easy-model-zoo

This is still under heavy development. APIs WILL still change.

Are you also frustrated by the installation process of different models? You are tired of Docker and C extensions failing while compiling. You just want to try out a new model? I agree! You just found the right place! Running deep learning models is easy now.

The only requirement of theses models is that they are pip installable.

You don't have a fancy GPU? Don't worry just run it on the CPU...

PRs are always welcome!

Installation

Simply run:

git clone https://github.com/SharifElfouly/easy-model-zoo
cd easy-model-zoo
pip3 install easy_model_zoo-0.2.4-py3-none-any.whl

Getting Started

from easy_model_zoo import ModelRunner

img_path = 'FULL PATH TO YOUR IMAGE'

device = 'GPU' # or CPU

# Choose a model from the list above
model_runner = ModelRunner('EfficientDet-d0', device)
model_runner.visualize(img_path, predictions)

Pre-trained Models

NOTE: You do NOT have to download the weights file yourself. The ModelRunner will do that for you. The links are just for convenience.

Benchmarks:

• All benchmarks include pre- and postprocessing.
• GPU used: GeForce GTX 1660
• CPU used: Intel(R) Core(TM) i5-9400F CPU @ 2.90GHz

Object Detection

For a full comparison with other Object Detection models see here.

Model Name	MS (GPU)	FPS (GPU)	MS (CPU)	FPS (CPU)	Cityscapes MIOU	Original Repo	Paper	Weights
EfficientDet-d0	41	24	22	4	33.8%	here	arxiv	efficientdet-d0.pth
EfficientDet-d1	54	18	478	2	39.6%	here	arxiv	efficientdet-d1.pth
EfficientDet-d2	83	12.1	768	1.3	43.0%	here	arxiv	efficientdet-d2.pth
EfficientDet-d3	133	7	1660	0.6	45.8%	here	arxiv	efficientdet-d3.pth
EfficientDet-d4	222	4	2984	0.34	49.4%	here	arxiv	efficientdet-d4.pth
EfficientDet-d5	500	2	6604	0.15	50.7%	here	arxiv	efficientdet-d5.pth
EfficientDet-d6	664	1.5	9248	0.11	51.7%	here	arxiv	efficientdet-d6.pth
EfficientDet-d7	763	1.31	13.751	0.07	53.7%	here	arxiv	efficientdet-d7.pth

Semantic Segmentation

Model Name	MS (GPU)	FPS (GPU)	MS (CPU)	FPS (CPU)	Cityscapes MIOU	Original Repo	Paper	Weights
Bisenet	37	50	613	1.63	74.7%	here	arxiv	bisenet.pth

Instance Segmentation

Model Name	MS (GPU)	FPS (GPU)	MS (CPU)	FPS (CPU)	Cityscapes MIOU	Original Repo	Paper	Weights
YOLACT-Resnet50	69	14	1397	0.72	28.2%	here	arxiv	yolact_resnet50_54_800000.pth

How to add a new Model?

Adding a new model is easy. Simply create a new directory inside easy_model_zoo with the name of your model. Define a new Model class that inherits from easy_model_zoo/model.py. For an example look at easy_model_zoo/bisenet/bisenet.py.

Just remember, it has to be pip installable.

License

Feel free to do what you want! Just don't blame me if it doesn't work ;)