Layer-Wise Deep Learning Network Encoding for Efficient Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning network models are too large to be stored in internal SRAM, leading to inefficient power consumption due to frequent access from external DRAM, and existing compression methods do not distinguish between layers, resulting in inefficient decoding processes.
Innovation Solution
A method and apparatus for encoding and decoding deep learning networks in layer units using network, layer, and sub-layer header information, including specific details for distinguishing layers, such as layer ID, type, size, and tile partitioning, to create a bitstream that efficiently compresses and decompresses the network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the deep learning network model is compressed without layer distinction, then the model size is reduced, but the decoding efficiency deteriorates since the whole model must be imported and decoded whenever every layer is processed
Solution Approach 1:
The patent segments the compressed deep learning network model into layer-specific units, where each layer is independently compressed and stored with its own metadata. This allows individual layers to be decoded and processed separately without requiring the entire model to be loaded, thus improving decoding efficiency while maintaining compression benefits.
2Quantity of substance
If the deep learning network model is stored in external DRAM, then the model capacity requirement is met, but the power consumption increases due to frequent access
Solution Approach 1:
By segmenting the model into compressed layer units with embedded metadata, the system can load only the necessary layers into internal SRAM when needed, rather than frequently accessing the entire model from external DRAM. This reduces the access rate to external memory and lowers power consumption.
Solution Approach 2:
The patent performs compression and metadata embedding in advance during the model preparation phase. This preliminary action enables efficient layer-by-layer decoding during inference, reducing the need for frequent external memory accesses and thereby reducing power consumption during operation.
3Productivity
If layer distinction information is added to the bitstream, then the decoding efficiency is improved, but the bitstream complexity increases
Solution Approach 1:
The patent divides the bitstream into separate layer units, each containing its own metadata and compression information. This segmentation allows the decoder to process layers independently using simple, repetitive decoding logic, improving efficiency without requiring complex overall bitstream handling.
Solution Approach 2:
Each layer unit in the bitstream contains localized metadata and compression parameters specific to that layer. This local quality approach allows each layer to be decoded with optimized, simple rules tailored to its specific characteristics, improving decoding efficiency while keeping the overall bitstream structure manageable.
Data Source
AI summary
Disclosed herein are a method and apparatus for encoding/decoding a deep learning network. According to an embodiment, the method for decoding a deep learning network may include decoding network header information regarding the deep learning network; decoding layer header information regarding a plurality of layers in the deep learning network; decoding layer data information regarding specific information of the plurality of layers; and obtaining the deep learning network and a plurality of layers in the deep learning network, and the layer header information includes layer distinction information associated with distinguishing the plurality of layers.