A TensorFlow wrapper that refuses to hide the wiring

What it does

TensorLayer is a deep learning and reinforcement learning library built on TensorFlow 2.0. It provides customizable neural layers and model abstractions, plus a large collection of community tutorials and applications. The project also maintains an RL zoo with both low-level and high-level APIs, and has been used by researchers at universities and companies including Google, Microsoft, and Alibaba.

The interesting bit

The library explicitly models its APIs on PyTorch’s transparency rather than Keras’s opacity—an odd marriage of PyTorch philosophy with TensorFlow infrastructure. The README now opens with a blazing redirect to TensorLayerX, a multi-backend successor that supports PyTorch, MindSpore, and others, suggesting the original TensorFlow-only library has been effectively superseded.

Key highlights

• Won 2017 Best Open Source Software from the ACM Multimedia Society
• Ships with extensive reinforcement learning tutorials and a dedicated RL zoo (RLzoo)
• Publishes benchmarks showing VGG16 training on a TITAN Xp within seconds of native TensorFlow 2.0 performance
• Documentation available in both English and Chinese, plus a published Chinese book
• Alpha support for running models on low-cost AI chips like the K210 via Sipeed Maxi-EMC

Caveats

• The README prominently redirects new users to TensorLayerX, suggesting this repository is no longer the primary focus of development
• The performance benchmark only covers VGG16 on a single GPU, so broader “zero-cost abstraction” claims are not fully substantiated in the provided sources

Verdict

TensorFlow users who need more flexibility than Keras and a strong reinforcement learning toolkit may still find value here, but new users should evaluate TensorLayerX first—its own maintainers clearly think it is the better starting point.