We are happy to announce today the release of Leaf 0.2 on which we have been working on for the last weeks. Leaf is a modular, performant, portable Machine Intelligence Framework. It is the Hacker's Machine Intelligence Framework, developed by software engineers.
You can install Leaf 0.2 and run examples, including popular Deep Neural Networks like Alexnet, Overfeat, VGG and more.
The release was mostly about finding an efficient and clean architecture, catching up with the performance level of other Machine Learning Frameworks. It shares concepts from the brilliant work done by the people behind Torch, Tensorflow, Caffe, Rust and numerous research papers. We have several large features under development, Leaf 0.2 gives us the platform to go on exploring new territory with Leaf 0.3.
Leaf 0.2 is one of the fastest Machine Intelligence Frameworks that exist today. Rust was a big help in developing the entire platform over the course of a few months. We achieved a very efficient GPU utilization and oriented our architecture close to Torch and achieved the distribution capabilities of Tensorflow, on a lower abstraction level. More information in the following sections.
More Benchmarks and comparisons, including Memory utilization, can be found on Deep Learning Benchmarks.
Leaf 0.2 uses Collenchyma for training and running models on CPUs, GPUs, FPGAs, etc. with OpenCL or CUDA or other Computation Languages, on various machines and operating systems, without the need to adapt your code what so ever. This makes deployment of models to servers, desktops, smartphones and later embedded devices very convenient.
With that abstraction and separation of algorithm representation and execution, we gain a nice Framework for distributed model execution, without relying on a symbolic, data-flow graph model like Tensorflow, which introduces performance and development overhead concerns.
Leaf 0.2 replaces special Network
objects with container layers
like the Sequential
layer. Where previously all weights were stored centrally
by the Network, each Layer is now responsible for managing its own weights.
This allows for more flexibility in expressing different network architectures.
It also enables better programmatic generation of networks by nesting container
layers where each container represents a common pattern in neural networks,
e.g. Convolution, Pooling and ReLU following each other.
We had 9 individual contributors, which made Leaf 0.2 possible. Thank you so much for your contribution, when Leaf wasn't even executable, yet. And thank you for everyone who took the time to engage with us on Gitter and Github.
- Maximilian Goisser (@hobofan)
- Michael Hirn (@mjhirn)
- Ewan Higgs (ehiggs)
- Florian Gilcher (@argorak)
- Paul Dib (pdib)
- David Irvine (dirvine)
- Pascal Hertleif (killercup)
- Kyle Schmit (kschmit90)
- Sébastien Lerique (wehlutyk)
More about Leaf and Autumn | Follow on Twitter: @autumn_eng