A Flax (Linen) implementation of ResNet (He et al. 2015), Wide ResNet (Zagoruyko & Komodakis 2016), ResNeXt (Xie et al. 2017), ResNet-D (He et al. 2020), and ResNeSt (Zhang et al. 2020). The code is modular so you can mix and match the various stem, residual, and bottleneck implementations. The models are tested to have the same intermediate activations and outputs as the torch.hub implementations, except ResNeSt-50 Fast, whose activations don't match exactly but the final accuracy does.