{"id":12640,"date":"2023-02-01T21:05:21","date_gmt":"2023-02-01T13:05:21","guid":{"rendered":"http:\/\/139.9.1.231\/?p=12640"},"modified":"2023-02-03T09:36:56","modified_gmt":"2023-02-03T01:36:56","slug":"pytorch-onnx","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2023\/02\/01\/pytorch-onnx\/","title":{"rendered":"PyTorch \u8f6c ONNX \u8be6\u89e3"},"content":{"rendered":"\n

\u8f6c\u81ea\uff1a\u6a21\u578b\u90e8\u7f72\u5165\u95e8\u6559\u7a0b\uff08\u4e09\uff09\uff1aPyTorch \u8f6c ONNX \u8be6\u89e3<\/a><\/p>\n\n\n\n

\u6a21\u578b\u8f6c\u6362\u5de5\u5177<\/strong>\uff1a https:\/\/convertmodel.com\/<\/a><\/p>\n\n\n\n

ONNX \u662f\u76ee\u524d\u6a21\u578b\u90e8\u7f72\u4e2d\u6700\u91cd\u8981\u7684\u4e2d\u95f4\u8868\u793a\u4e4b\u4e00\u3002\u5b66\u61c2\u4e86 ONNX \u7684\u6280\u672f\u7ec6\u8282\uff0c\u5c31\u80fd\u89c4\u907f\u5927\u91cf\u7684\u6a21\u578b\u90e8\u7f72\u95ee\u9898\u3002
\u5728\u628a PyTorch \u6a21\u578b\u8f6c\u6362\u6210 ONNX \u6a21\u578b\u65f6\uff0c\u6211\u4eec\u5f80\u5f80\u53ea\u9700\u8981\u8f7b\u677e\u5730\u8c03\u7528\u4e00\u53e5torch.onnx.export<\/code>\u5c31\u884c\u4e86\u3002\u8fd9\u4e2a\u51fd\u6570\u7684\u63a5\u53e3\u770b\u4e0a\u53bb\u7b80\u5355\uff0c\u4f46\u5b83\u5728\u4f7f\u7528\u4e0a\u8fd8\u6709\u7740\u8bf8\u591a\u7684\u201c\u6f5c\u89c4\u5219\u201d\u3002\u5728\u8fd9\u7bc7\u6559\u7a0b\u4e2d\uff0c\u6211\u4eec\u4f1a\u8be6\u7ec6\u4ecb\u7ecd PyTorch \u6a21\u578b\u8f6c ONNX \u6a21\u578b\u7684\u539f\u7406\u53ca\u6ce8\u610f\u4e8b\u9879\u3002\u9664\u6b64\u4e4b\u5916\uff0c\u6211\u4eec\u8fd8\u4f1a\u4ecb\u7ecd PyTorch \u4e0e ONNX \u7684\u7b97\u5b50\u5bf9\u5e94\u5173\u7cfb\uff0c\u4ee5\u6559\u4f1a\u5927\u5bb6\u5982\u4f55\u5904\u7406 PyTorch \u6a21\u578b\u8f6c\u6362\u65f6\u53ef\u80fd\u4f1a\u9047\u5230\u7684\u7b97\u5b50\u652f\u6301\u95ee\u9898\u3002<\/p>\n\n\n\n

torch.onnx.export<\/code> \u7ec6\u89e3<\/h2>\n\n\n\n


\u5728\u8fd9\u4e00\u8282\u91cc\uff0c\u6211\u4eec\u5c06\u8be6\u7ec6\u4ecb\u7ecd PyTorch \u5230 ONNX \u7684\u8f6c\u6362\u51fd\u6570\u2014\u2014 torch.onnx.export<\/code>\u3002\u6211\u4eec\u5e0c\u671b\u5927\u5bb6\u80fd\u591f\u66f4\u52a0\u7075\u6d3b\u5730\u4f7f\u7528\u8fd9\u4e2a\u6a21\u578b\u8f6c\u6362\u63a5\u53e3\uff0c\u5e76\u901a\u8fc7\u4e86\u89e3\u5b83\u7684\u5b9e\u73b0\u539f\u7406\u6765\u66f4\u597d\u5730\u5e94\u5bf9\u8be5\u51fd\u6570\u7684\u62a5\u9519\uff08\u7531\u4e8e\u6a21\u578b\u90e8\u7f72\u7684\u517c\u5bb9\u6027\u95ee\u9898\uff0c\u90e8\u7f72\u590d\u6742\u6a21\u578b\u65f6\u8be5\u51fd\u6570\u65f6\u5e38\u4f1a\u62a5\u9519\uff09\u3002<\/p>\n\n\n\n


\u8ba1\u7b97\u56fe\u5bfc\u51fa\u65b9\u6cd5<\/h3>\n\n\n\n

TorchScript<\/a> \u662f\u4e00\u79cd\u5e8f\u5217\u5316\u548c\u4f18\u5316 PyTorch \u6a21\u578b\u7684\u683c\u5f0f\uff0c\u5728\u4f18\u5316\u8fc7\u7a0b\u4e2d\uff0c\u4e00\u4e2atorch.nn.Module<\/code>\u6a21\u578b\u4f1a\u88ab\u8f6c\u6362\u6210 TorchScript \u7684 torch.jit.ScriptModule<\/code>\u6a21\u578b\u3002\u73b0\u5728\uff0c TorchScript \u4e5f\u88ab\u5e38\u5f53\u6210\u4e00\u79cd\u4e2d\u95f4\u8868\u793a\u4f7f\u7528\u3002\u6211\u4eec\u5728\u5176\u4ed6\u6587\u7ae0\u4e2d\u5bf9 TorchScript \u6709\u8be6\u7ec6\u7684\u4ecb\u7ecd\uff08https:\/\/zhuanlan.zhihu.com\/p\/486914187<\/a>\uff09\uff0c\u8fd9\u91cc\u4ecb\u7ecd TorchScript \u4ec5\u7528\u4e8e\u8bf4\u660e PyTorch \u6a21\u578b\u8f6c ONNX\u7684\u539f\u7406\u3002
torch.onnx.export<\/code>\u4e2d\u9700\u8981\u7684\u6a21\u578b\u5b9e\u9645\u4e0a\u662f\u4e00\u4e2atorch.jit.ScriptModule<\/code>\u3002\u800c\u8981\u628a\u666e\u901a PyTorch \u6a21\u578b\u8f6c\u4e00\u4e2a\u8fd9\u6837\u7684 TorchScript \u6a21\u578b\uff0c\u6709\u8ddf\u8e2a\uff08trace\uff09\u548c\u8bb0\u5f55\uff08script\uff09\u4e24\u79cd\u5bfc\u51fa\u8ba1\u7b97\u56fe\u7684\u65b9\u6cd5\u3002\u5982\u679c\u7ed9torch.onnx.export<\/code>\u4f20\u5165\u4e86\u4e00\u4e2a\u666e\u901a PyTorch \u6a21\u578b\uff08torch.nn.Module<\/code>)\uff0c\u90a3\u4e48\u8fd9\u4e2a\u6a21\u578b\u4f1a\u9ed8\u8ba4\u4f7f\u7528\u8ddf\u8e2a\u7684\u65b9\u6cd5\u5bfc\u51fa\u3002\u8fd9\u4e00\u8fc7\u7a0b\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n

\"\"\/<\/figure>\n\n\n\n

\u56de\u5fc6\u4e00\u4e0b\u6211\u4eec\u7b2c\u4e00\u7bc7\u6559\u7a0b<\/a>\u77e5\u8bc6\uff1a\u8ddf\u8e2a\u6cd5\u53ea\u80fd\u901a\u8fc7\u5b9e\u9645\u8fd0\u884c\u4e00\u904d\u6a21\u578b\u7684\u65b9\u6cd5\u5bfc\u51fa\u6a21\u578b\u7684\u9759\u6001\u56fe\uff0c\u5373\u65e0\u6cd5\u8bc6\u522b\u51fa\u6a21\u578b\u4e2d\u7684\u63a7\u5236\u6d41\uff08\u5982\u5faa\u73af\uff09\uff1b\u8bb0\u5f55\u6cd5\u5219\u80fd\u901a\u8fc7\u89e3\u6790\u6a21\u578b\u6765\u6b63\u786e\u8bb0\u5f55\u6240\u6709\u7684\u63a7\u5236\u6d41\u3002\u6211\u4eec\u4ee5\u4e0b\u9762\u8fd9\u6bb5\u4ee3\u7801\u4e3a\u4f8b\u6765\u770b\u4e00\u770b\u8fd9\u4e24\u79cd\u8f6c\u6362\u65b9\u6cd5\u7684\u533a\u522b\uff1a<\/p>\n\n\n\n

import torch \n \nclass Model(torch.nn.Module): \n    def __init__(self, n): \n        super().__init__() \n        self.n = n \n        self.conv = torch.nn.Conv2d(3, 3, 3) \n \n    def forward(self, x): \n        for i in range(self.n): \n            x = self.conv(x) \n        return x \n \n \nmodels = [Model(2), Model(3)] \nmodel_names = ['model_2', 'model_3'] \n \nfor model, model_name in zip(models, model_names): \n    dummy_input = torch.rand(1, 3, 10, 10) \n    dummy_output = model(dummy_input) \n    model_trace = torch.jit.trace(model, dummy_input) \n    model_script = torch.jit.script(model) \n \n    # \u8ddf\u8e2a\u6cd5\u4e0e\u76f4\u63a5 torch.onnx.export(model, ...)\u7b49\u4ef7 <\/em>\n    torch.onnx.export(model_trace, dummy_input, f'{model_name}_trace.onnx', example_outputs=dummy_output) \n    # \u8bb0\u5f55\u6cd5\u5fc5\u987b\u5148\u8c03\u7528 torch.jit.sciprt <\/em>\n    torch.onnx.export(model_script, dummy_input, f'{model_name}_script.onnx', example_outputs=dummy_output) <\/code><\/pre>\n\n\n\n
\u5728\u8fd9\u6bb5\u4ee3\u7801\u91cc\uff0c\u6211\u4eec\u5b9a\u4e49\u4e86\u4e00\u4e2a\u5e26\u5faa\u73af\u7684\u6a21\u578b\uff0c\u6a21\u578b\u901a\u8fc7\u53c2\u6570n<\/code>\u6765\u63a7\u5236\u8f93\u5165\u5f20\u91cf\u88ab\u5377\u79ef\u7684\u6b21\u6570\u3002\u4e4b\u540e\uff0c\u6211\u4eec\u5404\u521b\u5efa\u4e86\u4e00\u4e2an=2<\/code>\u548cn=3<\/code>\u7684\u6a21\u578b\u3002\u6211\u4eec\u628a\u8fd9\u4e24\u4e2a\u6a21\u578b\u5206\u522b\u7528\u8ddf\u8e2a\u548c\u8bb0\u5f55\u7684\u65b9\u6cd5\u8fdb\u884c\u5bfc\u51fa\u3002
\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u7531\u4e8e\u8fd9\u91cc\u7684\u4e24\u4e2a\u6a21\u578b\uff08model_trace<\/code>, model_script<\/code>)\u662f TorchScript \u6a21\u578b\uff0cexport<\/code>\u51fd\u6570\u5df2\u7ecf\u4e0d\u9700\u8981\u518d\u8fd0\u884c\u4e00\u904d\u6a21\u578b\u4e86\u3002\uff08\u5982\u679c\u6a21\u578b\u662f\u7528\u8ddf\u8e2a\u6cd5\u5f97\u5230\u7684\uff0c\u90a3\u4e48\u5728\u6267\u884ctorch.jit.trace<\/code>\u7684\u65f6\u5019\u5c31\u8fd0\u884c\u8fc7\u4e00\u904d\u4e86\uff1b\u800c\u7528\u8bb0\u5f55\u6cd5\u5bfc\u51fa\u65f6\uff0c\u6a21\u578b\u4e0d\u9700\u8981\u5b9e\u9645\u8fd0\u884c\uff09\u53c2\u6570\u4e2d\u7684dummy_input<\/code>\u548cdummy_output<\/code>`\u4ec5\u4ec5\u662f\u4e3a\u4e86\u83b7\u53d6\u8f93\u5165\u548c\u8f93\u51fa\u5f20\u91cf\u7684\u7c7b\u578b\u548c\u5f62\u72b6\u3002
\u8fd0\u884c\u4e0a\u9762\u7684\u4ee3\u7801\uff0c\u6211\u4eec\u628a\u5f97\u5230\u7684 4 \u4e2a onnx \u6587\u4ef6\u7528 Netron \u53ef\u89c6\u5316\uff1a<\/p>\n\n\n\n
\"\"<\/figure><\/div>\n\n\n\n
\u9996\u5148\u770b\u8ddf\u8e2a\u6cd5\u5f97\u5230\u7684 ONNX \u6a21\u578b\u7ed3\u6784\u3002\u53ef\u4ee5\u770b\u51fa\u6765\uff0c\u5bf9\u4e8e\u4e0d\u540c\u7684 n<\/code>,ONNX \u6a21\u578b\u7684\u7ed3\u6784\u662f\u4e0d\u4e00\u6837\u7684\u3002<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
\u800c\u7528\u8bb0\u5f55\u6cd5\u7684\u8bdd\uff0c\u6700\u7ec8\u7684 ONNX \u6a21\u578b\u7528 Loop<\/code> \u8282\u70b9\u6765\u8868\u793a\u5faa\u73af\u3002\u8fd9\u6837\u54ea\u6015\u5bf9\u4e8e\u4e0d\u540c\u7684 n<\/code>\uff0cONNX \u6a21\u578b\u4e5f\u6709\u540c\u6837\u7684\u7ed3\u6784\u3002<\/p>\n\n\n\n
\u672c\u6587\u4f7f\u7528\u7684 PyTorch \u7248\u672c\u662f 1.8.2\u3002\u636e\u53cd\u9988\uff0c\u5176\u4ed6\u7248\u672c\u7684 PyTorch \u53ef\u80fd\u4f1a\u5f97\u5230\u4e0d\u4e00\u6837\u7684\u7ed3\u679c\u3002<\/p><\/blockquote>\n\n\n\n
\u7531\u4e8e\u63a8\u7406\u5f15\u64ce\u5bf9\u9759\u6001\u56fe\u7684\u652f\u6301\u66f4\u597d\uff0c\u901a\u5e38\u6211\u4eec\u5728\u6a21\u578b\u90e8\u7f72\u65f6\u4e0d\u9700\u8981\u663e\u5f0f\u5730\u628a PyTorch \u6a21\u578b\u8f6c\u6210 TorchScript \u6a21\u578b\uff0c\u76f4\u63a5\u628a PyTorch \u6a21\u578b\u7528 torch.onnx.export<\/code> \u8ddf\u8e2a\u5bfc\u51fa\u5373\u53ef\u3002\u4e86\u89e3\u8fd9\u90e8\u5206\u7684\u77e5\u8bc6\u4e3b\u8981\u662f\u4e3a\u4e86\u5728\u6a21\u578b\u8f6c\u6362\u62a5\u9519\u65f6\u80fd\u591f\u66f4\u597d\u5730\u5b9a\u4f4d\u95ee\u9898\u662f\u5426\u53d1\u751f\u5728 PyTorch \u8f6c TorchScript \u9636\u6bb5\u3002<\/p>\n\n\n\n
\u53c2\u6570\u8bb2\u89e3<\/h3>\n\n\n\n
\u4e86\u89e3\u5b8c\u8f6c\u6362\u51fd\u6570\u7684\u539f\u7406\u540e\uff0c\u6211\u4eec\u6765\u8be6\u7ec6\u4ecb\u7ecd\u4e00\u4e0b\u8be5\u51fd\u6570\u7684\u4e3b\u8981\u53c2\u6570\u7684\u4f5c\u7528\u3002\u6211\u4eec\u4e3b\u8981\u4f1a\u4ece\u5e94\u7528\u7684\u89d2\u5ea6\u6765\u4ecb\u7ecd\u6bcf\u4e2a\u53c2\u6570\u5728\u4e0d\u540c\u7684\u6a21\u578b\u90e8\u7f72\u573a\u666f\u4e2d\u5e94\u8be5\u5982\u4f55\u8bbe\u7f6e\uff0c\u800c\u4e0d\u4f1a\u53bb\u5217\u51fa\u6bcf\u4e2a\u53c2\u6570\u7684\u6240\u6709\u8bbe\u7f6e\u65b9\u6cd5\u3002\u8be5\u51fd\u6570\u8be6\u7ec6\u7684 API \u6587\u6863\u53ef\u53c2\u8003\uff1a torch.onnx \u2012 PyTorch 1.11.0 documentation<\/a>
torch.onnx.export<\/code> \u5728 torch.onnx.__init__.py<\/code>\u6587\u4ef6\u4e2d\u7684\u5b9a\u4e49\u5982\u4e0b\uff1a<\/p>\n\n\n\n
def export(model, args, f, export_params=True, verbose=False, training=TrainingMode.EVAL, \n           input_names=None, output_names=None, aten=False, export_raw_ir=False, \n           operator_export_type=None, opset_version=None, _retain_param_name=True, \n           do_constant_folding=True, example_outputs=None, strip_doc_string=True, \n           dynamic_axes=None, keep_initializers_as_inputs=None, custom_opsets=None, \n           enable_onnx_checker=True, use_external_data_format=False): <\/code><\/pre>\n\n\n\n
\u524d\u4e09\u4e2a\u5fc5\u9009\u53c2\u6570\u4e3a\u6a21\u578b\u3001\u6a21\u578b\u8f93\u5165\u3001\u5bfc\u51fa\u7684 onnx \u6587\u4ef6\u540d\uff0c\u6211\u4eec\u5bf9\u8fd9\u51e0\u4e2a\u53c2\u6570\u5df2\u7ecf\u5f88\u719f\u6089\u4e86\u3002\u6211\u4eec\u6765\u7740\u91cd\u770b\u4e00\u4e0b\u540e\u9762\u7684\u4e00\u4e9b\u5e38\u7528\u53ef\u9009\u53c2\u6570\u3002<\/p>\n\n\n\n
export_params<\/strong><\/p>\n\n\n\n
\u6a21\u578b\u4e2d\u662f\u5426\u5b58\u50a8\u6a21\u578b\u6743\u91cd\u3002\u4e00\u822c\u4e2d\u95f4\u8868\u793a\u5305\u542b\u4e24\u5927\u7c7b\u4fe1\u606f\uff1a\u6a21\u578b\u7ed3\u6784\u548c\u6a21\u578b\u6743\u91cd\uff0c\u8fd9\u4e24\u7c7b\u4fe1\u606f\u53ef\u4ee5\u5728\u540c\u4e00\u4e2a\u6587\u4ef6\u91cc\u5b58\u50a8\uff0c\u4e5f\u53ef\u4ee5\u5206\u6587\u4ef6\u5b58\u50a8\u3002ONNX \u662f\u7528\u540c\u4e00\u4e2a\u6587\u4ef6\u8868\u793a\u8bb0\u5f55\u6a21\u578b\u7684\u7ed3\u6784\u548c\u6743\u91cd\u7684\u3002
\u6211\u4eec\u90e8\u7f72\u65f6\u4e00\u822c\u90fd\u9ed8\u8ba4\u8fd9\u4e2a\u53c2\u6570\u4e3a True\u3002\u5982\u679c onnx \u6587\u4ef6\u662f\u7528\u6765\u5728\u4e0d\u540c\u6846\u67b6\u95f4\u4f20\u9012\u6a21\u578b\uff08\u6bd4\u5982 PyTorch \u5230 Tensorflow\uff09\u800c\u4e0d\u662f\u7528\u4e8e\u90e8\u7f72\uff0c\u5219\u53ef\u4ee5\u4ee4\u8fd9\u4e2a\u53c2\u6570\u4e3a False\u3002<\/p>\n\n\n\n
input_names, output_names<\/strong><\/p>\n\n\n\n
\u8bbe\u7f6e\u8f93\u5165\u548c\u8f93\u51fa\u5f20\u91cf\u7684\u540d\u79f0\u3002\u5982\u679c\u4e0d\u8bbe\u7f6e\u7684\u8bdd\uff0c\u4f1a\u81ea\u52a8\u5206\u914d\u4e00\u4e9b\u7b80\u5355\u7684\u540d\u5b57\uff08\u5982\u6570\u5b57\uff09\u3002
ONNX \u6a21\u578b\u7684\u6bcf\u4e2a\u8f93\u5165\u548c\u8f93\u51fa\u5f20\u91cf\u90fd\u6709\u4e00\u4e2a\u540d\u5b57\u3002\u5f88\u591a\u63a8\u7406\u5f15\u64ce\u5728\u8fd0\u884c ONNX \u6587\u4ef6\u65f6\uff0c\u90fd\u9700\u8981\u4ee5\u201c\u540d\u79f0-\u5f20\u91cf\u503c\u201d\u7684\u6570\u636e\u5bf9\u6765\u8f93\u5165\u6570\u636e\uff0c\u5e76\u6839\u636e\u8f93\u51fa\u5f20\u91cf\u7684\u540d\u79f0\u6765\u83b7\u53d6\u8f93\u51fa\u6570\u636e\u3002\u5728\u8fdb\u884c\u8ddf\u5f20\u91cf\u6709\u5173\u7684\u8bbe\u7f6e\uff08\u6bd4\u5982\u6dfb\u52a0\u52a8\u6001\u7ef4\u5ea6\uff09\u65f6\uff0c\u4e5f\u9700\u8981\u77e5\u9053\u5f20\u91cf\u7684\u540d\u5b57\u3002
\u5728\u5b9e\u9645\u7684\u90e8\u7f72\u6d41\u6c34\u7ebf\u4e2d\uff0c\u6211\u4eec\u90fd\u9700\u8981\u8bbe\u7f6e\u8f93\u5165\u548c\u8f93\u51fa\u5f20\u91cf\u7684\u540d\u79f0\uff0c\u5e76\u4fdd\u8bc1 ONNX \u548c\u63a8\u7406\u5f15\u64ce\u4e2d\u4f7f\u7528\u540c\u4e00\u5957\u540d\u79f0\u3002<\/p>\n\n\n\n
opset_version<\/strong><\/p>\n\n\n\n
\u8f6c\u6362\u65f6\u53c2\u8003\u54ea\u4e2a ONNX \u7b97\u5b50\u96c6\u7248\u672c\uff0c\u9ed8\u8ba4\u4e3a 9\u3002\u540e\u6587\u4f1a\u8be6\u7ec6\u4ecb\u7ecd PyTorch \u4e0e ONNX \u7684\u7b97\u5b50\u5bf9\u5e94\u5173\u7cfb\u3002<\/p>\n\n\n\n
dynamic_axes<\/strong><\/p>\n\n\n\n
\u6307\u5b9a\u8f93\u5165\u8f93\u51fa\u5f20\u91cf\u7684\u54ea\u4e9b\u7ef4\u5ea6\u662f\u52a8\u6001\u7684\u3002
\u4e3a\u4e86\u8ffd\u6c42\u6548\u7387\uff0cONNX \u9ed8\u8ba4\u6240\u6709\u53c2\u4e0e\u8fd0\u7b97\u7684\u5f20\u91cf\u90fd\u662f\u9759\u6001\u7684\uff08\u5f20\u91cf\u7684\u5f62\u72b6\u4e0d\u53d1\u751f\u6539\u53d8\uff09\u3002\u4f46\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u6211\u4eec\u53c8\u5e0c\u671b\u6a21\u578b\u7684\u8f93\u5165\u5f20\u91cf\u662f\u52a8\u6001\u7684\uff0c\u5c24\u5176\u662f\u672c\u6765\u5c31\u6ca1\u6709\u5f62\u72b6\u9650\u5236\u7684\u5168\u5377\u79ef\u6a21\u578b\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u9700\u8981\u663e\u5f0f\u5730\u6307\u660e\u8f93\u5165\u8f93\u51fa\u5f20\u91cf\u7684\u54ea\u51e0\u4e2a\u7ef4\u5ea6\u7684\u5927\u5c0f\u662f\u53ef\u53d8\u7684\u3002
\u6211\u4eec\u6765\u770b\u4e00\u4e2adynamic_axes<\/code>\u7684\u8bbe\u7f6e\u4f8b\u5b50\uff1a<\/p>\n\n\n\n
import torch \n \nclass Model(torch.nn.Module): \n    def __init__(self): \n        super().__init__() \n        self.conv = torch.nn.Conv2d(3, 3, 3) \n \n    def forward(self, x): \n        x = self.conv(x) \n        return x \n \n \nmodel = Model() \ndummy_input = torch.rand(1, 3, 10, 10) \nmodel_names = ['model_static.onnx',  \n'model_dynamic_0.onnx',  \n'model_dynamic_23.onnx'] \n \ndynamic_axes_0 = { \n    'in' : [0], \n    'out' : [0] \n} \ndynamic_axes_23 = { \n    'in' : [2, 3], \n    'out' : [2, 3] \n} \n \ntorch.onnx.export(model, dummy_input, model_names[0],  \ninput_names=['in'], output_names=['out']) \ntorch.onnx.export(model, dummy_input, model_names[1],  \ninput_names=['in'], output_names=['out'], dynamic_axes=dynamic_axes_0) \ntorch.onnx.export(model, dummy_input, model_names[2],  \ninput_names=['in'], output_names=['out'], dynamic_axes=dynamic_axes_23) <\/code><\/pre>\n\n\n\n
\u9996\u5148\uff0c\u6211\u4eec\u5bfc\u51fa 3 \u4e2a ONNX \u6a21\u578b\uff0c\u5206\u522b\u4e3a\u6ca1\u6709\u52a8\u6001\u7ef4\u5ea6\u3001\u7b2c 0 \u7ef4\u52a8\u6001\u3001\u7b2c 2 \u7b2c 3 \u7ef4\u52a8\u6001\u7684\u6a21\u578b\u3002
\u5728\u8fd9\u4efd\u4ee3\u7801\u91cc\uff0c\u6211\u4eec\u662f\u7528\u5217\u8868\u7684\u65b9\u5f0f\u8868\u793a\u52a8\u6001\u7ef4\u5ea6\uff0c\u4f8b\u5982\uff1a<\/p>\n\n\n\n
dynamic_axes_0 = { \n    'in' : [0], \n    'out' : [0] \n} <\/code><\/pre>\n\n\n\n

\u7531\u4e8e ONNX \u8981\u6c42\u6bcf\u4e2a\u52a8\u6001\u7ef4\u5ea6\u90fd\u6709\u4e00\u4e2a\u540d\u5b57\uff0c\u8fd9\u6837\u5199\u7684\u8bdd\u4f1a\u5f15\u51fa\u4e00\u6761 UserWarning\uff0c\u8b66\u544a\u6211\u4eec\u901a\u8fc7\u5217\u8868\u7684\u65b9\u5f0f\u8bbe\u7f6e\u52a8\u6001\u7ef4\u5ea6\u7684\u8bdd\u7cfb\u7edf\u4f1a\u81ea\u52a8\u4e3a\u5b83\u4eec\u5206\u914d\u540d\u5b57\u3002\u4e00\u79cd\u663e\u5f0f\u6dfb\u52a0\u52a8\u6001\u7ef4\u5ea6\u540d\u5b57\u7684\u65b9\u6cd5\u5982\u4e0b\uff1a<\/p>\n\n\n\n
dynamic_axes_0 = { \n    'in' : {0: 'batch'}, \n    'out' : {0: 'batch'} \n} <\/code><\/pre>\n\n\n\n
\u7531\u4e8e\u5728\u8fd9\u4efd\u4ee3\u7801\u91cc\u6211\u4eec\u6ca1\u6709\u66f4\u591a\u7684\u5bf9\u52a8\u6001\u7ef4\u5ea6\u7684\u64cd\u4f5c\uff0c\u56e0\u6b64\u7b80\u5355\u5730\u7528\u5217\u8868\u6307\u5b9a\u52a8\u6001\u7ef4\u5ea6\u5373\u53ef\u3002
\u4e4b\u540e\uff0c\u6211\u4eec\u7528\u4e0b\u9762\u7684\u4ee3\u7801\u6765\u770b\u4e00\u770b\u52a8\u6001\u7ef4\u5ea6\u7684\u4f5c\u7528\uff1a<\/p>\n\n\n\n
import onnxruntime \nimport numpy as np \n \norigin_tensor = np.random.rand(1, 3, 10, 10).astype(np.float32) \nmult_batch_tensor = np.random.rand(2, 3, 10, 10).astype(np.float32) \nbig_tensor = np.random.rand(1, 3, 20, 20).astype(np.float32) \n \ninputs = [origin_tensor, mult_batch_tensor, big_tensor] \nexceptions = dict() \n \nfor model_name in model_names: \n    for i, input in enumerate(inputs): \n        try: \n            ort_session = onnxruntime.InferenceSession(model_name) \n            ort_inputs = {'in': input} \n            ort_session.run(['out'], ort_inputs) \n        except Exception as e: \n            exceptions[(i, model_name)] = e \n            print(f'Input[{i}] on model {model_name} error.') \n        else: \n            print(f'Input[{i}] on model {model_name} succeed.') <\/code><\/pre>\n\n\n\n
\u6211\u4eec\u5728\u6a21\u578b\u5bfc\u51fa\u8ba1\u7b97\u56fe\u65f6\u7528\u7684\u662f\u4e00\u4e2a\u5f62\u72b6\u4e3a(1, 3, 10, 10)<\/code>\u7684\u5f20\u91cf\u3002\u73b0\u5728\uff0c\u6211\u4eec\u6765\u5c1d\u8bd5\u4ee5\u5f62\u72b6\u5206\u522b\u662f(1, 3, 10, 10), (2, 3, 10, 10), (1, 3, 20, 20)<\/code>\u4e3a\u8f93\u5165\uff0c\u7528ONNX Runtime\u8fd0\u884c\u4e00\u4e0b\u8fd9\u51e0\u4e2a\u6a21\u578b\uff0c\u770b\u770b\u54ea\u4e9b\u60c5\u51b5\u4e0b\u4f1a\u62a5\u9519\uff0c\u5e76\u4fdd\u5b58\u5bf9\u5e94\u7684\u62a5\u9519\u4fe1\u606f\u3002\u5f97\u5230\u7684\u8f93\u51fa\u4fe1\u606f\u5e94\u8be5\u5982\u4e0b\uff1a<\/p>\n\n\n\n
Input[0] on model model_static.onnx succeed. \nInput[1] on model model_static.onnx error. \nInput[2] on model model_static.onnx error. \nInput[0] on model model_dynamic_0.onnx succeed. \nInput[1] on model model_dynamic_0.onnx succeed. \nInput[2] on model model_dynamic_0.onnx error. \nInput[0] on model model_dynamic_23.onnx succeed. \nInput[1] on model model_dynamic_23.onnx error. \nInput[2] on model model_dynamic_23.onnx succeed. <\/code><\/pre>\n\n\n\n
\u53ef\u4ee5\u770b\u51fa\uff0c\u5f62\u72b6\u76f8\u540c\u7684(1, 3, 10, 10)<\/code>\u7684\u8f93\u5165\u5728\u6240\u6709\u6a21\u578b\u4e0a\u90fd\u6ca1\u6709\u51fa\u9519\u3002\u800c\u5bf9\u4e8ebatch\uff08\u7b2c 0 \u7ef4\uff09\u6216\u8005\u957f\u5bbd\uff08\u7b2c 2\u30013\u7ef4\uff09\u4e0d\u540c\u7684\u8f93\u5165\uff0c\u53ea\u6709\u5728\u8bbe\u7f6e\u4e86\u5bf9\u5e94\u7684\u52a8\u6001\u7ef4\u5ea6\u540e\u624d\u4e0d\u4f1a\u51fa\u9519\u3002\u6211\u4eec\u53ef\u4ee5\u9519\u8bef\u4fe1\u606f\u4e2d\u627e\u51fa\u662f\u54ea\u4e9b\u7ef4\u5ea6\u51fa\u4e86\u95ee\u9898\u3002\u6bd4\u5982\u6211\u4eec\u53ef\u4ee5\u7528\u4ee5\u4e0b\u4ee3\u7801\u67e5\u770binput[1]<\/code>\u5728model_static.onnx<\/code>\u4e2d\u7684\u62a5\u9519\u4fe1\u606f\uff1a<\/p>\n\n\n\n
print(exceptions[(1, 'model_static.onnx')]) \n \n# output <\/em>\n# [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Got invalid dimensions for input: in for the following indices index: 0 Got: 2 Expected: 1 Please fix either the inputs or the model. <\/em><\/code><\/pre>\n\n\n\n
\u8fd9\u6bb5\u62a5\u9519\u544a\u8bc9\u6211\u4eec\u540d\u5b57\u53ebin<\/code>\u7684\u8f93\u5165\u7684\u7b2c 0 \u7ef4\u4e0d\u5339\u914d\u3002\u672c\u6765\u8be5\u7ef4\u7684\u957f\u5ea6\u5e94\u8be5\u4e3a 1\uff0c\u4f46\u6211\u4eec\u7684\u8f93\u5165\u662f 2\u3002\u5b9e\u9645\u90e8\u7f72\u4e2d\uff0c\u5982\u679c\u6211\u4eec\u78b0\u5230\u4e86\u7c7b\u4f3c\u7684\u62a5\u9519\uff0c\u5c31\u53ef\u4ee5\u901a\u8fc7\u8bbe\u7f6e\u52a8\u6001\u7ef4\u5ea6\u6765\u89e3\u51b3\u95ee\u9898\u3002<\/p>\n\n\n\n
\u4f7f\u7528\u63d0\u793a<\/h3>\n\n\n\n
\u901a\u8fc7\u5b66\u4e60\u4e4b\u524d\u7684\u77e5\u8bc6\uff0c\u6211\u4eec\u57fa\u672c\u638c\u63e1\u4e86 torch.onnx.export<\/code>\u51fd\u6570\u7684\u90e8\u5206\u5b9e\u73b0\u539f\u7406\u548c\u53c2\u6570\u8bbe\u7f6e\u65b9\u6cd5\uff0c\u8db3\u4ee5\u5b8c\u6210\u7b80\u5355\u6a21\u578b\u7684\u8f6c\u6362\u4e86\u3002\u4f46\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u4f7f\u7528\u8be5\u51fd\u6570\u8fd8\u4f1a\u8e29\u5f88\u591a\u5751\u3002\u8fd9\u91cc\u6211\u4eec\u6a21\u578b\u90e8\u7f72\u56e2\u961f\u628a\u5728\u5b9e\u6218\u4e2d\u79ef\u7d2f\u7684\u4e00\u4e9b\u7ecf\u9a8c\u5206\u4eab\u7ed9\u5927\u5bb6\u3002<\/p>\n\n\n\n
\u4f7f\u6a21\u578b\u5728 ONNX \u8f6c\u6362\u65f6\u6709\u4e0d\u540c\u7684\u884c\u4e3a<\/strong><\/p>\n\n\n\n
\u6709\u4e9b\u65f6\u5019\uff0c\u6211\u4eec\u5e0c\u671b\u6a21\u578b\u5728\u5bfc\u51fa\u81f3 ONNX \u65f6\u6709\u4e00\u4e9b\u4e0d\u540c\u7684\u884c\u4e3a\u6a21\u578b\u5728\u76f4\u63a5\u7528 PyTorch \u63a8\u7406\u65f6\u6709\u4e00\u5957\u903b\u8f91\uff0c\u800c\u5728\u5bfc\u51fa\u7684ONNX\u6a21\u578b\u4e2d\u6709\u53e6\u4e00\u5957\u903b\u8f91\u3002\u6bd4\u5982\uff0c\u6211\u4eec\u53ef\u4ee5\u628a\u4e00\u4e9b\u540e\u5904\u7406\u7684\u903b\u8f91\u653e\u5728\u6a21\u578b\u91cc\uff0c\u4ee5\u7b80\u5316\u9664\u8fd0\u884c\u6a21\u578b\u4e4b\u5916\u7684\u5176\u4ed6\u4ee3\u7801\u3002torch.onnx.is_in_onnx_export()<\/code>\u53ef\u4ee5\u5b9e\u73b0\u8fd9\u4e00\u4efb\u52a1\uff0c\u8be5\u51fd\u6570\u4ec5\u5728\u6267\u884c torch.onnx.export()<\/code>\u65f6\u4e3a\u771f\u3002\u4ee5\u4e0b\u662f\u4e00\u4e2a\u4f8b\u5b50\uff1a<\/p>\n\n\n\n
import torch \n \nclass Model(torch.nn.Module): \n    def __init__(self): \n        super().__init__() \n        self.conv = torch.nn.Conv2d(3, 3, 3) \n \n    def forward(self, x): \n        x = self.conv(x) \n        if torch.onnx.is_in_onnx_export(): \n            x = torch.clip(x, 0, 1) \n        return x <\/code><\/pre>\n\n\n\n

\u8fd9\u91cc\uff0c\u6211\u4eec\u4ec5\u5728\u6a21\u578b\u5bfc\u51fa\u65f6\u628a\u8f93\u51fa\u5f20\u91cf\u7684\u6570\u503c\u9650\u5236\u5728[0, 1]\u4e4b\u95f4\u3002\u4f7f\u7528 is_in_onnx_export<\/code>\u786e\u5b9e\u80fd\u8ba9\u6211\u4eec\u65b9\u4fbf\u5730\u5728\u4ee3\u7801\u4e2d\u6dfb\u52a0\u548c\u6a21\u578b\u90e8\u7f72\u76f8\u5173\u7684\u903b\u8f91\u3002\u4f46\u662f\uff0c\u8fd9\u4e9b\u4ee3\u7801\u5bf9\u53ea\u5173\u5fc3\u6a21\u578b\u8bad\u7ec3\u7684\u5f00\u53d1\u8005\u548c\u7528\u6237\u6765\u8bf4\u5f88\u4e0d\u53cb\u597d\uff0c\u7a81\u5140\u7684\u90e8\u7f72\u903b\u8f91\u4f1a\u964d\u4f4e\u4ee3\u7801\u6574\u4f53\u7684\u53ef\u8bfb\u6027\u3002\u540c\u65f6\uff0cis_in_onnx_export<\/code>\u53ea\u80fd\u5728\u6bcf\u4e2a\u9700\u8981\u6dfb\u52a0\u90e8\u7f72\u903b\u8f91\u7684\u5730\u65b9\u90fd\u201c\u6253\u8865\u4e01\u201d\uff0c\u96be\u4ee5\u8fdb\u884c\u7edf\u4e00\u7684\u7ba1\u7406\u3002\u6211\u4eec\u4e4b\u540e\u4f1a\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528 MMDeploy \u7684\u91cd\u5199\u673a\u5236\u6765\u89c4\u907f\u8fd9\u4e9b\u95ee\u9898\u3002<\/p>\n\n\n\n
\u5229\u7528\u4e2d\u65ad\u5f20\u91cf\u8ddf\u8e2a\u7684\u64cd\u4f5c<\/strong><\/p>\n\n\n\n
PyTorch \u8f6c ONNX \u7684\u8ddf\u8e2a\u5bfc\u51fa\u6cd5\u662f\u4e0d\u662f\u4e07\u80fd\u7684\u3002\u5982\u679c\u6211\u4eec\u5728\u6a21\u578b\u4e2d\u505a\u4e86\u4e00\u4e9b\u5f88\u201c\u51fa\u683c\u201d\u7684\u64cd\u4f5c\uff0c\u8ddf\u8e2a\u6cd5\u4f1a\u628a\u67d0\u4e9b\u53d6\u51b3\u4e8e\u8f93\u5165\u7684\u4e2d\u95f4\u7ed3\u679c\u53d8\u6210\u5e38\u91cf\uff0c\u4ece\u800c\u4f7f\u5bfc\u51fa\u7684 ONNX \u6a21\u578b\u548c\u539f\u6765\u7684\u6a21\u578b\u6709\u51fa\u5165\u3002\u4ee5\u4e0b\u662f\u4e00\u4e2a\u4f1a\u9020\u6210\u8fd9\u79cd\u201c\u8ddf\u8e2a\u4e2d\u65ad\u201d\u7684\u4f8b\u5b50\uff1a<\/p>\n\n\n\n
class Model(torch.nn.Module): \n    def __init__(self): \n        super().__init__() \n \n    def forward(self, x): \n        x = x * x[0].item() \n        return x, torch.Tensor([i for i in x]) \n \nmodel = Model()       \ndummy_input = torch.rand(10) \ntorch.onnx.export(model, dummy_input, 'a.onnx') <\/code><\/pre>\n\n\n\n
\u5982\u679c\u4f60\u5c1d\u8bd5\u53bb\u5bfc\u51fa\u8fd9\u4e2a\u6a21\u578b\uff0c\u4f1a\u5f97\u5230\u4e00\u5927\u5806 warning\uff0c\u544a\u8bc9\u4f60\u8f6c\u6362\u51fa\u6765\u7684\u6a21\u578b\u53ef\u80fd\u4e0d\u6b63\u786e\u3002\u8fd9\u4e5f\u96be\u602a\uff0c\u6211\u4eec\u5728\u8fd9\u4e2a\u6a21\u578b\u91cc\u4f7f\u7528\u4e86.item()<\/code>\u628a torch \u4e2d\u7684\u5f20\u91cf\u8f6c\u6362\u6210\u4e86\u666e\u901a\u7684 Python \u53d8\u91cf\uff0c\u8fd8\u5c1d\u8bd5\u904d\u5386 torch \u5f20\u91cf\uff0c\u5e76\u7528\u4e00\u4e2a\u5217\u8868\u65b0\u5efa\u4e00\u4e2a torch \u5f20\u91cf\u3002\u8fd9\u4e9b\u6d89\u53ca\u5f20\u91cf\u4e0e\u666e\u901a\u53d8\u91cf\u8f6c\u6362\u7684\u903b\u8f91\u90fd\u4f1a\u5bfc\u81f4\u6700\u7ec8\u7684 ONNX \u6a21\u578b\u4e0d\u592a\u6b63\u786e\u3002
\u53e6\u4e00\u65b9\u9762\uff0c\u6211\u4eec\u4e5f\u53ef\u4ee5\u5229\u7528\u8fd9\u4e2a\u6027\u8d28\uff0c\u5728\u4fdd\u8bc1\u6b63\u786e\u6027\u7684\u524d\u63d0\u4e0b\u4ee4\u6a21\u578b\u7684\u4e2d\u95f4\u7ed3\u679c\u53d8\u6210\u5e38\u91cf\u3002\u8fd9\u4e2a\u6280\u5de7\u5e38\u5e38\u7528\u4e8e\u6a21\u578b\u7684\u9759\u6001\u5316\u4e0a\uff0c\u5373\u4ee4\u6a21\u578b\u4e2d\u6240\u6709\u7684\u5f20\u91cf\u5f62\u72b6\u90fd\u53d8\u6210\u5e38\u91cf\u3002\u5728\u672a\u6765\u7684\u6559\u7a0b\u4e2d\uff0c\u6211\u4eec\u4f1a\u5728\u90e8\u7f72\u5b9e\u4f8b\u4e2d\u8be6\u7ec6\u4ecb\u7ecd\u8fd9\u4e9b\u201c\u9ad8\u7ea7\u201d\u64cd\u4f5c\u3002<\/p>\n\n\n\n
\u4f7f\u7528\u5f20\u91cf\u4e3a\u8f93\u5165\uff08PyTorch\u7248\u672c < 1.9.0\uff09<\/strong><\/p>\n\n\n\n
\u6b63\u5982\u6211\u4eec\u7b2c\u4e00\u7bc7\u6559\u7a0b<\/a>\u6240\u5c55\u793a\u7684\uff0c\u5728\u8f83\u65e7(< 1.9.0)\u7684 PyTorch \u4e2d\u628a Python \u6570\u503c\u4f5c\u4e3a torch.onnx.export()<\/code>\u7684\u6a21\u578b\u8f93\u5165\u65f6\u4f1a\u62a5\u9519\u3002\u51fa\u4e8e\u517c\u5bb9\u6027\u7684\u8003\u8651\uff0c\u6211\u4eec\u8fd8\u662f\u63a8\u8350\u4ee5\u5f20\u91cf\u4e3a\u6a21\u578b\u8f6c\u6362\u65f6\u7684\u6a21\u578b\u8f93\u5165\u3002<\/p>\n\n\n\n
PyTorch \u5bf9 ONNX \u7684\u7b97\u5b50\u652f\u6301<\/strong><\/h2>\n\n\n\n
\u5728\u786e\u4fddtorch.onnx.export()<\/code>\u7684\u8c03\u7528\u65b9\u6cd5\u65e0\u8bef\u540e\uff0cPyTorch \u8f6c ONNX \u65f6\u6700\u5bb9\u6613\u51fa\u73b0\u7684\u95ee\u9898\u5c31\u662f\u7b97\u5b50\u4e0d\u517c\u5bb9\u4e86\u3002\u8fd9\u91cc\u6211\u4eec\u4f1a\u4ecb\u7ecd\u5982\u4f55\u5224\u65ad\u67d0\u4e2a PyTorch \u7b97\u5b50\u5728 ONNX \u4e2d\u662f\u5426\u517c\u5bb9\uff0c\u4ee5\u52a9\u5927\u5bb6\u5728\u78b0\u5230\u62a5\u9519\u65f6\u80fd\u66f4\u597d\u5730\u628a\u9519\u8bef\u5f52\u7c7b\u3002\u800c\u5177\u4f53\u6dfb\u52a0\u7b97\u5b50\u7684\u65b9\u6cd5\u6211\u4eec\u4f1a\u5728\u4e4b\u540e\u7684\u6587\u7ae0\u91cc\u4ecb\u7ecd\u3002
\u5728\u8f6c\u6362\u666e\u901a\u7684torch.nn.Module<\/code>\u6a21\u578b\u65f6\uff0cPyTorch \u4e00\u65b9\u9762\u4f1a\u7528\u8ddf\u8e2a\u6cd5\u6267\u884c\u524d\u5411\u63a8\u7406\uff0c\u628a\u9047\u5230\u7684\u7b97\u5b50\u6574\u5408\u6210\u8ba1\u7b97\u56fe\uff1b\u53e6\u4e00\u65b9\u9762\uff0cPyTorch \u8fd8\u4f1a\u628a\u9047\u5230\u7684\u6bcf\u4e2a\u7b97\u5b50\u7ffb\u8bd1\u6210 ONNX \u4e2d\u5b9a\u4e49\u7684\u7b97\u5b50\u3002\u5728\u8fd9\u4e2a\u7ffb\u8bd1\u8fc7\u7a0b\u4e2d\uff0c\u53ef\u80fd\u4f1a\u78b0\u5230\u4ee5\u4e0b\u60c5\u51b5\uff1a<\/p>\n\n\n\n