CN109598348A - A kind of image pattern obtains, model training method and system - Google Patents
A kind of image pattern obtains, model training method and system Download PDFInfo
- Publication number
- CN109598348A CN109598348A CN201710896985.4A CN201710896985A CN109598348A CN 109598348 A CN109598348 A CN 109598348A CN 201710896985 A CN201710896985 A CN 201710896985A CN 109598348 A CN109598348 A CN 109598348A
- Authority
- CN
- China
- Prior art keywords
- image
- image data
- processing
- cpu
- dedicated
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
- Studio Devices (AREA)
- Image Processing (AREA)
Abstract
The embodiment of the invention provides a kind of acquisition of image pattern, model training method and systems, image pattern acquisition methods are applied to the processing equipment communicated to connect with multiple mechanical arms, processing equipment includes multiple first dedicated cpus for reading image data and multiple second dedicated cpus for image real time transfer, camera is installed on each mechanical arm, this method comprises: obtaining the image data of each camera acquisition and storage in real time;Multiple first dedicated cpus recycle respectively executes reading process;Multiple second dedicated cpus carry out data processing according to preset data processing method, to the image data that multiple first dedicated cpus have been read, and obtain the corresponding image pattern of each image data.As it can be seen that processing equipment uses Stream Processing mode, image pattern can be obtained with real time processed images data, not need re -training machine learning model when obtaining new image data, greatly shorten the model training time.
Description
Technical field
The present invention relates to model training technical field, obtained more particularly to a kind of image pattern, model training method and
System.
Background technique
Requirement with the continuous popularization of the mechanical arm crawl technology based on data-driven, for mechanical arm crawl precision
It is higher and higher.The acquisition of image pattern is an important component in the mechanical arm crawl technology based on data-driven, only
Have and obtain a large amount of image pattern, the training of these image patterns could be passed through and obtain the accurately engineering for mechanical arm crawl
Practise model.
Currently, the acquisition modes of image pattern generally comprise: the camera collection image data by being installed on mechanical arm,
According to the needs of machine learning model training, image data is handled, for example, color of image variation, matrixing etc.
Reason, obtains image pattern.After obtaining great amount of images sample in the same way, it can train to obtain for mechanical arm crawl
Machine learning model.
Using above-mentioned image pattern acquisition modes, after machine learning model is completed in training, if mechanical arm installation is taken the photograph
As head collects new image data, then must be handled new image data to obtain corresponding image pattern, merge
Into above-mentioned great amount of images sample, the training of machine learning model, very cost model training time are then re-started.
Summary of the invention
The embodiment of the present invention be designed to provide a kind of image pattern obtain, model training method and system, to shorten
The model training time.Specific technical solution is as follows:
In a first aspect, being applied to logical with multiple mechanical arms the embodiment of the invention provides a kind of image pattern acquisition methods
Believe the processing equipment of connection, the processing equipment includes multiple first dedicated cpus for reading image data and for image
Multiple second dedicated cpus of data processing are equipped with camera on each mechanical arm, which comprises
The image data of each camera acquisition and storage are obtained in real time;
The multiple first dedicated cpu recycles respectively executes reading process, wherein the reading process are as follows: from being stored
Image data in read the image data of the first pre-set image quantity;
The multiple second dedicated cpu has read the multiple first dedicated cpu according to preset data processing method
Image data carry out data processing, obtain the corresponding image pattern of each image data, wherein the preset data processing
Mode is preset according to the needs of machine learning model training.
Optionally, the multiple second dedicated cpu is divided into N group, wherein first group of CPU includes the of the first preset quantity
Two dedicated cpus, N group CPU include the second dedicated cpu of N preset quantity;
The default processing mode includes: first group of CPU according to the first pre-set image processing mode, to the multiple first
The image data that dedicated cpu has been read is handled, and corresponding first processing result of each image data is obtained, and N group CPU is obtained
(N-1) processing result is taken, and acquired processing result is handled using N pre-set image processing mode, is obtained every
The corresponding image pattern of a image data.
Optionally, first group of CPU is according to the first pre-set image processing mode, to the multiple first dedicated cpu
The image data of reading is handled, and corresponding first processing result of each image data is obtained, and N group CPU obtains (N-1)
Processing result, and acquired processing result is handled using N pre-set image processing mode, obtain each image data
The step of corresponding image pattern, comprising:
First group of CPU is according to the first pre-set image processing mode, the picture number read to the multiple first dedicated cpu
According to being handled, corresponding first processing result of each image data is obtained, first processing result is put into the first result
Queue, N group CPU obtains (N-1) processing result from (N-1) result queue, and uses N pre-set image processing mode
Acquired processing result is handled, the corresponding image pattern of each image data is obtained, described image sample is put into
Final result queue.
Optionally, the default processing mode includes:
The multiple second dedicated cpu is preset processing mode using pre-set image and has been read the multiple first dedicated cpu
The image data taken carries out data processing, obtains the corresponding image numerical example of each image data, wherein the pre-set image is pre-
If processing mode includes: the first pre-set image processing mode, the second pre-set image processing mode to N pre-set image processing side
Formula.
Optionally, described the step of obtaining the image data and storage that each camera acquires in real time, comprising:
The image data of each camera acquisition is obtained in real time;
Described image data are stored, and adjust the storage order of all image datas.
In a first aspect, being applied to and multiple mechanical arm communication links the embodiment of the invention provides a kind of model training method
The processing equipment connect, the processing equipment include for reading multiple first dedicated cpus of image data, at image data
Multiple second dedicated cpus and GPU of reason are equipped with camera on each mechanical arm, which comprises
The image data of each camera acquisition and storage are obtained in real time;
The multiple first dedicated cpu recycles respectively executes reading process, wherein the reading process are as follows: from being stored
Image data in read the image data of the first pre-set image quantity;
The multiple second dedicated cpu has read the multiple first dedicated cpu according to preset data processing method
Image data carry out data processing, obtain the corresponding image pattern of each image data, wherein the preset data processing
Mode is preset according to the needs of machine learning model training;
The GPU obtains described image sample, is joined using acquired image pattern to the machine learning model
Number training.
Optionally, the quantity of the GPU is multiple;
The GPU obtains described image sample, is joined using acquired image pattern to the machine learning model
The step of number training, comprising:
Each GPU obtains the image pattern of default sample size, using acquired image pattern to the machine learning
Model is trained, and the training result based on remaining GPU, adjusts the parameter of the machine learning model.
Optionally, the multiple second dedicated cpu is divided into N group, wherein first group of CPU includes the of the first preset quantity
Two dedicated cpus, N group CPU include the second dedicated cpu of N preset quantity;
The default processing mode includes: first group of CPU according to the first pre-set image processing mode, to the multiple first
The image data that dedicated cpu has been read is handled, and corresponding first processing result of each image data is obtained, and N group CPU is obtained
(N-1) processing result is taken, and acquired processing result is handled using N pre-set image processing mode, is obtained every
The corresponding image pattern of a image data.
Optionally, first group of CPU is according to the first pre-set image processing mode, to the multiple first dedicated cpu
The image data of reading is handled, and corresponding first processing result of each image data is obtained, and N group CPU obtains (N-1)
Processing result, and acquired processing result is handled using N pre-set image processing mode, obtain each image data
The step of corresponding image pattern, comprising:
First group of CPU is according to the first pre-set image processing mode, the picture number read to the multiple first dedicated cpu
According to being handled, corresponding first processing result of each image data is obtained, first processing result is put into the first result
Queue, N group CPU obtains (N-1) processing result from (N-1) result queue, and uses N pre-set image processing mode
Acquired processing result is handled, the corresponding image pattern of each image data is obtained, described image sample is put into
Final result queue;
The GPU obtains described image sample, is joined using acquired image pattern to the machine learning model
The step of number training, comprising:
The GPU obtains described image sample from the final result queue;
Parameter training is carried out to the machine learning model using acquired image pattern.
Optionally, the default processing mode includes:
The multiple second dedicated cpu is preset processing mode using pre-set image and has been read the multiple first dedicated cpu
The image data taken carries out data processing, obtains the corresponding image numerical example of each image data, wherein the pre-set image is pre-
If processing mode includes: the first pre-set image processing mode, the second pre-set image processing mode to N pre-set image processing side
Formula.
Optionally, described the step of obtaining the image data and storage that each camera acquires in real time, comprising:
The image data of each camera acquisition is obtained in real time;
Described image data are stored, and adjust the storage order of all image datas.
The third aspect, the embodiment of the invention provides a kind of image pattern obtain system, the system comprises with multiple machines
The storage equipment and processing equipment of tool arm communication connection, the processing equipment include for reading multiple the first of image data specially
With CPU and for multiple second dedicated cpus of image real time transfer, camera is installed on each mechanical arm, in which:
The storage equipment, for obtaining image data and the storage of each camera acquisition in real time;
The multiple first dedicated cpu is respectively used to circulation and executes reading process, wherein the reading process are as follows: from institute
The image data of the first pre-set image quantity is read in the image data of storage;
The multiple second dedicated cpu is used for according to preset data processing method, to the multiple first dedicated cpu
The image data read carries out data processing, obtains the corresponding image pattern of each image data, wherein the preset number
It is preset according to the needs of machine learning model training according to processing mode.
Optionally, the multiple second dedicated cpu is divided into N group, wherein first group of CPU includes the of the first preset quantity
Two dedicated cpus, N group CPU include the second dedicated cpu of N preset quantity;
First group of CPU, for having read the multiple first dedicated cpu according to the first pre-set image processing mode
The image data taken is handled, and corresponding first processing result of each image data is obtained;
The N group CPU, for obtaining (N-1) processing result, and using N pre-set image processing mode to being obtained
The processing result taken is handled, and the corresponding image pattern of each image data is obtained.
Optionally, first group of CPU is specifically used for according to the first pre-set image processing mode, dedicated to the multiple first
The image data that CPU has been read is handled, and obtains corresponding first processing result of each image data, described first is handled
As a result it is put into the first result queue;
N group CPU is specifically used for obtaining (N-1) processing result from (N-1) result queue, and default using N
Image procossing mode handles acquired processing result, obtains the corresponding image pattern of each image data, will be described
Image pattern is put into final result queue.
Optionally, the multiple second dedicated cpu, for presetting processing mode to the multiple first using pre-set image
The image data that dedicated cpu has been read carries out data processing, obtains the corresponding image numerical example of each image data, wherein institute
Stating pre-set image and presetting processing mode includes: that the first pre-set image processing mode, the second pre-set image processing mode are default to N
Image procossing mode.
Optionally, the storage equipment, specifically for obtaining the image data of each camera acquisition in real time;Described in storage
Image data, and adjust the storage order of all image datas.
Fourth aspect, the embodiment of the invention provides a kind of model training systems, the system comprises with multiple mechanical arms
The storage equipment and processing equipment of communication connection, the processing equipment include dedicated for reading multiple the first of image data
CPU, multiple second dedicated cpus and GPU for image real time transfer are equipped with camera on each mechanical arm, in which:
The storage equipment, for obtaining image data and the storage of each camera acquisition in real time;
The multiple first dedicated cpu is respectively used to circulation and executes reading process, wherein the reading process are as follows: from institute
The image data of the first pre-set image quantity is read in the image data of storage;
The multiple second dedicated cpu is used for according to preset data processing method, to the multiple first dedicated cpu
The image data read carries out data processing, obtains the corresponding image pattern of each image data, wherein the preset number
It is preset according to the needs of machine learning model training according to processing mode;
The GPU, for obtaining described image sample, using acquired image pattern to the machine learning model into
Row parameter training.
Optionally, the quantity of the GPU is multiple;
The GPU obtains the image pattern of default sample size specifically for each GPU, utilizes acquired image sample
This is trained the machine learning model, and the training result based on remaining GPU, adjusts the ginseng of the machine learning model
Number.
Optionally, the multiple second dedicated cpu is divided into N group, wherein first group of CPU includes the of the first preset quantity
Two dedicated cpus, N group CPU include the second dedicated cpu of N preset quantity;
First group of CPU, for having read the multiple first dedicated cpu according to the first pre-set image processing mode
The image data taken is handled, and corresponding first processing result of each image data is obtained
The N group CPU, for obtaining (N-1) processing result, and using N pre-set image processing mode to being obtained
The processing result taken is handled, and the corresponding image pattern of each image data is obtained.
Optionally, first group of CPU is specifically used for according to the first pre-set image processing mode, dedicated to the multiple first
The image data that CPU has been read is handled, and obtains corresponding first processing result of each image data, described first is handled
As a result it is put into the first result queue;
N group CPU is specifically used for obtaining (N-1) processing result from (N-1) result queue, and default using N
Image procossing mode handles acquired processing result, obtains the corresponding image pattern of each image data, will be described
Image pattern is put into final result queue;
The GPU is specifically used for the GPU from the final result queue, obtains described image sample;Using being obtained
The image pattern taken carries out parameter training to the machine learning model.
Optionally, the multiple second dedicated cpu is specifically used for presetting processing mode to the multiple using pre-set image
The image data that first dedicated cpu has been read carries out data processing, obtains the corresponding image numerical example of each image data,
In, it includes: the first pre-set image processing mode, the second pre-set image processing mode to that the pre-set image, which presets processing mode,
N pre-set image processing mode.
Optionally, the storage equipment, specifically for obtaining the image data of each camera acquisition in real time;Described in storage
Image data, and adjust the storage order of all image datas.
The embodiment of the invention also provides a kind of processing equipment, including processor, communication interface, memory, communication bus,
Multiple first dedicated cpus and multiple second dedicated cpus, wherein processor, communication interface, memory, multiple first dedicated cpus
Mutual communication, the processing equipment and multiple mechanical arm communication links are completed by communication bus with multiple second dedicated cpus
It connects, camera is installed on each mechanical arm;
Memory, for storing computer program;
Processor when for executing the program stored on memory, obtains the picture number of each camera acquisition in real time
According to and store;
Multiple first dedicated cpus are respectively used to circulation and execute reading process, wherein the reading process are as follows: from being stored
Image data in read the image data of the first pre-set image quantity;
Multiple second dedicated cpus, for having read the multiple first dedicated cpu according to preset data processing method
The image data taken carries out data processing, obtains the corresponding image pattern of each image data, wherein at the preset data
Reason mode is preset according to the needs of machine learning model training.
The embodiment of the invention also provides another processing equipments, including processor, communication interface, memory, communication are always
Line, multiple first dedicated cpus, multiple second dedicated cpus and GPU, wherein processor, communication interface, memory, multiple first
Dedicated cpu and multiple second dedicated cpus complete mutual communication, the processing equipment and multiple mechanical arms by communication bus
It communicates to connect, camera is installed on each mechanical arm;
Memory, for storing computer program;
Processor when for executing the program stored on memory, obtains the image data of each camera acquisition simultaneously
Storage;
Multiple first dedicated cpus are respectively used to circulation and execute reading process, wherein the reading process are as follows: from being stored
Image data in read the image data of the first pre-set image quantity;
Multiple second dedicated cpus, for having read the multiple first dedicated cpu according to preset data processing method
The image data taken carries out data processing, obtains the corresponding image pattern of each image data, wherein at the preset data
Reason mode is preset according to the needs of machine learning model training;
GPU joins the machine learning model using acquired image pattern for obtaining described image sample
Number training.
The embodiment of the invention also provides a kind of computer readable storage medium, the computer readable storage medium memory
Computer program is contained, the computer program realizes above-mentioned image pattern acquisition methods step when being executed by processor.
The embodiment of the invention also provides another computer readable storage medium, in the computer readable storage medium
It is stored with computer program, the computer program realizes above-mentioned model training method step when being executed by processor.
In scheme provided in an embodiment of the present invention, processing equipment obtains the image data of each camera acquisition in real time and deposits
Storage, multiple first dedicated cpus recycles execution reading process respectively, multiple second dedicated cpus according to preset data processing method,
Data processing is carried out to the image data that multiple first dedicated cpus have been read, obtains the corresponding image pattern of each image data.
As it can be seen that processing equipment uses Stream Processing mode, the image data of each camera acquisition can be handled in real time, obtains image sample
This does not need re -training machine learning model when obtaining new image data, contracts significantly for machine learning model training
The short model training time.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is a kind of flow chart of image pattern acquisition methods provided by the embodiment of the present invention;
Fig. 2 is a kind of schematic diagram that processing mode is preset in embodiment illustrated in fig. 1;
Fig. 3 is a kind of flow chart of model training method provided by the embodiment of the present invention;
Fig. 4 is a kind of structural schematic diagram of image pattern acquisition system provided by the embodiment of the present invention;
Fig. 5 is a kind of structural schematic diagram of model training systems provided by the embodiment of the present invention;
Fig. 6 is the structural schematic diagram of the first processing equipment provided by the embodiment of the present invention;
Fig. 7 is the structural schematic diagram of second of processing equipment provided by the embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
In order to shorten the training time of machine learning model, the embodiment of the invention provides a kind of image pattern acquisition methods
And system, model training method and system, processing equipment and computer readable storage medium.
A kind of image pattern acquisition methods are provided for the embodiments of the invention first below to be introduced.
A kind of image pattern acquisition methods provided by the embodiment of the present invention can be applied to and multiple mechanical arm communication links
The processing equipment connect, the processing equipment may include multiple first dedicated cpu (Central for reading image data
Processing Unit, central processing unit) and for image real time transfer multiple second dedicated cpus.Pacify on each mechanical arm
Equipped with camera, to acquire image data.
As shown in Figure 1, a kind of image pattern acquisition methods, applied to the processing equipment communicated to connect with multiple mechanical arms,
The processing equipment includes multiple first dedicated cpus for reading image data and for multiple the second of image real time transfer
Dedicated cpu is equipped with camera on each mechanical arm, which comprises
S101 obtains the image data of each camera acquisition and storage in real time;
S102, the multiple first dedicated cpu recycles respectively executes reading process, wherein the reading process are as follows: from institute
The image data of the first pre-set image quantity is read in the image data of storage;
S103, the multiple second dedicated cpu is according to preset data processing method, to the multiple first dedicated cpu
The image data read carries out data processing, obtains the corresponding image pattern of each image data, wherein the preset number
It is preset according to the needs of machine learning model training according to processing mode.
As it can be seen that processing equipment obtains the picture number of each camera acquisition in real time in scheme provided in an embodiment of the present invention
According to and store, multiple first dedicated cpus recycle respectively execute reading process, multiple second dedicated cpus according to preset data at
Reason mode carries out data processing to the image data that multiple first dedicated cpus have been read, obtains the corresponding figure of each image data
Decent.Processing equipment uses Stream Processing mode, can handle the image data of each camera acquisition in real time, and acquisition is used for
The image pattern of machine learning model training, does not need re -training machine learning model when obtaining new image data, greatly
It is big to shorten the model training time.
In above-mentioned steps S101, processing equipment can obtain the image data of each camera acquisition in real time, and will figure
As data are stored, for example, can store in hard disk.It is understood that the camera for being each installed on mechanical arm is equal
In acquisition image data, what processing equipment stored is all image datas of camera acquisition.
In above-mentioned steps S102, multiple first dedicated cpus can recycle execution reading process respectively, wherein the reading
Processing are as follows: the image data of the first pre-set image quantity is read from the image data stored.That is, each first is special
The image data of the first pre-set image quantity is constantly read from the image data stored with CPU.
Multiple second dedicated cpus can then read multiple first dedicated cpus according to preset data processing method
Image data carries out data processing, obtains the corresponding image pattern of each image data.Wherein, the preset data processing method
Be according to machine learning model training needs it is preset, for example, may include conversion process, pixel of the image data to matrix
The image procossings modes such as value processing, characteristic area calibration processing, are not specifically limited herein.
Wherein, above-mentioned first pre-set image quantity can be according to processing capacity and the image data institute of the first dedicated cpu
The factors such as the memory accounted for determine, for example, can be 10,15,20 etc., be not specifically limited herein.In one embodiment,
The image data for the first pre-set image quantity that one dedicated cpu is read from the image data stored, with farthest full
The processing capacity of each second dedicated cpu of foot, utilizes the resource of each second dedicated cpu farthest.
For example, each second dedicated cpu can handle 5 image datas simultaneously, 10 the second dedicated cpus are shared, then
The image data for the first pre-set image quantity that first dedicated cpu is read from the image data stored every time is at least 50
It is a, the processing capacity of each second dedicated cpu can be made full use of in this way, maximize resource utilization.
It should be noted that above-mentioned steps S101- step S103 is not necessarily according to the suitable of step S101- step S103
Sequence executes, but is performed simultaneously step S101- step S103, forms the mode of Stream Processing, as long as each camera acquisition figure
As data, then processing equipment can obtain the image data of each camera acquisition and storage, at the same time, step in real time
S102 and step S103 is also being carried out, and reaches the state of dynamic equilibrium, can handle the image of each camera acquisition in real time
Data continuously obtain the image pattern for machine learning model training, and then can constantly training machine learn
Model greatly shortens the model training time.
Data processing is promoted as a kind of embodiment of the embodiment of the present invention in order to further increase resource utilization
Speed, above-mentioned multiple second dedicated cpus can be divided into N group, wherein first group of CPU includes the second dedicated of the first preset quantity
CPU, N group CPU include the second dedicated cpu of N preset quantity.
Correspondingly, above-mentioned default processing mode may include: first group of CPU according to the first pre-set image processing mode, it is right
The image data that the multiple first dedicated cpu has been read is handled, and the corresponding first processing knot of each image data is obtained
Fruit, N group CPU obtain (N-1) processing result, and using N pre-set image processing mode to acquired processing result into
Row processing, obtains the corresponding image pattern of each image data.
The processing mode as needed for image data may be a variety of different processing modes, so in order to meet quick standard
Above-mentioned multiple second dedicated cpus can be divided into N group according to the processing mode needed for image data by process demand really, and one
As in the case of, it is identical as the quantity of the type of required processing mode to the number of packet of the second dedicated cpu, that is to say, that this
Each group of CPU of sample can carry out a kind of processing to image data.For example, processing mode needed for image data is 5 kinds, then just
Above-mentioned multiple second dedicated cpus can be divided into 5 groups, each group of CPU carries out a kind of processing to image data.
For the quantity of the second dedicated cpu included by each group of CPU, that is, the first preset quantity, the second present count
Amount ... N preset quantity, can according to the specific type of the first pre-set image processing mode to N pre-set image processing mode and
The factors such as required memory are determined.If memory needed for certain pre-set image processing mode is larger, corresponding
The quantity of two dedicated cpus can be more, corresponding if memory needed for certain pre-set image processing mode is smaller
The quantity of second dedicated cpu can be less.
For example, it needs to include 4 kinds to the processing mode that image data carries out, i.e. the first pre-set image processing mode is extremely
4th pre-set image processing mode, memory needed for 4 kinds of processing modes successively increases, in order to guarantee that all second dedicated cpus exist
Carry out data processing when resource utilization highest, according to needed for the processing capacity of the second dedicated cpu and every kind of processing mode in
It deposits and is calculated, the quantitative proportion of the second dedicated cpu needed for four kinds of processing modes is 1:2:3:4, then, if second is dedicated
The quantity of CPU is 80, then first group of CPU may include 8 the second dedicated cpus, second group of CPU may include 16 second
Dedicated cpu, third group CPU may include 24 the second dedicated cpus, and the 4th group of CPU may include 36 the second dedicated cpus.
It should be noted that the processing that above-mentioned N group CPU carries out the processing result handled by previous group CPU
It carrying out simultaneously, that is to say, that the image data that first group of CPU has constantly read multiple first dedicated cpus is handled,
The first processing result is obtained, at the same time, second group of CPU to N group CPU is also in the processing knot handled from upper one group of CPU
Processing result is obtained in fruit, and the processing result of acquisition is handled using pre-set image processing mode.
As it can be seen that multiple second dedicated cpus are divided into N group in the present embodiment, it is ensured that steadily carried out to image data
Processing, the processing of image data can be carried out with highest processing speed by guaranteeing each second dedicated cpu, improve the utilization of resources
Rate can also shorten the time of following model training.
As a kind of embodiment of the embodiment of the present invention, above-mentioned first group of CPU according to the first pre-set image processing mode,
The image data read to the multiple first dedicated cpu is handled, and obtains corresponding first processing of each image data
As a result, N group CPU obtains (N-1) processing result, and using N pre-set image processing mode to acquired processing result
The step of being handled, obtaining each image data corresponding image pattern may include:
First group of CPU is according to the first pre-set image processing mode, the picture number read to the multiple first dedicated cpu
According to being handled, corresponding first processing result of each image data is obtained, first processing result is put into the first result
Queue, N group CPU obtains (N-1) processing result from (N-1) result queue, and uses N pre-set image processing mode
Acquired processing result is handled, the corresponding image pattern of each image data is obtained, described image sample is put into
Final result queue.
Since multiple second dedicated cpus are divided into N group, and data processing is carried out simultaneously, so in order to make process flow more
Smoothly stablize, avoids the occurrence of chaotic and lead to handle mistake, as shown in Fig. 2, first group of CPU is according to the first pre-set image processing side
Formula, the image data read to multiple first dedicated cpus are handled, and obtain the corresponding first processing knot of each image data
After fruit, which can be put into the first result queue, in turn, second group of CPU can be from the first result queue
The first processing result is obtained, and is handled according to first processing result of the second pre-set image processing mode to acquisition, according to this
Analogize, N group CPU can then obtain (N-1) processing result from (N-1) result queue, and using at N pre-set image
Reason mode handles acquired processing result, obtains the corresponding image pattern of each image data, then by these figures
Decent is put into final result queue.It is subsequent to carry out machine learning model using the image pattern in final result queue
Training.
As it can be seen that processing result can be put into corresponding processing result queue by every group of CPU in the present embodiment, next group
CPU can obtain processing result from processing result queue and be handled, and process flow can be made more smoothly to stablize, avoided
It causes confusion and causes to handle mistake.
As a kind of embodiment of the embodiment of the present invention, above-mentioned default processing mode may include:
The multiple second dedicated cpu is preset processing mode using pre-set image and has been read the multiple first dedicated cpu
The image data taken carries out data processing, obtains the corresponding image numerical example of each image data, wherein the pre-set image is pre-
If processing mode includes: the first pre-set image processing mode, the second pre-set image processing mode to N pre-set image processing side
Formula.
In one embodiment, multiple second dedicated cpus can preset processing mode to multiple the using pre-set image
The image data that one dedicated cpu has been read carries out data processing, obtains the corresponding image numerical example of each image data.
Wherein, it may include: the first pre-set image processing mode, at the second pre-set image that pre-set image, which presets processing mode,
Reason mode is to N pre-set image processing mode.That is, what each second dedicated cpu can read the first dedicated cpu
Image data carries out the default processing of N kind.For example, pre-set image processing mode shares 5 kinds, i.e. the first pre-set image processing mode is extremely
5th pre-set image processing mode, then after each second dedicated cpu obtains the image data that the first dedicated cpu has been read, just
The image data can be handled according to the first pre-set image processing mode, then to obtained processing result according to second
Pre-set image processing mode is handled, and so on, it is finally handled according to the 5th pre-set image processing mode, obtains figure
Decent.
After obtaining above-mentioned image pattern, image pattern can be put into final result queue by the second dedicated cpu, so as to subsequent
Used when machine learning model training.
As a kind of embodiment of the embodiment of the present invention, the above-mentioned image data for obtaining each camera acquisition in real time is simultaneously
The step of storage may include:
The image data of each camera acquisition is obtained in real time;Described image data are stored, and adjust all image datas
Storage order.
Processing equipment can obtain the image data of each camera acquisition in real time, then store acquired picture number
According to, and adjust the storage order of all image datas.Specifically, since the quantity of mechanical arm is generally more, camera acquisition
Image data quantity it is also just more, when the first dedicated cpu reads image data, be also stored in processing equipment and do not read
The image data taken, while processing equipment is also constantly obtaining image data, in turn, the image that processing equipment is got in storage
After data, the storage order of the image data and the stored image data not being read that can will acquire is adjusted.
In this way, the first dedicated cpu when reading image data, can read the picture number that processing equipment has just stored
According to, can also read current time pre-processing device storage image data, in this way, if the camera shooting installed on mechanical arm
The factors such as the acquisition environment of head change, and may include acquisition ring in the image pattern handled by the second dedicated cpu
The corresponding image pattern of data acquired after the variation of the factors such as border, in turn, the mechanics obtained using the training of these image patterns
The needs of Current mechanical arm use environment and situation can be met by practising model, play the role of new and old image pattern mixing, can
To improve the accuracy for the rote learning model that training obtains.
In order to shorten the training time of machine learning model, the embodiment of the invention also provides a kind of model training methods.
A kind of model training method is provided for the embodiments of the invention below to be introduced.
A kind of model training method provided by the embodiment of the present invention can be applied to and multiple mechanical arms communication connections
Processing equipment, the processing equipment may include for reading multiple first dedicated cpus of image data, for image real time transfer
Multiple second dedicated cpus and GPU (Graphics Processing Unit, graphics processor).It is installed on each mechanical arm
There is camera, to acquire image data.
As shown in figure 3, a kind of model training method, described applied to the processing equipment communicated to connect with multiple mechanical arms
Processing equipment includes for reading multiple first dedicated cpus of image data, for the multiple second dedicated of image real time transfer
CPU and GPU is equipped with camera on each mechanical arm, which comprises
S301 obtains the image data of each camera acquisition and storage in real time;
S302, the multiple first dedicated cpu recycles respectively executes reading process, wherein the reading process are as follows: from institute
The image data of the first pre-set image quantity is read in the image data of storage;
S303, the multiple second dedicated cpu is according to preset data processing method, to the multiple first dedicated cpu
The image data read carries out data processing, obtains the corresponding image pattern of each image data, wherein the preset number
It is preset according to the needs of machine learning model training according to processing mode;
S304, the GPU obtain described image sample, using acquired image pattern to the machine learning model into
Row parameter training.
As it can be seen that processing equipment obtains the picture number of each camera acquisition in real time in scheme provided in an embodiment of the present invention
According to and store, multiple first dedicated cpus recycle respectively execute reading process, multiple second dedicated cpus according to preset data at
Reason mode carries out data processing to the image data that multiple first dedicated cpus have been read, obtains the corresponding figure of each image data
Decent, GPU obtains image pattern, carries out parameter training to machine learning model using acquired image pattern.Processing is set
It is standby to use Stream Processing mode, the image data of each camera acquisition can be handled in real time, obtained and be used for machine learning model
Trained image pattern does not need re -training machine learning model when obtaining new image data, greatly shortens model instruction
Practice the time.
In above-mentioned steps S301, processing equipment can obtain the image data of each camera acquisition in real time, and will figure
As data are stored, for example, can store in hard disk.It is understood that the camera for being each installed on mechanical arm is equal
In acquisition image data, what processing equipment stored is all image datas of camera acquisition.
In above-mentioned steps S302, multiple first dedicated cpus can recycle execution reading process respectively, wherein the reading
Processing are as follows: the image data of the first pre-set image quantity is read from the image data stored.That is, each first is special
The image data of the first pre-set image quantity is constantly read from the image data stored with CPU.
Multiple second dedicated cpus can then read multiple first dedicated cpus according to preset data processing method
Image data carries out data processing, obtains the corresponding image pattern of each image data.Wherein, the preset data processing method
Be according to machine learning model training needs it is preset, for example, may include conversion process, pixel of the image data to matrix
The image procossings modes such as value processing, characteristic area calibration processing, are not specifically limited herein.
Wherein, above-mentioned first pre-set image quantity can be according to processing capacity and the image data institute of the first dedicated cpu
The factors such as the memory accounted for determine, for example, can be 10,15,20 etc., be not specifically limited herein.In one embodiment,
The image data for the first pre-set image quantity that one dedicated cpu is read from the image data stored, with farthest full
The processing capacity of each second dedicated cpu of foot, utilizes the resource of each second dedicated cpu farthest.
For example, each second dedicated cpu can handle 5 image datas simultaneously, 10 the second dedicated cpus are shared, then
The image data for the first pre-set image quantity that first dedicated cpu is read from the image data stored every time is at least 50
It is a, the processing capacity of each second dedicated cpu can be made full use of in this way, maximize resource utilization.
In above-mentioned steps S304, the image pattern that available second dedicated cpu of GPU is handled, using acquired
Image pattern parameter training is carried out to the machine learning model that constructs in advance, and then required machine learning mould can be obtained
Type.
It should be noted that above-mentioned steps S301- step S304 is not necessarily according to the suitable of step S301- step S303
Sequence executes, but is performed simultaneously step S301- step S304, forms the mode of Stream Processing, as long as each camera acquisition figure
As data, then processing equipment can obtain the image data of each camera acquisition and storage, at the same time, step in real time
S302- step S304 is also being carried out, and reaches the state of dynamic equilibrium, can handle the picture number of each camera acquisition in real time
According to, continuously image pattern of the acquisition for machine learning model training, and constantly training machine learning model, significantly
Shorten the model training time.
As a kind of embodiment of the embodiment of the present invention, the quantity of above-mentioned GPU can be multiple;
Above-mentioned GPU obtains described image sample, is joined using acquired image pattern to the machine learning model
Trained step is counted, may include:
Each GPU obtains the image pattern of default sample size, using acquired image pattern to the machine learning
Model is trained, and the training result based on remaining GPU, adjusts the parameter of the machine learning model.
Processing equipment may include multiple GPU, in this case, the figure of each available default sample size of GPU
Decent, and machine learning model is trained using acquired image pattern.Wherein, default sample size can basis
The factors such as the processing capacity of GPU and the data volume of each image pattern are determined, for example, can be 5,10,15 etc.,
This is not specifically limited.Certainly, the quantity for the image pattern that each GPU is obtained is most can make full use of the resource of GPU
It is good, it can be further improved resource utilization in this way, shorten the model training time.
Due to above-mentioned multiple GPU simultaneously machine learning model is trained, so each GPU to machine learning model into
When row training, the parameter of machine learning model can be adjusted based on the training result of remaining GPU.It can pass through between multiple GPU
Communication bus and communication interface etc. are communicated, and to know the training result of remaining GPU, and adjust machine based on the training result
The parameter of learning model.
For example, processing equipment includes 4 GPU, then 4 GPU obtain image pattern simultaneously, while the figure obtained is utilized
The training of decent progress machine learning model, and communicate with each other, the training result of remaining GPU is obtained, in turn, adjusts engineering
Practise the parameter of model.
As it can be seen that processing equipment includes multiple GPU in the present embodiment, multiple GPU simultaneously instruct machine learning model
Practice, further increase model training efficiency, shortens the model training time.
Data processing is promoted as a kind of embodiment of the embodiment of the present invention in order to further increase resource utilization
Speed, above-mentioned multiple second dedicated cpus can be divided into N group, wherein first group of CPU includes the second dedicated of the first preset quantity
CPU, N group CPU include the second dedicated cpu of N preset quantity.
Correspondingly, above-mentioned default processing mode includes: first group of CPU according to the first pre-set image processing mode, to described
The image data that multiple first dedicated cpus have been read is handled, and corresponding first processing result of each image data is obtained, the
N group CPU obtain (N-1) processing result, and using N pre-set image processing mode to acquired processing result at
Reason, obtains the corresponding image pattern of each image data.
The processing mode as needed for image data may be a variety of different processing modes, so in order to meet quick standard
Above-mentioned multiple second dedicated cpus can be divided into N group according to the processing mode needed for image data by process demand really, and one
As in the case of, it is identical as the quantity of the type of required processing mode to the number of packet of the second dedicated cpu, that is to say, that this
Each group of CPU of sample can carry out a kind of processing to image data.For example, processing mode needed for image data is 5 kinds, then just
Above-mentioned multiple second dedicated cpus can be divided into 5 groups, each group of CPU carries out a kind of processing to image data.
For the quantity of the second dedicated cpu included by each group of CPU, that is, the first preset quantity, the second present count
Amount ... N preset quantity, can according to the specific type of the first pre-set image processing mode to N pre-set image processing mode and
The factors such as required memory are determined.If memory needed for certain pre-set image processing mode is larger, corresponding
The quantity of two dedicated cpus can be more, corresponding if memory needed for certain pre-set image processing mode is smaller
The quantity of second dedicated cpu can be less.
For example, it needs to include 4 kinds to the processing mode that image data carries out, i.e. the first pre-set image processing mode is extremely
4th pre-set image processing mode, memory needed for 4 kinds of processing modes successively increases, in order to guarantee that all second dedicated cpus exist
Carry out data processing when resource utilization highest, according to needed for the processing capacity of the second dedicated cpu and every kind of processing mode in
It deposits and is calculated, the quantitative proportion of the second dedicated cpu needed for four kinds of processing modes is 1:2:3:4, then, if second is dedicated
The quantity of CPU is 80, then first group of CPU may include 8 the second dedicated cpus, second group of CPU may include 16 second
Dedicated cpu, third group CPU may include 24 the second dedicated cpus, and the 4th group of CPU may include 36 the second dedicated cpus.
It should be noted that the processing that above-mentioned N group CPU carries out the processing result handled by previous group CPU
It carrying out simultaneously, that is to say, that the image data that first group of CPU has constantly read multiple first dedicated cpus is handled,
The first processing result is obtained, at the same time, second group of CPU to N group CPU is also in the processing knot handled from upper one group of CPU
Processing result is obtained in fruit, and the processing result of acquisition is handled using pre-set image processing mode.
As it can be seen that multiple second dedicated cpus are divided into N group in the present embodiment, it is ensured that steadily carried out to image data
Processing, the processing of image data can be carried out with highest processing speed by guaranteeing each second dedicated cpu, improve the utilization of resources
Rate shortens the time of model training.
As a kind of embodiment of the embodiment of the present invention, above-mentioned first group of CPU according to the first pre-set image processing mode,
The image data read to the multiple first dedicated cpu is handled, and obtains corresponding first processing of each image data
As a result, N group CPU obtains (N-1) processing result, and using N pre-set image processing mode to acquired processing result
The step of being handled, obtaining each image data corresponding image pattern may include:
First group of CPU is according to the first pre-set image processing mode, the picture number read to the multiple first dedicated cpu
According to being handled, corresponding first processing result of each image data is obtained, first processing result is put into the first result
Queue, N group CPU obtains (N-1) processing result from (N-1) result queue, and uses N pre-set image processing mode
Acquired processing result is handled, the corresponding image pattern of each image data is obtained, described image sample is put into
Final result queue.
Since multiple second dedicated cpus are divided into N group, and data processing is carried out simultaneously, so in order to make process flow more
Smoothly stablize, avoids the occurrence of chaotic and lead to handle mistake, first group of CPU is according to the first pre-set image processing mode, to multiple
The image data that first dedicated cpu has been read is handled, can be with after obtaining corresponding first processing result of each image data
First processing result is put into the first result queue, in turn, from second group of CPU can obtain first from the first result queue
Reason as a result, and handled according to first processing result of the second pre-set image processing mode to acquisition, and so on, N group
CPU can then obtain (N-1) processing result from (N-1) result queue, and using N pre-set image processing mode to institute
The processing result of acquisition is handled, and is obtained the corresponding image pattern of each image data, is then put into these image patterns
Final result queue.
Correspondingly, above-mentioned GPU obtains described image sample, using acquired image pattern to the machine learning model
The step of progress parameter training, may include:
The GPU obtains described image sample from the final result queue;Using acquired image pattern to institute
It states machine learning model and carries out parameter training.
N group CPU handled acquired processing result using N pre-set image processing mode, and will be obtained
Each image pattern is put into final result queue, and in turn, GPU can obtain image pattern, in turn from the final result queue
Parameter training is carried out to machine learning model using acquired image pattern.
As it can be seen that processing result can be put into corresponding processing result queue by every group of CPU in the present embodiment, next group
CPU can obtain processing result from processing result queue and be handled, and GPU can obtain image from final result queue
Sample carries out the parameter training of machine learning model, process flow can be made more smoothly to stablize, and avoids the occurrence of chaotic and causes
Handle mistake.
As a kind of embodiment of the embodiment of the present invention, above-mentioned default processing mode includes:
The multiple second dedicated cpu is preset processing mode using pre-set image and has been read the multiple first dedicated cpu
The image data taken carries out data processing, obtains the corresponding image numerical example of each image data, wherein the pre-set image is pre-
If processing mode includes: the first pre-set image processing mode, the second pre-set image processing mode to N pre-set image processing side
Formula.
In one embodiment, multiple second dedicated cpus can preset processing mode to multiple the using pre-set image
The image data that one dedicated cpu has been read carries out data processing, obtains the corresponding image numerical example of each image data.
Wherein, it may include: the first pre-set image processing mode, at the second pre-set image that pre-set image, which presets processing mode,
Reason mode is to N pre-set image processing mode.That is, what each second dedicated cpu can read the first dedicated cpu
Image data carries out the default processing of N kind.For example, pre-set image processing mode shares 5 kinds, i.e. the first pre-set image processing mode is extremely
5th pre-set image processing mode, then after each second dedicated cpu obtains the image data that the first dedicated cpu has been read, just
The image data can be handled according to the first pre-set image processing mode, then to obtained processing result according to second
Pre-set image processing mode is handled, and so on, it is finally handled according to the 5th pre-set image processing mode, obtains figure
Decent.
After obtaining above-mentioned image pattern, image pattern can be put into final result queue by the second dedicated cpu, so as into
When the training of row machine learning model, GPU can obtain image pattern from final result queue, join to machine learning model
Number adjustment.
As a kind of embodiment of the embodiment of the present invention, the above-mentioned image data for obtaining each camera acquisition in real time is simultaneously
The step of storage may include:
The image data of each camera acquisition is obtained in real time;Described image data are stored, and adjust all image datas
Storage order.
Processing equipment can obtain the image data of each camera acquisition in real time, then store acquired picture number
According to, and adjust the storage order of all image datas.Specifically, since the quantity of mechanical arm is generally more, camera acquisition
Image data quantity it is also just more, when the first dedicated cpu reads image data, be also stored in processing equipment and do not read
The image data taken, while processing equipment is also constantly obtaining image data, in turn, the image that processing equipment is got in storage
After data, the storage order of the image data and the stored image data not being read that can will acquire is adjusted.
In this way, the first dedicated cpu when reading image data, can read the picture number that processing equipment has just stored
According to the image data that processing equipment between current time stores can also be read, in this way, if the camera shooting installed on mechanical arm
The factors such as the acquisition environment of head change, and may include acquisition ring in the image pattern handled by the second dedicated cpu
The corresponding image pattern of data acquired after the variation of the factors such as border, in turn, the mechanics obtained using the training of these image patterns
The needs of Current mechanical arm use environment and situation can be met by practising model, play the role of new and old image pattern mixing, can
To improve the accuracy for the rote learning model that training obtains.
Corresponding to above-mentioned image pattern acquisition methods, the embodiment of the invention also provides a kind of image patterns to obtain system.
It it is provided for the embodiments of the invention a kind of image pattern below obtains system and be introduced.
As shown in figure 4, a kind of image pattern obtains system, the system comprises the storages communicated to connect with multiple mechanical arms
Equipment 410 and processing equipment 420, the processing equipment 420 include multiple first dedicated cpus 4201 for reading image data
With multiple second dedicated cpus 4202 for image real time transfer, camera is installed on each mechanical arm, in which:
The storage equipment 410, for obtaining image data and the storage of each camera acquisition in real time;
The multiple first dedicated cpu 4201 is respectively used to circulation and executes reading process, wherein the reading process are as follows:
The image data of the first pre-set image quantity is read from the image data stored;
The multiple second dedicated cpu 4202 is used for according to preset data processing method, dedicated to the multiple first
The image data that CPU has been read carries out data processing, obtains the corresponding image pattern of each image data, wherein described default
Data processing method be according to machine learning model training needs it is preset.
As it can be seen that storage equipment obtains the picture number of each camera acquisition in real time in scheme provided in an embodiment of the present invention
According to and store, multiple first dedicated cpus recycle respectively execute reading process, multiple second dedicated cpus according to preset data at
Reason mode carries out data processing to the image data that multiple first dedicated cpus have been read, obtains the corresponding figure of each image data
Decent.Processing equipment uses Stream Processing mode, can handle the image data of each camera acquisition in real time, and acquisition is used for
The image pattern of machine learning model training, does not need re -training machine learning model when obtaining new image data, greatly
It is big to shorten the model training time.
As a kind of embodiment of the embodiment of the present invention, above-mentioned multiple second dedicated cpus can be divided into N group, wherein the
One group of CPU includes the second dedicated cpu of the first preset quantity, and N group CPU includes the second dedicated cpu of N preset quantity;
Above-mentioned first group of CPU, for having read the multiple first dedicated cpu according to the first pre-set image processing mode
The image data taken is handled, and corresponding first processing result of each image data is obtained;
Above-mentioned N group CPU, for obtaining (N-1) processing result, and using N pre-set image processing mode to being obtained
The processing result taken is handled, and the corresponding image pattern of each image data is obtained.
As a kind of embodiment of the embodiment of the present invention, above-mentioned first group of CPU is specifically used for according to the first pre-set image
Processing mode, the image data read to the multiple first dedicated cpu are handled, and it is corresponding to obtain each image data
First processing result is put into the first result queue by the first processing result;
N group CPU is specifically used for obtaining (N-1) processing result from (N-1) result queue, and default using N
Image procossing mode handles acquired processing result, obtains the corresponding image pattern of each image data, will be described
Image pattern is put into final result queue.
As a kind of embodiment of the embodiment of the present invention, above-mentioned multiple second dedicated cpus, for pre- using pre-set image
If processing mode carries out data processing to the image data that the multiple first dedicated cpu has been read, each image data is obtained
Corresponding image numerical example, wherein it includes: the first pre-set image processing mode, second that the pre-set image, which presets processing mode,
Pre-set image processing mode is to N pre-set image processing mode.
As a kind of embodiment of the embodiment of the present invention, above-mentioned storage equipment, specifically for obtaining each camera shooting in real time
The image data of head acquisition;Described image data are stored, and adjust the storage order of all image datas.
Storage equipment can obtain the image data of each camera acquisition in real time, then store acquired image data,
And adjust the storage order of all image datas.Specifically, since the quantity of mechanical arm is generally more, the figure of camera acquisition
As the quantity of data is also more, when the first dedicated cpu reads image data, stores in equipment and be also stored with the figure not being read
As data, while storing equipment and also constantly obtaining image data, in turn, the image data that storage equipment is got in storage
Afterwards, the storage order of the image data that can be will acquire and the stored image data not being read is adjusted.
In this way, the first dedicated cpu when reading image data, can read the image that storage equipment has just stored,
The image data that equipment storage is stored before current time can also be read, in this way, if the camera installed on mechanical arm
Factors such as acquisition environment when changing, may include acquisition ring in the image pattern handled by the second dedicated cpu
The corresponding image pattern of data acquired after the variation of the factors such as border, in turn, the mechanics obtained using the training of these image patterns
The needs of Current mechanical arm use environment and situation can be met by practising model, play the role of new and old image pattern mixing, can
To improve the accuracy for the rote learning model that training obtains.
Corresponding to above-mentioned model training method, the embodiment of the invention also provides a kind of model training systems.
A kind of model training systems are provided for the embodiments of the invention below to be introduced.
As shown in figure 5, a kind of model training systems, the system comprises the storage equipment communicated to connect with multiple mechanical arms
510 and processing equipment 520, the processing equipment 520 includes for reading multiple first dedicated cpus 5201 of image data, using
In being equipped with camera on multiple second dedicated cpus 5202 and GPU5203 of image real time transfer, each mechanical arm, in which:
The storage equipment 510, for obtaining image data and the storage of each camera acquisition in real time;
The multiple first dedicated cpu 5201 is respectively used to circulation and executes reading process, wherein the reading process are as follows:
The image data of the first pre-set image quantity is read from the image data stored;
The multiple second dedicated cpu 5202 is used for according to preset data processing method, dedicated to the multiple first
The image data that CPU has been read carries out data processing, obtains the corresponding image pattern of each image data, wherein described default
Data processing method be according to machine learning model training needs it is preset;
The GPU5203, for obtaining described image sample, using acquired image pattern to the machine learning mould
Type carries out parameter training.
As it can be seen that storage equipment obtains the picture number of each camera acquisition in real time in scheme provided in an embodiment of the present invention
According to and store, multiple first dedicated cpus recycle respectively execute reading process, multiple second dedicated cpus according to preset data at
Reason mode carries out data processing to the image data that multiple first dedicated cpus have been read, obtains the corresponding figure of each image data
Decent, GPU obtains image pattern, carries out parameter training to machine learning model using acquired image pattern.Processing is set
It is standby to use Stream Processing mode, the image data of each camera acquisition can be handled in real time, obtained and be used for machine learning model
Trained image pattern does not need re -training machine learning model when obtaining new image data, greatly shortens model instruction
Practice the time.
As a kind of embodiment of the embodiment of the present invention, the quantity of above-mentioned GPU can be multiple;
The GPU obtains the image pattern of default sample size specifically for each GPU, utilizes acquired image sample
This is trained the machine learning model, and the training result based on remaining GPU, adjusts the ginseng of the machine learning model
Number.
As a kind of embodiment of the embodiment of the present invention, above-mentioned multiple second dedicated cpus are divided into N group, wherein first group
CPU includes the second dedicated cpu of the first preset quantity, and N group CPU includes the second dedicated cpu of N preset quantity;
Above-mentioned first group of CPU, for having read the multiple first dedicated cpu according to the first pre-set image processing mode
The image data taken is handled, and corresponding first processing result of each image data is obtained
Above-mentioned N group CPU, for obtaining (N-1) processing result, and using N pre-set image processing mode to being obtained
The processing result taken is handled, and the corresponding image pattern of each image data is obtained.
As a kind of embodiment of the embodiment of the present invention, above-mentioned first group of CPU is specifically used for according to the first pre-set image
Processing mode, the image data read to the multiple first dedicated cpu are handled, and it is corresponding to obtain each image data
First processing result is put into the first result queue by the first processing result;
N group CPU is specifically used for obtaining (N-1) processing result from (N-1) result queue, and default using N
Image procossing mode handles acquired processing result, obtains the corresponding image pattern of each image data, will be described
Image pattern is put into final result queue;
The GPU is specifically used for the GPU from the final result queue, obtains described image sample;Using being obtained
The image pattern taken carries out parameter training to the machine learning model.
As a kind of embodiment of the embodiment of the present invention, above-mentioned multiple second dedicated cpus are specifically used for using default figure
Data processing is carried out to the image data that the multiple first dedicated cpu has been read as presetting processing mode, obtains each image
The corresponding image numerical example of data, wherein the pre-set image preset processing mode include: the first pre-set image processing mode,
Second pre-set image processing mode is to N pre-set image processing mode.
As a kind of embodiment of the embodiment of the present invention, above-mentioned storage equipment, specifically for obtaining each camera shooting in real time
The image data of head acquisition;Described image data are stored, and adjust the storage order of all image datas.
Storage equipment can obtain the image data of each camera acquisition in real time, then store acquired image data,
And adjust the storage order of all image datas.Specifically, since the quantity of mechanical arm is generally more, the figure of camera acquisition
As the quantity of data is also more, when the first dedicated cpu reads image data, stores in equipment and be also stored with the figure not being read
As data, while storing equipment and also constantly obtaining image data, in turn, the image data that storage equipment is got in storage
Afterwards, the storage order of the image data that can be will acquire and the stored image data not being read is adjusted.
In this way, the first dedicated cpu when reading image data, can read the image that storage equipment has just stored,
The image data that equipment storage is stored before current time can also be read, in this way, if the camera installed on mechanical arm
Factors such as acquisition environment when changing, may include acquisition ring in the image pattern handled by the second dedicated cpu
The corresponding image pattern of data acquired after the variation of the factors such as border, in turn, the mechanics obtained using the training of these image patterns
The needs of Current mechanical arm use environment and situation can be met by practising model, play the role of new and old image pattern mixing, can
To improve the accuracy for the rote learning model that training obtains.
The embodiment of the invention also provides a kind of processing equipments, as shown in fig. 6, include processor 601, communication interface 602,
Memory 603, communication bus 604, multiple first dedicated cpus 605 and multiple second dedicated cpus 606, wherein processor 601,
Communication interface 602, memory 603, multiple first dedicated cpus 605 and multiple second dedicated cpus 606 are complete by communication bus 604
At mutual communication, the processing equipment and multiple mechanical arms are communicated to connect, and are equipped with camera on each mechanical arm;
Memory 603, for storing computer program;
Processor 601 when for executing the program stored on memory 603, obtains each camera acquisition in real time
Image data simultaneously stores;
Multiple first dedicated cpus 605 are respectively used to circulation and execute reading process, wherein the reading process are as follows: from institute
The image data of the first pre-set image quantity is read in the image data of storage;
Multiple second dedicated cpus 606 are used for according to preset data processing method, to the multiple first dedicated cpu
The image data of reading carries out data processing, obtains the corresponding image pattern of each image data, wherein the preset data
Processing mode is preset according to the needs of machine learning model training.
As it can be seen that processing equipment obtains the picture number of each camera acquisition in real time in scheme provided in an embodiment of the present invention
According to and store, multiple first dedicated cpus recycle respectively execute reading process, multiple second dedicated cpus according to preset data at
Reason mode carries out data processing to the image data that multiple first dedicated cpus have been read, obtains the corresponding figure of each image data
Decent.Processing equipment uses Stream Processing mode, can handle the image data of each camera acquisition in real time, and acquisition is used for
The image pattern of machine learning model training, does not need re -training machine learning model when obtaining new image data, greatly
It is big to shorten the model training time.
The communication bus that above-mentioned processing equipment is mentioned can be Peripheral Component Interconnect standard (Peripheral Component
Interconnect, PCI) bus or expanding the industrial standard structure (Extended Industry Standard
Architecture, EISA) bus etc..The communication bus can be divided into address bus, data/address bus, control bus etc..For just
It is only indicated with a thick line in expression, figure, it is not intended that an only bus or a type of bus.
Communication interface is for the communication between above-mentioned processing equipment and other equipment.
Memory may include random access memory (Random Access Memory, RAM), also may include non-easy
The property lost memory (Non-Volatile Memory, NVM), for example, at least a magnetic disk storage.Optionally, memory may be used also
To be storage device that at least one is located remotely from aforementioned processor.
Above-mentioned processor can be general processor, including central processing unit (Central Processing Unit,
CPU), network processing unit (Network Processor, NP) etc.;It can also be digital signal processor (Digital Signal
Processing, DSP), it is specific integrated circuit (Application Specific Integrated Circuit, ASIC), existing
It is field programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic device, discrete
Door or transistor logic, discrete hardware components.
Wherein, above-mentioned multiple second dedicated cpus can be divided into N group, wherein first group of CPU includes the first preset quantity
Second dedicated cpu, N group CPU include the second dedicated cpu of N preset quantity;
The default processing mode includes: first group of CPU according to the first pre-set image processing mode, to the multiple first
The image data that dedicated cpu has been read is handled, and corresponding first processing result of each image data is obtained, and N group CPU is obtained
(N-1) processing result is taken, and acquired processing result is handled using N pre-set image processing mode, is obtained every
The corresponding image pattern of a image data.
Wherein, above-mentioned first group of CPU has read the multiple first dedicated cpu according to the first pre-set image processing mode
The image data taken is handled, and corresponding first processing result of each image data is obtained, and N group CPU is obtained at (N-1)
Reason as a result, and acquired processing result is handled using N pre-set image processing mode, obtain each image data pair
The step of image pattern answered, may include:
First group of CPU is according to the first pre-set image processing mode, the picture number read to the multiple first dedicated cpu
According to being handled, corresponding first processing result of each image data is obtained, first processing result is put into the first result
Queue, N group CPU obtains (N-1) processing result from (N-1) result queue, and uses N pre-set image processing mode
Acquired processing result is handled, the corresponding image pattern of each image data is obtained, described image sample is put into
Final result queue.
Wherein, above-mentioned default processing mode may include:
The multiple second dedicated cpu is preset processing mode using pre-set image and has been read the multiple first dedicated cpu
The image data taken carries out data processing, obtains the corresponding image numerical example of each image data, wherein the pre-set image is pre-
If processing mode includes: the first pre-set image processing mode, the second pre-set image processing mode to N pre-set image processing side
Formula.
Wherein, above-mentioned the step of obtaining the image data and storage that each camera acquires in real time, comprising:
The image data of each camera acquisition is obtained in real time;
Described image data are stored, and adjust the storage order of all image datas.
The embodiment of the invention also provides another processing equipments, as shown in fig. 7, comprises processor 701, communication interface
702, memory 703, communication bus 704, multiple first dedicated cpus 705, multiple second dedicated cpus 706 and GPU707, wherein
Processor 701, communication interface 702, memory 703, multiple first dedicated cpus 705 and multiple second dedicated cpus 706 pass through logical
Letter bus 704 completes mutual communication, and the processing equipment and multiple mechanical arms are communicated to connect, are equipped on each mechanical arm
Camera;
Memory 703, for storing computer program;
Processor 701 when for executing the program stored on memory 703, obtains the image of each camera acquisition
Data simultaneously store;
Multiple first dedicated cpus 705 are respectively used to circulation and execute reading process, wherein the reading process are as follows: from institute
The image data of the first pre-set image quantity is read in the image data of storage;
Multiple second dedicated cpus 706 are used for according to preset data processing method, to the multiple first dedicated cpu
The image data of reading carries out data processing, obtains the corresponding image pattern of each image data, wherein the preset data
Processing mode is preset according to the needs of machine learning model training;
GPU707, for obtaining described image sample, using acquired image pattern to the machine learning model into
Row parameter training.
As it can be seen that processing equipment obtains the picture number of each camera acquisition in real time in scheme provided in an embodiment of the present invention
According to and store, multiple first dedicated cpus recycle respectively execute reading process, multiple second dedicated cpus according to preset data at
Reason mode carries out data processing to the image data that multiple first dedicated cpus have been read, obtains the corresponding figure of each image data
Decent, GPU obtains image pattern, carries out parameter training to machine learning model using acquired image pattern.Processing is set
It is standby to use Stream Processing mode, the image data of each camera acquisition can be handled in real time, obtained and be used for machine learning model
Trained image pattern does not need re -training machine learning model when obtaining new image data, greatly shortens model instruction
Practice the time.
The communication bus that above-mentioned processing equipment is mentioned can be Peripheral Component Interconnect standard (Peripheral Component
Interconnect, PCI) bus or expanding the industrial standard structure (Extended Industry Standard
Architecture, EISA) bus etc..The communication bus can be divided into address bus, data/address bus, control bus etc..For just
It is only indicated with a thick line in expression, figure, it is not intended that an only bus or a type of bus.
Communication interface is for the communication between above-mentioned processing equipment and other equipment.
Memory may include random access memory (Random Access Memory, RAM), also may include non-easy
The property lost memory (Non-Volatile Memory, NVM), for example, at least a magnetic disk storage.Optionally, memory may be used also
To be storage device that at least one is located remotely from aforementioned processor.
Above-mentioned processor can be general processor, including central processing unit (Central Processing Unit,
CPU), network processing unit (Network Processor, NP) etc.;It can also be digital signal processor (Digital Signal
Processing, DSP), it is specific integrated circuit (Application Specific Integrated Circuit, ASIC), existing
It is field programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic device, discrete
Door or transistor logic, discrete hardware components.
Wherein, the quantity of above-mentioned GPU can be multiple;
The GPU obtains described image sample, is joined using acquired image pattern to the machine learning model
The step of number training, comprising:
Each GPU obtains the image pattern of default sample size, using acquired image pattern to the machine learning
Model is trained, and the training result based on remaining GPU, adjusts the parameter of the machine learning model.
Wherein, above-mentioned multiple second dedicated cpus can be divided into N group, wherein first group of CPU includes the first preset quantity
Second dedicated cpu, N group CPU include the second dedicated cpu of N preset quantity;
The default processing mode includes: first group of CPU according to the first pre-set image processing mode, to the multiple first
The image data that dedicated cpu has been read is handled, and corresponding first processing result of each image data is obtained, and N group CPU is obtained
(N-1) processing result is taken, and acquired processing result is handled using N pre-set image processing mode, is obtained every
The corresponding image pattern of a image data.
Wherein, above-mentioned first group of CPU has read the multiple first dedicated cpu according to the first pre-set image processing mode
The image data taken is handled, and corresponding first processing result of each image data is obtained, and N group CPU is obtained at (N-1)
Reason as a result, and acquired processing result is handled using N pre-set image processing mode, obtain each image data pair
The step of image pattern answered, may include:
First group of CPU is according to the first pre-set image processing mode, the picture number read to the multiple first dedicated cpu
According to being handled, corresponding first processing result of each image data is obtained, first processing result is put into the first result
Queue, N group CPU obtains (N-1) processing result from (N-1) result queue, and uses N pre-set image processing mode
Acquired processing result is handled, the corresponding image pattern of each image data is obtained, described image sample is put into
Final result queue;
Above-mentioned GPU obtains described image sample, is joined using acquired image pattern to the machine learning model
Trained step is counted, may include:
The GPU obtains described image sample from the final result queue;
Parameter training is carried out to the machine learning model using acquired image pattern.
Wherein, above-mentioned default processing mode may include:
The multiple second dedicated cpu is preset processing mode using pre-set image and has been read the multiple first dedicated cpu
The image data taken carries out data processing, obtains the corresponding image numerical example of each image data, wherein the pre-set image is pre-
If processing mode includes: the first pre-set image processing mode, the second pre-set image processing mode to N pre-set image processing side
Formula.
Wherein, above-mentioned the step of obtaining the image data and storage that each camera acquires in real time, may include:
The image data of each camera acquisition is obtained in real time;
Described image data are stored, and adjust the storage order of all image datas.
The embodiment of the invention also provides a kind of computer readable storage medium, the computer readable storage medium memory
Computer program is contained, the computer program performs the steps of when being executed by processor
The image data of each camera acquisition and storage are obtained in real time;
The multiple first dedicated cpu recycles respectively executes reading process, wherein the reading process are as follows: from being stored
Image data in read the image data of the first pre-set image quantity;
The multiple second dedicated cpu has read the multiple first dedicated cpu according to preset data processing method
Image data carry out data processing, obtain the corresponding image pattern of each image data, wherein the preset data processing
Mode is preset according to the needs of machine learning model training.
As it can be seen that when computer program is executed by processor, being obtained in real time each in scheme provided by the embodiment of the present invention
The image data and storage, multiple first dedicated cpus of camera acquisition recycle execution reading process respectively, and multiple second is dedicated
CPU carries out data processing according to preset data processing method, to the image data that multiple first dedicated cpus have been read, and obtains
The corresponding image pattern of each image data.Processing equipment uses Stream Processing mode, can handle each camera in real time and adopt
The image data of collection obtains the image pattern for machine learning model training, and weight is not needed when obtaining new image data
New training machine learning model, greatly shortens the model training time.
Wherein, above-mentioned multiple second dedicated cpus can be divided into N group, wherein first group of CPU includes the first preset quantity
Second dedicated cpu, N group CPU include the second dedicated cpu of N preset quantity;
The default processing mode includes: first group of CPU according to the first pre-set image processing mode, to the multiple first
The image data that dedicated cpu has been read is handled, and corresponding first processing result of each image data is obtained, and N group CPU is obtained
(N-1) processing result is taken, and acquired processing result is handled using N pre-set image processing mode, is obtained every
The corresponding image pattern of a image data.
Wherein, above-mentioned first group of CPU has read the multiple first dedicated cpu according to the first pre-set image processing mode
The image data taken is handled, and corresponding first processing result of each image data is obtained, and N group CPU is obtained at (N-1)
Reason as a result, and acquired processing result is handled using N pre-set image processing mode, obtain each image data pair
The step of image pattern answered, may include:
First group of CPU is according to the first pre-set image processing mode, the picture number read to the multiple first dedicated cpu
According to being handled, corresponding first processing result of each image data is obtained, first processing result is put into the first result
Queue, N group CPU obtains (N-1) processing result from (N-1) result queue, and uses N pre-set image processing mode
Acquired processing result is handled, the corresponding image pattern of each image data is obtained, described image sample is put into
Final result queue.
Wherein, above-mentioned default processing mode may include:
The multiple second dedicated cpu is preset processing mode using pre-set image and has been read the multiple first dedicated cpu
The image data taken carries out data processing, obtains the corresponding image numerical example of each image data, wherein the pre-set image is pre-
If processing mode includes: the first pre-set image processing mode, the second pre-set image processing mode to N pre-set image processing side
Formula.
Wherein, above-mentioned the step of obtaining the image data and storage that each camera acquires in real time, comprising:
The image data of each camera acquisition is obtained in real time;
Described image data are stored, and adjust the storage order of all image datas.
The embodiment of the invention also provides another computer readable storage medium, in the computer readable storage medium
It is stored with computer program, the computer program performs the steps of when being executed by processor
The image data of each camera acquisition and storage are obtained in real time;
The multiple first dedicated cpu recycles respectively executes reading process, wherein the reading process are as follows: from being stored
Image data in read the image data of the first pre-set image quantity;
The multiple second dedicated cpu has read the multiple first dedicated cpu according to preset data processing method
Image data carry out data processing, obtain the corresponding image pattern of each image data, wherein the preset data processing
Mode is preset according to the needs of machine learning model training;
The GPU obtains described image sample, is joined using acquired image pattern to the machine learning model
Number training.
As it can be seen that when computer program is executed by processor, being obtained in real time each in scheme provided by the embodiment of the present invention
The image data and storage, multiple first dedicated cpus of camera acquisition recycle execution reading process respectively, and multiple second is dedicated
CPU carries out data processing according to preset data processing method, to the image data that multiple first dedicated cpus have been read, and obtains
The corresponding image pattern of each image data, GPU obtains image pattern, using acquired image pattern to machine learning model
Carry out parameter training.Processing equipment uses Stream Processing mode, can handle the image data of each camera acquisition in real time, obtain
It takes in the image pattern of machine learning model training, does not need re -training machine learning mould when obtaining new image data
Type greatly shortens the model training time.
Wherein, the quantity of above-mentioned GPU can be multiple;
The GPU obtains described image sample, is joined using acquired image pattern to the machine learning model
The step of number training, comprising:
Each GPU obtains the image pattern of default sample size, using acquired image pattern to the machine learning
Model is trained, and the training result based on remaining GPU, adjusts the parameter of the machine learning model.
Wherein, above-mentioned multiple second dedicated cpus can be divided into N group, wherein first group of CPU includes the first preset quantity
Second dedicated cpu, N group CPU include the second dedicated cpu of N preset quantity;
The default processing mode includes: first group of CPU according to the first pre-set image processing mode, to the multiple first
The image data that dedicated cpu has been read is handled, and corresponding first processing result of each image data is obtained, and N group CPU is obtained
(N-1) processing result is taken, and acquired processing result is handled using N pre-set image processing mode, is obtained every
The corresponding image pattern of a image data.
Wherein, above-mentioned first group of CPU has read the multiple first dedicated cpu according to the first pre-set image processing mode
The image data taken is handled, and corresponding first processing result of each image data is obtained, and N group CPU is obtained at (N-1)
Reason as a result, and acquired processing result is handled using N pre-set image processing mode, obtain each image data pair
The step of image pattern answered, may include:
First group of CPU is according to the first pre-set image processing mode, the picture number read to the multiple first dedicated cpu
According to being handled, corresponding first processing result of each image data is obtained, first processing result is put into the first result
Queue, N group CPU obtains (N-1) processing result from (N-1) result queue, and uses N pre-set image processing mode
Acquired processing result is handled, the corresponding image pattern of each image data is obtained, described image sample is put into
Final result queue;
Above-mentioned GPU obtains described image sample, is joined using acquired image pattern to the machine learning model
Trained step is counted, may include:
The GPU obtains described image sample from the final result queue;
Parameter training is carried out to the machine learning model using acquired image pattern.
Wherein, above-mentioned default processing mode may include:
The multiple second dedicated cpu is preset processing mode using pre-set image and has been read the multiple first dedicated cpu
The image data taken carries out data processing, obtains the corresponding image numerical example of each image data, wherein the pre-set image is pre-
If processing mode includes: the first pre-set image processing mode, the second pre-set image processing mode to N pre-set image processing side
Formula.
Wherein, above-mentioned the step of obtaining the image data and storage that each camera acquires in real time, may include:
The image data of each camera acquisition is obtained in real time;
Described image data are stored, and adjust the storage order of all image datas.
It should be noted that for above system, processing equipment and computer readable storage medium embodiment, due to
It is substantially similar to embodiment of the method, so being described relatively simple, related place is referring to the part explanation of embodiment of the method
It can.
Need further exist for explanation, herein, relational terms such as first and second and the like be used merely to by
One entity or operation are distinguished with another entity or operation, without necessarily requiring or implying these entities or operation
Between there are any actual relationship or orders.Moreover, the terms "include", "comprise" or its any other variant meaning
Covering non-exclusive inclusion, so that the process, method, article or equipment for including a series of elements not only includes that
A little elements, but also including other elements that are not explicitly listed, or further include for this process, method, article or
The intrinsic element of equipment.In the absence of more restrictions, the element limited by sentence "including a ...", is not arranged
Except there is also other identical elements in the process, method, article or apparatus that includes the element.
Each embodiment in this specification is all made of relevant mode and describes, same and similar portion between each embodiment
Dividing may refer to each other, and each embodiment focuses on the differences from other embodiments.Especially for system reality
For applying example, since it is substantially similar to the method embodiment, so being described relatively simple, related place is referring to embodiment of the method
Part explanation.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the scope of the present invention.It is all
Any modification, equivalent replacement, improvement and so within the spirit and principles in the present invention, are all contained in protection scope of the present invention
It is interior.
Claims (10)
1. a kind of image pattern acquisition methods, which is characterized in that applied to the processing equipment communicated to connect with multiple mechanical arms, institute
State processing equipment include multiple first dedicated cpus for reading image data and for image real time transfer multiple second specially
With CPU, camera is installed on each mechanical arm, which comprises
The image data of each camera acquisition and storage are obtained in real time;
The multiple first dedicated cpu recycles respectively executes reading process, wherein the reading process are as follows: from the figure stored
Image data as reading the first pre-set image quantity in data;
The multiple second dedicated cpu is according to preset data processing method, the figure read to the multiple first dedicated cpu
As data progress data processing, the corresponding image pattern of each image data is obtained, wherein the preset data processing method
It is preset according to the needs of machine learning model training.
2. the method as described in claim 1, which is characterized in that the multiple second dedicated cpu is divided into N group, wherein first group
CPU includes the second dedicated cpu of the first preset quantity, and N group CPU includes the second dedicated cpu of N preset quantity;
The default processing mode includes: first group of CPU according to the first pre-set image processing mode, dedicated to the multiple first
The image data that CPU has been read is handled, and obtains corresponding first processing result of each image data, and N group CPU obtains the
(N-1) processing result, and acquired processing result is handled using N pre-set image processing mode, obtain each figure
As the corresponding image pattern of data.
3. method according to claim 2, which is characterized in that first group of CPU according to the first pre-set image processing mode,
The image data read to the multiple first dedicated cpu is handled, and obtains corresponding first processing of each image data
As a result, N group CPU obtains (N-1) processing result, and using N pre-set image processing mode to acquired processing result
The step of being handled, obtaining each image data corresponding image pattern, comprising:
The image data that first group of CPU has been read according to the first pre-set image processing mode, to the multiple first dedicated cpu into
Row processing, obtains corresponding first processing result of each image data, first processing result is put into the first result queue,
N group CPU obtains (N-1) processing result from (N-1) result queue, and using N pre-set image processing mode to institute
The processing result of acquisition is handled, and obtains the corresponding image pattern of each image data, described image sample is put into finally
Result queue.
4. the method as described in claim 1, which is characterized in that the default processing mode includes:
The multiple second dedicated cpu presets what processing mode had read the multiple first dedicated cpu using pre-set image
Image data carries out data processing, obtains the corresponding image numerical example of each image data, wherein the default place of the pre-set image
Reason mode includes: the first pre-set image processing mode, the second pre-set image processing mode to N pre-set image processing mode.
5. method according to any of claims 1-4, which is characterized in that the figure for obtaining each camera acquisition in real time
The step of as data and storing, comprising:
The image data of each camera acquisition is obtained in real time;
Described image data are stored, and adjust the storage order of all image datas.
6. a kind of model training method, which is characterized in that applied to the processing equipment communicated to connect with multiple mechanical arms, the place
Managing equipment includes for reading multiple first dedicated cpus of image data, for multiple second dedicated cpus of image real time transfer
And GPU, camera is installed on each mechanical arm, which comprises
The image data of each camera acquisition and storage are obtained in real time;
The multiple first dedicated cpu recycles respectively executes reading process, wherein the reading process are as follows: from the figure stored
Image data as reading the first pre-set image quantity in data;
The multiple second dedicated cpu is according to preset data processing method, the figure read to the multiple first dedicated cpu
As data progress data processing, the corresponding image pattern of each image data is obtained, wherein the preset data processing method
It is preset according to the needs of machine learning model training;
The GPU obtains described image sample, carries out parameter instruction to the machine learning model using acquired image pattern
Practice.
7. a kind of image pattern obtains system, which is characterized in that the system comprises the storages communicated to connect with multiple mechanical arms
Equipment and processing equipment, the processing equipment include multiple first dedicated cpus for reading image data and for picture number
According to multiple second dedicated cpus of processing, camera is installed on each mechanical arm, in which:
The storage equipment, for obtaining image data and the storage of each camera acquisition in real time;
The multiple first dedicated cpu is respectively used to circulation and executes reading process, wherein the reading process are as follows: from being stored
Image data in read the image data of the first pre-set image quantity;
The multiple second dedicated cpu, for having read the multiple first dedicated cpu according to preset data processing method
The image data taken carries out data processing, obtains the corresponding image pattern of each image data, wherein at the preset data
Reason mode is preset according to the needs of machine learning model training.
8. a kind of model training systems, which is characterized in that the system comprises the storage equipment communicated to connect with multiple mechanical arms
And processing equipment, the processing equipment include for reading multiple first dedicated cpus of image data, for image real time transfer
Multiple second dedicated cpus and GPU, camera is installed on each mechanical arm, in which:
The storage equipment, for obtaining image data and the storage of each camera acquisition in real time;
The multiple first dedicated cpu is respectively used to circulation and executes reading process, wherein the reading process are as follows: from being stored
Image data in read the image data of the first pre-set image quantity;
The multiple second dedicated cpu, for having read the multiple first dedicated cpu according to preset data processing method
The image data taken carries out data processing, obtains the corresponding image pattern of each image data, wherein at the preset data
Reason mode is preset according to the needs of machine learning model training;
The GPU joins the machine learning model using acquired image pattern for obtaining described image sample
Number training.
9. a kind of processing equipment, which is characterized in that specially including processor, communication interface, memory, communication bus, multiple first
With CPU and multiple second dedicated cpus, wherein processor, communication interface, memory, multiple first dedicated cpus and multiple second
Dedicated cpu completes mutual communication by communication bus, and the processing equipment and multiple mechanical arms communicate to connect, each machinery
Camera is installed on arm;
Memory, for storing computer program;
Processor when for executing the program stored on memory, obtains the image data of each camera acquisition simultaneously in real time
Storage;
Multiple first dedicated cpus are respectively used to circulation and execute reading process, wherein the reading process are as follows: from the figure stored
Image data as reading the first pre-set image quantity in data;
Multiple second dedicated cpus, for having been read to the multiple first dedicated cpu according to preset data processing method
Image data carries out data processing, obtains the corresponding image pattern of each image data, wherein the preset data processing side
Formula is preset according to the needs of machine learning model training.
10. a kind of processing equipment, which is characterized in that specially including processor, communication interface, memory, communication bus, multiple first
With CPU, multiple second dedicated cpus and GPU, wherein processor, communication interface, memory, multiple first dedicated cpus and multiple
Second dedicated cpu completes mutual communication by communication bus, and the processing equipment and multiple mechanical arms communicate to connect, each
Camera is installed on mechanical arm;
Memory, for storing computer program;
Processor when for executing the program stored on memory, obtains the image data of each camera acquisition and storage;
Multiple first dedicated cpus are respectively used to circulation and execute reading process, wherein the reading process are as follows: from the figure stored
Image data as reading the first pre-set image quantity in data;
Multiple second dedicated cpus, for having been read to the multiple first dedicated cpu according to preset data processing method
Image data carries out data processing, obtains the corresponding image pattern of each image data, wherein the preset data processing side
Formula is preset according to the needs of machine learning model training;
GPU carries out parameter instruction to the machine learning model using acquired image pattern for obtaining described image sample
Practice.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710896985.4A CN109598348B (en) | 2017-09-28 | 2017-09-28 | Image sample acquisition and model training method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710896985.4A CN109598348B (en) | 2017-09-28 | 2017-09-28 | Image sample acquisition and model training method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109598348A true CN109598348A (en) | 2019-04-09 |
CN109598348B CN109598348B (en) | 2023-06-06 |
Family
ID=65955444
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710896985.4A Active CN109598348B (en) | 2017-09-28 | 2017-09-28 | Image sample acquisition and model training method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109598348B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111787185A (en) * | 2020-08-04 | 2020-10-16 | 成都云图睿视科技有限公司 | Method for real-time processing of multi-path camera data under VPU platform |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103067633A (en) * | 2011-10-20 | 2013-04-24 | 东友科技股份有限公司 | Working method capable of accelerating working speed and image processing equipment suitable for the same |
WO2015192812A1 (en) * | 2014-06-20 | 2015-12-23 | Tencent Technology (Shenzhen) Company Limited | Data parallel processing method and apparatus based on multiple graphic procesing units |
CN105205169A (en) * | 2015-10-12 | 2015-12-30 | 中国电子科技集团公司第二十八研究所 | Distributed image index and retrieval method |
CN105677752A (en) * | 2015-12-30 | 2016-06-15 | 深圳先进技术研究院 | Streaming computing and batch computing combined processing system and method |
CN106534784A (en) * | 2016-11-22 | 2017-03-22 | 苏州航天系统工程有限公司 | Acquisition analysis storage statistical system for video analysis data result set |
CN106603690A (en) * | 2016-12-27 | 2017-04-26 | 东华互联宜家数据服务有限公司 | Data analysis device, data analysis processing system and data analysis method |
CN106873945A (en) * | 2016-12-29 | 2017-06-20 | 中山大学 | Data processing architecture and data processing method based on batch processing and Stream Processing |
CN107067365A (en) * | 2017-04-25 | 2017-08-18 | 中国石油大学(华东) | The embedded real-time video stream processing system of distribution and method based on deep learning |
-
2017
- 2017-09-28 CN CN201710896985.4A patent/CN109598348B/en active Active
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103067633A (en) * | 2011-10-20 | 2013-04-24 | 东友科技股份有限公司 | Working method capable of accelerating working speed and image processing equipment suitable for the same |
WO2015192812A1 (en) * | 2014-06-20 | 2015-12-23 | Tencent Technology (Shenzhen) Company Limited | Data parallel processing method and apparatus based on multiple graphic procesing units |
US20160321777A1 (en) * | 2014-06-20 | 2016-11-03 | Tencent Technology (Shenzhen) Company Limited | Data parallel processing method and apparatus based on multiple graphic processing units |
CN105205169A (en) * | 2015-10-12 | 2015-12-30 | 中国电子科技集团公司第二十八研究所 | Distributed image index and retrieval method |
CN105677752A (en) * | 2015-12-30 | 2016-06-15 | 深圳先进技术研究院 | Streaming computing and batch computing combined processing system and method |
CN106534784A (en) * | 2016-11-22 | 2017-03-22 | 苏州航天系统工程有限公司 | Acquisition analysis storage statistical system for video analysis data result set |
CN106603690A (en) * | 2016-12-27 | 2017-04-26 | 东华互联宜家数据服务有限公司 | Data analysis device, data analysis processing system and data analysis method |
CN106873945A (en) * | 2016-12-29 | 2017-06-20 | 中山大学 | Data processing architecture and data processing method based on batch processing and Stream Processing |
CN107067365A (en) * | 2017-04-25 | 2017-08-18 | 中国石油大学(华东) | The embedded real-time video stream processing system of distribution and method based on deep learning |
Non-Patent Citations (4)
Title |
---|
欧阳建权 等: "基于Storm的在线序列极限学习机的气象预测模型", 《计算机研究与发展》 * |
欧阳建权 等: "基于Storm的在线序列极限学习机的气象预测模型", 《计算机研究与发展》, vol. 54, no. 8, 31 August 2017 (2017-08-31), pages 1736 - 1743 * |
章清锋: "基于分布式流计算框架的动态光场采集与绘制系统", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
章清锋: "基于分布式流计算框架的动态光场采集与绘制系统", 《中国优秀硕士学位论文全文数据库 信息科技辑》, 15 October 2015 (2015-10-15), pages 19 - 46 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111787185A (en) * | 2020-08-04 | 2020-10-16 | 成都云图睿视科技有限公司 | Method for real-time processing of multi-path camera data under VPU platform |
CN111787185B (en) * | 2020-08-04 | 2023-09-05 | 成都云图睿视科技有限公司 | Method for processing multi-path camera data in real time under VPU platform |
Also Published As
Publication number | Publication date |
---|---|
CN109598348B (en) | 2023-06-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Martin et al. | Augmented reality in education 4.0 | |
CN110263921A (en) | A kind of training method and device of federation's learning model | |
US8347044B2 (en) | Multi-processor based programmable logic controller and method for operating the same | |
JP6833916B2 (en) | Data processing equipment, artificial intelligence chips and electronic devices | |
CN104881666B (en) | A kind of real-time bianry image connected component labeling implementation method based on FPGA | |
CN111008561B (en) | Method, terminal and computer storage medium for determining quantity of livestock | |
CN109740657A (en) | A kind of training method and equipment of the neural network model for image data classification | |
CN103098077A (en) | Real-time video frame pre-processing hardware | |
JP2022520390A (en) | Image recognition method, recognition model training method and its equipment, and computer program | |
CN109407603A (en) | A kind of method and device of control mechanical arm crawl object | |
CN112541584B (en) | Deep neural network model parallel mode selection method | |
CN110969600A (en) | Product defect detection method and device, electronic equipment and storage medium | |
CN106373112A (en) | Image processing method, image processing device and electronic equipment | |
CN110293549A (en) | Mechanical arm control method, device and neural network model training method, device | |
CN113724128A (en) | Method for expanding training sample | |
CN112748941A (en) | Feedback information-based target application program updating method and device | |
CN109670578A (en) | Neural network first floor convolution layer data processing method, device and computer equipment | |
CN109741244A (en) | Picture Generation Method and device, storage medium and electronic equipment | |
CN109598348A (en) | A kind of image pattern obtains, model training method and system | |
CN109741296A (en) | Product quality detection method and device | |
CN106375658B (en) | A kind of very high-precision image processing VLSI verification method | |
CN107483868A (en) | Processing method, FPGA and the laser television of VBO signals | |
EP3306462B1 (en) | Display device, and display signal input system and display signal input method thereof | |
CN115690592B (en) | Image processing method and model training method | |
CN105702233B (en) | A kind of image data takes window method and device |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |