traiNNer-redux

Overview

A modern community fork of BasicSR and traiNNer-redux.

Usage Instructions

Initial Setup

1. Install Python if it's not already installed. Python 3.12 is recommended, and Python 3.11 is also supported. Python 3.13 is not supported yet.
2. Clone the repository:
   • To use the git command line, navigate to where you want to install traiNNer-redux, and enter this command (install git first if it's not already installed):

     git clone https://github.com/the-database/traiNNer-redux.git
     

   • To use a GUI for git, follow the instructions for that git client. For GitHub Desktop, for example, click on the green Code button near the top of this page, click Open with GitHub Desktop and follow the instructions.
3. For Windows users, double click install.bat, and for Linux users, from the terminal in the traiNNer-redux folder run chmod +x install.sh && ./install.sh to install all Python dependencies to a new virtual environment. The install.sh script is tested on Ubuntu and may need adjustments to work on other Linux distros.

Training a Model

Refer to the wiki for a full training guide and benchmarks.

Do a quick test run

The repository comes with several configs that are ready to use out of the box, as well as a tiny dataset for testing purposes only. To confirm that your PC can run the training software successfully, run the following command from the traiNNer-redux folder:

venv\Scripts\activate
python train.py --auto_resume -opt ./options/train/SPAN/SPAN.yml

You should see the following output within a few minutes, depending on your GPU speed:

...
2024-07-02 21:40:56,593 INFO: Model [SRModel] is created.
2024-07-02 21:40:56,668 INFO: Start training from epoch: 0, iter: 0
2024-07-02 21:41:17,816 INFO: [4x_SP..][epoch:  0, iter:     100, lr:(1.000e-04,)] [performance: 4.729] [eta: 14:11:33] l_g_mssim: 1.0000e+00 l_g_percep: 3.5436e+00 l_g_hsluv: 4.3935e-01 l_g_gan: 2.4346e+00 l_g_total: 7.4175e+00 l_d_real: 2.4136e-01 out_d_real: 2.9309e+00 l_d_fake: 5.2773e-02 out_d_fake: -2.4104e+01

The last line shows the progress of training after 100 iterations. If you get this far without any errors, your PC is able to train successfully. Press ctrl+C to end the training run.

Set up config file

1. Navigate to traiNNer-redux/options/train, select the architecture you want to train, and open the yml file in that folder in a text editor. A text editor that supports YAML syntax highlighting is recommended, such as VS Code or Notepad++. For example, to train SPAN, open traiNNer-redux/options/train/SPAN/SPAN.yml.
2. At the top of the file, set the name to the name of the model you want to train. Give it a unique name so you can differentiate it from other training runs.
3. Set the scale depending on what scale you want to train the model on. 2x doubles the width and height of the image, for example. Not all architectures support all scales. Supported scales appear next to the scale in a comment, so # 2, 4 means the architecture only supports a scale of 2 or 4.
4. Set the paths to your dataset HR and LR images, at dataroot_gt and dataroot_lq under the train: section. The HR images and LR images should match in numer of images and filenames. For each matching LR/HR pair, the image resolutions should work with the selected scale, so if a scale of 2 is selected then each HR must be 2x the resolution of its matching LR image.
5. If you want to enable validation during training, set val_enabled to true and set the paths to your validation HR and LR images, at dataroot_gt and dataroot_lq under the val section.
6. If you want to use a pretrain model, set the path of the pretrain model at pretrain_network_g and remove the # to uncomment that line.

Run command to start training

Run the following command to start training. Change ./options/train/arch/config.yml to point to the config file you set up in the previous step.

venv\Scripts\activate
python train.py --auto_resume -opt ./options/train/arch/config.yml

For example, to train with the SPAN config:

venv\Scripts\activate
python train.py --auto_resume -opt ./options/train/SPAN/SPAN.yml

To pause training, press ctrl+C or close the command window. To resume training, run the same command that was used to start training. The --auto_resume flag will resume training from when it was paused.

Test models

Models are saved in the safetensors format to traiNNer-redux/experiments/<name>/models, where name is whatever was used in the config file. chaiNNer can be used to run most models. If you want to run the model on images during training to monitor the progress of the model, set up validation in the config file, and find the validation results in traiNNer-redux/experiments/<name>/visualization.

The test script can also be used to test trained models, which is required to test models with architectures not yet supported by chaiNNer. For example, to test SPANPlus model, open the config file at ./options/test/SPANPlus/SPANPlus.yml, and update the following:

1. Edit the dataroot_lq option to point to a folder that contains the images you want to run the model on.
2. Make sure the options under network_g match the options under network_g in the training config file that you used. For example, if you trained SPANPlus_STS, then set the type to SPANPlus_STS under network_g in the test config file as well.
3. Update pretrain_network_g to point to the path of the model you want to test.

Then run this command to run the model on the images as specified in the config file:

venv\Scripts\activate
python test.py -opt ./options/test/SPANPlus/SPANPlus.yml

Convert models to ONNX

If you want to convert your PyTorch models to ONNX format, you can use the convert_to_onnx.py script to do so. First install the additional dependencies for ONNX:

venv\Scripts\activate
pip install .[onnx]

Then open a config file corresponding to the architecture of the model you trained. For example, if you trained SPANPlus, open the config file at ./options/onnx/SPANPlus/SPANPlus.yml, and update the following:

1. Set the name to the name of your model, the ONNX filename will include this name.
2. Make sure the setting under network_g match the settings you used to train your model.
3. Set pretrain_network_g to point to the path of your .safetensors or .pth model that you want to convert.
4. Set the options in the onnx section of the config file as needed.

Then run this command to do the conversion (make sure the path points to the .yml file you edited):

venv\Scripts\activate
python convert_to_onnx.py -opt ./options/onnx/SPANPlus/SPANPlus.yml

The converted onnx files will be saved to the onnx directory.

Contributing

Please see the contributing page for more info on how to contribute.

Resources

• OpenModelDB: Repository of AI upscaling models, which can be used as pretrain models to train new models. Models trained with this repo can be submitted to OMDB.
• chaiNNer: General purpose tool for AI upscaling and image processing, models trained with this repo can be run on chaiNNer. chaiNNer can also assist with dataset preparation.
• WTP Dataset Destroyer: Tool to degrade high quality images, which can be used to prepare the low quality images for the training dataset.
• helpful-scripts: Collection of scripts written to improve experience training AI models.
• Enhance Everything! Discord Server: Get help training a model, share upscaling results, submit your trained models, and more.

License and Acknowledgement

traiNNer-redux is released under the Apache License 2.0. See LICENSE for individual licenses and acknowledgements.

• This repository is a fork of traiNNer-redux which itself is a fork of BasicSR.
• Network architectures are imported from Spandrel.
• The SPANPlus architecture is from umzi2/SPANPlus which is a modification of SPAN.
• The TSCUNet architecture is from aaf6aa/SCUNet which is a modification of SCUNet, and parts of the training code for TSCUNet are adapted from TSCUNet_Trainer.
• Several enhancements reference implementations from Corpsecreate/neosr and its original repo neosr.
• Members of the Enhance Everything Discord server: Corpsecreate, joeyballentine, Kim2091.