Built this: Modern DCGAN reproducibility study
A DCGAN (Radford et al., 2015) reproduction that keeps the all-convolutional architecture but swaps in LSGAN loss, spectral normalisation and label smoothing, plus an analysis toolkit for filters and feature maps.
On 28 March 2026 I committed Modern-DCGAN-Reproducibility, a study of the DCGAN paper (Radford et al., 2015), implemented in PyTorch with Jupyter notebooks.
What changed from the paper
The README compares the paper’s choices with this implementation:
| component | 2015 paper | this repo |
|---|---|---|
| loss | binary cross-entropy | LSGAN (least squares) |
| discriminator normalisation | batch norm | spectral normalisation |
| initialisation | Normal(0, 0.02) | Xavier/Kaiming hybrid |
| regularisation | none | discriminator dropout (0.3) |
| labels | hard 0/1 | label smoothing (0.1/0.9) |
The all-convolutional generator and discriminator from the paper are kept.
Analysis layer
A GANAnalyzer class uses PyTorch hooks to:
- plot the RGB filters of the discriminator’s first layer, after a 25-epoch CIFAR-10 run
- show mid-layer generator feature maps (e.g. at 16×16)
- draw histograms of D(x) versus D(G(z)), to watch for mode collapse
It also produces a 4-panel training dashboard.
Where it connects
I wrote it before reaching the generative AI phase. When the phase arrives in June 2026, the GANs and DCGAN notes come back to it.
#gan#generative-ai#pytorch#reproducibility
Sources
- SahajIVVIX-1/Modern-DCGAN-Reproducibility (GitHub) — GitHub
- First commit c855489 — GitHub
- Radford, Metz & Chintala (2015), Unsupervised Representation Learning with Deep Convolutional GANs — arXiv
This is a summary written from the linked public sources. Figures and numbers are as reported there.