RealWaste TV3 E6/E8 inference bundle

Private artifact repository maintained by Thangtb (TV3) for Project 11. Contains the three already trained models used in E6/E8. No images or dataset splits are uploaded. This is a custom Python/Joblib pipeline; the project src package and pinned dependencies are required. Hosted inference is not configured.

Models and evaluation

Role File Pipeline
Global E6 winner / E8 best models/hsv_none_random_forest.joblib HSV joint 16³, scaler without centering, sparse CSR, RF 200 trees
E8 HOG models/hog8_none_svm.joblib HOG cell8, scaler, PCA128, SVC RBF C=10
E8 LBP models/lbp_none_random_forest.joblib Uniform LBP P8 R1/2, scaler, RF 200 trees

All three use preprocessing preset none: resize raw RGB to 128×128, preserving the saved preprocessing configuration. No fit occurs during inference. The custom InferenceModel stores the feature extractor, preprocessing dataclass, reducer, classifier, class names and trial metadata. Dependency versions and source hashes are in provenance.json (HOG/LBP) and resume_2.json (HSV), not inside the Joblib metadata.

RealWaste: 4752 images, fixed train/validation/test 3322/716/714, 9 classes, seed42. Scaler/PCA/codebooks/classifiers fit train only; 154 validation trials selected the global winner. Test was evaluated after selection was locked.

  • HSV+RF validation macro-F1: 0.722181051317698.
  • HSV+RF test macro-F1: 0.7839002959894169.
  • HSV+RF test accuracy: 0.7871148459383753.
  • E8 uses these three frozen models, 14 transformations, all 714 test images. Full-image background bias remains a limitation; this result does not establish robustness on new backgrounds or cropped/masked objects.

selection.json is the original unmodified selection record. It contains metadata for all 33 family winners; only the three entries in e8_models are included here. The remaining 30 model files are intentionally absent.

Integrity and environment

Verify SHA256 from checksums.sha256 before loading. The HSV file must match 5bf3385816d0d51fb5a5a54079321851a58c566859d7e0c70843415cd0268598. Use the immutable Hub commit recorded in the project's config.yaml and handoff.

Runtime used: Python3.12.3, NumPy2.1.3, pandas2.2.3, OpenCV4.10.0.84, scikit-image0.24.0, scikit-learn1.5.2, SciPy1.14.1, joblib1.4.2, threadpoolctl3.7.0. Install the project requirements.txt for the Hub loader/CLI too. The source checkout is https://github.com/canhtungdz/trash-classify, TV3 branch feat/tv3-e6-e8; source delivery before Hub integration was commit 46cf3efee3dd2a993c4be6fc7d45580470aac99d.

Input/output and use

Input is RGB uint8 H×W×3. Batch API accepts a sequence of RGB arrays of different sizes. Do not preprocess or resize before calling the saved pipeline, except for the explicitly documented E8 protocol (raw resize128 → transform → saved pipeline). Outputs are class names in prediction order, not class IDs: Cardboard, Food Organics, Glass, Metal, Miscellaneous Trash, Paper, Plastic, Textile Trash, Vegetation.

from src.classify import load_model_from_hub

# revision is the immutable Hub commit from config.yaml, not a moving branch name.
model = load_model_from_hub(
    repo_id=registry['repo_id'],
    filename=registry['models']['hsv'],
    revision=registry['revision'],
    expected_sha256=registry['sha256']['hsv'],
)
label = model.predict_one(img_rgb)

Authenticate with hf auth login using an account with access to this private repo. Never share a token in source code. Joblib uses pickle and can execute code: only load this trusted bundle after checksum verification. The loader downloads/cache files once at load time; predictions reuse the loaded pipeline.

For E8, use the project's prediction_adapters_from_hub or src.classify e8 --model-source huggingface. The TV1 evaluator/geometry modules remain on TV1's separate branch until the team integrates them; they are not copied into this bundle.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support