its-zion-18/sign-image-autogluon-predictor

AutoGluon Sign Identification Predictor

This repository contains a trained MultiModalPredictor from the AutoGluon library, which was trained to identify signs from images. Which can also be found in the files and versions section under AutoML_for_Neural_Networks

Dataset

The model was trained on the ecopus/sign_identification dataset. The augmented split was used for training and validation, while the original split was used for the final evaluation of the model's performance.

Evaluation Results

The final performance of the best model on the original dataset is as follows:

• Accuracy: 1.0000
• Weighted F1: 1.0000

Files in this Repository

• autogluon_image_predictor.pkl: The trained MultiModalPredictor pickled using cloudpickle.
• autogluon_image_predictor_dir.zip: The zipped native AutoGluon predictor directory for portability.

Potential Errors

The augmented split in the ecopus/sign_identification dataset is specifically designed to be an artificially expanded version of the original split. The images in the augmented set are simple transformations—like rotations, flips, or slight color changes—of the images in the original set. The code then trains the model on a portion of the augmented data (df_aug_train) and evaluates it on the original data (df_orig). Because the model was trained on data that is derived directly from the evaluation data, it's not actually seeing truly "new" information during the final test. Which could be leading to data leakage and overfitting