This repository contains the TensorFlow implementation of Complement Objective Training introduced in the following paper:
COMPLEMENT OBJECTIVE TRAINING.
Hao-Yun Chen, Pei-Hsin Wang, Chun-Hao Liu, Shih-Chieh Chang, Jia-Yu Pan, Yu-Ting Chen, Wei Wei, Da-Cheng Juan.
https://openreview.net/forum?id=HyM7AiA5YX
- Python 3.5
- tensorflow 1.13.2
- Pillow
Please find details dependency in requirements.txt
To prepare dataset, please find the folder structure below:
<path to data dir>
├── test
└── train
├── 14
│ ├── test_14_Left.tif
│ ├── test_14_Right.tif
│ └── test_14_Top.tif
├── 255
│ ├── test_255__left.tiff
│ ├── test_255_right.tiff
│ └── test_255_top.tiff
├── <label id in int>
│ ├── <images of label id>.tiff
│ ├ ..........
├ ...........
├── 37
│ ├── test_37_left.tiff
│ ├── test_37_right.tiff
│ └── test_37_top.tiff
└── 76
├── test_76_left.tif
├── test_76_right.tif
└── test_76_top.tif
To convert dataset into tfrecords please find the following commands:
python sem_to_tfrecord.py \
--raw_data_dir=${DATA_DIR}
For getting baseline results
python main.py \
--data_dir=${DATA_DIR} \
--model_dir=${TRAIN_DIR} \
--train_epochs=200 \
--epochs_per_eval=10 \
--batch_size=128 \
--resnet_size=101
For training via Complement objective
python main.py \
--data_dir=${DATA_DIR} \
--model_dir=${TRAIN_DIR} \
--train_epochs=200 \
--epochs_per_eval=10 \
--batch_size=128 \
--COT \
--resnet_size=101
Please find the detail runscript under run.bash
The following table shows the best test accuracy in a 200-epoch training session.
| Model | Baseline | COT |
|---|---|---|
| ResNet-110 | 93.49% | 93.68% |
The CIFAR-10 reimplementation of COT is adapted from the tensorflow/models repository by tensorflow.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent