♢TensorFlow

Developed by Google brain team, the most mainstream deep learning framework has the best ecosystem. However, it is difficult to get started, the API is complex and changeable, and debugging is difficult.

♢PyTorch

Facebook deep learning library is oriented to academia, with simple model construction, easy debugging and excellent ecosystem. But it is not suitable for industrial product level applications.

♢MXNet

The lightweight, portable and flexible deep learning library developed by DMLC has good memory and video memory optimization, and has great potential in the future. Amazon and Baidu support. However, the official documents are not very detailed, and there are confusing parts.

♢Keras

Tensorflow's advanced API can also use theano or cntk as the backend. Focus on user friendliness, can quickly design and build models, and turn ideas into results. It is most suitable for novices. But it is the slowest, highly encapsulated and has poor flexibility.

♢Theano

Theano is the first widely adopted deep learning framework. At present, the development and maintenance of theano project has been terminated.

♢CNTK

Microsoft Research open source deep learning framework.