Docs / Demo projects
Data curation
From a raw pile of images to a training-ready set: quality, duplicates, embeddings, label checks, a review grid, a split and a before and after report.
assets/images/site/demo-data-curation.jpg1600 × 900 pxThe Review grid of a fruit image set: cards with flags for blurred, duplicate and wrong label, a duplicate shown next to the image kept, and a before and after report of the classes.Two public image datasets cleaned with the same blocks: broken, blurred and duplicated pictures and wrong labels are flagged, reviewed in a grid, and the rest is split into training and validation sets with a report of what changed.
The flows
| Flow | What it does |
|---|---|
| 1 · Fruit: download and curate | FruitNet (good, bad and mixed quality of six fruits) downloaded once, checked, de-duplicated, labels checked with CLIP, split and reported. |
| 2 · Zebras: read from Hugging Face and curate | A sample of GZCD (Grevy's zebras, Kenya) read at an exact commit, checked, de-duplicated and sorted by CLIP into zebras, giraffes and empty frames. |
| 3 · Same step, any accelerator | The same images embedded on the CPU and on an NVIDIA GPU, images per second side by side; add a Tenstorrent worker for a third. |
What you need
Any worker; embeddings are faster on an NVIDIA GPU or a Tenstorrent card. The worker needs Run code from projects turned on (Computing > your worker > Worker settings).
Try it
- Open Projects > Demos and choose Data curation. Read its readme.
- Open flow 1 in the Blueprint and press Run: it runs as it is on your own workers. Switch to the View page to watch it. Its runs and results are yours only.
- To change it, choose Clone into my workspace: a copy goes into your workspace, with its readme, blueprints, the code of its blocks and its files, saved to a GitHub repository of yours (link GitHub first, under Profile > GitHub).
The blocks behind it
A model is only as good as the images it learns from. The curation blocks check every image of a set on your worker, flag the ones to look at, let a person decide in a grid, and write a clean set split into training and validation, with a report of what was removed and why. The images stay on the worker; only small previews and tables reach EsaSim.
Build the flow yourself
- Bring the images in. Use a Dataset block for a folder your worker shares, Hugging Face dataset for a dataset on the Hub at a fixed commit, or Download files for a link, a Zenodo record, a Mendeley Data dataset or a GitHub release. Folder names are often the labels (apple_good, zebra).
- Check them. Connect the images to Image quality check, Embed images, Find duplicates and Check labels. Each writes a table with one row per flagged image and the reason.
- Look at the flagged ones. Image previews packs small previews of the flagged images (and a sample of the rest) with what the tables say about each. Connect it to a Review grid view.
- Decide. In the Review grid, keep, remove or relabel each image, then Save decisions as
review.csv. - Write the clean set. Curate and split removes the flagged images (your decisions win over the checks), balances the classes and writes
train/andval/with a manifest. Send it to Save as dataset or Publish to Hugging Face.
What each check finds
| Block | Flags | How |
|---|---|---|
| Image quality check | broken, small, blurred, dark, bright, flat | Sharpness (variance of the Laplacian) against the typical sharpness of the same folder, brightness, contrast and the share of nearly black or white pixels. A few hundred images per second on any worker. |
| Embed images | (no flags: vectors for the others) | CLIP ViT-B/32 or DINOv2 small turn each image into a vector; on a Tenstorrent card, an NVIDIA GPU or the CPU. With CLIP, Prompts give each image the label that fits it best, with no training. |
| Find duplicates | duplicate, with the image kept | Perceptual hashes catch copies, resized or recompressed versions and small crops; with embeddings and a Similarity (0.97 for near copies, 0.9 for the same scene), images that look the same join the group. Each group keeps the largest, sharpest image. |
| Check labels | label (with the label suggested), outlier | The most similar images vote with their labels; an image whose own label loses the vote is flagged. A confusion table shows which labels are mixed up. |
| Embedding map | (a map) | UMAP or PCA lays the images out on a plane, with their flags, for a Chart of points: clusters, classes that mix, odd images. |
The Review grid
Each image is a card with its picture, name, label, the label it looks like, a badge per flag and the reason. Filter by decision, flag, label or name. Click a picture for a large view; a duplicate shows the image kept next to it.
| Key | Action |
|---|---|
| K / R | Keep or remove |
| L | Give another label |
| U | Undo |
| Arrows, Enter | Move between cards, enlarge one |
Decisions stay on your computer until you clear them. Save decisions gives review.csv (path, decision, label). Keep it in a dataset and connect it through a Dataset block to the Review input of Curate and split, then run again: your decisions override the checks.
What Curate and split writes
train/<class>/…andval/<class>/…, split by class with Validation %, at most Most per class each, the same way every time for the same Seed.manifest.csv, and the licence and credit files of the source.- A report (
curation.md,curation.jsonand CSV tables for charts): images in, removed by reason, kept, and the classes before and after. removed.csv: every image removed and why.
Sources and licences
Hugging Face dataset resolves a branch or tag to its commit, so the run records exactly which version it read, and Most files takes a repeatable sample for a demo-sized set. Download files keeps downloads in the worker's cache, reads only the files it keeps out of large zips when the server allows it, and writes a manifest with each file's source and checksum. Both write the licence next to the files. Publish to Hugging Face makes each publish one commit, with a card naming the licence and the credit. Respect the licence of the data you use and publish.
Save your Hugging Face token once in Profile > Keys: blocks that read private datasets or publish use it when you run a flow.