Unlock instant, AI-driven research and patent intelligence for your innovation.

Quantification method, device, electronic equipment and storage medium for feature map data

A quantization method and data storage format technology, applied in the field of devices, quantification methods of feature map data, electronic equipment and storage media, can solve the problems of reduced transmission efficiency and affecting the overall processing performance of computing hardware platforms, etc.

Active Publication Date: 2021-12-07
BEIJING BAIDU NETCOM SCI & TECH CO LTD
View PDF5 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

In the existing dynamic quantization method, Float 16 feature map data is used to interact between DDR and FPGA. Under the condition of limited DDR communication bandwidth, Float 16 feature map data will be more efficient than Int8 feature map data. It is reduced by half and affects the overall processing performance of the CNN edge computing hardware platform

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Quantification method, device, electronic equipment and storage medium for feature map data
  • Quantification method, device, electronic equipment and storage medium for feature map data
  • Quantification method, device, electronic equipment and storage medium for feature map data

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0049] figure 1 It is a schematic flowchart of the method for quantifying feature map data provided in Embodiment 1 of the present invention. The method can be executed by a quantization device or electronic device for feature map data, and the device or electronic device can be implemented by software and / or hardware. This device or electronic equipment can be integrated in any intelligent equipment with network communication function. Such as figure 1 As shown, the quantification method of feature map data may include the following steps:

[0050] S101. Acquire the original feature map data of each 8-bit integer fixed-point Int8 in at least one input line corresponding to the current cycle in the DDR according to the predetermined data storage format of the DDR, and convert the original feature map data of each Int8 into a corresponding The corresponding input feature map data.

[0051] In a specific embodiment of the present invention, the electronic device can obtain th...

Embodiment 2

[0062] figure 2 It is a schematic flowchart of the method for quantifying feature map data provided by Embodiment 2 of the present invention. Such as figure 2 As shown, the quantification method of feature map data may include the following steps:

[0063] S201. Acquire the original feature map data of each Int8 in at least one input row corresponding to the current cycle in the DDR according to the predetermined data storage format of the DDR, and convert the original feature map data of each Int8 into a corresponding input feature map. data.

[0064] In a specific embodiment of the present invention, the electronic device can obtain the original feature map data of each Int8 in at least one input row corresponding to the current cycle in the DDR according to the predetermined DDR data storage format, and store the original feature map data of each Int8 The graph data is transformed into its corresponding input feature map data. Specifically, if the input feature map da...

Embodiment 3

[0080] Figure 5 It is a schematic diagram of the first structure of the quantization device for feature map data provided by Embodiment 3 of the present invention. Such as Figure 5 As shown, the device for quantifying feature map data according to the embodiment of the present invention may include: an acquisition module 501, a processing module 502, a determination module 503, and a conversion module 504; wherein,

[0081] The acquisition module 501 is configured to acquire the original feature map data of each 8-bit integer fixed-point Int8 in at least one input row corresponding to the current period in the DDR according to a predetermined data storage format of the DDR, and convert each Int8 The original feature map data of is converted to the corresponding input feature map data;

[0082] The processing module 502 is configured to process each input feature map data through the current operation unit of the convolutional neural network, obtain each output feature map ...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The embodiment of the invention discloses a quantization method, device, electronic equipment and storage medium of feature map data. The method includes: according to the predetermined data storage format of the DDR, acquiring the original feature map data of each Int8 in at least one input line corresponding to the current cycle in the DDR, and converting the original feature map data of each Int8 into its corresponding Input feature map data; process each input feature map data through the current operation unit of the convolutional neural network, obtain each output feature map data in the output line corresponding to the current cycle, and convert each output feature map data into its corresponding The output feature map data of Float16; determine the target quantization parameter of the output row according to the output feature map data of each Float16; use the target quantization parameter to convert the output feature map data of each Float16 into the target feature map data of Int8, and convert the target feature map data of each Int8 Feature map data and target quantization parameters are stored in DDR.

Description

technical field [0001] Embodiments of the present invention relate to the technical field of artificial intelligence, and in particular, to a method, device, electronic device, and storage medium for quantifying feature map data. Background technique [0002] At present, the model of Convolutional Neural Network (CNN for short) is usually trained in 32-bit single-precision floating-point (Float) format, and then the model is transplanted to the edge computing device for model inference deployment to meet high-performance Scenario requirements such as computing and low processing latency. [0003] In edge computing devices with extremely limited resources and costs, for example, ARM FPGA-based edge computing boards, the original feature map (Feature Map) data is usually reduced to 16-bit floating point (Half) or 8-bit integer fixed-point (Int8) in order to improve computing performance and data throughput. Using the original Float 32 feature map data for CNN data processing...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Patents(China)
IPC IPC(8): G06K9/62
CPCG06F18/214
Inventor 朱琳韩布和曲春雨陈振王天飞张红光喻友平吴甜
Owner BEIJING BAIDU NETCOM SCI & TECH CO LTD