This study proposes and implements a method for mapping supported Keras convolutional neural network models for classification to a weighted graph at the level of individual neurons. The reviewed literature lacks a systematic procedure covering the diverse connectivity patterns found in convolutional architectures. Mapping rules are defined for standard, grouped, dilated, depthwise, and separable convolutional layers, as well as pooling layers and layers of the classification head. The implementation is verified using structural invariants and validated for topology and weights using small reference models. Practical applicability and scalability are evaluated on models based on AlexNet, VGG11, and MobileNetV1 at two input sizes, and a selected set of complex-network metrics is computed for the resulting graphs. The results show that the method consistently maps supported Keras convolutional neural network models to large weighted graphs suitable for further topological analysis.
