Mobile Technology

AI Benchmark: Deep Learning Neural Networks on Android Smartphones

With the wide use of the mechanical man OS, variety of standard deep learning frameworks were ported to the current platform, together with Torch, Deep learning, TensorFlow (Mobile, Lite, Caffe, Caffe2, MXNet, NNabla, etc. Nowadays, the foremost unremarkably used are 3 of them: Tensorflow Mobile, Tensorflow fat-free and Caffe2 that are described below.

Tensorflow Mobile

Tensorflow is an ASCII text file deep learning library for analysis and development free by Google in 2015. TensorFlow’s programming model may be described as a directed graph that defines the relation between the input and output (target) variables. The graph itself consists of a group of nodes representing varied operators applied consecutively to the input file (e.g., convolutional, pooling, LSTM layers, etc.) that are process a deep learning model and therefore the corresponding dataflow computation.

Tensorflow Lite

TensorFlow lite was presented in late 2017, as a successor of the TF Mobile library. With Google, it provides higher performance and a smaller binary size because of optimized kernels, pre-fused activations, and fewer dependencies. Equally to TF Mobile, a general TensorFlow pre-trained model is, in theory, regenerate to TensorFlow light.

It should be noted, but, that TensorFlow lite is in developer preview at the instant and encompasses a variety of considerable limitations. First of all, it supports solely a restricted set of operators, lacking the complete support of, e.g., image resizing, batch and instance standardization, LSTM units, some applied math functions or maybe straightforward mathematical operations like mathematical process or argmax. Officially, Google guarantees solely 3 models to work: the Inception-V3, MobileNet, and good Reply SSL algorithmic rule, with some modifications it’s potential to run a variety of different deep learning models. Second issue considerations the abstract thought time and also the quantity of consumed RAM.

Since the ByteBuffer format isn’t supported for the network’s output, these 2 values are up to 2× higher compared to TF Mobile for image-to-image translation issues. Finally, stability is another concern — this official version may not work cleanly with a variety of models and mobile devices, though a number of the problems are already resolved in the nightly TF fat-free version. Whereas several of those issues can most likely be overcome in the forthcoming library releases, presently they create the utilization of TensorFlow lite difficult for several existing deep learning issues.

Caffe2

Caffe2 is another ASCII text file deep learning a framework, originally developed at UC Berkeley by Yangqing Jia and discharged in 2013. Its 1st unofficial golem port appeared succeeding year and in 2017, with Facebook’s unharness of the successor, Caffe2, its mobile version for iOS and golem platforms was also conferred. Caffe2 is employing a programming model almost like TensorFlow’s, with static machine graphs and nodes representing varied operators. In keeping with the Caffe2 GitHub repository, the speed of its mobile library is usually corresponding to that of TensorFlow nonfat  (175ms vs. 158ms for the SqueezeNet model on flower 821 SoC). Equally to TensorFlow, acceleration for Caffe2 models is also supported by all proprietary SDKs (SNPE, HiAI, NeuroPilot, and ArmNN), but NNAPI support continues to be in development and isn’t absolutely integrated nonetheless.

4,173 total views, 11 views today

Leave a Reply

Your email address will not be published. Required fields are marked *