Error using activations with GoogLeNet in R2017b

The error using activations() with GoogLeNet in MATLAB R2017b usually stems from three issues: GoogLeNet is a DAGNetwork requiring exact layer names (like 'pool5-7x7_s1'), input images are not resized to [224, 224, 3], or the resulting 4D feature array is not flattened using squeeze().

Primary Root Causes in MATLAB R2017b

1. DAGNetwork Layer Name Conflict: Unlike linear networks such as AlexNet or VGG-16, GoogLeNet contains parallel branched Inception modules. Passing integer indices or generic layer names like 'fc' throws an error. You must pass exact string identifiers such as 'pool5-7x7_s1' (global average pooling) or 'loss3-classifier' (final classification layer).
2. Color Channel & Dimension Mismatch: GoogLeNet strictly enforces an input size of 224x224x3. If your test image is grayscale (2D) or not preprocessed, activations() will throw a dimension mismatch error. Grayscale images must be replicated across three channels using cat(3, img, img, img).
3. Unflattened 4D Output Tensor: Extracting features from 'pool5-7x7_s1' returns a 4D numeric array of shape [1, 1, 1024, N]. If passed directly into downstream classifiers such as fitcsvm(), MATLAB throws an invalid predictor matrix error unless converted to [1024, N] using squeeze().

Executable MATLAB Code: Correct GoogLeNet Feature Extraction

% =========================================================================
% Feature Extraction via activations() with GoogLeNet in MATLAB R2017b
% =========================================================================

% 1. Load the pre-trained GoogLeNet model
% Requires the "Deep Learning Toolbox Model for GoogLeNet Network" support package
net = googlenet; 

% 2. Inspect valid layer names in the DAGNetwork
% layerNames = {net.Layers.Name}'; % Uncomment to inspect all 144 layers

% 3. Read and preprocess test image
rawImage = imread('peppers.png');

% Ensure image has 3 channels (convert grayscale if necessary)
if size(rawImage, 3) == 1
    rawImage = cat(3, rawImage, rawImage, rawImage);
end

% Resize to exact GoogLeNet input dimensions [224, 224, 3]
inputSize = net.Layers(1).InputSize(1:2); % [224, 224]
processedImage = imresize(rawImage, inputSize);

% 4. Extract activations from the global pooling layer
% Valid feature layer in R2017b: 'pool5-7x7_s1' (produces 1024-dimensional feature vector)
featureLayer = 'pool5-7x7_s1'; 

try
    rawFeatures = activations(net, processedImage, featureLayer, ...
        'OutputAs', 'channels', ...
        'ExecutionEnvironment', 'auto'); % Uses GPU if available, else CPU
    
    % Squeeze the [1, 1, 1024] array into a flat 1D feature vector [1024, 1]
    featureVector = squeeze(rawFeatures);
    
    fprintf('Success! Extracted %d-dimensional feature vector.\n', length(featureVector));
    
catch ME
    fprintf('Error during activations extraction: %s\n', ME.message);
end

% 5. Batch Feature Extraction from an imageDatastore (Recommended for datasets)
% imds = imageDatastore('path_to_images', 'IncludeSubfolders', true);
% augmentedImds = augmentedImageDatastore(inputSize, imds, 'ColorPreprocessing', 'gray2rgb');
% batchFeatures = activations(net, augmentedImds, featureLayer, 'OutputAs', 'rows');
% size(batchFeatures); % Returns [numImages x 1024] directly, ready for SVM training

Key Diagnostic Checklist

• Support Package Check: In R2017b, if googlenet returns an unrecognized function error, open Add-On Explorer and install Deep Learning Toolbox Model for GoogLeNet Network.
• Using 'OutputAs', 'rows': In R2017b, adding the name-value pair 'OutputAs', 'rows' automatically flattens 4D tensors into a standard 2D observation matrix [N, numFeatures], eliminating manual squeeze() and reshape() operations.
• Layer Verification: If testing with alternative backbones like Inception-v3 or ResNet-50, verify the final pooling layer identifier using net.Layers(end-2).Name.