// the find
soumith/dcgan.torch
A torch implementation of http://arxiv.org/abs/1511.06434
The original Torch7 implementation of the DCGAN paper (Radford et al. 2015), by Soumith Chintala, one of PyTorch's creators. It's for anyone who wants to reproduce the actual paper results — face/bedroom generation, latent space walks, vector arithmetic on faces — rather than a generic GAN training template.
Ships the real artifacts from the paper, not just training code: pretrained CelebA and LSUN bedroom checkpoints, a generate.lua for sampling, and arithmetic.lua for the famous 'smiling woman - neutral woman + neutral man' latent arithmetic demo. Dataset loading supports folder/LSUN(LMDB)/ImageNet with async workers via nThreads, and the LSUN preprocessing pipeline (index generation, cropping) is actually included rather than assumed. As a reference for the DCGAN architecture conventions (strided/fractionally-strided convs, batchnorm placement, LeakyReLU in D) it's about as authoritative as it gets.
It's Torch7/Lua, a stack the ecosystem abandoned around 2017 in favor of PyTorch (built by this same author) — getting CUDA/cuDNN/luarocks working on a current machine is real archaeology, not a pip install. The LSUN loader is hardcoded to the bedroom class and requires editing a source line to change categories, which is a rough edge for anything beyond the documented example. No tests, no CI, and the last push in 2021 looks like housekeeping rather than actual maintenance — five years dormant on a dead framework means treat this as a historical reference, not something to build a current project on.