ForkPoint demo — where does the INT8 conversion leave the FP32 model?

Two versions of the same model — PathoCNN (PathMNIST, 9-class histology) as shipped FP32, and its static INT8 (QDQ) quantization — run side by side in your browser on 24 probe images. ForkPoint compares the activation at every graph node the two versions share and shows where they diverge, and which inputs flip their prediction. Nothing is uploaded; the images never leave your device.

Read this first. The probe set is selected: test rows 128–149 (22 consecutive) plus rows 286 and 338 — the two inputs (out of the library demo's 240 probes) whose predictions flip under INT8. So this page shows 2 flips in 24 by construction; the honest flip rate on the full probe set is 2/240, with accuracy 0.9000 → 0.8958 (full record in the ForkPoint repo). The demo models are the shipped graphs with intermediate tensors additionally declared as outputs so the browser can read them; weights and computation are unchanged (logits verified bit-identical).

Loading…

Aggregate over the 24 probes

Waiting for the models…

NodeOpMean rel. divergence (live)Recorded (Python)

Divergence = relative L2 between the two models' activations at that node, averaged over the probes. The fork point — the earliest node over the 0.01 threshold — is highlighted.

One probe at a time

Waiting…

Models: PathoCNN FP32 + its INT8 quantization (model card, CC BY 4.0; data: PathMNIST / MedMNIST v2, CC BY 4.0). Library: ForkPoint. Not a diagnostic tool.