pytorch - 有谁知道为什么 SHAP 的 Deep Explainer 在 ResNet-50 预训练模型上失败？

Question

我已经训练了一个 ResNet-50 模型，我正在尝试实现可解释性方法 SHAP。我发现这对其他人来说是个问题 https://github.com/slundberg/shap/issues/1479 .

我的代码严格遵循 https://github.com/slundberg/shap/blob/master/notebooks/image_examples/image_classification/PyTorch%20Deep%20Explainer%20MNIST%20example.ipynb 中 SHAP pytorch 示例中的说明。 :

def main():
    output_format = Pupils()

    model_path = ".../trained_model_30_epochs"
    ellipse_overlay_path = ".../ellipse_overlay.png"

    path_to_dataset = [DatasetPath(path='...',
                                   image_type='png')]

    path_to_dataset_to_explain = [DatasetPath(path='...',
                                   image_type='png')]

    dataset = RealDataset(path_to_dataset,
                          output_format.ground_truth_from_annotations,
                          transform=get_transforms(is_training=False, is_synthetic=False),
                          load_into_ram=True
                          )

    dataset_to_explain = RealDataset(path_to_dataset_to_explain,
                                     output_format.ground_truth_from_annotations,
                                     transform=get_transforms(is_training=False, is_synthetic=False),
                                     load_into_ram=True
                                     )
    dataloader = DataLoader(dataset,
                            batch_size=100,
                            shuffle=False)

    dataloader_to_explain = DataLoader(dataset_to_explain,
                            batch_size=32,
                            shuffle=False)

    resnet = models.resnet50(pretrained=True)
    resnet.fc = torch.nn.Linear(2048, 5)
    resnet.load_state_dict(torch.load(model_path))
    resnet.eval()

    batch = next(iter(dataloader))
    images = batch['image'].expand(-1, 3, -1, -1)

    # select a set of background examples to take an expectation over
    background = images[:100]  # dataset[0]['image']
    # background = background.expand(3, -1, -1).unsqueeze(0)

    batch = next(iter(dataloader_to_explain))
    test_images = batch['image'].expand(-1, 3, -1, -1) # dataset_to_explain[1]['image']
    # test_images = test_images.expand(3, -1, -1).unsqueeze(0)

    e = shap.DeepExplainer(resnet, background)
    shap_values = e.shap_values(test_images)

    shap_numpy = [np.swapaxes(np.swapaxes(s, 1, -1), 1, 2) for s in shap_values]
    test_numpy = np.swapaxes(np.swapaxes(test_images[:3].numpy(), 1, -1), 1, 2)

    # plot the feature attributions
    shap.image_plot(shap_numpy, -test_numpy)

回溯错误是:

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

有人知道是什么原因造成的吗？我已经检查代码很多次了，但我找不到哪里出错了……在训练有素的网络架构上，张量大小不应该不匹配。

请注意，我有意删除了任何个人识别路径 :)。

非常感谢任何输入!

Answer 1

我认为出现这个错误是因为ResNet在FC层需要softmax。尝试这样做:

resnet = models.resnet50(pretrained=True)
resnet.fc = torch.nn.Linear(2048, 5)
resnet.load_state_dict(torch.load(model_path))
# add this code above, after load model
resnet.fc = nn.Sequential(
    resnet.fc,
    nn.Softmax(1),
)
resnet.eval()