ImageClassification

Indoor vs outdoor image classification using TensorFlow/Keras and EfficientNet.

What was improved

• Consolidated repeated inference logic into a reusable module: image_classification/.
• Refactored all scripts into robust CLIs with argument validation.
• Removed deprecated training APIs (fit_generator) and added validation + early stopping.
• Fixed evaluation correctness (classification_report now receives y_true, y_pred in proper order).
• Cleaned dependency definitions and added a practical .gitignore.

Project structure

• train.py: model training
• pred.py: single-image inference
• run_evaluation.py: class-folder evaluation + CSV report
• unit_test.py: quick benchmark check on one indoor and one outdoor image
• image_classification/inference.py: shared prediction utilities

Installation

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Dataset layout

train.py expects:

images/
  indoor/
    img1.jpg
    ...
  outdoor/
    img2.jpg
    ...

Train

python train.py \
  --train-dir ./images \
  --output-dir ./training_1 \
  --epochs 10 \
  --batch-size 16

Model output is saved to ./training_1/saved_model.keras.

Inference

python pred.py \
  --input ./test_data/indoor/benchmark_in.jpg \
  --model-path ./training_1/saved_model.keras \
  --threshold 0.5

Evaluation

python run_evaluation.py \
  --class_1 ./test_data/indoor \
  --class_2 ./test_data/outdoor \
  --model-path ./training_1/saved_model.keras \
  --out-path ./evaluation.csv

Quick benchmark check

python unit_test.py \
  --indoor ./test_data/indoor/benchmark_in.jpg \
  --outdoor ./test_data/outdoor/benchmark_out.jpg \
  --model-path ./training_1/saved_model.keras

This command returns non-zero if predictions do not match expected classes.