CN110516805A - The training duration prediction method and device of training pattern - Google Patents
The training duration prediction method and device of training pattern Download PDFInfo
- Publication number
- CN110516805A CN110516805A CN201910782983.1A CN201910782983A CN110516805A CN 110516805 A CN110516805 A CN 110516805A CN 201910782983 A CN201910782983 A CN 201910782983A CN 110516805 A CN110516805 A CN 110516805A
- Authority
- CN
- China
- Prior art keywords
- training
- duration
- prediction
- amendment
- training duration
- 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
- 238000012549 training Methods 0.000 title claims abstract description 464
- 238000000034 method Methods 0.000 title claims abstract description 80
- 238000001914 filtration Methods 0.000 claims abstract description 27
- 230000007787 long-term memory Effects 0.000 claims description 11
- 230000003252 repetitive effect Effects 0.000 description 19
- 230000008569 process Effects 0.000 description 11
- 238000012512 characterization method Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000000151 deposition Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000004321 preservation Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000001960 triggered effect Effects 0.000 description 1
Classifications
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
-
- 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
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Filters That Use Time-Delay Elements (AREA)
Abstract
The present invention provides a kind of training duration prediction method and device of training pattern, comprising: when training pattern is iterated it is trained when, determine training pattern current iteration training training duration；It obtains corresponding with current iteration training first and predicts training duration；By in the training duration input LSTM model that training is completed in advance, obtains the second of the output of LSTM model and predict training duration；It predicts that training duration and preset Kalman filtering algorithm predict that training duration is modified to second according to first, obtains revised amendment and predict training duration, amendment is predicted that training duration is saved into preset tables of data；Amendment is predicted that training duration is determined as the prediction training duration of the next iteration training of training pattern.Present invention application LSTM model obtains the prediction training duration that training pattern carries out next iteration training, and application Kalman filtering algorithm is modified prediction training duration, to obtain more accurately predicting training duration.
Description
Technical field
The present invention relates to field of computer technology, in particular to the training duration prediction method and dress of a kind of training pattern
It sets.
Background technique
With popularizing for artificial intelligence, more and more deep learning training patterns start to emerge.Deep learning instruction
Practice model to need to be trained before starting to apply, so that deep learning training pattern reaches preset accuracy.Depth
Training pattern is practised when being trained, needs to predict the training duration of training pattern, the existing training duration to training pattern
When being predicted, mostly by duration used in each iteration before being averaging, multiplied by number of iterations on the basis of an iteration
General duration used is found out, the obtained time is i.e. to the prediction duration of training pattern.
Through inventor the study found that the prediction technique of existing prediction training pattern is widely used in single machine single deck tape-recorder and does not have
The simple forecast of unusual condition, when applied in the multimachine mostly distributed training pattern of card, multimachine blocks in the presence of many sudden more
Problem, such as a certain node are restarted, and a certain node GPU falls the special statuss such as card, are had very to the prediction result of training duration
Big interference reduces the accuracy of the training duration to training pattern prediction.
Summary of the invention
In view of this, the present invention provides a kind of training duration prediction method of training pattern, can be predicted using the present invention
The prediction training duration that next iteration training is carried out in training pattern, can be improved the accuracy for the prediction duration that prediction obtains.
The present invention also provides a kind of training time premeauring device of training pattern, to guarantee above method realization in practice and
Using.
A kind of training duration prediction method of training pattern, comprising:
When training pattern is iterated it is trained when, determine the training pattern current iteration training training duration；
It obtains corresponding with current iteration training first and predicts training duration；
By in the trained duration input shot and long term memory network LSTM model that training is completed in advance, the LSTM is obtained
The second of model output predicts training duration；
Training duration is predicted according to described first, and the preset Kalman filtering algorithm of application trains second prediction
Duration is modified, and is obtained amendment corresponding with the second prediction training duration and is predicted training duration, the amendment is predicted
Training duration is saved into preset tables of data；
When the amendment is predicted that training duration is determined as the prediction training of the next iteration training of the training pattern
It is long.
Above-mentioned method, optionally, the acquisition corresponding with current iteration training first predict training duration, packet
It includes:
Preset tables of data in database is obtained, and determines the iteration mark of the current iteration training；
Prediction corresponding with iteration mark in tables of data training duration is determined as the first prediction training duration.
Above-mentioned method, it is optionally, described by the trained duration input shot and long term memory network that training is completed in advance
In LSTM model, obtain the LSTM model output second predicts training duration, comprising:
Obtain the characteristic parameter of the trained duration；
Using the duration prediction method in the LSTM model, the characteristic parameter is analyzed, is obtained and the spy
Levy the corresponding prediction duration of parameter；
The prediction duration is determined as the second prediction training duration.
Above-mentioned method, it is optionally, described to predict training duration according to described first, and apply preset Kalman filtering
Algorithm predicts that training duration is modified to described second, obtains amendment prediction corresponding with the second prediction training duration and instructs
Practice duration, comprising:
Obtain the duration category of amendment weight and the second prediction training duration that described first predicts training duration
Property；
According to the amendment weight, the duration for predicting training duration to described second using preset Kalman filtering algorithm
Attribute is modified, and is obtained amendment corresponding with the second prediction training duration and is predicted training duration.
Above-mentioned method, optionally, further includes:
Generate prompt information corresponding with amendment prediction training duration；
The prompt information is sent to preset display equipment and is shown.
A kind of training time premeauring device of training pattern, comprising:
First determination unit, for determining the current iteration of the training pattern when training pattern is iterated trained
Trained training duration；
Acquiring unit predicts training duration for obtaining corresponding with current iteration training first；
Output unit, for the shot and long term memory network LSTM model completed to be trained in the trained duration input in advance
In, obtain the LSTM model output second predicts training duration；
Amending unit, for predicting training duration according to described first, and the preset Kalman filtering algorithm of application is to institute
It states the second prediction training duration to be modified, obtains amendment corresponding with the second prediction training duration and predict training duration,
The amendment is predicted that training duration is saved into preset tables of data；
Second determination unit, for the amendment to be predicted that training duration is determined as the next iteration of the training pattern
Trained prediction training duration.
Above-mentioned device, optionally, the acquiring unit, comprising:
First obtains subelement, for obtaining preset tables of data in database, and determines the current iteration training
Iteration mark；
First determines subelement, for determining prediction corresponding with iteration mark in tables of data training duration
Training duration is predicted for first.
Above-mentioned device, optionally, the output unit, comprising:
Second obtains subelement, for obtaining the characteristic parameter of the trained duration；
Subelement is analyzed, for the duration prediction method in the application LSTM model, the characteristic parameter is divided
Analysis, obtains prediction duration corresponding with the characteristic parameter；
First determines subelement, for the prediction duration to be determined as the second prediction training duration.
Above-mentioned device, optionally, the amending unit, comprising:
Third obtains subelement, for obtaining the amendment weight and described second of the first prediction training duration in advance
Survey the duration attribute of training duration；
Revise subelemen is used for according to the amendment weight, pre- to described second using preset Kalman filtering algorithm
The duration attribute for surveying training duration is modified, when obtaining amendment prediction training corresponding with the second prediction training duration
It is long.
Above-mentioned device, optionally, further includes:
Generation unit, for generating prompt information corresponding with amendment prediction training duration；
Display unit is shown for sending the prompt information to preset display equipment.
Compared with prior art, the invention has the following advantages that
The present invention provides a kind of training duration prediction methods of training pattern, comprising: when training pattern is iterated instruction
When practicing, the training duration of the current iteration training of the training pattern is determined；It obtains and the current iteration trained corresponding the
One predicts training duration；By in the trained duration input shot and long term memory network LSTM model that training is completed in advance, obtain
The second of the LSTM model output predicts training duration；Training duration is predicted according to described first, and applies preset karr
Graceful filtering algorithm predicts that training duration is modified to described second, obtains amendment corresponding with the second prediction training duration
It predicts training duration, the amendment is predicted that training duration is saved into preset tables of data；When training is predicted in the amendment
The prediction training duration of the long next iteration training for being determined as the training pattern.Present invention application LSTM model prediction training
Model carries out the prediction training duration of next iteration training, and application Kalman filtering algorithm repairs prediction training duration
Just, to obtain more accurately predicting training duration, the accuracy of the prediction duration obtained to prediction is improved.
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
The embodiment of invention for those of ordinary skill in the art without creative efforts, can also basis
The attached drawing of offer obtains other attached drawings.
Fig. 1 is a kind of method flow diagram of the training duration prediction method of training pattern provided in an embodiment of the present invention；
Fig. 2 is a kind of another method process of the training duration prediction method of training pattern provided in an embodiment of the present invention
Figure；
Fig. 3 is a kind of another method process of the training duration prediction method of training pattern provided in an embodiment of the present invention
Figure；
Fig. 4 is a kind of structure drawing of device of the training time premeauring device of training pattern provided in an 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 this application, the terms "include", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion,
So that the process, method, article or equipment for including a series of elements not only includes those elements, but also including not having
The other element being expressly recited, or further include for elements inherent to such a process, method, article, or device.Do not having
There is the element limited in the case where more limiting by sentence "including a ...", it is not excluded that in the mistake including the element
There is also other identical elements in journey, method, article or equipment.
The present invention can be used in numerous general or special purpose computing device environment or configurations.Such as: personal computer, service
Device computer, handheld device or portable device, laptop device, multiprocessing stout man including any of the above stout man are set
Standby distributed computing environment etc..
The embodiment of the invention provides a kind of training duration prediction method of training pattern, this method can be applied a variety of
Artificial intelligence platform or application in system platform, such as based on cluster arranging system (Kubernetes, K8s)
In the system platform in TensorFlow open source software library, executing subject can be server or processor in system, institute
The method flow diagram of method is stated as shown in Figure 1, specifically including:
S101, when training pattern is iterated trained, determine the training pattern current iteration training training when
It is long.
In method provided in an embodiment of the present invention, the training pattern is the training pattern that user selectes, the trained mould
It is trained in an iterative manner when type is trained.When the training pattern is trained according to preset training method
When, it is to carry out an iteration training that training pattern, which executes a training method, and training pattern executes a training method and spent
The time taken is training duration.
S102, the first prediction training duration corresponding with current iteration training is obtained.
In method provided in an embodiment of the present invention, the repetitive exercise carried out every time has corresponding prediction training duration, needs
It is noted that prediction when training pattern is iterated trained for the first time is a length of pre-set when training, it is iterated for the first time
Trained prediction training duration can be determined according to the corresponding model parameter of training pattern and the data parameters being trained.
S103, the trained duration input is trained in advance in the shot and long term memory network LSTM model completed, obtains institute
It states the second of the output of LSTM model and predicts training duration.
In method provided in an embodiment of the present invention, shot and long term memory network (Long Short-Term Memory, LSTM)
Model is that training is completed in advance, and when being trained to LSTM model, the training dataset of application is the list of each training pattern
The total duration that secondary iteration duration and training pattern are spent after the completion of being trained, the data that training data is concentrated input
It is trained in LSTM model, until the weighted value of LSTM model output result meets preset demand, then the LSTM
Model training is completed.In the LSTM model that training duration input training is completed, the LSTM model is triggered according to training in advance
Process, the trained duration is analyzed, with output second predict training duration.
It should be noted that the trained duration, there are corresponding characteristic parameter, the characteristic parameter can characterize
Training pattern in repetitive exercise difference training progresses, such as train when it is 2 minutes a length of, wherein time interval are as follows: 0 second is extremely
15 seconds training progresses are to read data, and 16 seconds to 60s training progresses are to carry out first time data conversion, 61 seconds to 120 seconds
Training progress be carry out first time data conversion, complete an iteration training.
S104, training duration is predicted according to described first, and the preset Kalman filtering algorithm of application is pre- to described second
It surveys training duration to be modified, obtains amendment corresponding with the second prediction training duration and predict training duration, repaired described
It is positive to predict that training duration is saved into preset tables of data.
In method provided in an embodiment of the present invention, gets corresponding with current iteration training first and predicts training duration,
Pre-set Kalman filtering algorithm predicts training duration according to first in system, predicts that training duration is repaired to second
Just, so that the amendment obtained after amendment predicts that training duration is more accurate；It is in store in the tables of data to be iterated training
Training duration is predicted in the amendment that exports afterwards, and the mode of preservation can be and successively be saved according to being iterated trained sequence,
It can also be saved according to the iteration mark of repetitive exercise.
S105, the amendment being predicted to, training duration is determined as the prediction instruction of the next iteration training of the training pattern
Practice duration.
In method provided in an embodiment of the present invention, Kalman filtering algorithm is predicted to obtain after training duration is modified to second
The prediction training duration of a length of next iteration training when the amendment training arrived；Assuming that current iteration is trained for A, next iteration
It is trained for B；A length of T1 when the prediction training that repetitive exercise A is obtained is executed, when using Kalman filtering algorithm to prediction training
Long T1 is modified, and is obtained amendment and is predicted training duration T2；Then the amendment predicts that training duration T2 is corresponding with repetitive exercise B
Prediction training duration.
It is required when by being iterated trained to training pattern using LSTM model in method provided in an embodiment of the present invention
Training duration predicted, obtain the required training time that training pattern carries out next iteration training, i.e. training pattern
Carry out the prediction training duration of next iteration training；Obtained prediction training duration is repaired using Kalman filtering algorithm
Just, so that revised prediction training duration is more accurate.
In method provided in an embodiment of the present invention, when prediction training pattern carries out the training duration of next iteration training,
Need to obtain the prediction training duration of current iteration training, the process of acquisition is as described below:
Preset tables of data in database is obtained, and determines the iteration mark of the current iteration training；
Prediction corresponding with iteration mark in tables of data training duration is determined as the first prediction training duration.
In method provided in an embodiment of the present invention, the iteration mark of current iteration training can changing for current iteration training
Algebraic value, for example, current iteration training characterization training pattern is iterated training for the first time when iterative numerical is 1；Iterative numerical
When being 2, current iteration training characterization training pattern is iterated training for the second time；When iterative numerical is 3, current iteration training table
Sign training pattern third time is iterated training, and so on, when iterative numerical is N, current iteration training characterization training pattern
N-th is iterated training, and N is positive integer.
It should be noted that in store training pattern is iterated the prediction training duration that training obtains in tables of data, protect
It can identify and be saved with the iteration of repetitive exercise when depositing, therefore the present invention can be monitored the repetitive exercise of training pattern；
It should be noted that iteration mark can characterize training pattern, which time is iterated training；Such as it is preserved in tables of data pre-
Survey training duration 1；Predict training duration 2；Predict training duration 3；Wherein, predict that training duration 1 carries out the 1st time with training pattern
Repetitive exercise is corresponding, and it is corresponding to predict that training duration 2 and training pattern carry out the 2nd repetitive exercise, predict training duration 3 and
It is corresponding that training pattern carries out the 3rd repetitive exercise；When iteration is identified as 2, characterizes training pattern the 2nd time and be iterated instruction
Practice, and will predict that training duration 2 is determined as described first and predicts training duration.
It, can be true after the completion of the current iteration training in training pattern executes in method provided in an embodiment of the present invention
Determine the training duration of current iteration, while the prediction training that training pattern carries out next iteration training is obtained according to training duration
Duration, specific process is as shown in Fig. 2, be described as follows described:
S201, the characteristic parameter for obtaining the trained duration.
In method provided in an embodiment of the present invention, by the training duration input LSTM model of current iteration training, trigger
The LSTM model parses training duration, obtains the characteristic parameter of the trained duration；It should be noted that the spy
Sign parameter can characterize training progress of the training pattern when executing current iteration training, can also include training pattern because of system
The data that middle other factors influence, such as the data that can also be interfered comprising other factors in characteristic parameter, such as: it changes currently
For in the training duration of parameter, within the 6th second to the 40th second period, training pattern receives dry when being iterated trained
It disturbs.
S202, using the duration prediction method in the LSTM model, the characteristic parameter is analyzed, is obtained and institute
State the corresponding prediction duration of characteristic parameter.
In method provided in an embodiment of the present invention, there is the duration prediction side of repetitive exercise in the LSTM model that training is completed
Method analyzes characteristic parameter using duration prediction method, obtains the prediction duration predicted according to the characteristic parameter.
S203, the prediction duration is determined as to the second prediction training duration.
Method provided in an embodiment of the present invention, the next iteration training of a length of pair of training pattern when the second prediction is trained
Original predictive trains duration.
In method provided in an embodiment of the present invention, by applying preset LSTM model, training pattern carries out current iteration
After training, can be obtained training pattern carry out next iteration training prediction training duration, by with the application of the invention, can and
When to training pattern be iterated training needed for duration prediction and record, the data of record can be used as it is subsequent to training mould
The data that the training of type is analyzed.
In method provided in an embodiment of the present invention, training pattern is obtained using LSTM model prediction and is iterated instruction next time
After experienced prediction training duration, also needs to be modified prediction training duration, to improve the accuracy for predicting training duration, have
The process of body is as shown in figure 3, be described as follows described:
S301, obtain it is described first predict training duration amendment weight and it is described second predict training duration when
Long attribute.
In method provided in an embodiment of the present invention, first predicts that the amendment weight of training duration can be trained according to the first prediction
Duration attribute in duration is determined；The amendment weight is a numerical value, and amendment weight is located in preset numerical intervals.
The amendment weight for predicting training duration according to first predicts that training duration is modified to second.
S302, according to the amendment weight, predict training duration to described second using preset Kalman filtering algorithm
Duration attribute be modified, obtain predicting the trained duration of training duration corresponding amendment prediction with described second.
In method provided in an embodiment of the present invention, using Kalman filtering algorithm to the duration category of the second prediction training duration
Property the process that is modified be also the process for being adjusted, optimizing to duration attribute so that described second predicts training duration
Amendment weight is located in preset numerical intervals, obtains amendment corresponding with the second prediction training duration and predicts training duration, institute
It states amendment and predicts that training duration is iterated trained training duration closer to training pattern next time.
S303, prompt information corresponding with amendment prediction training duration is generated.
In method provided in an embodiment of the present invention, training duration is predicted according to the amendment, generates corresponding prompt information,
It may include the information of training pattern in the prompt information, be iterated trained number and carried out next iteration instruction
Experienced prediction training duration.
S304, it the prompt information is sent to preset display equipment shows.
In method provided in an embodiment of the present invention, it will be prompted to information and be shown in preset display equipment to user.
In method provided in an embodiment of the present invention, current iteration is executed to training pattern by application Kalman filtering algorithm
The prediction training duration obtained after training is modified, and obtains more accurately correcting prediction training duration.
The method for implementing to provide for the present invention, is described in detail herein, is described as follows described:
Assuming that user needs that the repetitive exercise process of training pattern D is monitored and is predicted the prediction instruction of training pattern
Practice duration, training pattern D is iterated training, there are corresponding tables of data S, the tables of data S to use by the training pattern D
Duration is trained in storing prediction corresponding to the repetitive exercise of training pattern；When each prediction training saved in the tables of data
It is long to constitute time series corresponding with the training pattern.It should be noted that having pre-saved prediction in the tables of data S
Training duration D1 determines according to the various parameters of training pattern and predicts training duration D1, the prediction predict training duration D1 and
It is corresponding that training pattern D is iterated training for the first time.
Training pattern D is iterated training for the first time, obtains training pattern D and is iterated trained training duration B1 for the first time,
And obtain the training of the prediction in tables of data S duration D1；In the LSTM model that training duration B1 input is completed with preparatory training, obtain
The prediction training duration B2 exported to LSTM model；According to prediction training duration D1, and apply Kalman filtering algorithm pair
It predicts that training duration B2 is modified, obtains revised prediction training duration D2, and prediction training duration D2 is saved
In tables of data S, it should be noted that the prediction training duration D2 is opposite with training pattern D second of repetitive exercise of progress
It answers.It should be noted that the training duration of the m times repetitive exercise of application training model of the present invention, is predicted in LSTM model
The prediction training duration of the m+1 times repetitive exercise of training pattern is obtained, m is positive integer；It can will predict that training duration is first protected
There are in tables of data, prediction training duration is modified, amendment is obtained and predicts training duration, and will correct and predict training duration
Corresponding prediction training duration in replacement data table；It can also will first predict that training duration is modified, by revised prediction
Training duration is stored in preset tables of data.
It should be noted that LSTM model can be made by chain structure using in method provided in an embodiment of the present invention
Incoming information remains unchanged outflow before, further through the structure of " door ", including forgets door, input gate, out gate, information is allowed to select
Selecting property passes through, so that addition or removal information, solve the problems, such as that gradient disappears and gradient is exploded.With the application of the invention, can be to instruction
Practice when model is iterated trained every time and is monitored, it can also be according to the prediction training of the training of the next iteration to training pattern
Duration in platform calculation power and resource is allocated or training duration according to training pattern is charged.
Corresponding with method described in Fig. 1, the embodiment of the invention also provides a kind of training duration of training pattern is pre-
Device is surveyed, for the specific implementation of method in Fig. 1, the training duration prediction of the training pattern of this law sold to you embodiment body is filled
Setting can be applied in terminal or equipment, and concrete structure diagram is as shown in figure 4, specifically include as follows:
First determination unit 401, for when training pattern is iterated trained, determining that the current of the training pattern changes
The training duration of generation training；
Acquiring unit 402 predicts training duration for obtaining corresponding with current iteration training first；
Output unit 403, for the shot and long term memory network LSTM mould completed to be trained in the trained duration input in advance
In type, obtain the LSTM model output second predicts training duration；
Amending unit 404 for predicting training duration according to described first, and applies preset Kalman filtering algorithm pair
Described second predicts that training duration is modified, when obtaining amendment prediction training corresponding with the second prediction training duration
It is long, the amendment is predicted that training duration is saved into preset tables of data；
Second determination unit 405, for the amendment to be predicted that training duration is determined as the training pattern next time
The prediction training duration of repetitive exercise.
The embodiment of the present invention provide device in, when training pattern is iterated it is trained when, determine working as the training pattern
The training duration of preceding repetitive exercise；It obtains corresponding with current iteration training first and predicts training duration；By the training
In the duration input shot and long term memory network LSTM model that training is completed in advance, the second prediction of the LSTM model output is obtained
Training duration；Training duration is predicted according to described first, and the preset Kalman filtering algorithm of application instructs second prediction
Practice duration to be modified, obtains amendment corresponding with the second prediction training duration and predict training duration, in advance by the amendment
Training duration is surveyed to save into preset tables of data；The amendment is predicted that training duration is determined as the next of the training pattern
The prediction training duration of secondary repetitive exercise.Present invention application LSTM model obtains training pattern and carries out the pre- of next iteration training
Training duration is surveyed, and application Kalman filtering algorithm is modified prediction training duration, to obtain more accurately predicting instructing
Practice duration.
In device provided in an embodiment of the present invention, acquiring unit 402 in device be can be set are as follows:
First obtains subelement, for obtaining preset tables of data in database, and determines the current iteration training
Iteration mark；
First determines subelement, for determining prediction corresponding with iteration mark in tables of data training duration
Training duration is predicted for first.
In device provided in an embodiment of the present invention, output unit 403 in device be can be set are as follows:
Second obtains subelement, for obtaining the characteristic parameter of the trained duration；
Subelement is analyzed, for the duration prediction method in the application LSTM model, the characteristic parameter is divided
Analysis, obtains prediction duration corresponding with the characteristic parameter；
First determines subelement, for the prediction duration to be determined as the second prediction training duration.
In device provided in an embodiment of the present invention, amending unit 404 in device be can be set are as follows:
Third obtains subelement, for obtaining the amendment weight and described second of the first prediction training duration in advance
Survey the duration attribute of training duration；
Revise subelemen is used for according to the amendment weight, pre- to described second using preset Kalman filtering algorithm
The duration attribute for surveying training duration is modified, when obtaining amendment prediction training corresponding with the second prediction training duration
It is long.
In device provided in an embodiment of the present invention, described device be may be arranged as:
Generation unit, for generating prompt information corresponding with amendment prediction training duration；
Display unit is shown for sending the prompt information to preset display equipment.
All the embodiments in this specification are described in a progressive manner, 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 or
For system embodiment, since it is substantially similar to the method embodiment, so describing fairly simple, related place is referring to method
The part of embodiment illustrates.System and system embodiment described above is only schematical, wherein the conduct
The unit of separate part description may or may not be physically separated, component shown as a unit can be or
Person may not be physical unit, it can and it is in one place, or may be distributed over multiple network units.It can root
According to actual need that some or all of the modules therein is selected to achieve the purpose of the solution of this embodiment.Ordinary skill
Personnel can understand and implement without creative efforts.
Professional further appreciates that, unit described in conjunction with the examples disclosed in the embodiments of the present disclosure
And algorithm steps, can be realized with electronic hardware, computer software, or a combination of the two, in order to clearly demonstrate hardware and
The interchangeability of software generally describes each exemplary composition and step according to function in the above description.These
Function is implemented in hardware or software actually, the specific application and design constraint depending on technical solution.Profession
Technical staff can use different methods to achieve the described function each specific application, but this realization is not answered
Think beyond the scope of this invention.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, of the invention
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest scope of cause.
Claims (10)
1. a kind of training duration prediction method of training pattern characterized by comprising
When training pattern is iterated it is trained when, determine the training pattern current iteration training training duration；
It obtains corresponding with current iteration training first and predicts training duration；
By in the trained duration input shot and long term memory network LSTM model that training is completed in advance, the LSTM model is obtained
The second of output predicts training duration；
Training duration is predicted according to described first, and the preset Kalman filtering algorithm of application predicts training duration to described second
It is modified, obtains amendment corresponding with the second prediction training duration and predict training duration, training is predicted into the amendment
Duration is saved into preset tables of data；
The amendment is predicted that training duration is determined as the prediction training duration of the next iteration training of the training pattern.
2. the method according to claim 1, wherein the acquisition trains corresponding first with the current iteration
Predict training duration, comprising:
Preset tables of data in database is obtained, and determines the iteration mark of the current iteration training；
Prediction corresponding with iteration mark in tables of data training duration is determined as the first prediction training duration.
3. the method according to claim 1, wherein described train completion for the trained duration input in advance
In shot and long term memory network LSTM model, obtain the LSTM model output second predicts training duration, comprising:
Obtain the characteristic parameter of the trained duration；
Using the duration prediction method in the LSTM model, the characteristic parameter is analyzed, obtains joining with the feature
The corresponding prediction duration of number；
The prediction duration is determined as the second prediction training duration.
4. the method according to claim 1, wherein described predict training duration according to described first, and applying
Preset Kalman filtering algorithm predicts that training duration is modified to described second, obtains predicting training duration with described second
Training duration is predicted in corresponding amendment, comprising:
Obtain the duration attribute of amendment weight and the second prediction training duration that described first predicts training duration；
According to the amendment weight, the duration attribute for predicting training duration to described second using preset Kalman filtering algorithm
It is modified, obtains amendment corresponding with the second prediction training duration and predict training duration.
5. the method according to claim 1, wherein further include:
Generate prompt information corresponding with amendment prediction training duration；
The prompt information is sent to preset display equipment and is shown.
6. a kind of training time premeauring device of training pattern characterized by comprising
First determination unit, for when training pattern is iterated trained, determining the current iteration training of the training pattern
Training duration；
Acquiring unit predicts training duration for obtaining corresponding with current iteration training first；
Output unit, for obtaining in the trained duration input shot and long term memory network LSTM model that training is completed in advance
Training duration is predicted to the LSTM model exports second；
Amending unit, for predicting training duration according to described first, and the preset Kalman filtering algorithm of application is to described the
Two predict that training duration is modified, and obtain amendment corresponding with the second prediction training duration and predict training duration, by institute
It states amendment and predicts that training duration is saved into preset tables of data；
Second determination unit, for the amendment to be predicted that training duration is determined as the next iteration training of the training pattern
Prediction training duration.
7. device according to claim 6, which is characterized in that the acquiring unit, comprising:
First obtains subelement, for obtaining preset tables of data in database, and determines the iteration of the current iteration training
Mark；
First determines subelement, for the trained duration of prediction corresponding with iteration mark in the tables of data to be determined as the
One predicts training duration.
8. device according to claim 6, which is characterized in that the output unit, comprising:
Second obtains subelement, for obtaining the characteristic parameter of the trained duration；
Subelement is analyzed, for the duration prediction method in the application LSTM model, the characteristic parameter is analyzed, is obtained
To prediction duration corresponding with the characteristic parameter；
First determines subelement, for the prediction duration to be determined as the second prediction training duration.
9. device according to claim 6, which is characterized in that the amending unit, comprising:
Third obtains subelement, for obtaining the amendment weight and the second prediction instruction that described first predicts training duration
Practice the duration attribute of duration；
Revise subelemen, for being instructed to second prediction using preset Kalman filtering algorithm according to the amendment weight
The duration attribute for practicing duration is modified, and is obtained amendment corresponding with the second prediction training duration and is predicted training duration.
10. device according to claim 6, which is characterized in that further include:
Generation unit, for generating prompt information corresponding with amendment prediction training duration；
Display unit is shown for sending the prompt information to preset display equipment.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910782983.1A CN110516805A (en) | 2019-08-23 | 2019-08-23 | The training duration prediction method and device of training pattern |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910782983.1A CN110516805A (en) | 2019-08-23 | 2019-08-23 | The training duration prediction method and device of training pattern |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110516805A true CN110516805A (en) | 2019-11-29 |
Family
ID=68627257
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910782983.1A Pending CN110516805A (en) | 2019-08-23 | 2019-08-23 | The training duration prediction method and device of training pattern |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110516805A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115953738A (en) * | 2023-03-02 | 2023-04-11 | 上海燧原科技有限公司 | Monitoring method, device, equipment and medium for image recognition distributed training |
| CN116680619A (en) * | 2023-07-28 | 2023-09-01 | 江西中医药大学 | Method and device for predicting decoction time classification, electronic equipment and storage medium |
| CN116720544A (en) * | 2023-08-04 | 2023-09-08 | 浪潮电子信息产业股份有限公司 | Model training time-consuming prediction method, device and system based on heterogeneous computing system |
-
2019
- 2019-08-23 CN CN201910782983.1A patent/CN110516805A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115953738A (en) * | 2023-03-02 | 2023-04-11 | 上海燧原科技有限公司 | Monitoring method, device, equipment and medium for image recognition distributed training |
| CN116680619A (en) * | 2023-07-28 | 2023-09-01 | 江西中医药大学 | Method and device for predicting decoction time classification, electronic equipment and storage medium |
| CN116720544A (en) * | 2023-08-04 | 2023-09-08 | 浪潮电子信息产业股份有限公司 | Model training time-consuming prediction method, device and system based on heterogeneous computing system |
| CN116720544B (en) * | 2023-08-04 | 2023-11-07 | 浪潮电子信息产业股份有限公司 | Model training time-consuming prediction method, device and system based on heterogeneous computing system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110516805A (en) | The training duration prediction method and device of training pattern | |
| Khan et al. | A comprehensive study of commonly practiced heavy & light weight software methodologies | |
| CN103593479B (en) | Application program menu recommendation method and device | |
| US8631384B2 (en) | Creating a test progression plan | |
| CN109034660A (en) | Based on the determination method and relevant apparatus of the risk control strategy of prediction model | |
| US10546507B2 (en) | Recommending a set of learning activities based on dynamic learning goal adaptation | |
| US8463174B2 (en) | System and methods for sequencing learning objects based upon query and user relevancy | |
| US8843878B1 (en) | Quality software development process | |
| CN110990691A (en) | Online course recommendation method and device and computer storage medium | |
| EP1637944A2 (en) | Bootstrap data methodology for sequential hybrid model building | |
| CN109903100A (en) | A kind of customer churn prediction technique, device and readable storage medium storing program for executing | |
| CN109300069A (en) | Acquisition methods, device and the electronic equipment of user's learning path model | |
| CN108182448A (en) | A kind of selection method and relevant apparatus for marking strategy | |
| US20140002266A1 (en) | Methods and Apparatus for Muscle Memory Training | |
| US20100162185A1 (en) | Electronic circuit design | |
| CN106550208A (en) | Video method for splitting, equipment and video analytic system | |
| CN106557414A (en) | Application testing system and method | |
| US20140316846A1 (en) | Estimating financial risk based on non-financial data | |
| CN109961160A (en) | A kind of power grid future operation trend predictor method and system based on trend parameter | |
| CN111080025A (en) | Learning feature data processing method and device and electronic equipment | |
| CN110008880A (en) | A kind of model compression method and device | |
| JP7245139B2 (en) | WORK SUPPORT DEVICE, WORK SUPPORT SYSTEM AND WORK SUPPORT METHOD | |
| Banks et al. | A heuristic incremental modeling approach to course timetabling | |
| Sarkar et al. | Adaptive E-learning using deterministic finite automata | |
| CN105786789A (en) | Method and device for computing text similarity degree |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191129 |
|
| RJ01 | Rejection of invention patent application after publication |