\u578b\u5206\u5e03:"<\/span>,<\/span> dict<\/span>(<\/span>ops)<\/span>)<\/span>
        \n# \u6253\u5370\u524d 8 \u4e2a\u8282\u70b9\u5c55\u793a\u8ba1\u7b97\u56fe<\/span>
        \nprint<\/span>(<\/span>"\u524d8\u4e2a\u8282\u70b9:"<\/span>)<\/span>
        \nfor<\/span> n in<\/span> m.<\/span>graph.<\/span>node[<\/span>:<\/span>8<\/span>]<\/span>:<\/span>
        \n print<\/span>(<\/span>" "<\/span>,<\/span> n.<\/span>op_type,<\/span> "in:"<\/span>,<\/span> list<\/span>(<\/span>n.<\/span>input<\/span>)<\/span>,<\/span> "->"<\/span>,<\/span> list<\/span>(<\/span>n.<\/span>output)<\/span>)<\/span><\/p>\n

        # ———- 2. \u91cd\u5efa torch \u6a21\u578b，\u5bf9\u6bd4\u63a8\u7406\u901f\u5ea6 ———-<\/span>
        \nclass<\/span> TinyCNN<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
        \n def<\/span> __init__<\/span>(<\/span>self)<\/span>:<\/span>
        \n super<\/span>(<\/span>)<\/span>.<\/span>__init__(<\/span>)<\/span>
        \n self.<\/span>conv1 =<\/span> nn.<\/span>Conv2d(<\/span>3<\/span>,<\/span> 16<\/span>,<\/span> 3<\/span>,<\/span> padding=<\/span>1<\/span>)<\/span>
        \n self.<\/span>bn1 =<\/span> nn.<\/span>BatchNorm2d(<\/span>16<\/span>)<\/span>
        \n self.<\/span>relu =<\/span> nn.<\/span>ReLU(<\/span>)<\/span>
        \n self.<\/span>pool =<\/span> nn.<\/span>MaxPool2d(<\/span>2<\/span>)<\/span>
        \n self.<\/span>conv2 =<\/span> nn.<\/span>Conv2d(<\/span>16<\/span>,<\/span> 32<\/span>,<\/span> 3<\/span>,<\/span> padding=<\/span>1<\/span>)<\/span>
        \n self.<\/span>bn2 =<\/span> nn.<\/span>BatchNorm2d(<\/span>32<\/span>)<\/span>
        \n self.<\/span>fc =<\/span> nn.<\/span>Linear(<\/span>32<\/span> *<\/span> 28<\/span> *<\/span> 28<\/span>,<\/span> 10<\/span>)<\/span>
        \n def<\/span> forward<\/span>(<\/span>self,<\/span> x)<\/span>:<\/span>
        \n x =<\/span> self.<\/span>pool(<\/span>self.<\/span>relu(<\/span>self.<\/span>bn1(<\/span>self.<\/span>conv1(<\/span>x)<\/span>)<\/span>)<\/span>)<\/span>
        \n x =<\/span> self.<\/span>pool(<\/span>self.<\/span>relu(<\/span>self.<\/span>bn2(<\/span>self.<\/span>conv2(<\/span>x)<\/span>)<\/span>)<\/span>)<\/span>
        \n x =<\/span> x.<\/span>view(<\/span>x.<\/span>size(<\/span>0<\/span>)<\/span>,<\/span> –<\/span>1<\/span>)<\/span>
        \n return<\/span> self.<\/span>fc(<\/span>x)<\/span><\/p>\n

        torch.<\/span>manual_seed(<\/span>0<\/span>)<\/span>
        \nmodel =<\/span> TinyCNN(<\/span>)<\/span>.<\/span>eval<\/span>(<\/span>)<\/span>
        \n# \u4e3a\u4e86\u516c\u5e73，\u7ed9 torch \u6a21\u578b\u4e5f\u505a torch.compile \u4e4b\u5916\u7684\u57fa\u7840\u4f18\u5316：\u5173\u6389\u68af\u5ea6<\/span>
        \nmodel =<\/span> model.<\/span>float<\/span>(<\/span>)<\/span>
        \nx =<\/span> torch.<\/span>randn(<\/span>1<\/span>,<\/span> 3<\/span>,<\/span> 112<\/span>,<\/span> 112<\/span>)<\/span><\/p>\n

        # warmup<\/span>
        \nwith<\/span> torch.<\/span>no_grad(<\/span>)<\/span>:<\/span>
        \n for<\/span> _ in<\/span> range<\/span>(<\/span>10<\/span>)<\/span>:<\/span>
        \n model(<\/span>x)<\/span>
        \n torch.<\/span>cuda.<\/span>synchronize(<\/span>)<\/span> if<\/span> torch.<\/span>cuda.<\/span>is_available(<\/span>)<\/span> else<\/span> None<\/span><\/p>\n

        def<\/span> bench<\/span>(<\/span>fn,<\/span> n=<\/span>200<\/span>)<\/span>:<\/span>
        \n t0 =<\/span> time.<\/span>perf_counter(<\/span>)<\/span>
        \n for<\/span> _ in<\/span> range<\/span>(<\/span>n)<\/span>:<\/span>
        \n fn(<\/span>)<\/span>
        \n if<\/span> torch.<\/span>cuda.<\/span>is_available(<\/span>)<\/span>:<\/span>
        \n torch.<\/span>cuda.<\/span>synchronize(<\/span>)<\/span>
        \n return<\/span> (<\/span>time.<\/span>perf_counter(<\/span>)<\/span> –<\/span> t0)<\/span> \/<\/span> n *<\/span> 1000<\/span> # ms<\/span><\/p>\n

        with<\/span> torch.<\/span>no_grad(<\/span>)<\/span>:<\/span>
        \n torch_ms =<\/span> bench(<\/span>lambda<\/span>:<\/span> model(<\/span>x)<\/span>,<\/span> 200<\/span>)<\/span><\/p>\n

        # ———- onnxruntime CPU ———-<\/span>
        \nsess =<\/span> ort.<\/span>InferenceSession(<\/span>"pt_onnx_blog\/tinycnn.onnx"<\/span>,<\/span> providers=<\/span>[<\/span>"CPUExecutionProvider"<\/span>]<\/span>)<\/span>
        \ninputs =<\/span> {<\/span>sess.<\/span>get_inputs(<\/span>)<\/span>[<\/span>0<\/span>]<\/span>.<\/span>name:<\/span> x.<\/span>numpy(<\/span>)<\/span>}<\/span>
        \nfor<\/span> _ in<\/span> range<\/span>(<\/span>10<\/span>)<\/span>:<\/span>
        \n sess.<\/span>run(<\/span>None<\/span>,<\/span> inputs)<\/span>
        \nort_ms =<\/span> bench(<\/span>lambda<\/span>:<\/span> sess.<\/span>run(<\/span>None<\/span>,<\/span> inputs)<\/span>,<\/span> 200<\/span>)<\/span><\/p>\n

        print<\/span>(<\/span>f"\\\\n=== \u63a8\u7406\u901f\u5ea6\u5bf9\u6bd4 (\u5355\u5f20 112×112, CPU) ==="<\/span><\/span>)<\/span>
        \nprint<\/span>(<\/span>f"PyTorch eager \u5e73\u5747: <\/span>{<\/span>torch_ms:<\/span>.3f<\/span>}<\/span><\/span> ms"<\/span><\/span>)<\/span>
        \nprint<\/span>(<\/span>f"ONNX Runtime \u5e73\u5747: <\/span>{<\/span>ort_ms:<\/span>.3f<\/span>}<\/span><\/span> ms"<\/span><\/span>)<\/span>
        \nprint<\/span>(<\/span>f"\u52a0\u901f\u6bd4: <\/span>{<\/span>torch_ms\/<\/span>ort_ms:<\/span>.2f<\/span>}<\/span><\/span>x"<\/span><\/span>)<\/span><\/p>\n

        # ———- \u6570\u5b57\u4e00\u81f4\u6027 ———-<\/span>
        \nwith<\/span> torch.<\/span>no_grad(<\/span>)<\/span>:<\/span>
        \n torch_out =<\/span> model(<\/span>x)<\/span>
        \nort_out =<\/span> sess.<\/span>run(<\/span>None<\/span>,<\/span> inputs)<\/span>[<\/span>0<\/span>]<\/span>
        \nprint<\/span>(<\/span>f"\\\\n=== \u8f93\u51fa\u4e00\u81f4\u6027 ==="<\/span><\/span>)<\/span>
        \nprint<\/span>(<\/span>"torch \u8f93\u51fa shape:"<\/span>,<\/span> torch_out.<\/span>shape,<\/span> "| ort \u8f93\u51fa shape:"<\/span>,<\/span> ort_out.<\/span>shape)<\/span>
        \ndiff =<\/span> np.<\/span>abs<\/span>(<\/span>torch_out.<\/span>numpy(<\/span>)<\/span> –<\/span> ort_out)<\/span>
        \nprint<\/span>(<\/span>f"\u6700\u5927\u7edd\u5bf9\u8bef\u5dee: <\/span>{<\/span>diff.<\/span>max<\/span>(<\/span>)<\/span>:<\/span>.2e<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

        \n

        \u603b\u7ed3：\u4e00\u5f20\u8868\u8bfb\u61c2<\/h3>\n\n\u95ee\u9898\u7b54\u6848\u9a8c\u8bc1\u65b9\u5f0f<\/tr>\n\n\n\n\n\n\n
        .pt \u662f\u4ec0\u4e48<\/td>\nZIP \u5916\u58f3 + Pickle \u5185\u6838 + \u72ec\u7acb\u5f20\u91cf\u6587\u4ef6<\/td>\n\u8bfb\u9b54\u6570 PK\\\\x03\\\\x04，zipfile \u770b data.pkl<\/td>\n<\/tr>\n
        .onnx \u662f\u4ec0\u4e48<\/td>\nProtobuf \u5e8f\u5217\u5316\u7684\u8ba1\u7b97\u56fe（\u7b97\u5b50 + \u6743\u91cd）<\/td>\n\u8bfb\u6587\u4ef6\u5934 \\\\x08\\\\n，onnx \u5e93\u6253\u5370\u56fe<\/td>\n<\/tr>\n
        \u4e3a\u4ec0\u4e48\u80fd\u8f6c<\/td>\n\u6a21\u578b\u672c\u6765\u5c31\u6709\u8ba1\u7b97\u56fe，\u5bfc\u51fa=trace \u4e00\u904d + \u7ffb\u8bd1\u7b97\u5b50 + \u56fa\u5316\u6743\u91cd<\/td>\ntorch.onnx.export，\u56fe\u91cc BN \u88ab\u7194\u8fdb Conv<\/td>\n<\/tr>\n
        \u4e3a\u4ec0\u4e48\u9700\u8981\u4e2d\u95f4\u6001<\/td>\nONNX \u662f\u8de8\u6846\u67b6\u3001\u8de8\u786c\u4ef6\u7684"\u901a\u7528 IR"，\u5404\u5f15\u64ce\u518d\u8f6c\u79c1\u6709\u683c\u5f0f<\/td>\n\u753b\u51fa PT\u2192ONNX\u2192TRT\/RKNN \u94fe\u8def<\/td>\n<\/tr>\n
        \u4e3a\u4ec0\u4e48 ONNX \u66f4\u5feb<\/td>\n\u2460 \u65e0 Python \u5f00\u9500 \u2461 \u56fe\u7ea7\u7b97\u5b50\u878d\u5408 \u2462 \u9759\u6001\u56fe\u5185\u5b58\u89c4\u5212<\/td>\n\u5b9e\u6d4b 5.11\u00d7；\u5355\u56fe\u4f18\u5316 17.9\u00d7<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

        \u6700\u540e\u4e00\u53e5\u5927\u767d\u8bdd：.pt \u5b58\u7684\u662f"\u53c2\u6570\u548c\u7c7b"，.onnx \u5b58\u7684\u662f"\u4e00\u5f20\u9759\u6001\u7b97\u5b50\u56fe + \u53c2\u6570"\u3002\u80fd\u8f6c\u662f\u56e0\u4e3a\u6a21\u578b\u672c\u8d28\u5c31\u662f\u4e00\u5f20\u56fe；\u9700\u8981\u4e2d\u95f4\u6001\u662f\u56e0\u4e3a ONNX \u662f\u5404\u79cd\u786c\u4ef6\u5f15\u64ce\u90fd\u8ba4\u7684"\u901a\u7528\u8bed\u8a00"；\u53d8\u5feb\u4e0d\u662f\u56e0\u4e3a\u6587\u4ef6\u6362\u4e86\u4e2a\u683c\u5f0f，\u800c\u662f\u4ea4\u7ed9\u4e86 ONNX Runtime \u8fd9\u79cd\u7eaf C++ \u7684\u3001\u5e26\u56fe\u4f18\u5316\u7684\u9759\u6001\u63a8\u7406\u5f15\u64ce\u53bb\u6267\u884c\u3002<\/p>\n

        \u5f15\u7533\u5f69\u86cb：\u5982\u679c\u4f60\u60f3\u8ba9 ONNX \u518d\u5feb，\u901a\u5e38\u4f1a\u518d\u7528 onnxsim（onnx-simplifier）\u628a\u56fe\u518d\u64b8\u4e00\u904d（\u62d3\u6251\u6392\u5e8f\u3001\u5e38\u91cf\u6298\u53e0\u3001\u6d88\u9664\u5197\u4f59 reshape \u7b49），\u7136\u540e\u4ea4\u7ed9 TensorRT \u751f\u6210\u5f15\u64ce\u3002\u90a3\u53c8\u662f\u53e6\u4e00\u7bc7\u6545\u4e8b\u4e86\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

        \u52a8\u624b\u8dd1\u8fc7\u624d\u6562\u5199\u3002\u672c\u6587\u6240\u6709\u7ed3\u8bba\u90fd\u6765\u81ea\u6211\u5728\u81ea\u5df1\u73af\u5883\u91cc\u4eb2\u624b\u8dd1\u51fa\u6765\u7684\u771f\u5b9e\u8f93\u51fa，\u4e00\u4e2a\u63a8\u6f14\u3001\u4e00\u4e2a\u62cd\u8111\u888b\u7684\u7ed3\u8bba\u90fd\u6ca1\u6709\u3002\u73af\u5883：torch 2.10.0cu128 \/ onnx 1.21.0 \/ onnxruntime 1.26.0\u3002\u5f15\u8a00：\u4e09\u4e2a\u6bcf\u5929\u90fd\u60f3\u95ee\u7684\u95ee\u9898
        \n\u505a\u63a8\u7406\u90e8\u7f72\u7684\u4eba，\u51e0\u4e4e\u5929\u5929\u548c .pt\u3001.onnx \u6253\u4ea4\u9053\u3002\u4f46\u5927\u591a\u6570\u4eba\u662f\\”\u4f1a\u590d\u5236\u547d\u4ee4，\u4e0d\u61c2\u539f\u7406\\”：
        \n\u4e3a\u4ec0\u4e48 torch.onnx.export \u4e00\u4e0b&#xf<\/p>\n","protected":false},"author":2,"featured_media":91889,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[10187,9477,10489,152,2094,86],"topic":[],"class_list":["post-91890","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-codingplan","tag-onnx","tag-onnxruntime","tag-pytorch","tag-2094","tag-86"],"yoast_head":"\n\u4e3a\u4ec0\u4e48 .pt \u80fd\u8f6c onnx\uff1fonnx \u5230\u5e95\u662f\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u90e8\u7f72\u90fd\u8981\u5148\u8fc7\u4e00\u9053 ONNX\uff1f\u2014\u2014 \u4eb2\u624b\u8dd1\u4e00\u904d\uff0c\u5feb 5 \u500d\u7684\u539f\u56e0\u5168\u5728\u8fd9 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/91890.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u4e3a\u4ec0\u4e48 .pt \u80fd\u8f6c onnx\uff1fonnx \u5230\u5e95\u662f\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u90e8\u7f72\u90fd\u8981\u5148\u8fc7\u4e00\u9053 ONNX\uff1f\u2014\u2014 \u4eb2\u624b\u8dd1\u4e00\u904d\uff0c\u5feb 5 \u500d\u7684\u539f\u56e0\u5168\u5728\u8fd9 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u52a8\u624b\u8dd1\u8fc7\u624d\u6562\u5199\u3002\u672c\u6587\u6240\u6709\u7ed3\u8bba\u90fd\u6765\u81ea\u6211\u5728\u81ea\u5df1\u73af\u5883\u91cc\u4eb2\u624b\u8dd1\u51fa\u6765\u7684\u771f\u5b9e\u8f93\u51fa，\u4e00\u4e2a\u63a8\u6f14\u3001\u4e00\u4e2a\u62cd\u8111\u888b\u7684\u7ed3\u8bba\u90fd\u6ca1\u6709\u3002\u73af\u5883：torch 2.10.0cu128 \/ onnx 1.21.0 \/ onnxruntime 1.26.0\u3002\u5f15\u8a00：\u4e09\u4e2a\u6bcf\u5929\u90fd\u60f3\u95ee\u7684\u95ee\u9898 \u505a\u63a8\u7406\u90e8\u7f72\u7684\u4eba，\u51e0\u4e4e\u5929\u5929\u548c .pt\u3001.onnx \u6253\u4ea4\u9053\u3002\u4f46\u5927\u591a\u6570\u4eba\u662f"\u4f1a\u590d\u5236\u547d\u4ee4，\u4e0d\u61c2\u539f\u7406"： \u4e3a\u4ec0\u4e48 torch.onnx.export \u4e00\u4e0b&#xf\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/91890.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-08T09:24:50+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260808092448-6a76f5e07a500.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"6 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/91890.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/91890.html\",\"name\":\"\u4e3a\u4ec0\u4e48 .pt \u80fd\u8f6c onnx\uff1fonnx \u5230\u5e95\u662f\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u90e8\u7f72\u90fd\u8981\u5148\u8fc7\u4e00\u9053 ONNX\uff1f\u2014\u2014 \u4eb2\u624b\u8dd1\u4e00\u904d\uff0c\u5feb 5 \u500d\u7684\u539f\u56e0\u5168\u5728\u8fd9 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-08-08T09:24:50+00:00\",\"dateModified\":\"2026-08-08T09:24:50+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/91890.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/91890.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/91890.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4e3a\u4ec0\u4e48 .pt \u80fd\u8f6c onnx\uff1fonnx \u5230\u5e95\u662f\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u90e8\u7f72\u90fd\u8981\u5148\u8fc7\u4e00\u9053 ONNX\uff1f\u2014\u2014 \u4eb2\u624b\u8dd1\u4e00\u904d\uff0c\u5feb 5 \u500d\u7684\u539f\u56e0\u5168\u5728\u8fd9\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u4e3a\u4ec0\u4e48 .pt \u80fd\u8f6c onnx\uff1fonnx \u5230\u5e95\u662f\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u90e8\u7f72\u90fd\u8981\u5148\u8fc7\u4e00\u9053 ONNX\uff1f\u2014\u2014 \u4eb2\u624b\u8dd1\u4e00\u904d\uff0c\u5feb 5 \u500d\u7684\u539f\u56e0\u5168\u5728\u8fd9 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/91890.html","og_locale":"zh_CN","og_type":"article","og_title":"\u4e3a\u4ec0\u4e48 .pt \u80fd\u8f6c onnx\uff1fonnx \u5230\u5e95\u662f\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u90e8\u7f72\u90fd\u8981\u5148\u8fc7\u4e00\u9053 ONNX\uff1f\u2014\u2014 \u4eb2\u624b\u8dd1\u4e00\u904d\uff0c\u5feb 5 \u500d\u7684\u539f\u56e0\u5168\u5728\u8fd9 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u52a8\u624b\u8dd1\u8fc7\u624d\u6562\u5199\u3002\u672c\u6587\u6240\u6709\u7ed3\u8bba\u90fd\u6765\u81ea\u6211\u5728\u81ea\u5df1\u73af\u5883\u91cc\u4eb2\u624b\u8dd1\u51fa\u6765\u7684\u771f\u5b9e\u8f93\u51fa，\u4e00\u4e2a\u63a8\u6f14\u3001\u4e00\u4e2a\u62cd\u8111\u888b\u7684\u7ed3\u8bba\u90fd\u6ca1\u6709\u3002\u73af\u5883：torch 2.10.0cu128 \/ onnx 1.21.0 \/ onnxruntime 1.26.0\u3002\u5f15\u8a00：\u4e09\u4e2a\u6bcf\u5929\u90fd\u60f3\u95ee\u7684\u95ee\u9898 \u505a\u63a8\u7406\u90e8\u7f72\u7684\u4eba，\u51e0\u4e4e\u5929\u5929\u548c .pt\u3001.onnx \u6253\u4ea4\u9053\u3002\u4f46\u5927\u591a\u6570\u4eba\u662f\"\u4f1a\u590d\u5236\u547d\u4ee4，\u4e0d\u61c2\u539f\u7406\"： \u4e3a\u4ec0\u4e48 torch.onnx.export \u4e00\u4e0b&#xf","og_url":"https:\/\/www.wsisp.com\/helps\/91890.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-08-08T09:24:50+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260808092448-6a76f5e07a500.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"6 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/91890.html","url":"https:\/\/www.wsisp.com\/helps\/91890.html","name":"\u4e3a\u4ec0\u4e48 .pt \u80fd\u8f6c onnx\uff1fonnx \u5230\u5e95\u662f\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u90e8\u7f72\u90fd\u8981\u5148\u8fc7\u4e00\u9053 ONNX\uff1f\u2014\u2014 \u4eb2\u624b\u8dd1\u4e00\u904d\uff0c\u5feb 5 \u500d\u7684\u539f\u56e0\u5168\u5728\u8fd9 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-08-08T09:24:50+00:00","dateModified":"2026-08-08T09:24:50+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/91890.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/91890.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/91890.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u4e3a\u4ec0\u4e48 .pt \u80fd\u8f6c onnx\uff1fonnx \u5230\u5e95\u662f\u4ec0\u4e48\uff1f\u4e3a\u4ec0\u4e48\u90e8\u7f72\u90fd\u8981\u5148\u8fc7\u4e00\u9053 ONNX\uff1f\u2014\u2014 \u4eb2\u624b\u8dd1\u4e00\u904d\uff0c\u5feb 5 \u500d\u7684\u539f\u56e0\u5168\u5728\u8fd9"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/91890","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=91890"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/91890\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/91889"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=91890"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=91890"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=91890"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=91890"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}