Training an image upscaling model

Title: Training an Image Upscaling Model

Channel: Philip Hofmann

Description: This is a follow-up video to the previous one where I showed how to train a Real-ESRGAN Compact (SRVGGNet) model. In this video, I demonstrate how to train models of other networks like HAT, SwinIR, etc., using the same codebase. I encounter multiple errors during the process, which provides valuable insights into what could go wrong and how to fix it. I share my learnings and current understanding throughout the tutorial.

Tutorial:

1. TraiNNers (training code) (00:50)

   • Understand the basics of training code for image upscaling models.

2. TraiNNer-redux-FJ (02:37)

   • Familiarize yourself with TraiNNer-redux-FJ for model training.

3. Clone Code Repo (03:30)

   • Clone the code repository containing the necessary scripts and configurations.

4. Virtual Environment & Install Dependencies (04:40)

   • Set up a virtual environment and install the required dependencies for training the models.

5. Create New Config File (06:30)

   • Create a new configuration file for the specific model you want to train.

6. Adapting Config File (08:32)

   • Adjust the configuration settings based on the requirements of the model and your system specifications.

7. On-The-Fly Degradations - RealESRGANDataset (13:03)

   • Understand and implement on-the-fly degradations using RealESRGAN dataset for training.

8. Validation (47:00)

   • Validate the model training process to ensure accuracy and efficiency.

9. Training Outputs (1:25:00)

   • Monitor and analyze the training outputs to track the progress of the model.

10. Resume Training with Set Checkpoints in Paths (1:32:15)

    • Learn how to resume training from a specific checkpoint to continue the training process.

11. Releasing (Discord & openmodeldb.info) (1:43:35)

    • Share your trained models with the community on platforms like Discord and openmodeldb.info.

12. Upping Params by Training Remotely (Kaggle) (1:38:00)

    • Explore options for increasing model parameters by training remotely on platforms like Kaggle.

13. Onnxify (1:41:39)

    • Convert the trained model to ONNX format for compatibility with chaiNNer if using a new architecture.

14. Own GitHub Repo (1:48:05)

    • Create your own GitHub repository to manage and share your trained models and code.

15. Comparison Website (1:49:36)

    • Utilize a comparison website to showcase and compare your trained models.

16. Outro (1:51:14)

    • Summarize the key takeaways and insights from the video tutorial.

By following these steps, you will be able to train image upscaling models using different networks and troubleshoot common errors that may arise during the process.