eandujar/nnl_automl_model

Extrudability AutoGluon TIMM Classifier

This repository contains the best-performing TIMM image-classification model selected from a randomized model sweep using AutoGluon MultiModalPredictor.

Model

• Framework: AutoGluon MultiModal
• Image model library: TIMM
• Problem type: Binary classification
• Target: extrudable
• Selection metric: Balanced accuracy
• Best TIMM checkpoint: repghostnet_200
• Validation balanced accuracy: 0.500000

Model selection

Multiple pretrained TIMM image-classification backbones were evaluated using the same fixed training and validation splits.

Each candidate was trained using:

• AutoGluon MultiModalPredictor
• timm_image as the only model branch
• pretrained TIMM weights
• the same training dataset
• the same fixed validation dataset
• balanced accuracy as the selection metric

The checkpoint with the highest validation balanced accuracy was retained:

repghostnet_200

Only the winning predictor and the files required for inference are included in this repository.

Loading

from huggingface_hub import snapshot_download
from autogluon.multimodal import MultiModalPredictor

model_path = snapshot_download(
    repo_id="eandujar/nnl_automl_model"
)

predictor = MultiModalPredictor.load(model_path)