使用dlfeval实现3D模型Grad-CAM时AlexNet分组卷积维度不匹配问题

问题背景 #

在为输入尺寸227×227×7的3D成像模型实现Grad-CAM时，遇到如下异常：

• dlfeval外部调用predict可正常输出结果
• dlfeval内部执行相同predict代码时，触发分组卷积维度不匹配错误，报错指向conv2层输入通道96与权重通道48不匹配，但该层NumChannels设置为[48,48]

测试代码：

image_data = zeros(227,227,7);
load(['case_1.mat']); % DL model
dlnet = dag2dlnetwork(net_final);
dlInputData = dlarray(image_data);
prediction1 = predict(dlnet, dlInputData)
prediction2 = dlfeval(@gradcam, dlnet, dlInputData)
function prediction = gradcam(dlnet, dlInputData)
    prediction = predict(dlnet, dlInputData);
end

外部正常输出：

prediction1 = 
  1×2 single dlarray
    0.0301    0.9699

内部报错信息：

Error using dlnetwork/predict (line 658)
Execution failed during layer(s) 'conv2, relu2'.
Error in test2>gradcam (line 11)
    prediction = predict(dlnet, dlInputData);
Error in deep.internal.dlfeval (line 17)
[varargout{1:nargout}] = fun(x{:});
Error in deep.internal.dlfevalWithNestingCheck (line 19)
    [varargout{1:nargout}] = deep.internal.dlfeval(fun,varargin{:});
Error in dlfeval (line 31)
[varargout{1:nargout}] = deep.internal.dlfevalWithNestingCheck(fun,varargin{:});
Error in test2 (line 9)
prediction2 = dlfeval(@gradcam, dlnet, dlInputData)
Caused by:
    Error using deep.internal.dlarray.validateConvolutionChannelDimension (line 26)
    The size of the 'C' dimension of the input data (96) must be equal to the number of channels of weights (48,
    specified by the size of weights dimension number 3).

原因分析 #

问题源于dag2dlnetwork的转换逻辑缺陷：
旧AlexNet DAG模型的分组卷积是通过两个并行的48通道卷积层实现的，转换为dlnetwork后，普通predict模式下框架会自动处理并行分支的拼接逻辑；但在自动微分（dlfeval）模式下，框架无法识别这种隐式分组结构，将其视为单个卷积层，导致输入通道数与权重通道数校验失败。

解决方案 #

手动将并行分组卷积分支重构为单个带NumGroups参数的卷积层，让dlnetwork在自动微分时正确识别分组逻辑：

1. 定位并修改conv2层参数

% 找到conv2层的索引
conv2Idx = find(strcmp({dlnet.Layers.Name}, 'conv2'));
conv2Layer = dlnet.Layers(conv2Idx);

% 设置分组卷积参数：输入通道96，分组数2
conv2Layer.NumChannels = 96;
conv2Layer.NumGroups = 2;

% 合并原并行分支的权重：将两个48输出通道的权重合并为96输出通道
weightsBranch1 = conv2Layer.Weights(:, :, :, 1:48);
weightsBranch2 = conv2Layer.Weights(:, :, :, 49:96);
conv2Layer.Weights = cat(4, weightsBranch1, weightsBranch2);

% 合并原并行分支的偏置
biasBranch1 = conv2Layer.Bias(:, 1:48);
biasBranch2 = conv2Layer.Bias(:, 49:96);
conv2Layer.Bias = cat(2, biasBranch1, biasBranch2);

% 替换层并重新编译网络
dlnet.Layers(conv2Idx) = conv2Layer;
dlnet = dlnetwork(dlnet.Layers, dlnet.Connections);

2. 验证修改效果
   重新运行测试代码，此时dlfeval内部的predict可正常执行，不再触发维度不匹配错误。

若模型中存在其他类似分组卷积层（如conv4），需重复上述步骤逐一修改。

内容的提问来源于stack exchange，提问作者Jiren Li