The overall code framework is shown in the following figure. It mainly consists of four parts - Config, Data, Model and Network.

Let us take the train commad python train.py -opt options/train/train_esrgan.json for example. A sequence of actions will be done after this command.

• train.py is called.
• Reads the configuration (a json file) in options/train/train_esrgan.json, including the configurations for data loader, network, loss, training strategies and etc. The json file is processed by options/options.py.
• Creates the train and validation data loader. The data loader is constructed in data/__init__.py according to different data modes.
• Creates the model (is constructed in models/__init__.py according to differnt model types). A model mainly consists of two parts - [network structure] and [model defination, e.g., loss definition, optimization and etc]. The network is constructed in models/network.py and the detailed structures are in models/modules.
• Start to train the model. Other actions like logging, saving intermediate models, validation, updating learning rate and etc are also done during the training.

Moreover, there are utils and userful scripts. A detailed description is provided as follows.

1. Config
2. Data
3. Model
4. Network
5. Utils
6. Scripts

• json file is used to configure options and options/options.py will convert the json file to python dict.
• json file uses null for None; and supports // comments, i.e., in each line, contents after the // will be ignored.
• Supports debug mode, i.e, model name start with debug_ will trigger the debug mode.
• The configuration file and descriptions can be found in options.

• A separate data loader module. You can modify/create data loader to meet your own needs.
• Uses cv2 package to do image processing, which provides rich operations.
• Supports reading files from image folder or lmdb file. For faster IO during training, recommand to create lmdb dataset first. More details including lmdb format, creation and usage can be found in our lmdb wiki.
• data/util.py provides useful tools. For example, the MATLAB bicubic operation; rgb<-->ycbcr as MATLAB. We also provide MATLAB bicubic imresize wiki and Color conversion in SR wiki.
• Now, we convert the images to format NCHW, [0,1], RGB, torch float tensor.

• A model mainly consists of two parts - [network structure] and [model defination, e.g., loss definition, optimization and etc]. The network description is in the Network part.
• Based on the base_model.py, we define different models, e.g., SR_model.py, SRGAN_model.py, SRRaGAN_model.py and SFTGAN_ACD_model.py.

• The network is constructed in models/network.py and the detailed structures are in models/modules.
• We provide some useful blocks in block.py and it is flexible to construct your network structures with these pre-defined blocks.
• You can also easily write your own network architecture in a seperate file like sft_arch.py.

• logger.py provides logging service during training and testing.
• Support to use tensorboard to visualize and compare training loss, validation PSNR and etc. Installationand usage can be found here.
• progress_bar.py provides a progress bar which can print the progress.

Details can be found here.