This repository is developed by @penguin1214 and @Paper99.

Code structure is inspired by pytorch-cyclegan and BasicSR.

Notice

• num_workers in options/train/train.json should fix at 1. If you want to use multiple num_workers. You should firstly install hdf5 package with parallel settings. You can find solutions from here

Running now

• VDSR baseline : Every epoch takes 5 minutes. 80 epochs totally.

TODO

Zhen Li

Deadline: 7.10

• Check code for VDSR baseline
• VDSR : complete Kmeans train

Qilei Li

Deadline: 7.10

• Check K-means weight
• Carefully read VDSR
• Check network architecture of K-VDSR

Requirements

• Python 3.6
• Pytorch
• MatConvNet(It is included in scripts\Test, you should complie it firstly)

Code architecture:

-- data (对数据进行操作)

-- datasets (存放数据集)
-- DIV2K
-- VOC2012

-- models (算法模块)
models.modules --> modules within networks
models.modules.blocks --> basic blocks

models.modules.xxx_arch --> 特定网络的 building blocks 和 网络的完整结构
models --> solver classes
models.base_model --> base class solver
models.networks --> utilitie

-- options (输入参数)
-- train (训练参数)
-- test (测试参数)

-- experiments (存储结果)
-- train (训练结果)
-- test （测试结果）

-- scripts (运行脚本)

-- utils (其它)

-- tmp_deprecated (暂时弃用的文件)

-- train.py（训练代码）