An Inception module is a multi-branch convolutional block that combines different receptive-field sizes. Its dimension-reduced form applies 1×1 convolutions before 3×3 and 5×5 filters, reducing input channels and therefore parameters; another 1×1 convolution follows pooling before branch outputs are concatenated. section 11.2.2
The design connects local operations to an efficient image-classification architecture:
- GoogLeNet stacks nine Inception modules into a 22-layer network and uses global average pooling at the end. section 11.2.2
- Two auxiliary classifiers apply softmax to intermediate module outputs and add weighted auxiliary losses to the final loss. section 11.2.2
PyTorch implementation
A dimensionally reduced block can be implemented from scratch with separate 1×1, 3×3, and 5×5 branches, reduction convolutions, pooling projection, batch normalization, and ReLU. The implementation is primarily instructional because complete architectures are available through torchvision. section 11.2.2
