CN110991659B - Abnormal node identification method, device, electronic equipment and storage medium - Google Patents
Abnormal node identification method, device, electronic equipment and storage medium Download PDFInfo
- Publication number
- CN110991659B CN110991659B CN201911250256.7A CN201911250256A CN110991659B CN 110991659 B CN110991659 B CN 110991659B CN 201911250256 A CN201911250256 A CN 201911250256A CN 110991659 B CN110991659 B CN 110991659B
- Authority
- CN
- China
- Prior art keywords
- node
- deep learning
- learning model
- nodes
- identified
- 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.)
- Active
Links
- 230000002159 abnormal effect Effects 0.000 title claims abstract description 103
- 238000000034 method Methods 0.000 title claims abstract description 76
- 238000013136 deep learning model Methods 0.000 claims abstract description 214
- 238000012545 processing Methods 0.000 claims abstract description 143
- 238000012544 monitoring process Methods 0.000 claims abstract description 51
- 238000012360 testing method Methods 0.000 claims abstract description 42
- 230000008569 process Effects 0.000 claims abstract description 18
- 238000004891 communication Methods 0.000 claims description 18
- 230000015654 memory Effects 0.000 claims description 18
- 238000005457 optimization Methods 0.000 claims description 18
- 238000013528 artificial neural network Methods 0.000 claims description 13
- 238000004590 computer program Methods 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 9
- 230000011218 segmentation Effects 0.000 claims description 9
- 238000001514 detection method Methods 0.000 claims description 7
- 230000000007 visual effect Effects 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000010191 image analysis Methods 0.000 claims 4
- 230000001133 acceleration Effects 0.000 abstract description 8
- 238000010586 diagram Methods 0.000 description 5
- 230000009471 action Effects 0.000 description 3
- 238000010845 search algorithm Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 230000002093 peripheral effect Effects 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 230000017105 transposition Effects 0.000 description 2
- 244000025254 Cannabis sativa Species 0.000 description 1
- 241000282326 Felis catus Species 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000000802 evaporation-induced self-assembly Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 239000013307 optical fiber Substances 0.000 description 1
- 210000000056 organ Anatomy 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000012549 training Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- 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)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the invention provides an abnormal node identification method, an abnormal node identification device, electronic equipment and a storage medium, wherein the method comprises the following steps: the method comprises the steps of inputting feature data of a test image into a deep learning model to be identified, wherein the deep learning model to be identified comprises a plurality of nodes, monitoring processing time of designated nodes in the plurality of nodes in the process of processing the feature data by the deep learning model to be identified, wherein the processing time of the designated nodes is time of the designated nodes for processing received data, and determining the designated nodes as abnormal nodes when the processing time of the designated nodes is longer than a preset time threshold. By adopting the scheme provided by the embodiment of the invention, the abnormal nodes are identified from the plurality of nodes contained in the deep learning model, and after the abnormal nodes are identified, the abnormal nodes can be further processed, so that the deep learning model reasoning acceleration performance method can be studied deeply, and the running speed of the deep learning model can be accelerated.
Description
Technical Field
The present invention relates to the field of machine learning technologies, and in particular, to a method and apparatus for identifying abnormal nodes, an electronic device, and a storage medium.
Background
In the technical field of machine learning, the deep learning model is developed rapidly and is widely applied. Currently, in practical applications, there may be redundancy in the deep learning model, for example, there is redundancy in parameters and structures of the deep learning model, that is, there are redundant nodes or parameters in the model, which results in a longer running time of the model when the deep learning model is applied. In addition, the current deep learning model has the problem that part of node operation time is long, and part of node operation time is long, so that the operation speed of the deep learning model is low. Therefore, the optimization processing of the nodes with long running time in the deep learning model is beneficial to improving the running speed of the deep learning model. Before optimizing the nodes with long running time in the deep learning model, how to identify the nodes with long running time from the nodes contained in the deep learning model is important.
Disclosure of Invention
The embodiment of the invention aims to provide an abnormal node identification method which is used for realizing how to identify abnormal nodes from a plurality of nodes contained in a deep learning model. The specific technical scheme is as follows:
to achieve the above object, an embodiment of the present invention provides a method for identifying an abnormal node, including:
Inputting feature data of a test image into a deep learning model to be identified, wherein the deep learning model to be identified comprises a plurality of nodes;
monitoring the processing time length of a designated node in the plurality of nodes in the process of processing the feature data by the deep learning model to be identified, wherein the processing time length of the designated node is the time length of the designated node for processing the received data;
and when the processing time length of the designated node is greater than a preset time length threshold value, determining that the designated node is an abnormal node.
Further, the monitoring the processing duration of a designated node in the plurality of nodes includes:
monitoring an input time point when a designated node in the plurality of nodes receives data to be processed, and an output time point when the received data is processed;
and calculating the difference value of the output time point minus the input time point as the processing time length of the designated node.
Further, the monitoring the processing duration of a designated node in the plurality of nodes includes:
monitoring the time length from the start of inputting the characteristic data into the deep learning model to be identified to the time length when each node in the plurality of nodes receives the data to be processed as the arrival time length;
And calculating the arrival time of the next node of the designated node, and subtracting the difference value of the arrival time of the designated node to be used as the processing time of the designated node.
Further, the calculating the arrival time of the next node of the designated node, subtracting the difference value of the arrival time of the designated node, includes:
when a plurality of next nodes exist in the designated node, selecting the next node with the minimum arrival duration from the plurality of next nodes;
and calculating the arrival time of the next node with the minimum arrival time, and subtracting the difference value of the arrival time of the appointed node.
Further, the monitoring the processing duration of a designated node in the plurality of nodes includes:
monitoring a time period from the start of inputting the characteristic data into the deep learning model to be identified to the completion of processing the received data by each of the plurality of nodes as an output time period;
and calculating the output time length of the designated node, and subtracting the difference value of the output time length of the node before the designated node as the processing time length of the designated node.
Further, the calculating the output duration of the designated node, subtracting the difference value of the output duration of the node before the designated node, includes:
When a plurality of previous nodes exist in the designated node, selecting the previous node with the largest output duration from the plurality of previous nodes;
and calculating the output time length of the designated node, and subtracting the difference value of the output time length of the previous node with the largest selected output time length.
Further, the deep learning model to be identified is a model obtained by optimizing an original deep learning model based on a high-performance neural network reasoning engine TensorRT; or alternatively
The deep learning model to be identified is a model obtained by optimizing an original deep learning model based on an open visual reasoning and neural network optimization tool OpenVINO.
In order to achieve the above object, an embodiment of the present invention further provides an abnormal node identification apparatus, including:
the input module is used for inputting the characteristic data of the test image into a deep learning model to be identified, and the deep learning model to be identified comprises a plurality of nodes;
the monitoring module is used for monitoring the processing time length of a designated node in the plurality of nodes in the process of processing the feature data by the deep learning model to be identified, wherein the processing time length of the designated node is the time length of the designated node for processing the received data;
And the determining module is used for determining the designated node as an abnormal node when the processing time length of the designated node is greater than a preset time length threshold value.
Further, the monitoring module includes:
the monitoring sub-module is used for monitoring the input time point when a designated node in the plurality of nodes receives the data to be processed and the output time point when the received data is processed;
and the calculating sub-module is used for calculating the difference value of the output time point minus the input time point as the processing time length of the designated node.
Further, the monitoring module includes:
the monitoring sub-module is used for monitoring the time length from the start of inputting the characteristic data into the deep learning model to be identified to the time length from the time when each node in the plurality of nodes receives the data to be processed as the arrival time length;
and the calculation sub-module is used for calculating the arrival time of the next node of the designated node and subtracting the difference value of the arrival time of the designated node to be used as the processing time of the designated node.
Further, the computing submodule is specifically configured to select, when a plurality of next nodes exist in the designated node, a next node with a minimum arrival duration from the plurality of next nodes; and calculating the arrival time of the next node with the minimum arrival time, and subtracting the difference value of the arrival time of the appointed node.
Further, the monitoring module includes:
the monitoring sub-module is used for monitoring the duration from the start of inputting the characteristic data into the deep learning model to be identified to the completion of processing the received data by each node in the plurality of nodes, and taking the duration as the output duration;
and the calculating sub-module is used for calculating the output time length of the designated node and subtracting the difference value of the output time length of the node before the designated node as the processing time length of the designated node.
Further, the computing submodule is specifically configured to select a previous node with a maximum output duration from the plurality of previous nodes when the designated node has the plurality of previous nodes; and calculating the output time length of the designated node, and subtracting the difference value of the output time length of the previous node with the largest selected output time length.
Further, the deep learning model to be identified is a model obtained by optimizing an original deep learning model based on a high-performance neural network reasoning engine TensorRT; or alternatively
The deep learning model to be identified is a model obtained by optimizing an original deep learning model based on an open visual reasoning and neural network optimization tool OpenVINO.
In order to achieve the above object, an embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, where the processor, the communication interface, and the memory complete communication with each other through the communication bus;
a memory for storing a computer program;
and the processor is used for realizing any abnormal node identification method step when executing the program stored in the memory.
To achieve the above object, an embodiment of the present invention provides a computer-readable storage medium having stored therein a computer program which, when executed by a processor, implements any of the above-described abnormal node identification method steps.
To achieve the above object, an embodiment of the present invention further provides a computer program product containing instructions, which when run on a computer, cause the computer to perform any of the above-mentioned abnormal node identification method steps.
The embodiment of the invention has the beneficial effects that:
according to the abnormal node identification method provided by the embodiment of the invention, the nodes of the deep learning model are obtained, the time length of processing the received data by the designated nodes is monitored in the process of processing the characteristic data of the test image by the deep learning model to be identified, and the designated nodes with the processing time length being greater than the preset time length threshold are determined to be abnormal nodes. By adopting the method provided by the embodiment of the invention, the appointed node with the processing time longer than the preset time threshold is used as the abnormal node by monitoring the time for processing the received data by the appointed node, so that the abnormal node is identified from a plurality of nodes contained in the deep learning model.
Of course, it is not necessary for any one product or method of practicing the invention to achieve all of the advantages set forth above at the same time.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below.
FIG. 1 is a first flowchart of a method for identifying abnormal nodes according to an embodiment of the present invention;
FIG. 2 is a second flowchart of a method for identifying abnormal nodes according to an embodiment of the present invention;
FIG. 3 is a third flowchart of an abnormal node identification method according to an embodiment of the present invention;
FIG. 4 is a fourth flowchart of an abnormal node identification apparatus according to an embodiment of the present invention;
FIG. 5 is a first block diagram of an abnormal node identification apparatus according to an embodiment of the present invention;
FIG. 6 is a second block diagram of an abnormal node identification apparatus according to an embodiment of the present invention;
fig. 7 is a schematic structural diagram of an electronic device according to an embodiment of the present invention;
fig. 8 is a schematic structural diagram of a part of nodes of a deep learning model according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be described below with reference to the accompanying drawings in the embodiments of the present invention.
Because the existing deep learning model has the problem that the operation speed of the deep learning model is low due to long operation time of part of nodes, in order to solve the technical problem, the embodiment of the invention provides an abnormal node identification method, as shown in fig. 1, which comprises the following steps:
step 101, inputting feature data of a test image into a deep learning model to be identified, wherein the deep learning model to be identified comprises a plurality of nodes.
Step 102, monitoring the processing time length of a designated node in the plurality of nodes in the process of processing the feature data by the deep learning model to be identified, wherein the processing time length of the designated node is the time length of the designated node for processing the received data.
And step 103, determining the designated node as an abnormal node when the processing time of the designated node is longer than a preset time threshold.
By adopting the method provided by the embodiment of the invention, the appointed node with the processing time longer than the preset time threshold is used as the abnormal node by monitoring the time for processing the received data by the appointed node, so that the abnormal node is identified from a plurality of nodes contained in the deep learning model.
In the application of the deep learning model, nodes with processing time length larger than a preset time length threshold value may cause the running speed of the deep learning model to be reduced, and the application of the deep learning model is affected. And therefore, the node with the processing time length larger than the preset time length threshold value in the nodes of the deep learning model is used as the abnormal node. By adopting the method provided by the embodiment of the invention, the abnormal node of the deep learning model can be identified, and after the abnormal node of the deep learning model is identified, the abnormal node can be further processed, so that the deep learning model reasoning acceleration performance method can be studied deeply, the deep learning model is optimized, and the running speed of the deep learning model is accelerated.
The method and the device for identifying the abnormal node provided by the embodiment of the invention are described in detail by a specific embodiment.
The embodiment of the invention discloses an abnormal node identification method, which can comprise the following steps as shown in fig. 2:
step 201, inputting feature data of a test image into a deep learning model to be identified, wherein the deep learning model to be identified comprises a plurality of nodes.
In the embodiment of the invention, the deep learning model is formed by mutually connecting a plurality of nodes. Each node represents a specific output function, called an activation function, such as Sigmoid (logistic regression function). Each node is used for receiving data to be processed from the connected upper-level node, processing the data to be processed and outputting the processed data to the connected lower-level node.
In the embodiment of the invention, the deep learning model to be identified can be a model obtained by optimizing an original deep learning model based on TensorRT (high performance neural network reasoning engine); or, the deep learning model to be identified can also be a model obtained by optimizing the original deep learning model based on OpenVINO (open visual reasoning and neural network optimization tool); alternatively, the deep learning model to be identified may be an original deep learning model that has not undergone any optimization processing.
At present, the model obtained by optimizing the original deep learning model based on TensorRT and the model obtained by optimizing the original deep learning model based on OpenVINO can both improve the running speed of the model on the original deep learning model. However, the existing optimization mode has the limitation that the optimization is unreasonable and the speed of the optimized model is improved less. For example, when the OpenVINO is used to optimize the original deep learning model, since the default input channel is in the NCHW format, and the default channel used by the TensorFlow (artificial intelligence learning system) is in the NHWC format, the optimized model needs to be transposed, where the NCHW format indicates that the storage order of the input data in the memory is NCHW (batch channels height width, the number of pixels in the bulk channel number height direction, and the number of pixels in the width direction), and the NHWC format indicates that the storage order of the input data in the memory is NHWC (batch height width channels, the number of pixels in the bulk height direction, and the number of pixels in the width direction). And the transposition time of the three-dimensional convolution is high in expenditure, so that the transposition operation seriously affects the running speed of the optimized model. Therefore, the deep learning model obtained after optimization based on the original deep learning model can be used as the deep learning model to be identified, and the abnormal node can be identified.
In this step, a plurality of node names of the deep learning model to be identified may also be obtained, for example, after the original deep learning model is optimized by using OpenVINO, optimized deep learning model structure information to be identified may be obtained, where the deep learning model structure information to be identified includes names of all nodes of the deep learning model to be identified.
Step 202, monitoring an input time point when a designated node in the plurality of nodes receives data to be processed, and an output time point when processing of the received data is completed.
In the embodiment of the invention, when the abnormal node identification is performed on the deep learning model to be identified, the abnormal node identification can be performed on all the nodes of the deep learning model to be identified, and the abnormal node identification can be performed only on the nodes in part of the parameter layers of the deep learning model to be identified. When abnormal node identification is performed on nodes in a part of the parameter layers of the deep learning model to be identified, the nodes in the part of the parameter layers can be determined as designated nodes, and further, the abnormal nodes can be identified from the designated nodes.
In this step, a plurality of node names of the deep learning model to be identified may be obtained based on the structural information of the deep learning model to be identified, and further, a designated node may be selected from the obtained plurality of nodes, and the selected designated node may be one designated node, may be a plurality of designated nodes, or may be all the nodes of the deep learning model to be identified.
In this step, the designated node receives the data characterizing the test image feature, and after receiving the data characterizing the test image feature, the designated node may perform corresponding processing on the received data to obtain processed data, and the designated node may output the processed data to the next designated node.
For the designated nodes, according to the processing sequence of the data representing the characteristics of the test image, the input time point of each designated node receiving the data to be processed and the output time point of the received data after processing are recorded in sequence. The method comprises the steps of marking data representing the characteristics of the test image, judging whether the designated node finishes processing the received data, further recording the output time point of the designated node for finishing the received data processing, and determining that the designated node finishes processing the received data when the marked data representing the characteristics of the test image is changed compared with the received data of the designated node, and recording the time point of the marked data representing the characteristics of the test image as the output time point of the designated node. For example, in one possible implementation, the data characterizing the test image features that the designated node a receives to be processed are a, b, c, d, e and f, where f is the label data, and when it is monitored that f changes to f ′ Indicating that the processing of the received data a, b, c, d, e and f by the designated node a is complete,can record data f to f ′ Is the output time point corresponding to the designated node a.
In one possible implementation manner, after the feature data of the test image is input into the deep learning model to be identified, an input time point when each node of the deep learning model to be identified receives data to be processed and an output time point when processing the received data is completed may be recorded by using logging. Furthermore, the input time point when the appointed node in the deep learning model to be identified receives the data to be processed and the output time point when the received data is processed can be obtained, so that the appointed node in the deep learning model to be identified is monitored.
In step 203, the difference obtained by subtracting the input time point from the output time point is calculated as the processing time length of the designated node.
In this step, the difference of the output time point minus the input time point of the designated node represents the length of time it takes for the designated node to process the data characterizing the test image feature.
In one possible embodiment, the point in time at which node a receives the data characterizing the test image feature to be processed is designated as t 1 Designating the point in time when node A completes processing the received data as t 2 Calculate (t 2 -t 1 ) As the processing time length of the designated node a, the length of time taken for the designated node a to process the data representing the test image feature is represented.
And 204, determining the designated node as an abnormal node when the processing time of the designated node is longer than a preset time threshold.
In this step, the preset duration threshold may be specifically set according to different deep learning models to be identified and different performances of running the deep learning model to be identified, and different preset duration thresholds may be set for different designated nodes.
In one possible implementation manner, 5 times of the processing duration of the convolution node B specified in the deep learning model to be identified may be set as a preset duration threshold, for each specified node, it is determined whether the processing duration of the specified node is greater than 5 times of the processing duration of the convolution node B, and when the processing duration of the specified node is greater than 5 times of the processing duration of the convolution node B, the specified node may be determined as an abnormal node.
By adopting the method provided by the embodiment of the invention, the input time point of the data to be processed received by the designated node and the output time point of the received data after processing are monitored, the difference value of the input time point subtracted by the output time point is calculated and used as the processing time of the designated node, and the designated node with the processing time longer than the preset time threshold is used as the abnormal node, so that the abnormal node is identified from a plurality of nodes contained in the deep learning model. And after the abnormal nodes of the deep learning model are identified, the abnormal nodes can be further processed, so that the deep learning model reasoning acceleration performance method can be studied deeply, the deep learning model is optimized in a targeted manner, the running speed of the deep learning model is increased, and the optimization efficiency of the model is improved.
In still another embodiment of the present invention, as shown in fig. 3, the method for identifying abnormal nodes provided in the embodiment of the present invention may include the following steps:
step 301, inputting feature data of a test image into a deep learning model to be identified, wherein the deep learning model to be identified comprises a plurality of nodes.
This step is the same as step 201 described above, and will not be described here again.
In step 302, a time period from inputting feature data into the deep learning model to be identified until each node in the plurality of nodes receives data to be processed is monitored as an arrival time period.
In this step, the plurality of nodes of the deep learning model to be identified receive data representing the characteristics of the test image, and after each of the plurality of nodes receives the data representing the characteristics of the test image, the plurality of nodes can perform corresponding processing on the received data to obtain processed data, and can output the processed data to the next node.
In this step, the time when the feature data is input into the deep learning model to be identified may be taken as an initial time, and for a plurality of nodes, according to a processing sequence of data representing the feature of the test image, time points when each node receives the data to be processed are sequentially recorded, where the initial time may be set according to a specific application scenario, for example, the initial time may be set to zero.
In one possible implementation manner, after the feature data of the test image is input into the deep learning model to be identified, a logging/info function in the deep learning model to be identified may be used to record a time point when the designated node of the deep learning model to be identified receives the data to be processed.
Step 303, determining whether the designated node corresponds to only one next node, if yes, executing step 304a, and if no, executing step 304b.
Step 304a, calculating the arrival time of the next node of the designated node, and subtracting the difference value of the arrival time of the designated node as the processing time of the designated node.
In this step, when the designated node corresponds to only one next node, the arrival time of the next node of the designated node may be calculated, and the difference between the arrival time of the designated node and the arrival time of the designated node may be subtracted, where the obtained difference represents the length of time consumed by the designated node in the processing procedure of the data characterizing the test image feature, and the obtained difference may be used as the processing time of the designated node.
In one possible implementation, the duration of time that the designated node C receives the data to be processed is t C The next node corresponding to the designated node C is only designated node D, and the duration of the designated node D receiving the data to be processed is t D Can calculate (t D -t C ) As the processing time length of the designated node C, the length of time taken for the designated node C to process the data representing the test image feature is represented.
Step 304b, calculating the arrival time of the next node with the minimum arrival time in the plurality of next nodes, and subtracting the difference value of the arrival time of the designated node as the processing time of the designated node.
In this step, when a plurality of next nodes exist in the designated node, a next node with the minimum arrival time length can be selected from the plurality of next nodes, the arrival time length of the selected next node with the minimum arrival time length is calculated, and the difference value of the arrival time lengths of the designated nodes is subtracted as the processing time length of the designated node.
In one possible implementation, the duration of time that the designated node E receives the data to be processed is t E Designating node E to correspond to a plurality of next nodes: designated node F 1 Designated node F 2 Designated node F 3 And designated node F 4 And specify node F 1 The time length of receiving the data to be processed is t F1 Designated node F 2 The time length of receiving the data to be processed is t F2 Designated node F 3 The time length of receiving the data to be processed is t F3 Designated node F 4 The time length of receiving the data to be processed is t F4 Comparison of t F1 、t F2 、t F3 And t F4 Selecting the smallest min { t } F1 ,t F2 ,t F3 ,t F4 Can calculate (min { t } F1 ,t F2 ,t F3 ,t F4 }-t E ) As the processing time length of the designated node E, the length of time taken for the designated node E to process the data representing the test image feature is represented.
And 305, determining the designated node as an abnormal node when the processing time of the designated node is longer than a preset time threshold.
This step is the same as step 204 described above, and will not be described here again.
By adopting the method provided by the embodiment of the invention, the arrival time of the designated node is monitored, the processing time of the designated node is calculated, and the designated node with the processing time longer than the preset time threshold is used as the abnormal node, so that the abnormal node is identified from a plurality of nodes contained in the deep learning model. And after the abnormal nodes of the deep learning model are identified, the abnormal nodes can be further processed, so that the deep learning model reasoning acceleration performance method can be researched deeply, the deep learning model is optimized, and the running speed of the deep learning model is accelerated.
In still another embodiment of the present invention, as shown in fig. 4, the method for identifying abnormal nodes provided in the embodiment of the present invention may include the following steps:
In step 401, feature data of the test image is input into a deep learning model to be identified, wherein the deep learning model to be identified comprises a plurality of nodes.
This step is the same as step 201 described above, and will not be described here again.
Step 402, monitoring a time period from inputting feature data into the deep learning model to be identified to completion of processing of the received data by each of the plurality of nodes as an output time period.
In this step, the plurality of nodes of the deep learning model to be identified receive data representing the features of the test image, and after each of the plurality of nodes receives the data representing the features of the test image, the received data may be processed accordingly to obtain processed data, and each node outputs the duration of the processed data to the next node as the corresponding output duration.
In this step, the time when the feature data is input into the deep learning model to be identified may be taken as an initial time, and for a plurality of nodes, according to the processing sequence of the data characterizing the feature of the test image, the time point when each node finishes processing the received data is sequentially recorded, where the initial time may be set according to a specific application scenario, for example, the initial time may be set to zero.
In one possible implementation, after the feature data of the test image is input into the deep learning model to be identified, a logging/info function in the deep learning model to be identified may be used to record a point in time when the processing of the received data by the designated node of the deep learning model to be identified is completed.
Step 403, determining whether the designated node corresponds to only one previous node, if yes, executing step 404a, and if no, executing step 404b.
Step 404a, calculating the output time length of the designated node, and subtracting the difference value of the previous node output time length of the designated node as the processing time length of the designated node.
In this step, when the designated node corresponds to only one previous node, the output duration of the designated node may be calculated, and the difference between the output durations of the previous nodes of the designated node may be subtracted, where the obtained difference represents the length of time consumed by the designated node in the processing procedure of the data characterizing the test image feature, and the obtained difference may be used as the processing duration of the designated node.
In one possible implementation, the duration of time that the node G finishes processing the received data is specified to be t H The previous node corresponding to the designated node H is only designated node G, and the time length of the designated node G for processing the received data is t G Can calculate (t H -t G ) As the processing time length of the designated node H, the length of time taken for the designated node H to process the data representing the test image feature is represented.
Step 404b, calculating the output time length of the designated node, and subtracting the difference value of the output time lengths of the previous nodes with the largest output time length from the plurality of previous nodes as the processing time length of the designated node.
In this step, when the designated node has a plurality of previous nodes, a previous node with the largest output duration may be selected from the plurality of previous nodes, the output duration of the designated node is calculated, and the difference between the output durations of the previous nodes with the largest selected output durations is subtracted as the processing duration of the designated node.
In one possible implementation, the duration of time that the node M finishes processing the received data is specified to be t M The designated node M corresponds to a plurality of previous nodes: designated node L 1 Designated node L 2 Designated node L 3 Designated node L 4 And designated node L 5 And specifies node L 1 The received data processing is completed for a time period of t L1 Designated node L 2 The received data processing is completed for a time period of t L2 Designated node L 3 The received data processing is completed for a time period of t L3 Designated node L 4 The received data processing is completed for a time period of t L4 Designated node L 5 The received data processing is completed for a time period of t L5 Comparison of t L1 、t L2 、t L3 、t L4 And t L5 Selecting the maximum max { t }, among the magnitudes L1 ,t L2 ,t L3 ,t L4 ,t L5 Can calculate (t) M -max{t L1 ,t L2 ,t L3 ,t L4 ,t L5 }) as a processing time length of the designated node M, the length of time it takes for the designated node M to process data representing the test image feature.
And step 405, determining the designated node as an abnormal node when the processing time of the designated node is longer than a preset time threshold.
This step is the same as step 204 described above, and will not be described here again.
By adopting the method provided by the embodiment of the invention, the output time length of the designated node is monitored, the processing time length of the designated node is calculated, and the designated node with the processing time length being greater than the preset time length threshold is used as the abnormal node, so that the abnormal node is identified from a plurality of nodes contained in the deep learning model. And after the abnormal nodes of the deep learning model are identified, the abnormal nodes can be further processed, so that the deep learning model reasoning acceleration performance method can be researched deeply, the deep learning model is optimized, and the running speed of the deep learning model is accelerated.
In the embodiment of the invention, after determining the abnormal node of the deep learning model to be identified, the identified abnormal node can be processed by adopting the following method:
the first way is: when the identified abnormal node is a redundant node for the deep learning model, the identified abnormal node may be selected for deletion. And deleting the abnormal nodes to obtain a new deep learning model, and applying the new deep learning model to the applicable scene.
The second way is: when the identified abnormal node is not redundant for the deep learning model, the node which consumes less time but can execute the same function can be selected to replace the identified abnormal node, a new deep learning model is obtained after the node is changed, whether the new deep learning model operates normally can be detected, and if the obtained new deep learning model operates normally and the operating speed is improved, the new deep learning model can be applied to an applicable scene.
The scheme provided by the embodiment of the invention can also identify abnormal nodes aiming at the optimized deep learning model. For example, the deep learning model is optimized by using a development tool such as OpenVINO, tensorRT, the running speed of the optimized model is accelerated, but after the deep learning model is optimized by using two tools such as OpenVINO and TensorRT, the running time of part of nodes of the optimized deep learning model is longer than that before the optimization. Aiming at the problems, the scheme provided by the embodiment of the invention can be adopted, the optimized deep learning model is used as the deep learning model to be identified, the abnormal nodes of the model are identified through the scheme provided by the embodiment of the invention, and after the abnormal nodes are identified, the study of the deep learning model reasoning acceleration performance method can be further advanced through further processing the abnormal nodes.
In one possible implementation, 398 short videos of 10 to 20 seconds in duration are tested using the deep learning model, with a test time of 1190 seconds using the original deep learning model. Since the default channel for training the original deep learning model is NHWC, and the default input channel for optimizing the original deep learning model is NCHW, when the original deep learning model is optimized by using the OpenVINO tool, the default channel NHWC of the original deep learning model needs to be converted into the NCHW channel. When converting a default channel of an original deep learning model into an NCHW channel, introducing a node with a node name of 3 Dtransfer, and optimizing the original deep learning model by an OpenVINO tool to obtain a first optimization model, wherein the first optimization model comprises the node 3 Dtransfer (transposed node), and the test time of using the first optimization model for testing is 427 seconds. And using the first optimization model as a deep learning model to be identified, and identifying abnormal nodes by using the scheme provided by the embodiment of the invention, wherein the processing time length of the node 3 Dtransose of the deep learning model to be identified is longer than the set threshold time length. After the abnormal node 3 Dtransfer is identified, a second optimization model is obtained when the original deep learning model is optimized by using an OpenVINO tool on the premise that the node 3 Dtransfer is not introduced, and the second optimization model is used for testing, wherein the testing time is 300 seconds.
Therefore, aiming at the optimized deep learning model, after the abnormal nodes of the deep learning model are identified, the abnormal nodes can be further processed, so that the deep learning model is deeply optimized, the deep learning model reasoning acceleration performance method is deeply researched, and the running speed of the deep learning model is further accelerated.
The deep learning model in the embodiment of the invention may be DNN (Deep Neural Network ), which may specifically include: CNN (Convolutional Neural Network ), RNN (Recurrent Neural Network, recurrent neural network), LSTM (Long Short Term Memory, long term memory network).
The deep learning model in the embodiment of the invention can be specifically used for:
object classification: object classification is based on the object recognition problem of classification tasks, i.e. a computer finds out which of these data are the desired objects from given data. For example, cat and dog classification or flower and grass classification;
and (3) target detection: the specific position of the target to be detected can be determined from the current image by target detection, and the application of the target detection is very wide and is often applied to power system detection, medical image detection and the like;
Target segmentation: the target segmentation is to segment the region of a specific target in an image, in the field of deep learning, the research direction of the target segmentation is mainly divided into two types of semantic segmentation and instance segmentation, wherein the semantic segmentation is to classify each pixel point in the image, judge which pixels in the image belong to which target, and the instance segmentation is to judge which pixels belong to the target and which pixels belong to the first target and which pixels belong to the second target, and the key in the medical image at present is to segment human organs;
and (3) voice recognition: the objective of speech recognition is to transmit a segment of natural language to a computer in the form of acoustic signals, which are understood and responded by the computer, and the application scenario of speech recognition may be: the driving navigation software guides the road for the driver and broadcasts the road condition through a voice recognition technology;
automatic driving: in the automatic driving technique, a deep learning model may be used to identify vehicle driving environmental conditions.
In the embodiment of the invention, aiming at each node of the deep learning model, each node of the deep learning model can be identified by adopting DFS (DepthFirst Search, depth-first search of the graph) and BFS (Breadth First Search, breadth-first search algorithm). As shown in FIG. 8, the nodes a1-a19 are part of the deep learning model, where node a1 is the root node.
The nodes a1-a19 can be found out by a breadth-first search algorithm, the breadth-first algorithm can traverse from the root node to sequentially traverse the next level node adjacent to the previous node, the traversed node does not need secondary traversal, and the specific node traversing steps can be as follows:
the traversal may begin with root node a1, followed by traversing node a2 adjacent to root node a1, followed by traversing nodes adjacent to node a2 in turn: node a13, node a17, node a19, node a3; then traversing the nodes adjacent to node a 13: node a10 traverses the nodes adjacent to node a 17: node a16 traverses nodes adjacent to node a 19: node a15 traverses the nodes adjacent to node a 3: node a4; then traversing the nodes adjacent to node a 10: node a9 traverses nodes adjacent to node a 16: node a14 node a18 traverses the nodes adjacent to node a 15: node a8 traverses nodes adjacent to node a 4: node a6; then traversing the nodes adjacent to node a 14: node a12 traverses nodes adjacent to node a 18: node a7 then traverses the nodes adjacent to node a 12: node a11. Nodes a1-a19 of the deep learning model are found.
For any node of the deep learning model, a previous level node adjacent to the node is a father node of the node, a next level node adjacent to the node is a child node of the node, the breadth-first search algorithm can search the node of the deep learning model, and the father node of the child node of the deep learning model can be determined according to the adjacent relation between the nodes.
The node depth of the deep learning model node is the path from the root node to the node, i.e. the number of nodes passed from the root node to the node, plus 2, wherein the depth of the root node is 1. After each node of the deep learning model is found, the found nodes can be ordered according to the node depth, and the nodes of the deep learning model can be generalized. As shown in fig. 8, the node a1 is a root node having a depth of 1, the node a2 having a maximum depth of 2, the node a3 having a maximum depth of 3, the node a4 having a maximum depth of 4, the node a6 having a maximum depth of 5, the node a7 having a maximum depth of 6, the node a8 having a maximum depth of 5, the node a9 having a maximum depth of 10, the node a10 having a maximum depth of 9, the node a11 having a maximum depth of 7, the node a12 having a maximum depth of 6, the node a13 having a maximum depth of 8, the node a14 having a maximum depth of 5, the node a15 having a maximum depth of 4, the node a16 having a maximum depth of 4, the node a17 having a maximum depth of 3, the node a18 having a maximum depth of 5, and the node a19 having a maximum depth of 3. The nodes a1-a19 in fig. 8 may be ordered according to the maximum node depth, and the ordered nodes are sequentially: node a1, node a2, node a3, node a17, node a19, node a4, node a16, node a15, node a6, node a14, node a18, node a8, node a7, node a12, node a11, node a13, node a10, node a9.
By using BFS firstly, we can definitely determine the father node of the deep learning model node, and further generalize the node by DFS, on one hand, the node of the deep learning model is identified, and on the other hand, for each node, the time difference between the node and the father node is the processing duration corresponding to the node, and further according to the scheme provided by the embodiment of the invention, when the processing duration of the node is greater than the preset duration threshold, the node is determined to be an abnormal node.
Based on the same inventive concept, according to the method for identifying an abnormal node provided in the foregoing embodiment of the present invention, correspondingly, another embodiment of the present invention further provides an apparatus for identifying an abnormal node, a schematic structural diagram of which is shown in fig. 5, which specifically includes:
the input module 501 is configured to input feature data of a test image into a deep learning model to be identified, where the deep learning model to be identified includes a plurality of nodes;
the monitoring module 502 is configured to monitor a processing duration of a designated node in the plurality of nodes during a process of processing the feature data by the deep learning model to be identified, where the processing duration of the designated node is a duration of processing the received data by the designated node;
a determining module 503, configured to determine that the designated node is an abnormal node when the processing time period of the designated node is greater than the preset time period threshold.
Therefore, by adopting the device provided by the embodiment of the invention, the appointed node with the processing time longer than the preset time threshold is used as the abnormal node by monitoring the time for processing the received data by the appointed node, so that the abnormal node is identified from a plurality of nodes contained in the deep learning model. And after the abnormal nodes of the deep learning model are identified, the abnormal nodes can be further processed, so that the deep learning model reasoning acceleration performance method can be researched deeply, the deep learning model is optimized, and the running speed of the deep learning model is accelerated.
Further, as shown in fig. 6, the monitoring module 502 includes:
a monitoring submodule 601, configured to monitor an input time point when a designated node among the plurality of nodes receives data to be processed, and an output time point when processing of the received data is completed;
a calculating sub-module 602, configured to calculate a difference value of the output time point minus the input time point as a processing duration of the designated node.
Further, as shown in fig. 6, the monitoring module 502 includes:
a monitoring submodule 601, configured to monitor a duration from inputting feature data into the deep learning model to be identified to receiving data to be processed by each of a plurality of nodes as an arrival duration;
The calculation sub-module 602 is configured to calculate an arrival time of a next node of the designated node, and subtract a difference between the arrival time of the designated node and the arrival time of the next node as a processing time of the designated node.
Further, as shown in fig. 6, the calculating submodule 602 is specifically configured to select, when the designated node has a plurality of next nodes, a next node with a minimum arrival duration from the plurality of next nodes; and calculating the arrival time of the next node with the minimum arrival time, and subtracting the difference value of the arrival time of the designated node.
Further, as shown in fig. 6, the monitoring module 502 includes:
a monitoring submodule 601, configured to monitor a duration from when feature data is input into the deep learning model to be identified to when each of the plurality of nodes completes processing the received data, as an output duration;
the calculating sub-module 602 is configured to calculate an output duration of the designated node, and subtract a difference between an output duration of a node previous to the designated node as a processing duration of the designated node.
Further, as shown in fig. 6, the calculating submodule 602 is specifically configured to select, when the designated node has a plurality of previous nodes, a previous node with a maximum output duration from the plurality of previous nodes; and calculating the output time length of the designated node, and subtracting the difference value of the output time length of the previous node with the largest selected output time length.
Further, the deep learning model to be identified is a model obtained by optimizing an original deep learning model based on TensorRT; or the deep learning model to be identified is a model obtained by optimizing the original deep learning model based on OpenVINO.
Based on the same inventive concept, according to the abnormal node identification method provided by the above embodiment of the present invention, correspondingly, another embodiment of the present invention further provides an electronic device, referring to fig. 7, where the electronic device according to the embodiment of the present invention includes a processor 701, a communication interface 702, a memory 703 and a communication bus 704, and the processor 701, the communication interface 702 and the memory 703 complete communication with each other through the communication bus 704.
A memory 703 for storing a computer program;
the processor 701 is configured to execute the program stored in the memory 703, and implement the following steps:
inputting feature data of a test image into a deep learning model to be identified, wherein the deep learning model to be identified comprises a plurality of nodes;
monitoring the processing time length of a designated node in the plurality of nodes in the process of processing the feature data by the deep learning model to be identified, wherein the processing time length of the designated node is the time length of the designated node for processing the received data;
And when the processing time length of the designated node is greater than a preset time length threshold value, determining that the designated node is an abnormal node.
The communication bus mentioned above for the electronic devices may be a peripheral component interconnect standard (Peripheral Component Interconnect, PCI) bus or an extended industry standard architecture (Extended Industry Standard Architecture, EISA) bus, etc. The communication bus may be classified as an address bus, a data bus, a control bus, or the like. For ease of illustration, the figures are shown with only one bold line, but not with only one bus or one type of bus.
The communication interface is used for communication between the electronic device and other devices.
The Memory may include random access Memory (Random Access Memory, RAM) or may include Non-Volatile Memory (NVM), such as at least one disk Memory. Optionally, the memory may also be at least one memory device located remotely from the aforementioned processor.
The processor may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; but also digital signal processors (Digital Signal Processing, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), field programmable gate arrays (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
In yet another embodiment of the present invention, there is also provided a computer readable storage medium having stored therein a computer program which, when executed by a processor, implements the steps of any of the above-described abnormal node identification methods.
In yet another embodiment of the present invention, there is also provided a computer program product containing instructions that, when run on a computer, cause the computer to perform any of the abnormal node identification methods of the above embodiments.
In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed on a computer, produces a flow or function in accordance with embodiments of the present invention, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another, for example, by wired (e.g., coaxial cable, optical fiber, digital Subscriber Line (DSL)), or wireless (e.g., infrared, wireless, microwave, etc.). The computer readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains an integration of one or more available media. The usable medium may be a magnetic medium (e.g., floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid State Disk (SSD)), etc.
It is noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.
In this specification, each embodiment is described in a related manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments. In particular, for the apparatus, electronic device and storage medium embodiments, since they are substantially similar to the method embodiments, the description is relatively simple, and references to the parts of the description of the method embodiments are only needed.
The foregoing description is only of the preferred embodiments of the present invention and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.
Claims (14)
1. An abnormal node identification method, comprising:
inputting feature data of a test image into a deep learning model to be identified, wherein the deep learning model to be identified comprises a plurality of nodes; the deep learning model to be identified is a deep learning model for realizing an image analysis task, wherein the image analysis task comprises target classification, target detection or target segmentation;
monitoring the processing time length of a designated node in the plurality of nodes in the process of processing the feature data by the deep learning model to be identified, wherein the processing time length of the designated node is the time length of the designated node for processing the received data;
when the processing time length of the designated node is greater than a preset time length threshold value, determining that the designated node is an abnormal node;
deleting the abnormal node in the deep learning model to be identified when the abnormal node is redundant, so as to obtain a new deep learning model;
When the abnormal node is a non-redundant node, replacing the abnormal node in the deep learning model to be identified by using a target node to obtain a new deep learning model; wherein the target node is: and using a node which consumes less time and performs the same function as the abnormal node.
2. The method of claim 1, wherein the monitoring the processing time of a designated node of the plurality of nodes comprises:
monitoring an input time point when a designated node in the plurality of nodes receives data to be processed, and an output time point when the received data is processed;
and calculating the difference value of the output time point minus the input time point as the processing time length of the designated node.
3. The method of claim 1, wherein the monitoring the processing time of a designated node of the plurality of nodes comprises:
monitoring the time length from the start of inputting the characteristic data into the deep learning model to be identified to the time length when each node in the plurality of nodes receives the data to be processed as the arrival time length;
and calculating the arrival time of the next node of the designated node, and subtracting the difference value of the arrival time of the designated node to be used as the processing time of the designated node.
4. A method according to claim 3, wherein said calculating the arrival time of the next node of the specified node, subtracting the difference in arrival time of the specified node, comprises:
when a plurality of next nodes exist in the designated node, selecting the next node with the minimum arrival duration from the plurality of next nodes;
and calculating the arrival time of the next node with the minimum arrival time, and subtracting the difference value of the arrival time of the appointed node.
5. The method of claim 1, wherein the monitoring the processing time of a designated node of the plurality of nodes comprises:
monitoring a time period from the start of inputting the characteristic data into the deep learning model to be identified to the completion of processing the received data by each of the plurality of nodes as an output time period;
and calculating the output time length of the designated node, and subtracting the difference value of the output time length of the node before the designated node as the processing time length of the designated node.
6. The method of claim 5, wherein the calculating the output duration of the specified node, subtracting the difference in the output duration of the previous node of the specified node, comprises:
When a plurality of previous nodes exist in the designated node, selecting the previous node with the largest output duration from the plurality of previous nodes;
and calculating the output time length of the designated node, and subtracting the difference value of the output time length of the previous node with the largest selected output time length.
7. The method according to claim 1, wherein the deep learning model to be identified is a model obtained by optimizing an original deep learning model based on a high-performance neural network reasoning engine TensorRT; or,
the deep learning model to be identified is a model obtained by optimizing an original deep learning model based on an open visual reasoning and neural network optimization tool OpenVINO.
8. An abnormal node identification apparatus, comprising:
the input module is used for inputting the characteristic data of the test image into a deep learning model to be identified, and the deep learning model to be identified comprises a plurality of nodes; the deep learning model to be identified is a deep learning model for realizing an image analysis task, wherein the image analysis task comprises target classification, target detection or target segmentation;
the monitoring module is used for monitoring the processing time length of a designated node in the plurality of nodes in the process of processing the feature data by the deep learning model to be identified, wherein the processing time length of the designated node is the time length of the designated node for processing the received data;
The determining module is used for determining the designated node as an abnormal node when the processing time length of the designated node is greater than a preset time length threshold; deleting the abnormal node in the deep learning model to be identified when the abnormal node is redundant, so as to obtain a new deep learning model; when the abnormal node is a non-redundant node, replacing the abnormal node in the deep learning model to be identified by using a target node to obtain a new deep learning model; wherein the target node is: and using a node which consumes less time and performs the same function as the abnormal node.
9. The apparatus of claim 8, wherein the monitoring module comprises:
the monitoring sub-module is used for monitoring the input time point when a designated node in the plurality of nodes receives the data to be processed and the output time point when the received data is processed;
and the calculating sub-module is used for calculating the difference value of the output time point minus the input time point as the processing time length of the designated node.
10. The apparatus of claim 8, wherein the monitoring module comprises:
The monitoring sub-module is used for monitoring the time length from the start of inputting the characteristic data into the deep learning model to be identified to the time length from the time when each node in the plurality of nodes receives the data to be processed as the arrival time length;
and the calculation sub-module is used for calculating the arrival time of the next node of the designated node and subtracting the difference value of the arrival time of the designated node to be used as the processing time of the designated node.
11. The apparatus of claim 8, wherein the monitoring module comprises:
the monitoring sub-module is used for monitoring the duration from the start of inputting the characteristic data into the deep learning model to be identified to the completion of processing the received data by each node in the plurality of nodes, and taking the duration as the output duration;
and the calculating sub-module is used for calculating the output time length of the designated node and subtracting the difference value of the output time length of the node before the designated node as the processing time length of the designated node.
12. The apparatus of claim 8, wherein the deep learning model to be identified is a model obtained by optimizing an original deep learning model based on a high-performance neural network reasoning engine TensorRT; or alternatively
The deep learning model to be identified is a model obtained by optimizing an original deep learning model based on an open visual reasoning and neural network optimization tool OpenVINO.
13. The electronic equipment is characterized by comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory are communicated with each other through the communication bus;
a memory for storing a computer program;
a processor for carrying out the method steps of any one of claims 1-7 when executing a program stored on a memory.
14. A computer-readable storage medium, characterized in that the computer-readable storage medium has stored therein a computer program which, when executed by a processor, implements the method steps of any of claims 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911250256.7A CN110991659B (en) | 2019-12-09 | 2019-12-09 | Abnormal node identification method, device, electronic equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911250256.7A CN110991659B (en) | 2019-12-09 | 2019-12-09 | Abnormal node identification method, device, electronic equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110991659A CN110991659A (en) | 2020-04-10 |
| CN110991659B true CN110991659B (en) | 2024-03-08 |
Family
ID=70091350
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911250256.7A Active CN110991659B (en) | 2019-12-09 | 2019-12-09 | Abnormal node identification method, device, electronic equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110991659B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111695445A (en) * | 2020-05-25 | 2020-09-22 | 深圳丽泽智能科技有限公司 | Face recognition method, device, equipment and computer readable storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CA2433929A1 (en) * | 2003-07-16 | 2005-01-16 | George Fierlbeck | Artificial neural network structure optimizer |
| CN104331520A (en) * | 2014-11-28 | 2015-02-04 | 北京奇艺世纪科技有限公司 | Performance optimization method and device of Hadoop cluster and node state recognition method and device |
| CN106650928A (en) * | 2016-10-11 | 2017-05-10 | 广州视源电子科技股份有限公司 | Method and device for optimizing neural network |
| KR101897962B1 (en) * | 2017-11-27 | 2018-10-31 | 한국인터넷진흥원 | Method for performing machine learning for neural network model and apparatus thereof |
| CN109993287A (en) * | 2017-12-29 | 2019-07-09 | 北京中科寒武纪科技有限公司 | Processing with Neural Network method, computer system and storage medium |
| CN110175679A (en) * | 2019-05-29 | 2019-08-27 | 深圳前海微众银行股份有限公司 | A kind of method and device of monitoring model training |
-
2019
- 2019-12-09 CN CN201911250256.7A patent/CN110991659B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CA2433929A1 (en) * | 2003-07-16 | 2005-01-16 | George Fierlbeck | Artificial neural network structure optimizer |
| CN104331520A (en) * | 2014-11-28 | 2015-02-04 | 北京奇艺世纪科技有限公司 | Performance optimization method and device of Hadoop cluster and node state recognition method and device |
| CN106650928A (en) * | 2016-10-11 | 2017-05-10 | 广州视源电子科技股份有限公司 | Method and device for optimizing neural network |
| KR101897962B1 (en) * | 2017-11-27 | 2018-10-31 | 한국인터넷진흥원 | Method for performing machine learning for neural network model and apparatus thereof |
| CN109993287A (en) * | 2017-12-29 | 2019-07-09 | 北京中科寒武纪科技有限公司 | Processing with Neural Network method, computer system and storage medium |
| CN110175679A (en) * | 2019-05-29 | 2019-08-27 | 深圳前海微众银行股份有限公司 | A kind of method and device of monitoring model training |
Non-Patent Citations (1)
| Title |
|---|
| 深度神经网络压缩与加速综述;纪荣嵘等;计算机研究与发展(09);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110991659A (en) | 2020-04-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107122594B (en) | New energy vehicle battery health prediction method and system | |
| Dinkel et al. | Towards duration robust weakly supervised sound event detection | |
| CN109936582B (en) | Method and device for constructing malicious traffic detection model based on PU learning | |
| CN111914644A (en) | Dual-mode cooperation based weak supervision time sequence action positioning method and system | |
| WO2020028762A1 (en) | Neural network orchestration | |
| CN112183758A (en) | Method and device for realizing model training and computer storage medium | |
| CN115769229A (en) | Method and apparatus for training and testing object detection networks by detecting objects on images using attention-deficit hyperactivity disorder | |
| CN115640159A (en) | Micro-service fault diagnosis method and system | |
| CN115083171B (en) | Unmanned aerial vehicle fixed-point inspection method and system based on muck vehicle monitoring | |
| CN112784905A (en) | Data sample expansion method and device and electronic equipment | |
| CN112613617A (en) | Uncertainty estimation method and device based on regression model | |
| US11354796B2 (en) | Image identification and retrieval for component fault analysis | |
| CN110991659B (en) | Abnormal node identification method, device, electronic equipment and storage medium | |
| WO2022141489A1 (en) | Deep learning model reasoning method and apparatus, computer device, and storage medium | |
| JPWO2019189249A1 (en) | Learning devices, learning methods, and programs | |
| CN113821452B (en) | Intelligent test method for dynamically generating test case according to test performance of tested system | |
| CN111814813A (en) | Neural network training and image classification method and device | |
| CN114397306B (en) | Power grid grading ring hypercomplex category defect multi-stage model joint detection method | |
| CN113298822B (en) | Point cloud data selection method and device, equipment and storage medium | |
| CN114861858A (en) | Method, device and equipment for detecting road surface abnormal data and readable storage medium | |
| CN113469204A (en) | Data processing method, device, equipment and computer storage medium | |
| CA3106394C (en) | Selecting unlabeled data objects to be processed | |
| Urselmann et al. | Semi-automatic Pipeline for Large-Scale Dataset Annotation Task: A DMD Application | |
| CN114254763B (en) | Machine learning model repairing method, system, computer equipment and storage medium | |
| CN116168201B (en) | Lane line segmentation method and device without accurate data labeling |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |