ESRGAN-ONNX

Inference-only ONNX implementation of ESRGAN.
Requires minimal additional dependencies.

Installation

To install it as a package:

Choose a version of onnxruntime suitable for your hardware.
For CPU inference

pip install onnxruntime

For NVIDIA GPU

pip install onnxruntime-gpu

Install wheel from the releases.

pip install esrgan_onnx-0.1.0-py3-none-any.whl

Usage

The package only consists of 1 class, ESRGAN.
Object's positional arguments are onnxruntime session, tile size, padding, and manual set scale.
If manual scale is not set - on init it autodetects scale.
get function accepts numpy image array[height, width, channel].

Convert PyTorch models

I recommend chaiNNer
Download chaiNNer and run PyTorch2ONNX_convert.chn
Sample Compact model is in releases.

Test run

To determine optimal tile size for your image:

git clone https://github.com/fAIseh00d/ESRGAN-ONNX.git

Put sample or your own model in the script folder

cd ESRGAN-ONNX
pip install -r requirements-run.txt
python run.py

Please see parser arguments and tile_size list for reference.

Reference

1. ESRGAN
2. Model Database
3. Original ESRGAN-ONNX repo