tensorpack
A high-performance neural network training interface for TensorFlow, optimized for speed and research flexibility.
What is tensorpack?
Tensorpack is a neural network training interface built on TensorFlow that focuses on maximizing training speed and providing the flexibility required for advanced research. It achieves significant performance gains over alternatives like Keras while supporting scalable multi-GPU training and high-efficiency data loading. The framework includes reproducible implementations of state-of-the-art papers across computer vision, reinforcement learning, and NLP.
Researchers and deep learning practitioners who need fast, flexible, and reproducible training pipelines for TensorFlow, particularly those working on novel architectures or large-scale datasets.
Developers choose Tensorpack for its unmatched training speed, research-oriented design, and high-quality reproducible examples, avoiding the overhead and limitations of other high-level TensorFlow wrappers.
Overview
A Neural Net Training Interface on TensorFlow, with focus on speed + flexibility
Use Cases
Best For
- Training large-scale vision models like ResNet on ImageNet with maximum speed
- Implementing and reproducing state-of-the-art research papers faithfully
- Multi-GPU or distributed training of deep neural networks
- High-performance data loading and preprocessing for research workflows
- Building custom training pipelines with TensorFlow without wrapper limitations
- Reinforcement learning experiments with DQN variants or A3C on OpenAI Gym
Not Ideal For
- Projects requiring out-of-the-box, pre-built model architectures without custom implementation
- Teams heavily invested in TensorFlow 2 eager execution who prefer its native APIs over graph-mode compatibility
- Production environments needing long-term stability and guaranteed backward compatibility
- Educational settings or rapid prototyping where simplicity and extensive tutorials outweigh raw speed
Pros & Cons
Pros
Benchmarks show Tensorpack runs 1.2–5x faster than equivalent Keras code on common CNNs by optimizing TensorFlow usage with minimal overhead, as highlighted in its performance comparisons.
Offers off-the-shelf data-parallel multi-GPU and distributed training strategies, making it easy to scale models across hardware without custom code.
tensorpack.dataflow provides high-performance data processing in pure Python, offering more flexibility than symbolic approaches like tf.data for complex research workflows, as emphasized in the documentation.
Includes high-quality, faithful reproductions of state-of-the-art papers across vision, RL, and NLP, ensuring reliability for research comparisons and avoiding toy examples.
Cons
The project is not yet stable, requiring users to pin exact versions to avoid breaking changes, which adds risk for long-term or collaborative projects.
Tensorpack uses TensorFlow 1 compatibility mode for TF2, and the README notes that some examples are not migrated, limiting its appeal for TF2-native development.
Focused on efficiency and flexibility, it requires deeper understanding of TensorFlow graph mode and custom pipeline design, compared to more beginner-friendly wrappers like Keras.
Frequently Asked Questions
Related Projects
Tensorflow - Open source software library for numerical computation using data flow graphsAn Open Source Machine Learning Framework for Everyone
LudwigLow-code framework for building custom LLMs, neural networks, and other AI models
Sonnet - a library for constructing neural networks by Google's DeepMindTensorFlow-based neural network library
tflearnDeep learning library featuring a higher-level API for TensorFlow.
Found a gem we're missing?
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.