Original September floor-model components
The five original, unchanged September checkpoints used by the retained SFG combined floor pipeline. This release preserves the original models separately from October improvement experiments. Every checkpoint has a SHA-256 checksum in MANIFEST.json.
This initial release contains reusable model components and feature code. It is not yet a portable installation of the entire combined floor pipeline. Native polygon construction, all floor-completion stages, final ownership assignment, and the takeoff application are not included. Loading these weights alone will not reproduce the complete retained system's rooms. The weight files themselves are not distributed here; they are kept in a private repository. The table below documents what each component is. Full pipeline packaging and cloud validation are ongoing.
| File | Function |
|---|---|
roles-e070b.pt |
DeepLabV3 ResNet50 visual context: background, wall, fixture, annotation, hatch |
walls-e068.pt |
SegFormer B2 wall-context segmentation |
wall_v0_noband.pkl |
Original 36-feature vector wall classifier |
seal_v0.pkl |
Original paired doorway-seal classifier |
expanded_head.pkl |
Original September visual-context nonwall veto, including frozen wall classifier |
The role features supplied to the nonwall classifier are role_wall, role_fixture, role_annotation, role_hatch, and role_wall_frac. They are segment aggregates, not the raw five class probabilities. The veto threshold is 0.95; wall admission is 0.35; seal admission is 0.5. Floor ownership and physical wall-body exclusion are separate downstream operations.
Download and verify
hf download Kentucky-ai/september-floor-model --local-dir september-floor-model
cd september-floor-model
python load_original.py
pip install -r requirements.txt
The verification command only hashes the files. It does not execute a model.
Use components
from load_original import load_classifier, load_visual, score_named_features
wall_packet = load_classifier('wall_v0_noband')
required_names = wall_packet['features']
# X must contain the original segment features in the supplied name order.
# probabilities = score_named_features('wall_v0_noband', X, feature_names)
roles = load_visual('roles-e070b', device='cuda')
walls = load_visual('walls-e068', device='cuda')
Only load the pickle files from this trusted repository and verify their hashes first. Classifiers were serialized with scikit-learn 1.9.0. Visual inputs require the original preprocessing: RGB, ImageNet normalization, physical scale 32 pixels/foot, 1024-pixel tiles and 768 stride. Wall and role maps use different overlap aggregation; a raw image forward pass is not a full-sheet reproduction. Feature sources are included for inspection and integration. The portable checkpoint loaders passed a Lambda A10 smoke test: all three classifiers loaded and both visual networks loaded strictly and produced finite outputs. See CLOUD-VALIDATION.json. This validates checkpoint compatibility, not full-sheet accuracy or the complete combined pipeline.
Limits and data
No private drawings, labels, training arrays, exported project geometry, credentials, or customer review screenshots are distributed. Known pipeline failures include merged rooms, wall intrusion, and fixtures becoming separate floor regions. No independently measured 95% room-accuracy claim is made. Checkpoints are preserved as research/evaluation assets, not a certified takeoff system.
License
Public download does not imply unrestricted commercial use. The E068 component derives from NVIDIA SegFormer and retains its research/evaluation-only license. See LICENSE.md, licenses/NVIDIA-SegFormer.txt, and licenses/torchvision-BSD.txt.