This video demonstrates how to train an upscaling model using the NEOSr software. The speaker details the steps involved, from defining the model's purpose and preparing the dataset to training, releasing, and sharing the model on platforms like Discord and OpenModelZoo. A specific use case of upscaling game textures is used as an example throughout.
-
Training Steps: The process involves defining the model's purpose, preparing the dataset (including cleaning and deduplication), selecting the training software (NEOSr in this case), configuring training parameters, training the model, and finally releasing it to the community.
-
Dataset Preparation: Crucial steps include normalizing filenames, removing low-information images (e.g., single-color tiles), deduplicating images, and creating paired low-resolution datasets using techniques like downscaling, blurring, and JPEG compression.
-
NEOSr Software: NEOSr is a Python-based training framework with a comprehensive wiki, offering various network architectures for upscaling model training. It allows for pausing and resuming training, and provides visualizations and metrics for monitoring progress.
-
Model Release and Sharing: The speaker demonstrates releasing the trained model via GitHub, including providing interactive examples and incorporating the model into the OpenModelZoo. The importance of clear documentation and readily available examples for model users is highlighted.
-
Out-of-Memory (OOM) Errors: The video addresses potential OOM errors during training and suggests solutions like decreasing batch size or ground truth size to optimize VRAM usage.
