Research On Accelerating Conditional Random Fields By Coordinating CPU And GPU

Conditional Random Fields (CRFs) is one of the machine learning algorithms, andhas been applied to a wide range of classification tasks with various degrees of success.However, performing learning and inference process in serial CRFs with various featuresare computationally expensive and intractable. Paralleling CRFs on CPU or GraphicsProcessing Units (GPU) has been considered in many related researches desperately, butthe performances are not satisfactory in some respects.Considering the parallel characteristic of different devices and CRFs, a novelcoordinating framework based on CPU and GPU has been proposed to accelerate CRFs.In the coordinating framework, for the high request of memory space and conditionalbranch in the optimization of training stage, a new criterion is used to employ CPU to dealwith it, instead of the way in most recent methods, which parallelize optimization processon GPU with lose of precision and being unstable to some extent. Two-levelparallelization is also proposed for the relative compute-intensive tasks to maximizedparallelism at every step: possibilities of the state lattice are calculated at node level, andparallel computations are effectively executed among all possible paths in node matrix atsentence level. Moreover, an optimized coalesced memory-accessing manner is employedon GPU to alleviate the cost of thread accessing and improve the parallel efficiency.Experimental results show that when compared with a single task process on the CPU,the performance is much faster than single CPU computing, and the inference stage canreach more than15times speedup. On the other hand, compared with the implementationparallel totally on GPU, the implementation achieves up to50%improvement in the speedwith even higher accuracy.