azcrawlerpy-recaptcha-google

The default model for azcrawlerpy's recaptcha_v2 captcha solver: TesseraNet, one network for both grid kinds, as tesseranet_3x3.onnx and tesseranet_4x4.onnx (one graph per grid kind, from the same weights), cell_head.json naming each channel's category and threshold, and the provenance card tesseranet_3x3.json written when a model is promoted through captcha-training. azcrawlerpy[recaptcha] installs it; there is nothing to import.

It is published apart from azcrawlerpy so that code releases do not upload the weights again, and a retrained model ships without an azcrawlerpy release.

Versions

• Major: the contract azcrawlerpy relies on (file name, format, class names, input size). azcrawlerpy pins the major it reads, so a major bump needs an azcrawlerpy release that accepts it.
• Minor: retrained weights. Existing azcrawlerpy installs pick them up on upgrade.

Releasing a model

promote() in captcha-training writes the weights and card into azcrawlerpy_recaptcha_google/ in this checkout, only when they beat the model already here. Commit both, bump the minor version, check with uv sync --extra dev && uv run pytest, and publish with build.txt. It builds the wheel only: an sdist would upload the same weights a second time against the project's PyPI quota.