mniny-inception-net

This is my version of inception net for mnist. I use keras with tensorflow for this project. It was created as part of a competition between friends (I won with the lowest categorical crossentropy). I created this architecture after the one detailed in Going Deeper with Convolutions. Major changes made include downscaling, simplifying kernel sizes, and adding Batch Normalization. I also use this snapshot ensembles object (thanks, Somshubra Majumdar!).

Notable Results:

Results gained after approx. 24 hours of training on a NVIDIA GTX 1060 6GB GPU. Some of these results are without ensembling. No data augmentation is used.

---

ensemble

Test loss (categorical crossentropy): 0.0100604627563

error: 0.330001115799%

---

individual

Test loss: 0.0148106931992

error: 0.34%

---

individual

Test loss: 0.0129321537635

error: 0.34%

---

individual

Test loss: 0.0130156670398

error: 0.34%

---

Usage

Training

from mniny_inception_net import train

run = 0
while True:
    train(run)
    run += 1

Evaluation

ensemble

from mniny_inception_module import evaluate_ensemble

# To evaluate ensemble of all models in weights folder:
evaluate_ensemble(Best=False)

# To evaluate ensemble of best models per training session:
evaluate_ensemble()

individual

from mniny_inception_module import evaluate

#Evaluate all models in weights directory:
evaluate(eval_all=True)

# Evaluate 'Best' models in weights directory:
evaluate()

Requirements

• Tensorflow
• Keras

Setting up an Environment:

1. Install Anaconda
2. pip install tensorflow-gpu keras