This is a non-commercial project, which aims to simplify deep learning research and reproducibility. One of the current research bottlenecks is in the Experiment stage, during which researchers usually spend a lot of time on re-implementation of the models from the other researches. Two usual reasons are the lack of the public access code and the non-portability of the code.
This project aims to overcome those problems with the use of model translation (ONNX soon) and distributed (volunteer) computing.
At the current stage, we propose naive distributed computing for the single fully-connected MNIST model. Anyone may find it's weights here.
Users may run their fully-connected models (at the current stage after a bit of manual work) in the distributed manner. Servers that perform computations do not require deep learning framework installed, so anyone may easily use his or her device for the computations.
- rpyc 5.0.0+
- numpy 1.19+
- Use ONNX for the Computer-Processor communication
- Implement reliable distributed computation
- Change the architecture, adding the Resource Manager