CN110321961A - A kind of data processing method and device - Google Patents
A kind of data processing method and device Download PDFInfo
- Publication number
- CN110321961A CN110321961A CN201910615993.6A CN201910615993A CN110321961A CN 110321961 A CN110321961 A CN 110321961A CN 201910615993 A CN201910615993 A CN 201910615993A CN 110321961 A CN110321961 A CN 110321961A
- Authority
- CN
- China
- Prior art keywords
- matrix
- encoder
- encoded
- data processing
- obtains
- 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.)
- Pending
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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
The application provides a kind of data processing method and device, is used for encoder, and the encoder includes at least two coding layers；For each coding layer, which comprises receive matrix to be encoded；The Input matrix bull attention layer to be encoded is subjected to the calculating of bull attention, obtains submatrix；Submatrix input convolutional network layer is subjected to convolutional calculation, obtains encoder matrix；Export the encoder matrix.
Description
Technical field
This application involves field of computer technology, in particular to a kind of data processing method and device calculate equipment, storage
Medium and chip.
Background technique
In practical applications, when many times needing to carry out identification description to picture, such as needing to classify to picture,
Just need to identify the content in picture, e.g. scenery or animal or personage etc..
When picture is less, identification description manually can be carried out to picture.But with the network technology
Development, picture number sharply increase, and when needing to carry out mass picture identification description, manual processing mode becomes not cut excessively
It is practical.
So, how rapidly and accurately to carry out identification description to picture just becomes particularly important.
Summary of the invention
In view of this, the embodiment of the present application provide a kind of data processing method and device, calculate equipment, storage medium and
Chip, to solve technological deficiency existing in the prior art.
According to the embodiment of the present application in a first aspect, providing a kind of data processing method, comprising:
For encoder, the encoder includes at least two coding layers；
For each coding layer, which comprises
Receive matrix to be encoded；
The Input matrix bull attention layer to be encoded is subjected to the calculating of bull attention, obtains submatrix；
Submatrix input convolutional network layer is subjected to convolutional calculation, obtains encoder matrix；
Export the encoder matrix.
According to the second aspect of the embodiment of the present application, a kind of data processing equipment is provided, comprising:
For encoder, the encoder includes at least one coding layer；
For each coding layer, described device includes:
First receiving module is configured as receiving matrix to be encoded；
First computing module is configured as the Input matrix bull attention layer to be encoded carrying out bull attention meter
It calculates, obtains submatrix；
Second computing module is configured as inputting the submatrix into convolutional network layer progress convolutional calculation, be encoded
Matrix；
Output module is configured as exporting the encoder matrix.
According to the third aspect of the embodiment of the present application, a kind of electronic equipment is provided, comprising:
Memory, processor and storage on a memory and the computer instruction that can run on a processor, the processing
The step of device realizes one the method for any of the above when executing described instruction.
According to the fourth aspect of the embodiment of the present application, a kind of computer readable storage medium is provided, is stored with calculating
Machine instruction, when which is executed by processor the step of realization one the method for any of the above.
According to the 5th of the embodiment of the present application the aspect, a kind of chip is provided, computer instruction is stored with, the instruction quilt
The step of one the method for any of the above is realized when chip executes.
Data processing method provided by the present application and device receive matrix to be encoded；The Input matrix to be encoded is more
Head attention layer carries out the calculating of bull attention, obtains submatrix；Submatrix input convolutional network layer is subjected to convolution meter
It calculates, obtains encoder matrix；The encoder matrix is exported, is calculated by the bull attention in each coding layer, picture is paid close attention to
Global characteristics information, by the convolutional network layer in each coding layer to submatrix carry out multiple convolution operation, Ke Yigeng
The local feature information of good extraction picture, Transformer model may be used also while the carry out word processing of precise and high efficiency
Effectively to merge the global characteristics information and part characteristic information of picture, makes to identify that the accuracy of picture is obviously improved, make
Transformer model is obtained to describe quickly to generate more accurate picture description in task in picture recognition.
Detailed description of the invention
Fig. 1 is the structural schematic diagram of the calculating equipment of one embodiment of the application；
Fig. 2 is the flow diagram of the data processing method of one embodiment of the application；
Fig. 3 a is the flow diagram of the data processing method of one embodiment of the application；
Fig. 3 b is the schematic diagram for doing convolution algorithm in one embodiment of the application to submatrix；
Fig. 4 a~Fig. 4 b is the architecture diagram of the translation model of one embodiment of the application；
Fig. 5 is the block schematic illustration of the data processing equipment of one embodiment of the application.
Specific embodiment
Many details are explained in the following description in order to fully understand the application.But the application can be with
Much it is different from other way described herein to implement, those skilled in the art can be without prejudice to the application intension the case where
Under do similar popularization, therefore the application is not limited by following public specific implementation.
The term used in the application one or more embodiment be only merely for for the purpose of describing particular embodiments, and
It is not intended to be limiting the application one or more embodiment.The institute in the application one or more embodiment and the appended claims
The "an" of the singular used, " described " and "the" are also intended to including most forms, unless context clearly shows that it
His meaning.It is also understood that term "and/or" used in the application one or more embodiment refers to and includes one or more
A associated any or all of project listed may combine.
It will be appreciated that though may be described using term first, second etc. in the application one or more embodiment
Various information, but these information should not necessarily be limited by these terms.These terms are only used to for same type of information being distinguished from each other out.
For example, first can also be referred to as second in the case where not departing from the application one or more scope of embodiments, similarly,
Second can also be referred to as first.Depending on context, word as used in this " if " can be construed to " ...
When " or " when ... " or " in response to determination ".
Firstly, the vocabulary of terms being related to one or more embodiments of the invention explains.
Transformer: Google proposes a kind of translation model, is remembered with from the structure of attention model instead of shot and long term
Model achieves better achievement in translation duties.
From attention: attention mechanism is frequently used in the network structure using coder-decoder, and essence comes from
In human visual attention's mechanism.People's vision generally will not be that a scene is all seen when perceiving thing, and past
Past is that observation according to demand pays attention to specific a part.In terms of the level high from one, attention mechanism allows decoder from multiple
The part of needs is chosen in context vector, and then can indicate more information.
In this application, a kind of data processing method and device are provided, calculates equipment, storage medium and chip, under
It is described in detail one by one in the embodiment in face.
Fig. 1 shows the structural block diagram of the calculating equipment 100 according to one embodiment of the application.The portion of the calculating equipment 100
Part includes but is not limited to memory 110 and processor 120.Processor 120 is connected with memory 110 by bus 130, data
Library 150 is for saving data.
Calculating equipment 100 further includes access device 140, access device 140 enable calculate equipment 100 via one or
Multiple networks 160 communicate.The example of these networks includes public switched telephone network (PSTN), local area network (LAN), wide area network
(WAN), the combination of the communication network of personal area network (PAN) or such as internet.Access device 140 may include wired or wireless
One or more of any kind of network interface (for example, network interface card (NIC)), such as IEEE802.11 wireless local area
Net (WLAN) wireless interface, worldwide interoperability for microwave accesses (Wi-MAX) interface, Ethernet interface, universal serial bus (USB) connect
Mouth, cellular network interface, blue tooth interface, near-field communication (NFC) interface, etc..
In one embodiment of the application, unshowned other component in the above-mentioned component and Fig. 1 of equipment 100 is calculated
It can also be connected to each other, such as pass through bus.It should be appreciated that calculating device structure block diagram shown in FIG. 1 is merely for the sake of showing
The purpose of example, rather than the limitation to the application range.Those skilled in the art can according to need, and increase or replace other portions
Part.
Calculating equipment 100 can be any kind of static or mobile computing device, including mobile computer or mobile meter
Calculate equipment (for example, tablet computer, personal digital assistant, laptop computer, notebook computer, net book etc.), movement
Phone (for example, smart phone), wearable calculating equipment (for example, smartwatch, intelligent glasses etc.) or other kinds of shifting
Dynamic equipment, or the static calculating equipment of such as desktop computer or PC.Calculating equipment 100 can also be mobile or state type
Server.
Wherein, processor 120 can execute the step in method shown in Fig. 2.Fig. 2 shows according to one embodiment of the application
Data processing method schematic flow chart.The data processing method of the present embodiment is used for encoder, and decoder includes at least
Two coding layers.For each coding layer, the method includes the following steps 202 to step 208:
Step 202: receiving matrix to be encoded.
Wherein, for different decoding layers, receive matrix to be encoded and be different, for first coding layer, receive to
Encoder matrix is to receive initial matrix to be encoded；For removing other coding layers of first coding layer, matrix to be encoded is received
For the encoder matrix for receiving upper coding layer output.
Before receiving initial matrix to be encoded, further includes:
Receive picture to be identified；
By the picture to be identified by Processing with Neural Network trained in advance, picture feature matrix is obtained；
Picture feature matrix progress is position encoded, obtain initial matrix to be encoded.
By taking one " diver seabed observe green turtle " picture as an example, picture to be identified is received, and by the picture
It is input to convolutional neural networks model trained in advance, gets the eigenmatrix of picture, is matched for each picture feature matrix
The coding for setting a corresponding position obtains initial matrix to be encoded.
For first coding layer, initial matrix to be encoded is received.
For removing other coding layers of first coding layer, the encoder matrix of upper coding layer output is received.
Step 204: the Input matrix bull attention layer to be encoded being subjected to the calculating of bull attention, obtains sub- square
Battle array.
The Input matrix to be encoded received to bull is noticed that layer repeatedly calculate from attention, obtains submatrix.
The ability for facilitating enhancing concern holistic correlation is calculated from attention.
Step 206: submatrix input convolutional network layer being subjected to convolutional calculation, obtains encoder matrix.
By carrying out convolution algorithm to the submatrix in convolutional network layer, the feature of submatrix is extracted, to realize volume
Product network layer carries out local shape factor to the submatrix of input.
The convolutional calculation process of the submatrix is referring to following formula (1)；
CFN (x)=Conv (Relu (Conv (x))) (1)
Wherein, x represents the submatrix, and Conv represents convolutional calculation, and Relu represents activation primitive.
CFN represents the encoder matrix that the submatrix obtains after convolutional calculation.
Convolutional network layer receives submatrix, carries out a convolution algorithm to submatrix and obtains convolution matrix, to the convolution
Matrix does activation with activation primitive and activated matrix is calculated, and carries out second of convolution algorithm to the activated matrix and is encoded
Matrix, submatrix, by convolution algorithm twice and an activation primitive activation, can preferably extract son in convolutional network layer
The local feature of matrix.
Submatrix input convolutional network layer is subjected to convolutional calculation, obtains encoder matrix.Referring to Fig. 3 a, step 206
Include the following steps 302 to step 306:
Step 302: submatrix input convolutional network layer being subjected to convolution algorithm, obtains convolution matrix.
The submatrix obtained in step 204 is input to convolutional network layer, carries out a convolution algorithm, obtains convolution square
Battle array.
In embodiments herein, convolution algorithm is carried out to the submatrix of input in convolutional network layer and obtains the son
The depth of the characteristic information of matrix, convolution kernel is identical as the input port number of submatrix.
Fig. 3 b is the signal for doing convolution algorithm when extracting one characteristic information of submatrix in the embodiment of the present application to submatrix
Figure, wherein the size of submatrix A is 6 × 6 × 3, and the size of convolution kernel B is 3 × 3 × 3, because submatrix A has 3 channels, institute
With the depth of convolution kernel B be 3, convolution algorithm is carried out to sub- matrix A with convolution kernel B, by convolution kernel B in submatrix A according to from
Left-to-right, sequence sliding from top to bottom, sliding step 1, in sliding process, convolution kernel B is overlapped corresponding with submatrix A
Element does dot-product operation, and one 4 × 4 × 1 convolution eigenmatrix C is obtained after the completion of sliding.
Convolution algorithm is carried out to the submatrix using different convolution kernels, different convolution eigenmatrixes can be obtained,
The convolution eigenmatrix of acquisition constitutes convolution matrix.
Step 304: the convolution matrix being calculated with activation primitive, obtains activated matrix.
Wherein, activation primitive can be a variety of, such as sigmoid function, relu function, softmax function etc., if not
With activation primitive, each layer of output is all the linear function of upper layer input, no matter neural network how many layer, output is all input
Linear combination.If you are using, activation primitive introduces non-linear factor to neuron, allows neural network any
Any nonlinear function is approached, such neural network can be applied in numerous nonlinear models.
Step 306: convolution algorithm being carried out to the activated matrix, obtains encoder matrix.
The activated matrix obtained in step 304 is subjected to convolution algorithm again, obtains encoder matrix.
Step 208: exporting the encoder matrix.
The encoder matrix that the last one coding layer is exported is as the final encoder matrix of encoder；Or according to all codings
The encoder matrix of layer output is calculated, and the final encoder matrix of encoder is obtained.
For the encoder including multiple coding layers, the final encoder matrix of encoder can be according to the volume of all coding layers
Code matrix carries out fusion treatment and generates, and the mode of fusion can be and be equipped with weight for the encoder matrix of each coding layer, then
Summation generates final encoder matrix.
After exporting the encoder matrix, further includes:
Decoder receives the encoder matrix；
The decoder is decoded the encoder matrix, obtains the description information of the encoder matrix.
Specifically, decoder receives the encoder matrix of encoder output, and encoder matrix is decoded, is encoded
The description information of matrix is " diver observes green turtle in seabed ", to obtain the description information of picture to be identified.
Data processing method provided by the present application receives matrix to be encoded；The Input matrix bull to be encoded is paid attention to
Power layer carries out the calculating of bull attention, obtains submatrix；Submatrix input convolutional network layer is subjected to convolutional calculation, is obtained
Encoder matrix；The encoder matrix is exported, bull attention calculates the global characteristics information that can more pay close attention to picture, rolling up
Multiple convolution operation is carried out to submatrix in product network layer, the local feature information of picture can be preferably extracted, in each layer
It is calculated and convolutional calculation using bull attention simultaneously in coding layer, can preferably take into account the global characteristics information drawn game of picture
Portion's characteristic information makes to identify that the accuracy of picture is obviously improved, and generates more accurate picture description.
It is provided in order to make it easy to understand, Fig. 4 a~Fig. 4 b is shown based on Transformer model application the embodiment of the present application
Data processing method translation model architecture diagram.It, will when carrying out identification description to picture in embodiments herein
Picture to be identified obtains corresponding picture feature matrix, by the picture feature square by Processing with Neural Network trained in advance
Battle array is input in the encoder of Transformer model.
In the Transformer model of Fig. 4 a, encoder includes six coding layers, for each coding layer, b referring to fig. 4,
Including bull attention layer and convolutional network layer.It is calculated separately using bull attention and convolution, obtains encoder matrix.
For the first coding layer: receiving initial matrix to be encoded；Bull note is carried out according to the initial matrix to be encoded of input
Power of anticipating calculates, and obtains submatrix；Submatrix input convolutional network layer is subjected to convolutional calculation, obtains encoder matrix；Output
The encoder matrix.
For the second coding layer: receiving the encoder matrix of the first coding layer, carried out according to the encoder matrix of the first coding layer
Submatrix is calculated in bull attention；Submatrix input convolutional network layer is subjected to convolutional calculation, obtains encoder matrix；
Export the encoder matrix.
For third coding layer: receiving the encoder matrix of the second coding layer, carried out according to the encoder matrix of the second coding layer
Submatrix is calculated in bull attention；Submatrix input convolutional network layer is subjected to convolutional calculation, obtains encoder matrix；
Export the encoder matrix.
For the 4th coding layer: receiving the encoder matrix of third coding layer, carried out according to the encoder matrix of third coding layer
Submatrix is calculated in bull attention；Submatrix input convolutional network layer is subjected to convolutional calculation, obtains encoder matrix；
Export the encoder matrix.
For the 5th coding layer: receiving the encoder matrix of the 4th coding layer, carried out according to the encoder matrix of the 4th coding layer
Submatrix is calculated in bull attention；Submatrix input convolutional network layer is subjected to convolutional calculation, obtains encoder matrix；
Export the encoder matrix.
For the 6th coding layer: receiving the encoder matrix of the 5th coding layer, carried out according to the encoder matrix of the 5th coding layer
Submatrix is calculated in bull attention；Submatrix input convolutional network layer is subjected to convolutional calculation, obtains encoder matrix；
The encoder matrix is exported, and is exported the encoder matrix of the 6th coding layer as the final encoder matrix of encoder.
Decoder receives the encoder matrix of encoder output, and the encoder matrix is decoded operation, obtains described
The description information of encoder matrix, to obtain the description information of picture to be identified.
Transformer model provided by the present application is by multiple coding layers in encoder to the initial matrix to be encoded of input
It is repeatedly encoded, is calculated by the bull attention in each coding layer, the global characteristics information of picture is paid close attention to, in convolution
Multiple convolution operation is carried out to submatrix in network layer, can preferably extract the local feature information of picture, makes to identify picture
Accuracy be obviously improved, the characteristics of Transformer model is when carrying out word processing is that have jumping characteristic and neighborhood information,
Convolution algorithm focuses more on neighborhood information, can be with while the fusion of the two allows to the carry out word processing of precise and high efficiency
The global characteristics information of effective fusion picture and local characteristic information, so that Transformer model is described in picture recognition
More accurate picture description is quickly generated in task.
One embodiment of the application also provides a kind of data processing equipment, referring to Fig. 5, comprising:
First receiving module 502 is configured as receiving matrix to be encoded.
For first coding layer, first receiving module 502 is configured as receiving initial matrix to be encoded；
For removing other coding layers of first coding layer, first receiving module 502 is configured as receiving upper one
The encoder matrix of a coding layer output.
First computing module 504 is configured as the Input matrix bull attention layer to be encoded carrying out bull attention
Power calculates, and obtains submatrix.
Second computing module 506 is configured as inputting the submatrix into convolutional network layer progress convolutional calculation, be compiled
Code matrix.
Second computing module 506 is configured to inputting the submatrix into convolutional network layer progress convolution
Operation obtains convolution matrix；The convolution matrix is calculated with activation primitive, obtains activated matrix；To the activation square
Battle array carries out convolution algorithm, obtains encoder matrix.
Output module 508 is configured as exporting the encoder matrix.
The output module 508 is configured to the volume of the last one coding layer output in the encoder
Final encoder matrix of the code matrix as the encoder；Or calculated according to the encoder matrix that all coding layers export, it obtains
To the final encoder matrix of the encoder.
Second receiving module 510 is configured as receiving picture to be identified.
Picture processing module 512 is configured as obtaining the picture to be identified by Processing with Neural Network trained in advance
To picture feature matrix.
Position encoded module 514, is configured as carrying out in the picture feature matrix position encoded, obtains initial to be encoded
Matrix.
Third receiving module 516 is configured as receiving the encoder matrix.
Decoder module 518 is configured as being decoded the encoder matrix, obtains the description letter of the encoder matrix
Breath.
Data processing equipment provided by the present application effectively merges the global information and local message of picture to be identified, pole
The earth improves the accuracy of identification picture.
A kind of exemplary scheme of above-mentioned data processing equipment for the present embodiment.It should be noted that the data processing
The technical solution of the technical solution of device and above-mentioned data processing method belongs to same design, the technical side of data processing equipment
The detail content that case is not described in detail may refer to the description of the technical solution of above-mentioned data processing method.
One embodiment of the application also provides a kind of computer readable storage medium, is stored with computer instruction, the instruction
The step of data processing method as previously described is realized when being executed by processor.
A kind of exemplary scheme of above-mentioned computer readable storage medium for the present embodiment.It should be noted that this is deposited
The technical solution of the technical solution of storage media and above-mentioned data processing method belongs to same design, the technical solution of storage medium
The detail content being not described in detail may refer to the description of the technical solution of above-mentioned data processing method.
One embodiment of the application also provides a kind of chip, is stored with computer instruction, real when which is executed by chip
Now the step of data processing method as previously described.
The computer instruction includes computer program code, the computer program code can for source code form,
Object identification code form, executable file or certain intermediate forms etc..The computer-readable medium may include: that can carry institute
State any entity or device, recording medium, USB flash disk, mobile hard disk, magnetic disk, CD, the computer storage of computer program code
Device, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory),
Electric carrier signal, telecommunication signal and software distribution medium etc..It should be noted that the computer-readable medium include it is interior
Increase and decrease appropriate can be carried out according to the requirement made laws in jurisdiction with patent practice by holding, such as in certain jurisdictions of courts
Area does not include electric carrier signal and telecommunication signal according to legislation and patent practice, computer-readable medium.
It should be noted that for the various method embodiments described above, describing for simplicity, therefore, it is stated as a series of
Combination of actions, but those skilled in the art should understand that, the application is not limited by the described action sequence because
According to the application, certain steps can use other sequences or carry out simultaneously.Secondly, those skilled in the art should also know
It knows, embodiment described herein belongs to preferred embodiment, and related actions and modules might not all be this Shen
It please be necessary.
In the above-described embodiments, it all emphasizes particularly on different fields to the description of each embodiment, there is no the portion being described in detail in some embodiment
Point, it may refer to the associated description of other embodiments.
The application preferred embodiment disclosed above is only intended to help to illustrate the application.There is no detailed for alternative embodiment
All details are described, are not limited the invention to the specific embodiments described.It obviously, can according to present context
It makes many modifications and variations.The application chooses and specifically describes these embodiments, is the original in order to preferably explain the application
Reason and practical application, so that skilled artisan be enable to better understand and utilize the application.The application is only authorized
The limitation of sharp claim and its full scope and equivalent.
Claims (17)
1. a kind of data processing method, which is characterized in that be used for encoder, the encoder includes at least two coding layers；
For each coding layer, which comprises
Receive matrix to be encoded；
The Input matrix bull attention layer to be encoded is subjected to the calculating of bull attention, obtains submatrix；
Submatrix input convolutional network layer is subjected to convolutional calculation, obtains encoder matrix；
Export the encoder matrix.
2. data processing method as described in claim 1, which is characterized in that for first coding in the encoder
Layer,
Receiving matrix to be encoded includes: to receive initial matrix to be encoded.
3. data processing method as described in claim 1, which is characterized in that for removing first coding in the encoder
Other coding layers of layer,
Receiving matrix to be encoded includes: the encoder matrix for receiving upper coding layer output.
4. data processing method as described in claim 1, which is characterized in that
Submatrix input convolutional network layer is subjected to convolutional calculation, obtaining encoder matrix includes:
Submatrix input convolutional network layer is subjected to convolution algorithm, obtains convolution matrix；
The convolution matrix is calculated with activation primitive, obtains activated matrix；
Convolution algorithm is carried out to the activated matrix, obtains encoder matrix.
5. data processing method as described in claim 1, which is characterized in that
Exporting the encoder matrix includes:
Using the encoder matrix of the last one coding layer output in the encoder as the final encoder matrix of the encoder；
Or
It is calculated according to the encoder matrix that all coding layers export, obtains the final encoder matrix of the encoder.
6. data processing method as claimed in claim 2, which is characterized in that before receiving initial matrix to be encoded, further includes:
Receive picture to be identified；
By the picture to be identified by Processing with Neural Network trained in advance, picture feature matrix is obtained；
Picture feature matrix progress is position encoded, obtain initial matrix to be encoded.
7. data processing method as described in claim 1, which is characterized in that after exporting the encoder matrix, further includes:
Decoder receives the encoder matrix；
The decoder is decoded the encoder matrix, obtains the description information of the encoder matrix.
8. a kind of data processing equipment, which is characterized in that be used for encoder, the encoder includes at least one coding layer；
For each coding layer, described device includes:
First receiving module is configured as receiving matrix to be encoded；
First computing module is configured as the Input matrix bull attention layer to be encoded carrying out the calculating of bull attention,
Obtain submatrix；
Second computing module is configured as inputting the submatrix into convolutional network layer progress convolutional calculation, obtains encoder matrix；
Output module is configured as exporting the encoder matrix.
9. data processing equipment as claimed in claim 8, which is characterized in that
First receiving module is configured as receiving initial matrix to be encoded.
10. data processing equipment as claimed in claim 8, which is characterized in that for removing first volume in the encoder
Other coding layers of code layer,
First receiving module is configured as receiving the encoder matrix of coding layer output.
11. data processing equipment as claimed in claim 8, which is characterized in that
Second computing module is configured to inputting the submatrix into convolutional network layer progress convolution algorithm, obtain
To convolution matrix；The convolution matrix is calculated with activation primitive, obtains activated matrix；The activated matrix is rolled up
Product operation, obtains encoder matrix.
12. data processing equipment as claimed in claim 8, which is characterized in that
The output module is configured to make the encoder matrix of the last one coding layer output in the encoder
For the final encoder matrix of the encoder；Or calculated according to the encoder matrix that all coding layers export, obtain the volume
The final encoder matrix of code device.
13. data processing equipment as claimed in claim 9, which is characterized in that further include:
Second receiving module is configured as receiving picture to be identified；
Picture processing module is configured as the picture to be identified obtaining picture by Processing with Neural Network trained in advance
Eigenmatrix；
Position encoded module, is configured as carrying out in the picture feature matrix position encoded, obtains initial matrix to be encoded.
14. data processing equipment as claimed in claim 8, which is characterized in that after exporting the encoder matrix, further includes:
Third receiving module is configured as receiving the encoder matrix；
Decoder module is configured as being decoded the encoder matrix, obtains the description information of the encoder matrix.
15. a kind of calculating equipment including memory, processor and stores the calculating that can be run on a memory and on a processor
Machine instruction, which is characterized in that the processor realizes the step of claim 1-7 any one the method when executing described instruction
Suddenly.
16. a kind of computer readable storage medium, is stored with computer instruction, which is characterized in that the instruction is held by processor
The step of claim 1-7 any one the method is realized when row.
17. a kind of chip, is stored with computer instruction, which is characterized in that the instruction realizes claim when being executed by chip
The step of 1-7 any one the method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910615993.6A CN110321961A (en) | 2019-07-09 | 2019-07-09 | A kind of data processing method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910615993.6A CN110321961A (en) | 2019-07-09 | 2019-07-09 | A kind of data processing method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110321961A true CN110321961A (en) | 2019-10-11 |
Family
ID=68121602
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910615993.6A Pending CN110321961A (en) | 2019-07-09 | 2019-07-09 | A kind of data processing method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110321961A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116542290A (en) * | 2023-06-25 | 2023-08-04 | 城云科技（中国）有限公司 | Information prediction model construction method, device and application based on multi-source multi-dimensional data |
Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106023220A (en) * | 2016-05-26 | 2016-10-12 | 史方 | Vehicle exterior part image segmentation method based on deep learning |
| CN106407931A (en) * | 2016-09-19 | 2017-02-15 | 杭州电子科技大学 | Novel deep convolution neural network moving vehicle detection method |
| CN108681539A (en) * | 2018-05-07 | 2018-10-19 | 内蒙古工业大学 | A kind of illiteracy Chinese nerve interpretation method based on convolutional neural networks |
| WO2018217948A1 (en) * | 2017-05-23 | 2018-11-29 | Google Llc | Attention-based sequence transduction neural networks |
| CN109190131A (en) * | 2018-09-18 | 2019-01-11 | 北京工业大学 | A kind of English word and its capital and small letter unified prediction based on neural machine translation |
| CN109359309A (en) * | 2018-12-11 | 2019-02-19 | 成都金山互动娱乐科技有限公司 | A kind of interpretation method and device, the training method of translation model and device |
| CN109492232A (en) * | 2018-10-22 | 2019-03-19 | 内蒙古工业大学 | A kind of illiteracy Chinese machine translation method of the enhancing semantic feature information based on Transformer |
| CN109543824A (en) * | 2018-11-30 | 2019-03-29 | 腾讯科技（深圳）有限公司 | A kind for the treatment of method and apparatus of series model |
| US20190130273A1 (en) * | 2017-10-27 | 2019-05-02 | Salesforce.Com, Inc. | Sequence-to-sequence prediction using a neural network model |
| CN109710953A (en) * | 2018-12-29 | 2019-05-03 | 成都金山互动娱乐科技有限公司 | A kind of interpretation method and device calculate equipment, storage medium and chip |
| CN109726794A (en) * | 2017-10-27 | 2019-05-07 | 谷歌有限责任公司 | Image based on concern generates neural network |
| CN109740123A (en) * | 2018-12-21 | 2019-05-10 | 北京信息科技大学 | The method for generating competitive sports war communique using real time data |
| CN109859760A (en) * | 2019-02-19 | 2019-06-07 | 成都富王科技有限公司 | Phone robot voice recognition result bearing calibration based on deep learning |
| CN109949796A (en) * | 2019-02-28 | 2019-06-28 | 天津大学 | A kind of end-to-end framework Lhasa dialect phonetic recognition methods based on Tibetan language component |
-
2019
- 2019-07-09 CN CN201910615993.6A patent/CN110321961A/en active Pending
Patent Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106023220A (en) * | 2016-05-26 | 2016-10-12 | 史方 | Vehicle exterior part image segmentation method based on deep learning |
| CN106407931A (en) * | 2016-09-19 | 2017-02-15 | 杭州电子科技大学 | Novel deep convolution neural network moving vehicle detection method |
| WO2018217948A1 (en) * | 2017-05-23 | 2018-11-29 | Google Llc | Attention-based sequence transduction neural networks |
| US20190130273A1 (en) * | 2017-10-27 | 2019-05-02 | Salesforce.Com, Inc. | Sequence-to-sequence prediction using a neural network model |
| CN109726794A (en) * | 2017-10-27 | 2019-05-07 | 谷歌有限责任公司 | Image based on concern generates neural network |
| CN108681539A (en) * | 2018-05-07 | 2018-10-19 | 内蒙古工业大学 | A kind of illiteracy Chinese nerve interpretation method based on convolutional neural networks |
| CN109190131A (en) * | 2018-09-18 | 2019-01-11 | 北京工业大学 | A kind of English word and its capital and small letter unified prediction based on neural machine translation |
| CN109492232A (en) * | 2018-10-22 | 2019-03-19 | 内蒙古工业大学 | A kind of illiteracy Chinese machine translation method of the enhancing semantic feature information based on Transformer |
| CN109543824A (en) * | 2018-11-30 | 2019-03-29 | 腾讯科技（深圳）有限公司 | A kind for the treatment of method and apparatus of series model |
| CN109359309A (en) * | 2018-12-11 | 2019-02-19 | 成都金山互动娱乐科技有限公司 | A kind of interpretation method and device, the training method of translation model and device |
| CN109740123A (en) * | 2018-12-21 | 2019-05-10 | 北京信息科技大学 | The method for generating competitive sports war communique using real time data |
| CN109710953A (en) * | 2018-12-29 | 2019-05-03 | 成都金山互动娱乐科技有限公司 | A kind of interpretation method and device calculate equipment, storage medium and chip |
| CN109859760A (en) * | 2019-02-19 | 2019-06-07 | 成都富王科技有限公司 | Phone robot voice recognition result bearing calibration based on deep learning |
| CN109949796A (en) * | 2019-02-28 | 2019-06-28 | 天津大学 | A kind of end-to-end framework Lhasa dialect phonetic recognition methods based on Tibetan language component |
Non-Patent Citations (4)
| Title |
|---|
| ALEXANDER M. RUSH: "The Annotated Transformer", 《PROCEEDINGS OF WORKSHOP FOR NLP OPEN SOURCE SOFTWARE》 * |
| ASHISH VASWANI等: "Attention Is All You Need", 《ARXIV:1706.03762V5 [CS.CL]》 * |
| PATRICK VERGA 等: "Simultaneously Self-Attending to All Mentions for Full-Abstract Biological Relation Extraction", 《ARXIV:1802.10569V1 [CS.CL]》 * |
| SEBASTIAN SPRINGENBERG 等: "Image-to-Text Transduction with Spatial Self-Attention", 《ESANN 2018 PROCEEDINGS, EUROPEAN SYMPOSIUM ON ARTIFICIAL NEURAL NETWORKS, COMPUTATIONAL INTELLIGENCE AND MACHINE LEARNING》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116542290A (en) * | 2023-06-25 | 2023-08-04 | 城云科技（中国）有限公司 | Information prediction model construction method, device and application based on multi-source multi-dimensional data |
| CN116542290B (en) * | 2023-06-25 | 2023-09-08 | 城云科技（中国）有限公司 | Information prediction model construction method, device and application based on multi-source multi-dimensional data |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107632987B (en) | A kind of dialogue generation method and device | |
| CN109377544B (en) | Human face three-dimensional image generation method and device and readable medium | |
| CN108629743A (en) | Processing method, device, storage medium and the electronic device of image | |
| CN108230278B (en) | Image raindrop removing method based on generation countermeasure network | |
| CN108171701B (en) | Significance detection method based on U network and counterstudy | |
| CN108564611A (en) | A kind of monocular image depth estimation method generating confrontation network based on condition | |
| CN109785928A (en) | Diagnosis and treatment proposal recommending method, device and storage medium | |
| CN110060097A (en) | User behavior sequence of recommendation method based on attention mechanism and convolutional neural networks | |
| CN110348535A (en) | A kind of vision Question-Answering Model training method and device | |
| CN109977428A (en) | A kind of method and device that answer obtains | |
| CN107194426A (en) | A kind of image-recognizing method based on Spiking neutral nets | |
| CN109902723A (en) | Image processing method and device | |
| CN109657582A (en) | Recognition methods, device, computer equipment and the storage medium of face mood | |
| CN109859120A (en) | Image defogging method based on multiple dimensioned residual error network | |
| CN110059728A (en) | RGB-D image vision conspicuousness detection method based on attention model | |
| CN110321962A (en) | A kind of data processing method and device | |
| CN115713715B (en) | Human behavior recognition method and recognition system based on deep learning | |
| CN110222728A (en) | The training method of article discrimination model, system and article discrimination method, equipment | |
| CN112883227B (en) | Video abstract generation method and device based on multi-scale time sequence characteristics | |
| CN110148088A (en) | Image processing method, image rain removing method, device, terminal and medium | |
| CN108197427A (en) | Proteins subcellular location method and apparatus based on depth convolutional neural networks | |
| CN110634103A (en) | Image demosaicing method based on generation of countermeasure network | |
| CN114782694A (en) | Unsupervised anomaly detection method, system, device and storage medium | |
| CN110121719A (en) | Device, method and computer program product for deep learning | |
| CN110321961A (en) | A kind of data processing 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 |