Pytorch 有人知道为什么SHAP'；s的深度解释在ResNet-50预训练模型上失败？

我已经培训了一个ResNet-50模型，我正在尝试实现解释性方法SHAP。我发现这一直是中国其他人的一个问题

我的代码严格遵循中SHAP pytorch示例中的说明：

回溯错误为：

Traceback (most recent call last):
  File "/.../PycharmProjects/thesis/SHAP.py", line 133, in <module>
    main()
  File ".../PycharmProjects/thesis/SHAP.py", line 123, in main
    shap_values = e.shap_values(test_images)
  File ".../anaconda3/lib/python3.8/site-packages/shap/explainers/_deep/__init__.py", line 124, in shap_values
    return self.explainer.shap_values(X, ranked_outputs, output_rank_order, check_additivity=check_additivity)
  File ".../anaconda3/lib/python3.8/site-packages/shap/explainers/_deep/deep_pytorch.py", line 185, in shap_values
    sample_phis = self.gradient(feature_ind, joint_x)
  File ".../anaconda3/lib/python3.8/site-packages/shap/explainers/_deep/deep_pytorch.py", line 121, in gradient
    grad = torch.autograd.grad(selected, x,
  File ".../anaconda3/lib/python3.8/site-packages/torch/autograd/__init__.py", line 202, in grad
    return Variable._execution_engine.run_backward(
  File ".../anaconda3/lib/python3.8/site-packages/shap/explainers/_deep/deep_pytorch.py", line 226, in deeplift_grad
    return op_handler[module_type](module, grad_input, grad_output)
  File ".../anaconda3/lib/python3.8/site-packages/shap/explainers/_deep/deep_pytorch.py", line 358, in nonlinear_1d
    grad_output[0] * (delta_out / delta_in).repeat(dup0))
RuntimeError: The size of tensor a (512) must match the size of tensor b (2048) at non-singleton dimension 1

回溯（最近一次呼叫最后一次）：
文件“/…/PycharmProjects/thesis/SHAP.py”，第133行，在
main（）
文件“../PycharmProjects/thesis/SHAP.py”，第123行，主目录
形状值=e.形状值（测试图像）
文件“../anaconda3/lib/python3.8/site packages/shap/explainers/_deep/_init__.py”，第124行，在shap_值中
返回self.explainer.shap\u值（X，排列输出，输出排列顺序，检查可加性=检查可加性）
文件“../anaconda3/lib/python3.8/site packages/shap/explainers/\u deep/deep\u pytorch.py”，第185行，以shap\u值表示
示例\u phis=自梯度（特征\u ind，关节\u x）
文件“../anaconda3/lib/python3.8/site packages/shap/explainers/\u deep/deep\u pytorch.py”，第121行，渐变
梯度=火炬自动加载梯度（选定，x，
文件“../anaconda3/lib/python3.8/site packages/torch/autograd/_init__.py”，第202行，渐变
返回变量。\u执行\u引擎。向后运行\u(
文件“../anaconda3/lib/python3.8/site packages/shap/explainers/\u deep/deep\u pytorch.py”，第226行，在deeplift\u grad中
返回操作处理程序[模块类型]（模块、梯度输入、梯度输出）
文件“../anaconda3/lib/python3.8/site packages/shap/explainers/\u deep/deep\u pytorch.py”，第358行，非线性1d
梯度输出[0]*（增量输出/增量输入）。重复（dup0））
RuntimeError：张量a（512）的大小必须与张量b（2048）在非单态维度1的大小相匹配

有人知道是什么原因导致了这种情况吗？我已经检查了代码很多次，但我找不到哪里出错了……在经过培训的网络体系结构上，张量大小不应该不匹配

请注意，我故意删除了任何个人识别路径：）

---

非常感谢您的任何意见

我认为发生此错误是因为ResNet在FC层中需要softmax。尝试这样做：

resnet=models.resnet50（pretrained=True）
resnet.fc=torch.nn.线性（2048,5）
resnet.load\u state\u dict（火炬加载（模型路径））
#在加载模型之后，在上面添加此代码
resnet.fc=nn.Sequential(
resnet.fc，
nn.Softmax（1），
)
resnet.eval（）

Traceback (most recent call last):
  File "/.../PycharmProjects/thesis/SHAP.py", line 133, in <module>
    main()
  File ".../PycharmProjects/thesis/SHAP.py", line 123, in main
    shap_values = e.shap_values(test_images)
  File ".../anaconda3/lib/python3.8/site-packages/shap/explainers/_deep/__init__.py", line 124, in shap_values
    return self.explainer.shap_values(X, ranked_outputs, output_rank_order, check_additivity=check_additivity)
  File ".../anaconda3/lib/python3.8/site-packages/shap/explainers/_deep/deep_pytorch.py", line 185, in shap_values
    sample_phis = self.gradient(feature_ind, joint_x)
  File ".../anaconda3/lib/python3.8/site-packages/shap/explainers/_deep/deep_pytorch.py", line 121, in gradient
    grad = torch.autograd.grad(selected, x,
  File ".../anaconda3/lib/python3.8/site-packages/torch/autograd/__init__.py", line 202, in grad
    return Variable._execution_engine.run_backward(
  File ".../anaconda3/lib/python3.8/site-packages/shap/explainers/_deep/deep_pytorch.py", line 226, in deeplift_grad
    return op_handler[module_type](module, grad_input, grad_output)
  File ".../anaconda3/lib/python3.8/site-packages/shap/explainers/_deep/deep_pytorch.py", line 358, in nonlinear_1d
    grad_output[0] * (delta_out / delta_in).repeat(dup0))
RuntimeError: The size of tensor a (512) must match the size of tensor b (2048) at non-singleton dimension 1