nsfw-image-detection-ExecuTorch

nsfw_image_detection — ExecuTorch

A two-class image classifier: normal or nsfw. For filtering, moderation or parental controls that run on the phone, where sending the photograph to a server is the thing you were trying to avoid.

• Source: Falconsai/nsfw_image_detection — 86M parameters, ViT-Base/16, 12 layers, hidden 768, 224x224
• License: apache-2.0
• Input: pixel_values [1, 3, 224, 224] fp32 — RGB image resized to 224x224, scaled to [0,1], then normalised with mean=(0.5, 0.5, 0.5) std=(0.5, 0.5, 0.5)
• Output: logits [1, 2] — 0=normal, 1=nsfw. Softmax is left to the caller.

Variants

*Mac arm64, single process, median of 10. PyTorch eager fp32 on the same machine is 38.3 ms, so the Core ML build is 8.9x eager and the int8 build is the fastest portable one. fp16 is slower than fp32 here — XNNPACK emulates it — and is listed only because it halves the file.

How these were checked, and why the obvious check is not enough

Label agreement on ordinary photographs is free: the fp32 model calls all 24 of them normal at probability 1.000, and every build reproduces that to 0.0000. Correlation says 1.000000 for all four. None of that separates a good quantized build from one that only errs near the decision boundary — and the boundary is the only place the error changes an answer.

So each photograph is also walked along the gradient of the class margin until the margin crosses zero, bisected onto it, and the builds compared there. The number in the last column is how far the margin moves at that point. It answers the question a caller has: how confident does fp32 have to be before this build is guaranteed to agree with it?

Measured on this set, the closest photograph sits 7.88 logits from the decision, median 8.90. Every build's shift is smaller than that, so none of them could relabel any of these images — but the margin between them is real and only this test shows it.

The boundary inputs are perturbed photographs, not natural ones. They exercise the arithmetic where it decides something; they are not a claim about accuracy on real data.

What is not tested here

Every calibration photograph on this shelf is normal, and none were fetched for the other class. So what is verified is fidelity to the fp32 model, not the model's own accuracy on the class it was trained to find. If that matters for your use, evaluate the upstream model on your own data first; these files reproduce whatever it does.

Conversion

python convert/export_imgcls.py nsfw
python convert/check_imgcls.py nsfw int8

(conversion scripts: executorch-models)