Sample Image Classification

An implementation of image classification based on ResNet

Requirement

• Python 3.6.5
• Tensorflow 1.10.1
• Pillow 5.0.0
• numpy 1.14.5
• opencv-python 3.4.3.18

Dataset

• STL-10 dataset

Training

1. Download STL-10 dataset

2. Unzip the dataset zip file

tar -xvf ./path/to/file/stl10_binary.tar.gz

3. Train the model

python main.py --tr_img ./path/to/file/test_X.bin --tr_label ./path/to/file/test_y.bin --te_img ./path/to/file/train_X.bin --te_label ./path/to/file/train_y.bin

Test

python main.py --mode test --pre_trained_model ./model/model_best_on_val --te_img ./path/to/file/train_X.bin --te_label ./path/to/file/train_y.bin

Making grad cam

python main.py --mode grad_cam --pre_trained_model ./model/model_best_on_val --te_img ./path/to/file/train_X.bin --te_label ./path/to/file/train_y.bin

Details

• The number of parameters : 2,952,928 (less than 3M)
• Top 1 accuracy(test data) : 78.3%
• I used learning rate scheduling (49-th line in res_model.py)
• I used two data augmentation(The functions (data_aug(), extract_patch()) in util.py)

1. With random horizontal flip
2. With single scale jittering

• I used separate training set, validataion set, and test set (The function split_data() in util.py)

Experimental Results

For each train batch, these are the train accuracy and train loss graphs

If you want to see these results in tensorboard,

tensorboard --logdir=./log

log.txt in log directory : Training log file, which measured every epoch for total train data and validation data

Grad Cam Results

Correctly predicted samples

Input	Grad cam

Wrongly predicted samples

Input	Grad cam

Reference

[1] https://github.com/mttk/STL10

• The functions read_labels and read_all_images in util.py are borrowed from [1]

[2] https://github.com/cydonia999/Grad-CAM-in-TensorFlow

• The function save_cam in util.py is borrowed from [2]