Tensorpack is a neural network training interface on top of TensorFlow, built for training speed and research flexibility. A veteran project created in 2015, it has about 6,300 stars on GitHub, though its most recent updates were in 2023.
Tensorpack works with graph-mode TensorFlow and is deliberately not another model wrapper: you use ordinary TensorFlow symbolic functions — including Keras layers — inside it, while Tensorpack handles the training loop efficiently. The project reports training common CNNs 1.2 to 5 times faster than equivalent Keras code, with data-parallel multi-GPU and distributed training available off the shelf. Its dataflow library squeezes maximum throughput out of pure-Python data loading with automatic parallelization. The examples are a highlight: rather than toy demos, they faithfully reproduce well-known papers — ResNet and Mask/Faster R-CNN, MoCo, GAN variants, DQN and A3C reinforcement learning, and LSTM-CTC speech recognition. Note that Tensorpack targets the TF1 graph-mode world, using TF2’s compatibility mode. Tensorpack was created by Yuxin Wu and collaborators and became a standard reference during the TensorFlow 1.x era, when its training scripts were often the fastest public implementations available. Many of its examples are still used today to sanity-check reimplementations of classic models, even as the ecosystem has largely moved on to newer frameworks.
Tensorpack is completely free and open source under the Apache-2.0 license.
Deep learning researchers maintaining TensorFlow graph-mode codebases, and anyone studying how classic papers were actually trained, who values faithful reproductions over toy examples.