无法将torchvision FasterRCNN-ResNet50模型导出为ONNX文件的问题求助

我编写了一段简单的Python脚本，加载torchvision的fasterrcnn_resnet50_fpn_v2模型（使用默认权重），尝试通过torch.onnx.export和torch.onnx.dynamo_export两种方式将其导出为ONNX文件，但均出现错误。

脚本代码：

import torch
import torch.onnx

import torchvision

torch_model = torchvision.models.detection.fasterrcnn_resnet50_fpn_v2(weights='DEFAULT')
torch_model.eval()
torch_input = torch.randn(1, 3, 32, 32)

is_dynamo_export = False

if (is_dynamo_export):
    onnx_program = torch.onnx.dynamo_export(torch_model, torch_input)
    onnx_program.save("onnx_dynamo_export_ResNET50.onnx")        
else:
    torch.onnx.export(torch_model,               # model being run
                      torch_input,                         # model input (or a tuple for multiple inputs)
                      "onnx_export_ResNET50.onnx",   # where to save the model (can be a file or file-like object)
                      export_params=True,        # store the trained parameter weights inside the model file
                      opset_version=10,          # the ONNX version to export the model to
                      do_constant_folding=True,  # whether to execute constant folding for optimization
                      input_names = ['input'],   # the model's input names
                      output_names = ['output'], # the model's output names
                      dynamic_axes={'input' : {0 : 'batch_size'},    # variable length axes
                                    'output' : {0 : 'batch_size'}})  

错误信息：

1. 使用dynamo_export时：

File "C:\\tools\\Python311\\Lib\\site-packages\\torch\\onnx\\_internal\\exporter.py", line 1439, in dynamo_export
  raise OnnxExporterError(

torch.onnx.OnnxExporterError: Failed to export the model to ONNX. Generating SARIF report at 'report_dynamo_export.sarif'. SARIF is a standard format for the output of static analysis tools. SARIF logs can be loaded in VS Code SARIF viewer extension, or SARIF web viewer. Please report a bug on PyTorch Github.

2. 使用普通torch.onnx.export时：

torch.onnx.errors.SymbolicValueError: Unsupported: ONNX export of Pad in opset 9. The sizes of the padding must be constant. Please try opset version 11. [Caused by the value '535 defined in (%535 : int[] = prim::ListConstruct(%405, %534, %405, %533, %405, %532), scope: torchvision.models.detection.faster_rcnn.FasterRCNN::

注：两种导出方式对自定义简单模型均可正常工作，自定义模型代码：

import torch.nn as nn
import torch.nn.functional as F

class MyModel(nn.Module):

    def __init__(self):
        super(MyModel, self).__init__()
        self.conv1 = nn.Conv2d(1, 6, 5)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
        x = F.max_pool2d(F.relu(self.conv2(x)), 2)
        x = torch.flatten(x, 1)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

修复普通torch.onnx.export的报错 #

报错已明确提示Pad操作在opset 9/10不支持动态padding，需做以下调整：

• 升级opset版本：将opset_version=10改为opset_version=11或更高（推荐16以上，适配更多模型操作）。
• 修正输出配置：Faster RCNN的输出是包含boxes、labels、scores的字典，不能用单一的output作为输出名，需对应修改：

  output_names = ['boxes', 'labels', 'scores'],
  dynamic_axes={
      'input' : {0 : 'batch_size'},
      'boxes' : {0 : 'batch_size'},
      'labels' : {0 : 'batch_size'},
      'scores' : {0 : 'batch_size'}
  }
  

修复torch.onnx.dynamo_export的报错 #

Dynamo导出对检测类复杂模型的兼容性仍不完善，可尝试：

1. 升级PyTorch到最新稳定版，新版本通常修复了更多Dynamo导出的适配问题。
2. 优先使用调整后的普通torch.onnx.export方式，该方式对经典检测模型的支持更成熟。
3. 若必须使用Dynamo导出，添加动态形状配置参数：

   onnx_program = torch.onnx.dynamo_export(
       torch_model, 
       torch_input,
       export_options=torch.onnx.ExportOptions(dynamic_shapes=True)
   )
   

额外注意事项 #

• 检测模型在eval模式下默认返回字典，可修改forward方法转为tuple输出，适配ONNX要求：

  def modified_forward(self, x):
      outputs = self.__original_forward(x)
      return outputs['boxes'], outputs['labels'], outputs['scores']
  
  torch_model.__original_forward = torch_model.forward
  torch_model.forward = modified_forward.__get__(torch_model)
  

• 输入张量尺寸建议贴合实际检测场景，比如改用640x640这类常见尺寸，避免过小尺寸导致模型内部计算异常。

内容的提问来源于stack exchange，提问作者ghost113116