---
language: 
- en
tags:
- automl
- image-classification
- autogluon
- cmu-course
datasets:
- keerthikoganti/lipstick-image-dataset
metrics:
- type: accuracy
- type: f1
model-index:
- name: Lipstick Detection (Neural Network AutoML)
  results:
  - task:
      type: image-classification
      name: Binary Image Classification
    dataset:
      name: keerthikoganti/lipstick-image-dataset
      type: classification
      split: augmented
    metrics:
    - type: accuracy
      value: 1.00
    - type: f1
      value: 1.00
  - task:
      type: image-classification
      name: Binary Image Classification
    dataset:
      name: keerthikoganti/lipstick-image-dataset
      type: classification
      split: original
    metrics:
    - type: accuracy
      value: 0.93
    - type: f1
      value: 0.93
---

# Model Card for Lipstick Detection (Neural Network AutoML)

This model performs **binary classification** of images into **lipstick** (1) vs. **no lipstick** (0).  
It was trained with **AutoGluon Multimodal AutoML**, which automatically explored different **neural network backbones** (ResNet18, ResNet34, EfficientNet-B0) under a fixed budget with early stopping.  
The best-performing backbone selected was **EfficientNet-B0**.

---

## Model Details

### Model Description
- **Developed by:** Xinxuan Tang (CMU)  
- **Dataset curated by:** Keerthi Koganti (CMU)  
- **Model type:** AutoML neural network (best = EfficientNet-B0)  
- **Language(s):** N/A (image dataset)  
- **Finetuned from:** `timm/efficientnet_b0` pretrained weights  

### Model Sources
- **Repository:** [Hugging Face Model Repo](https://huggingface.co/)  
- **Dataset:** [keerthikoganti/lipstick-image-dataset](https://huggingface.co/datasets/keerthikoganti/lipstick-image-dataset)

---

## Uses

### Direct Use
- Educational practice in **binary image classification**.  
- Experimenting with AutoML search over neural architectures.  

### Downstream Use
- Could be adapted for **teaching transfer learning** workflows.  

### Out-of-Scope Use
- **Not suitable for real-world cosmetics applications**.  
- Not for deployment in automated decision-making or safety-critical contexts.  

---

## Bias, Risks, and Limitations

- **Small dataset**: limited original images, heavy reliance on synthetic augmentation.  
- **Domain bias**: images are from a single source/product and background setup.  
- **Synthetic augmentation**: does not capture real-world variation in lighting, product types, or diversity of appearances.  

### Recommendations
Use primarily for **teaching and demonstration** purposes.  
Do not generalize conclusions beyond this dataset.  

---

## How to Get Started with the Model

```python
from autogluon.multimodal import MultiModalPredictor
import pandas as pd

# Load trained predictor
predictor = MultiModalPredictor.load("autogluon_efficientnet_b0/")

# Run inference on a new image
test_data = pd.DataFrame([{"image": "example.jpg"}])
print(predictor.predict(test_data))