CarClassifier

This repository gathers the code for Stanford car image classification from the in-class Kaggle challenge.

Using EfficientNetB7 (noisy-student) with AutoAugmentation + MixUp and SGD + Lookahead optimizer + 1 cycle cosine annealing learning rate scheduler.

Without any ensemble models or extra training data.

Final submission score of 0.95920 which places me second on the final leaderboard.

Hardware

• Ubuntu 18.04.4 LTS
• Intel(R) Xeon(R) Gold 6154 CPU @ 3.00GHz
• 1x NVIDIA Tesla V100

Getting Started

Enviornment

virtualenv .
source bin/activate
pip3 install -r requirements.txt

Dataset Preparation

Join the competition and download the dataset.

cd CarClassifier
kaggle competitions download -c cs-t0828-2020-hw1
unzip cs-t0828-2020-hw1
python dataset.py

CarClassifier
  +- training_data/
  +- testing_data/
  +- training_labels.csv
  +- tmp/
  +- logs/
  +- cutmix_keras.py
  +- dataset.py
  +- EfficientNetB7.ipynb
  +- Normalization_testing.ipynb
  +- panda.jpg
  +- ResNet50.ipynb

Results

Model	Size	Batch	Methods	Testing Accuracy
ResNet50	256	16	Hflip only	0.89
EfficientNetB3	456	16	+ Rotate 10	0.915
EfficientNetB7	456	4		0.932
EfficientNetB7	456	4	AutoAugment	0.944
EfficientNetB7	456	4	wd1e-3 to 1e-4	0.949
EfficientNetB7	456	4	wd1e-4 to 1e-5	0.951
EfficientNetB7	456	4	+ RAdam	0.952
EfficientNetB7	456	4	+ SGD_Lookahead	0.954
EfficientNetB7	600	4	+ Cutout	0.956
EfficientNetB7 (noisy-student)	600	4	- Cutout + Mixup	0.9592
EfficientNetB7 (noisy-student)	600	4	+ Dropout 0.5	0.9594