hujie-frank/senet

This repository contains the official code and trained models for Squeeze-and-Excitation Networks (SENet), a research project from Momenta and the University of Oxford. The work was published at a major computer vision conference (CVPR 2018) and won first place at the ILSVRC 2017 image classification challenge, a large annual competition where teams try to correctly identify objects in photographs. The core idea is a small addition to existing image recognition network architectures. Standard image recognition networks process visual information channel by channel, where each channel captures a different type of pattern in an image. A Squeeze-and-Excitation block adds a short pathway that looks at all channels together, figures out which ones are most informative for a given image, and scales each channel's contribution up or down accordingly. This lets the network focus its attention on the most relevant features. The name comes from a "squeeze" step that compresses spatial information and an "excitation" step that recalibrates channel importance. The implementation uses Caffe, a deep learning framework designed for GPU-accelerated training. The repository includes configuration files for several model variants: SE-BN-Inception, SE-ResNet at multiple depths, SE-ResNeXt, and the flagship SENet-154. Pre-trained model weights for all of these are available for download. The table of results shows top-1 and top-5 error rates on the standard ImageNet benchmark. The README also notes two engineering choices made for efficiency: a custom combined layer that reduces memory use in networks with skip connections, and a rewritten global average pooling operation that runs faster on GPU than the version built into Caffe. Third-party re-implementations in TensorFlow, PyTorch, MXNet, and other frameworks are linked but not maintained in this repository.