About PhotoUpscaler

PhotoUpscaler is an educational project for DLS: https://stepik.org/lesson/2104043/step/1?unit=2134957

It implements a ML model for image SuperRes upscaling (2x scale factor). This project contains:

• pre-trained model weights
• notebooks for training and testing the model locally
• telegram bot implementation for interactive upscaling

Description of the implementation process is available here: Detailed Worklog

Inference Examples

Here are some examples comparing image patches scaled by my model, using traditional algorithms and by 3rd party model:

Usage

IMPORTANT: Model weights are stored using Git LFS (Large File Storage). Use git lfs install before cloning repo, otherwise pre-trained model weights will not be obtained.

Training the model

• Open the notebooks/train.ipynb
• Follow the instructions in the notebook

Testing the model

• Open the notebooks/test_patches.ipynb (for testing raw patches) or notebooks/test_inference.ipynb (for testing final image assembling)
• Follow the instructions in the notebook

Deploying a telegram bot

• Ensure model upscaler.ckpt file is presented in the checkpoints folder
• Obtain a new bot token from @BotFather
• Add it to the .env file using BOT_TOKEN= key
• Run bot service via docker-compose.yml