使用SHAP解释DeepLabV3+语义分割模型时遇max_evals过低错误求助

问题描述 #

想用SHAP测试多类别语义分割模型deeplab_v3plus的可解释性，分析哪些特征对语义分类贡献最大，但运行代码时触发如下错误：

ValueError: max_evals=500 is too low for the Permutation explainer, it must be at least 2 * num_features + 1 = 1854721!

当前test_img/*下只有3张图片，怀疑是图片数量不足导致的问题？

运行代码 #

import glob
from PIL import Image

import torch
from torchvision import transforms
from torchvision.utils import make_grid
import torchvision.transforms.functional as tf

from deeplab import deeplab_v3plus

import shap

def test(args):
    # make a video prez
    
    model = deeplab_v3plus('resnet101', num_classes=args.nclass, output_stride=16, pretrained_backbone=True)
    model.load_state_dict(torch.load(args.seg_file,map_location=torch.device('cpu'))) # because no gpu available on sandbox environnement
    model = model.to(args.device)
    model.eval()
    explainer = shap.Explainer(model)
    with torch.no_grad():
        for i, file in enumerate(args.img_folder):
            img = img2tensor(file, args)
    
            pred = model(img)
            print(explainer(img))

if __name__ == '__main__':
    class Arguments:
        def __init__(self):
            self.device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
            self.seg_file = "Model_Woodscape.pth"
            self.img_folder = glob.glob("test_img/*.png")
            self.mean = [0.485, 0.456, 0.406]
            self.std = [0.229, 0.224, 0.225]
            self.h, self.w = 483, 640
            self.nclass = 10
            self.cmap = {
                1: [128, 64, 128],  # "road",
                2: [69, 76, 11],    # "lanemarks",
                3: [0, 255, 0],     # "curb",
                4: [220, 20, 60],   # "person",
                5: [255, 0, 0],     # "rider",
                6: [0, 0, 142],     # "vehicles",
                7: [119, 11, 32],   # "bicycle",
                8: [0, 0, 230],     # "motorcycle",
                9: [220, 220, 0],   # "traffic_sign",
                0: [0, 0, 0]        # "void"
            }

    args = Arguments()
    test(args)

报错信息 #

(dee_env) jovyan@jupyter:~/use-cases/Scene_understanding/Code_Woodscape/deeplab_v3+$ python test_shap.py 
BILINEAR is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.BILINEAR instead.
Traceback (most recent call last):
  File "/home/jovyan/use-cases/Scene_understanding/Code_Woodscape/deeplab_v3+/test_shap.py", line 85, in <module>
    test(args)
  File "/home/jovyan/use-cases/Scene_understanding/Code_Woodscape/deeplab_v3+/test_shap.py", line 37, in test
    print(explainer(img))
  File "/home/jovyan/use-cases/Scene_understanding/Code_Woodscape/deeplab_v3+/dee_env/lib/python3.9/site-packages/shap/explainers/_permutation.py", line 82, in __call__
    return super().__call__(
  File "/home/jovyan/use-cases/Scene_understanding/Code_Woodscape/deeplab_v3+/dee_env/lib/python3.9/site-packages/shap/explainers/_explainer.py", line 266, in __call__
    row_result = self.explain_row(
  File "/home/jovyan/use-cases/Scene_understanding/Code_Woodscape/deeplab_v3+/dee_env/lib/python3.9/site-packages/shap/explainers/_permutation.py", line 164, in explain_row
    raise ValueError(f"max_evals={max_evals} is too low for the Permutation explainer, it must be at least 2 * num_features + 1 = {2 * len(inds) + 1}!")
ValueError: max_evals=500 is too low for the Permutation explainer, it must be at least 2 * num_features + 1 = 1854721!

原因分析 #

这个错误和测试图片数量无关，核心问题在于：

• 调用shap.Explainer(model)时，SHAP自动选择了Permutation Explainer，但该解释器是为表格数据设计的，完全不适合处理图像这种高维度输入。
• 你的输入是3通道483×640的图像，总特征数为3×483×640=927360，Permutation Explainer要求的最小max_evals为2×特征数+1=1854721，远大于默认值500，因此触发报错。

解决方法 #

使用SHAP专为深度学习图像模型设计的DeepExplainer或GradientExplainer，这类基于梯度的解释器计算效率高，适合图像任务。

1. 修改解释器初始化代码 #

替换原有的解释器初始化逻辑，加入背景数据集（用少量测试图即可，3-5张足够）：

def test(args):
    model = deeplab_v3plus('resnet101', num_classes=args.nclass, output_stride=16, pretrained_backbone=True)
    model.load_state_dict(torch.load(args.seg_file,map_location=torch.device('cpu')))
    model = model.to(args.device)
    model.eval()

    # 准备背景数据集：用3张测试图作为背景，用于估算模型期望输出
    background_imgs = []
    for file in args.img_folder[:3]:
        img = img2tensor(file, args)
        background_imgs.append(img)
    background = torch.cat(background_imgs, dim=0)

    # 使用DeepExplainer替代默认的Permutation Explainer
    explainer = shap.DeepExplainer(model, background)

    with torch.no_grad():
        for i, file in enumerate(args.img_folder):
            img = img2tensor(file, args)
            # 生成所有类别的SHAP值
            shap_values = explainer.shap_values(img)
            # shap_values是列表，每个元素对应一个类别的SHAP值，形状与输入img一致
            print(f"处理第{i+1}张图，生成了{len(shap_values)}个类别的SHAP值")

2. 补充缺失的img2tensor函数 #

你的代码中未实现img2tensor，补充如下：

def img2tensor(file, args):
    img = Image.open(file).convert('RGB')
    # 替换废弃的BILINEAR为Resampling.BILINEAR
    transform = transforms.Compose([
        transforms.Resize((args.h, args.w), interpolation=transforms.Resampling.BILINEAR),
        transforms.ToTensor(),
        transforms.Normalize(mean=args.mean, std=args.std)
    ])
    img_tensor = transform(img).unsqueeze(0).to(args.device)
    return img_tensor

3. 可选：可视化SHAP结果 #

添加可视化代码，直观展示每个类别的特征贡献：

import shap.plots

def denormalize(img, mean, std):
    """将标准化后的图像转回原始范围，用于可视化"""
    img = img.clone()
    for channel in range(img.size(0)):
        img[channel] = img[channel] * std[channel] + mean[channel]
    return img.permute(1,2,0).cpu().numpy()

# 在循环中添加可视化
with torch.no_grad():
    for i, file in enumerate(args.img_folder):
        img = img2tensor(file, args)
        shap_values = explainer.shap_values(img)
        # 可视化类别1（道路）的SHAP值
        original_img = denormalize(img.squeeze(0), args.mean, args.std)
        shap.plots.image(shap_values[1], original_img)

注意事项 #

• 背景数据集无需大量样本，3-5张即可满足需求。
• 确保背景数据和输入数据在同一设备（CPU/GPU）上。
• 如果使用GPU，注意控制单批次数据大小，避免显存溢出。

内容的提问来源于stack exchange，提问作者Revolucion for Monica