nmz0429/biwako

The largest freshwater lake. The home of onnx models.

The home of moduled onnx models

---

BiWAKO

Docs

Refer to the docs

This README is the copy of the front page of the docs site and some sections are not properly rendered.

Model/Feature Requests

I am more than welcome to work for adding new models or features! If you have a model in pytorch or tensorflow, feel free to create an issue or new discussion stating which model or what feature you want from this repository!

For model addition, links to the implementation or some working inference code definitely helps me to add the model and inference code.

0. Introduction

This repository offers

1. Models: Trained state-of-the-art models for various vision tasks in ONNXRuntime backend
2. No-Code Modules: Easy interface to use those models for both prediction and visualizing output. No coding is needed. The interface is universal among all models in this library.
3. Extentiability: Customizable modules to use it for applications such as realtime inference.

1. Installation

Install directly from this repository.

$cd BiWAKO
$pip install -e .

!!! warning Downloading from pip server is currently suspended in order to protect weight files. We will update it soon.

2. Usage

No matter which model you use, these interface is the same.

import BiWAKO

# 1. Initialize Model
model = BiWAKO.MiDAS(model="mono_depth_small")

# 2. Feed Image (accept cv2 image or path to the image)
prediction = model.predict(image_or_image_path)

# 3. Visiualize result as a cv2 image
result_img = model.render(prediction, image_or_image_path)

More specifically...

1. Instantiate model with BiWAKO.ModelName(weight). The ModelName and weight corresponding to the task you want to work on can be found at the table in the next section. Weight file is automaticaly downloaded.
2. call predict(image). image can be either path to the image or cv2 image array.
3. call render(prediction, image). prediction is the return value of predict() method and image is the same as above. Some model takes optional arguments to control details in the output.

2-2 High Level APIs

We also provides some APIs to even accelerate productions. See API page for further details/

4. Models

The following list is the current availability of models with weight variations.
Click the link at the model column for futher documentation.

Task	Model	Weights
Mono Depth Prediction	MiDAS	mono_depth_small mono_depth_large
Salient Object Detection	U2Net	mobile basic human_seg portrait
Super Resolution	RealESRGAN	super_resolution4864 super_resolution6464
Object Detection	YOLO2/YOLO	Please refer to docs for details
Emotion Prediction	FerPlus	ferplus8
Human Parsing	HumanParsing	human_attribute
Denoise	HINet	denoise_320_480
Face Detection	YuNet	yunet_120_160
Style Transfer	AnimeGAN	animeGAN512
Image Classification	ResNetV2	resnet18v2 resnet50v2 resnet101v2 resnet152v2
Human Portrait Segmentation	MODNet	modnet_256
Semantic Segmentation	FastSCNN	fast_scnn384 fast_scnn7681344
Diver's View Segmentation	SUIMNet	suim_net_3248 suim_rsb_72128 suim_vgg_25632 suim_vgg_72128

5. Deployment

It is extremely easy to use BiWAKO at application layer.

1. Real Time Prediction

Any model can be used in the same way to run real-time inference.

2. FastAPI Implementation

Like the above example, you can build simple Backend API for inference on web server. We have prepared sample deployment of the library with FastAPI.Read this for details.

3. Video Prediction

We also provides pre-defined video prediction API. Read this for details