Convolutional Neural Network in PyTorch

Implement the inference path of a small convolutional neural network used to classify frames from a Bosch AUTODOME IP starlight 7000i camera. Do not use PyTorch convolution or pooling helpers. Implement the tensor operations directly with Python lists so that padding, stride, channel aggregation, activation, and pooling behavior are explicit.

The network has one convolutional layer, ReLU activation, non-overlapping max pooling, flattening, and a fully connected output layer. Return the index of the class with the largest logit. If logits tie, return the smallest class index.

Formal Specification

Implement cnn_predict(image, kernels, biases, dense_weights, dense_bias, stride, padding, pool_size).

• image is a 3D list with shape [input_channels][height][width].
• kernels is a 4D list with shape [output_channels][input_channels][kernel_size][kernel_size].
• biases contains one value per output channel.
• dense_weights has shape [classes][flattened_features].
• dense_bias contains one value per class.
• Convolution uses zero padding and the specified stride.
• Apply ReLU before max pooling.
• Pooling uses square windows of size pool_size and stride pool_size.
• Return an integer class index.