METHOD AND DEVICE FOR COMPILING NEURAL NETWORKS Russian patent published in 2025 - IPC G06N3/10 G06F8/41 

FIELD: physics.

SUBSTANCE: invention relates to the field of deep learning, in particular to the method and device for compilation of neural networks. Method is executed by at least one processor and comprises steps of: a) obtaining a neural network model and transforming it into a unified format representing a directed graph, where the graph vertices are operations of the neural network model, b) determining the capabilities of the neural network model operations on the main or auxiliary computing device, c) separating the directed graph into a main subgraph and preprocessing and postprocessing subgraphs, d) generating a sequence of starting operations of the neural network model based on the interaction of the main subgraph sections with the preprocessing and postprocessing subgraphs, e) performing the operations of the neural network model of the main subgraph to optimize their execution, resulting in an optimized unified format, f) converting the optimized unified format into a high-level dialect, g) converting a high-level dialect into a common level dialect, h) converting a common level dialect into a low-level dialect, i) generating a binary data stream based on a low-level dialect and splitting the binary data stream into separate instructions for parallel processing on hardware units of the main computing device, j) executing instructions obtained in step i) on a main computing device, and performing operations from preprocessing and postprocessing subgraphs on an auxiliary computing device based on a start-up sequence generated at step d).

EFFECT: high efficiency of the neural network compilation process.

13 cl, 2 dwg