Google launches TpuGraphs training data set to improve the deep learning capabilities of AI models

Google recently launched a model training data set called TpuGraphs, which is mainly used to optimize compilers and improve artificial intelligence deep learning capabilities

▲ Image source Google Blog (the same below)

Google pointed out that current AI deep learning systems usually use TensorFlow, JAX, PyTorch and other frameworks for training. These frameworks mainly optimize models through heuristics of the underlying compiler. Applying the "learning cost model" in the relevant compiler can improve the performance of the compiler and enhance the deep learning capabilities of the final output model

The TpuGraphs data set launched by Google was learned by IT House that it is a "learning cost model". This data set mainly contains various open source deep learning programs, covering a variety of popular model architectures, such as ResNet, EfficientNet, Mask R-CNN and Transformer, etc.

Google claims that compared to industry competitors, Google’s TpuGraphs dataset is 770 times larger in terms of “average graph size” and 25 times larger in terms of “number of graphs”. Google claims that applying the TpuGraphs dataset can effectively solve the "scalability", "efficiency" and "quality" issues of the final output model

In addition, Google has also launched a model training method called GST (Graph Segment Training), which allows the training of large graph neural networks on devices with limited memory. It is said that this method can shorten the "end-to-end training time" of the model by three times and effectively improve the efficiency of model training

The above is the detailed content of Google launches TpuGraphs training data set to improve the deep learning capabilities of AI models. For more information, please follow other related articles on the PHP Chinese website!