You can deploy this model and start earning money today!

0408happyfeet/
bracelet-detector-automl

Name: bracelet-detector-automl
Availability: InStock

Views

No views yet

Bracelet vs. Not-Bracelet — AutoML (24-679)

Compact image classifier trained with AutoGluon MultiModal using Ray Tune (ASHA early stopping). This model predicts whether an image contains a bracelet.

Dataset

Source: samder03/2025-24679-image-dataset
Task: Binary image classification (bracelet vs. not-bracelet)
Splits used here: train≈70%, val≈20%, test≈10% — test from augmented (hold-out 10%).

Training / AutoML

Library: AutoGluon MultiModal (MultiModalPredictor)
Search space: mobilenetv3_small_100, efficientnet_b0, resnet18, swin_tiny_patch4_window7_224, deit_tiny_patch16_224
Tuned hparams: learning rate, weight decay, batch size, max epochs; ASHA for early stopping
Compute budget: trials = 12, time_limit = None
Seed: 42

Best trial (as recorded)

{}

Results (test on `original` split)

{
  "has bracelet mod": {
    "precision": 1.0,
    "recall": 1.0,
    "f1-score": 1.0,
    "support": 14.0
  },
  "no bracelet mod": {
    "precision": 1.0,
    "recall": 1.0,
    "f1-score": 1.0,
    "support": 19.0
  },
  "accuracy": 1.0,
  "macro avg": {
    "precision": 1.0,
    "recall": 1.0,
    "f1-score": 1.0,
    "support": 33.0
  },
  "weighted avg": {
    "precision": 1.0,
    "recall": 1.0,
    "f1-score": 1.0,
    "support": 33.0
  }
}

Usage

Download the zip artifact and load locally:

python

from autogluon.multimodal import MultiModalPredictor
import pandas as pd

predictor = MultiModalPredictor.load("predictor_native")  # unzip content to ./predictor_native first

# Predict a few images
df = pd.DataFrame({"image": ["path/to/image1.png", "path/to/image2.png"]})
preds = predictor.predict(df)
proba = predictor.predict_proba(df)
print(preds, proba.head())

Trained in a class assignment (24-679). Dataset license: MIT (see dataset card).

0408happyfeet/bracelet-detector-automl