// the find
zalandoresearch/fashion-mnist
A MNIST-like fashion product database. Benchmark :point_down:
A 70k-image, 28x28 grayscale clothing dataset (10 classes: shirts, shoes, bags, etc.) built as a drop-in replacement for MNIST. It's for anyone who needs a classification toy dataset that's slightly harder than MNIST without changing their existing pipeline code.
Genuine drop-in compatibility — same binary format, same train/test split sizes as original MNIST, so existing MNIST loaders work unchanged. It's already a built-in dataset in basically every ML framework (Keras, PyTorch, TF Datasets, HuggingFace, MXNet), so you often don't even need to clone this repo. Ships an actual benchmark harness (scikit-learn across 129 classifiers, no deep learning) with reproducible results, not just marketing claims, plus a large community-submitted table of CNN/ResNet/WRN results for comparison. It's meaningfully harder than MNIST (top CNNs land ~93-96% vs 99.7%), so it separates model quality better.
No commits since mid-2022 — fine for a static dataset, but the Dockerfile and benchmark runner are unmaintained if either breaks on newer dependencies. The community benchmark table is unverified and self-reported; the README admits results haven't been checked, so treat the numbers as anecdotal, not a leaderboard. The dataset itself is as simple as MNIST in every other respect (no bounding boxes, no color, no real-world noise), so it's already considered too easy for modern computer vision work and mostly useful as a quick sanity-check dataset, not a serious benchmark. Most of the actual utility now lives in other frameworks' loaders rather than this repo, so there's little reason to interact with the code here beyond grabbing the raw files.