In the stage of deep learning open source framework, foreign technology giants and scientific research institutions have been occupying the C position, but with the development of domestic technology enterprises and colleges and universities, open source framework has become more and more popular PaddlePaddle “Domestic deep learning open source framework” has gradually occupied a place.
On July 31 this year, ONEFLOW, a deep learning framework for first-class technology development of domestic start-ups, was officially opened on GitHub.
Small companies are small and sophisticated. They should go deep into a certain field and win the market by differentiation The main task is to solve the problem of large-scale distributed AI training.
Without central scheduling, the algorithm is efficient and fast
The whole ONEFLOW project lasted more than 1300 days and was completed by a team of more than 30 people. Yuan Jinhui, CEO of first-class technology, gave ONEFLOW 85 points. He thinks ONEFLOW has great advantages in ease of use, scalability and efficiency.
ONEFLOW mainly solves the problem of multi machine and multi card. It achieves the goal through static scheduling and streaming execution. It is divided into two steps: compiler and runtime. It runs in large-scale distributed training scenarios through decentralized scheduling. It does not need a central scheduler and can run asynchronously in the whole system.
Yuan Jinhui said that ONEFLOW, by exploring the scheduling rules of deep learning, enables the system to get a lot of scheduling rules and strategies in advance before the calculation takes place, which is conducive to CCB’s scheduling process.
By carrying out data transportation planning in advance, the flow execution extracts relevant data before the calculation, and carries out data transportation and calculation at the same time, which not only simplifies the process, but also improves the efficiency.
In the static analysis phase, ONEFLOW does not rank the tasks, but takes them as equally important tasks at the same time, so as to get the optimal scheme of overlapping transmission and calculation.
The details still need to be polished and submitted to the developer for inspection
ONEFLOW mainly provides services for to B customers, focusing on efficiency and scalability. Its document is divided into three levels, one is API document, the other is user model document, the third is design document. Yuan Jinhui said frankly that compared with mature tensorflow and pytorch, ONEFLOW still has many parts to improve, such as supporting dynamic diagrams, which will be gradually supplemented after the framework is open source.
In fact, according to the original plan, ONEFLOW should have been open source a year and a half ago, but because the details need to be debugged and polished, it has to be delayed for such a long time. After more than 1300 days of research and development, ONEFLOW finally came out.
Yuan Jinhui gave ONEFLOW 85 points, but there are still deficiencies in ease of use and completeness. The latter improvement needs to be gradually completed in the test of developers.
At present, the first-class technology has completed about 30 million round a financing. Onebrain, the commercial version of ONEFLOW, is also bringing more business possibilities to enterprises.