Run anya against real models, and fix the twelve things that broke

[Karthik777]

The tiny graphs in nbs/fixtures prove the plumbing. They say nothing about what a real export declares, and it turns out most real exports declare something anya read wrong. evals/e2e.py downloads eight repos, runs them over five COCO and Ultralytics photographs plus three generated sounds, and asserts on the answers. Everything below was found by that script failing.

The one that mattered most

Three of the four ONNX repos tested declare their input as ['batch_size', 'num_channels', 'height', 'width'] — every axis symbolic, and named. anya read the channel axis from the sizes, found no integers, guessed NHWC, and fed channels-last pixels into a channels-first graph:

Non-zero status code returned while running Conv node. C: 256 kernel channels: 3

chan_axis now reads the names when the sizes cannot say. This is the commonest shape optimum emits, so it is the difference between anya working on the Hub and not.

The rest

What	Was	Is
webnn/yolov8n size	224 from preprocessor_config.json, into a graph fixed at 640	the size the graph states wins; a config cannot override it
Xenova/segformer-… task	classify, because the output is named logits	segment; logits is what both call their output and so decides nothing
SegFormer's 150 classes on a 128×128 grid	read as 128 channels of a 128×150 image	channels-first, and the mask goes back onto the original picture
Xenova/dinov2-small	a 98,688-long "embedding"	257 tokens of 384 mean-pooled to one vector
timm crop_size 224 with size 256	a 256 crop	a 224 crop out of a 256 short side, timm's crop_pct
resample	always bilinear	as the config asks; bicubic moved top-1 on one of five photos
yolo box units	assumed pixels	read off the range: two of three yolo exports ship 0..1
SSD boxes	normalised against the padded input, then scaled as if it were not	through scale_boxes, like every other detector
ssdlite320's twelve raw heads	decoded to []	raises at load, naming task=
yamnet's 521 sigmoid scores	softmaxed a second time, so silence read 0.004	0.801, because being in 0..1 is the test, not summing to 1
yamnet's per-frame output	flattened, so the argmax crossed frames	frames averaged, which is what yamnet documents
labels.txt at the repo root, weights under onnx/	class_15 for a bus	its 80 COCO names
--conf=0.6	'0.6' into a numpy comparison	0.6
a tool that raises	a traceback where a chat expects JSON	{"error": …}, and a non-zero exit from the CLI
find_similar(…, exclude=…)	exclude into ONNX Runtime's session options	a call keyword, like run() already treated it
read_labels('model.onnx')	100,877 class names of binary	refuses anything but .txt, .csv, .json

That --conf one: tools.py has from __future__ import annotations, so every annotation reaches _coerce as a string and no CLI flag was ever converted. The old test passed the real int and so tested nothing the CLI does.

Evidence that the answers are right, not just different

Nothing here asserts against numbers written down by hand. Two detector families check each other, and the classifier checks itself against the transform its weights were evaluated with.

• anya's tensor for mobilenetv4_conv_small equals timm's torchvision eval transform to max|diff| 0.000000, and SegFormer's and DINOv2's own processors to 0.0175, which is one uint8 step over the ImageNet std.
• A quantised efficientdet and a float yolov8n, different architectures on different runtimes, now put the same boxes in the same places: the bus in bus.jpg to max|diff| 31px of an 800px-wide box, and the two cats in 000000039769.jpg to 13.5px and 11.7px. Before the SSD padding fix the nearer cat was 47px out in y.
• A 440Hz sine through yamnet is Sine wave at 0.996; silence is Silence; noise is Static.
• lraspp on VOC gives class 8, cat, 36.8% of a picture of two cats. SegFormer on ADE20K gives building, sidewalk, bus for a photo of a bus.

13/13 passed

The eval is re-runnable and clean-runnable, which took fixing twice: once because classify over a folder read the sorted tree a previous run had written into it, and once because find_similar did.

Structure

• Model._finish resolves the labels, the task and the Prep for every runtime. All three repeated the same eight lines, and each new prep keyword had to be threaded through three files. _own_labels, _guess_task and _mk_prep are the hooks; ONNX overrides none of them. This is what CLAUDE.md already said the architecture was.
• read_labels takes a dict and inverts label2id, so hub.config_labels is gone.
• with_hub_defaults returns prep keywords rather than a fixed 4-tuple, so crop_pct and resample needed no new slot anywhere.
• find_models searches each word after the phrase, in search order: the Hub matches search against the repo id, so bird classifier with runtime='litert' found nothing where bird found a repo, and sorting the merged results by downloads buried it again under repos matching classifier.

The environment

.cursor/environment.json runs .cursor/install.sh, which installs the dev group rather than restating it: eleven seconds on a fresh pod, two warm. Two draft builds off this branch both succeeded, and two fresh agents booted from them ran nbdev-export clean, nbdev-test green, evals/e2e.py twice at 13/13, the CLI, and the install script again for idempotence.

The eval takes about six seconds on a pod booted from the tested build rather than downloading 190MB, because the snapshot carries the warmed HuggingFace cache. It downloads unauthenticated; an HF_TOKEN secret is not needed but would remove any rate-limit risk on a busy pod.

Each runtime notebook and the index now show a real repo id, marked eval: false, with the answers the eval asserts on. nbdev-test is green and nbdev-export is clean. No fixture changed shape.