Model Card for keerthikoganti/nn-image-automll
Multiclass image classifier trained with AutoKeras/AutoML on the augmented split of a class dataset of student-collected images.
Model Details
Model Description
Multiclass image classifier trained with AutoKeras/AutoML on the augmented split of a class dataset of student-collected images.
The model predicts the correct label among several categories from tabularized image data.
Artifacts include (1) a zipped Keras/TensorFlow model file and (2) JSON metrics for reproducibility.
- Developed by: Keerthi Koganti
- Model type: AutoML Neural Network AutoKeras/TensorFlow, hyperparameter and architecture search
- Language(s) (NLP): English
- License: Carnegie Mellon
- Task: Multiclass image classification
-Target column: label
-Framework: TensorFlow / AutoKeras
-Repo Artifacts: best_model.zip, metrics.json
Model Sources
- Repository: Anyuhhh/2025-24679-image-dataset
Uses
Direct Use
Classroom demonstration of AutoML for image classification
Baseline experiments to compare manual CNNs vs AutoML-optimized models
Teaching reproducibility and Hugging Face model sharing
Out-of-Scope Use
Production deployment
Generalization beyond the course dataset
Medical, safety-critical, or sensitive decision-making
Bias, Risks, and Limitations
Small sample size: Original dataset is small (≈30 images) augmented to ≈300.
Synthetic augmentation: May inflate validation performance relative to external evaluation.
Class bias: Distribution is uneven across labels.
Label noise: Augmentation inherits any mislabels.
Recommendations
Report both augmented-test metrics and, if possible, metrics on the original split.
Treat results as didactic only.
When extending, run calibration checks and additional validation.
How to Get Started with the Model
Use the code below to get started with the model.
import pathlib, shutil, zipfile
import huggingface_hub as hf
import tensorflow as tf
REPO = "keerthikoganti/nn-image-automll"
ZIPNAME = "best_model.zip"
dest = pathlib.Path("hf_download")
dest.mkdir(exist_ok=True)
Download zipped model
zip_path = hf.hf_hub_download(
repo_id=REPO,
filename=ZIPNAME,
repo_type="model",
local_dir=str(dest),
local_dir_use_symlinks=False,
)
Extract
extract_dir = dest / "best_model_dir"
if extract_dir.exists():
shutil.rmtree(extract_dir)
extract_dir.mkdir(parents=True, exist_ok=True)
with zipfile.ZipFile(zip_path, "r") as zf:
zf.extractall(str(extract_dir))
Load model
model = tf.keras.models.load_model(str(extract_dir / "best_model.keras"))
Example: predictions
preds = model.predict(X) # X = batch of images preprocessed as during training
Training Details
Training Data
Dataset: Anyuhhh/2025-24679-image-dataset
Splits: 80/20 random split (random_state=42) on the augmented data (~300 images)
Target column: label
Training Procedure
Library: AutoKeras (KerasTuner for hyperparameter search)
Search space: conv layers, filter counts, kernel sizes, dropout rates, dense layers, learning rate, optimizer
Training budget: 15 trials, max 300s
Early stopping: patience=4 epochs
Training Hyperparameters
- Training regime: Hyperparameters (best model)
Convolutional blocks: 3
Filters per block: [32, 16, 16]
Dense units: 64
Learning rate: 0.0005
Dropout: Disabled on all blocks
Evaluation
Testing Data, Factors & Metrics
Testing Data
Held-out 20% of augmented split 60 images
Metrics
Accuracy: 0.62
Confusion matrix included
Environmental Impact
Hardware: Google Colab CPU runtime
Training wall-time: ≤ 10 minutes
Estimated emissions: negligible
Model Card Contact
