cvlab-dl 0.0.1

An unified laboratory/framework for Computer Vision research, development and deployment.

cvlab

An unified laboratory/framework for Computer Vision research, development and deployment. Here we cover the whole life cycle of CV models: building, training, evaluation, optimization and deployment.

We roughly target these platform/frameworks internally:

Python >= 3.6
PyTorch >= 1.6
TensorFlow >= 1.15
ONNX >= 1.7
TensorRT >= 6

Other dependencies could be found in requirements.txt.

About the project

This project empowers CV developers with several tool boxes, and we try our best to make each of them independent enough so that you can extract them from this project and use them in your own.

Curated models. Loads of new models are coming to the world every day, but only some of them are proven to be real applicable, these models are hand picked by our own experience during everyday work, covering multiple tasks including image classification, detection, segmentation, etc. We provide implementation of these models, pre-trained weights, and an nice guide to re-train/fine tune them on your own data. Utilizing our optimization and deployment tools, these models can be deployed and start creating real value for you.

You can absolutely build new models and train, test, evaluate, optimize, deploy them with exactly the same tool chain which already existed models used.

Deployment toolbox. Nice models are just the first step towards application, the work after training a model is often complicated and painful, such as model conversion from one platform to others, quantization and compressing, graph optimization for inference and so on. We provide several utilities about these tasks, hope they can be helpful.

Evaluation metrics. Without reasonable metrics we can't tell the performance of anything. We provide easy-to-use tools to measure the performance of your models with the most popular metrics of certain tasks. We believe by providing such unified tools can greatly improve the everyday work experience for us.

Miscellaneous tools. We provide a bunch of tools that are nice if you have them, so you don't need to write them over and over again, such as visualization tools, dataset adapters, loggers, etc.

LICENSE

MIT © AlanDecode