CN109634401B - Control method and electronic equipment - Google Patents
Control method and electronic equipment Download PDFInfo
- Publication number
- CN109634401B CN109634401B CN201811639238.3A CN201811639238A CN109634401B CN 109634401 B CN109634401 B CN 109634401B CN 201811639238 A CN201811639238 A CN 201811639238A CN 109634401 B CN109634401 B CN 109634401B
- Authority
- CN
- China
- Prior art keywords
- layer
- energy consumption
- neural network
- weight
- network model
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F1/00—Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
- G06F1/26—Power supply means, e.g. regulation thereof
- G06F1/32—Means for saving power
- G06F1/3203—Power management, i.e. event-based initiation of a power-saving mode
- G06F1/3234—Power saving characterised by the action undertaken
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Image Analysis (AREA)
Abstract
The application provides a control method, which comprises the following steps: acquiring each layer of a neural network model; according to a first calculation rule, calculating the energy consumption of each layer; selecting a first layer with energy consumption meeting a first condition, and pruning a first weight set in the first layer, wherein the first weight set comprises at least one weight; the first condition is: the first energy consumption of the first layer is not lower than the energy consumption of the layer with the lowest energy consumption. In the scheme, partial weights in the layers with high energy consumption are trimmed through the layered structure of the neural network model, so that the energy consumption of the neural network model is reduced, and the calculation accuracy of the neural network model is ensured due to trimming only partial weights in the layers.
Description
Technical Field
The present disclosure relates to the field of electronic devices, and more particularly, to a control method and an electronic device.
Background
CNN (Convolutional Neural Network ) has wide application from pattern-to-speech recognition to system state detection.
However, AI (Artificial Intelligence ) chips used in electronic devices consume large amounts of power for computing and storing CNN models for applications with large amounts of data such as video and audio, and therefore, a method for controlling the power consumption of CNN models is needed.
Disclosure of Invention
In view of the above, the present disclosure provides a control method, which solves the problem of large power consumption of the CNN model in the prior art.
In order to achieve the above object, the present disclosure provides the following technical solutions:
a control method, comprising:
acquiring each layer of a neural network model;
according to a first calculation rule, calculating the energy consumption of each layer;
selecting a first layer with energy consumption meeting a first condition, and pruning a first weight set in the first layer, wherein the first weight set comprises at least one weight;
the first condition is:
the energy consumption of the first layer is higher than the energy consumption of the layer with the lowest energy consumption.
Preferably, the method further includes calculating the energy consumption of each layer according to a first calculation rule, including:
according to the calculation times in any layer, calculating to obtain the calculation energy consumption of the layer;
according to the access times of the internal memory in the layer, calculating to obtain the storage energy consumption of the layer;
and obtaining the energy consumption of the layer according to the calculated energy consumption and the stored energy consumption.
Preferably, in the method, the pruning the first set of weights in the first layer includes:
And deleting the first weight set with the weight value in the first layer smaller than a first threshold value.
Preferably, in the above method, after deleting the first set of weights with the weight value in the first layer smaller than the first threshold, the method further includes:
processing preset input information according to the trimmed first neural network model to obtain a first output result;
according to the first output result analysis, obtaining output evaluation errors and/or output accuracy which do not meet specific conditions, cancelling deletion of a second weight set, wherein the first weight set at least comprises the second weight set;
modifying the weights in the second set of weights from a first value to a second value, the first value being greater than the second value.
Preferably, in the method, the deleting the second weight set is cancelled, including:
selecting a second weight set meeting important conditions from the first weight set, and deleting the second weight set;
maintaining the deletion of other weights in the first set of weights than the second set of weights.
Preferably, in the above method, after modifying the weights in the second weight set from the first value to the second value, the method further includes:
Acquiring a first weight in a second weight set in the first layer, and acquiring a first parameter corresponding to the first weight;
acquiring at least one second parameter associated with the first parameter in the first layer;
and adjusting the weight values of the first parameter and the at least one second parameter based on the correlation between the at least one second parameter and the first parameter, so that the output evaluation error and the output accuracy of the neural network model based on the adjusted weight values meet preset requirements.
Preferably, in the above method, after modifying the weights in the second weight set from the first value to the second value, the method further includes:
acquiring second weights in all second weight sets of the trimmed first neural network model, and acquiring third parameters corresponding to the second weights;
acquiring at least one fourth parameter associated with the third parameter in the trimmed first neural network model;
and adjusting the weight values of the third parameter and the at least one fourth parameter based on the correlation of the at least one fourth parameter and the third parameter so that the output evaluation error and the output accuracy of the neural network model based on the adjusted weight values meet preset requirements.
Preferably, in the method, after trimming the first weight set in the first layer, the method further includes:
judging whether an untrimmed layer exists according to a preset judging rule;
and based on the existence of the untrimmed layer, performing the energy consumption step of calculating each layer according to the first calculation rule in a return loop.
Preferably, the method for determining whether an untrimmed layer exists includes:
according to a first calculation rule, calculating the energy consumption of the trimmed first neural network model, wherein the energy consumption is the sum of the energy consumption of each layer in the trimmed first neural network model;
based on the energy consumption of the pruned first neural network model being below an energy consumption threshold, no untrimmed layers are present;
and based on the energy consumption of the trimmed first neural network model not being lower than an energy consumption threshold, an untrimmed layer exists.
An electronic device, comprising:
a housing;
the processor is used for acquiring each layer of the neural network model; according to a first calculation rule, calculating the energy consumption of each layer; selecting a first layer with energy consumption meeting a first condition, and pruning a first weight set in the first layer, wherein the first weight set comprises at least one weight; the first condition is: the energy consumption of the first layer is higher than the energy consumption of the layer with the lowest energy consumption.
As can be seen from the above technical solution, compared with the prior art, the present disclosure provides a control method, including: acquiring each layer of a neural network model; according to a first calculation rule, calculating the energy consumption of each layer; selecting a first layer with energy consumption meeting a first condition, and pruning a first weight set in the first layer, wherein the first weight set comprises at least one weight; the first condition is: the first energy consumption of the first layer is not lower than the energy consumption of the layer with the lowest energy consumption. In the scheme, partial weights in the layers with high energy consumption are trimmed through the layered structure of the neural network model, so that the energy consumption of the neural network model is reduced, and the calculation accuracy of the neural network model is ensured due to trimming only partial weights in the layers.
Drawings
In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings that are required to be used in the embodiments or the description of the prior art will be briefly described below, and it is apparent that the drawings in the following description are only embodiments of the present disclosure, and other drawings may be obtained according to the provided drawings without inventive effort to those of ordinary skill in the art.
FIG. 1 is a schematic diagram of a convolutional neural network model data processing process in the prior art;
fig. 2 is a schematic architecture diagram of a neural network model applied by the control method and the electronic device provided in the present application;
FIG. 3 is a schematic diagram of an architecture of sensor aggregate environmental scene awareness;
fig. 4 is a schematic diagram of an architecture provided in the present application in an application scenario;
fig. 5 is a flowchart of an embodiment 1 of a control method provided in the present application;
FIG. 6 is a flowchart of an embodiment 2 of a control method provided in the present application;
fig. 7 is a flowchart of an embodiment 3 of a control method provided in the present application;
fig. 8 is a flowchart of an embodiment 4 of a control method provided in the present application;
FIG. 9 is a flowchart of an embodiment 5 of a control method provided in the present application;
FIG. 10 is a flowchart of an embodiment 6 of a control method provided in the present application;
FIG. 11 is a flowchart of an embodiment 7 of a control method provided in the present application;
FIG. 12 is a flowchart of an embodiment 8 of a control method provided in the present application;
fig. 13 is a flowchart of an application scenario of a control method provided in the present application;
fig. 14 is a schematic structural diagram of an embodiment of an electronic device provided in the present application.
Detailed Description
The following description of the technical solutions in the embodiments of the present disclosure will be made clearly and completely with reference to the accompanying drawings in the embodiments of the present disclosure, and it is apparent that the described embodiments are only some embodiments of the present disclosure, not all embodiments. Based on the embodiments in this disclosure, all other embodiments that a person of ordinary skill in the art would obtain without making any inventive effort are within the scope of protection of this disclosure.
First, in the prior art, in applications with large data size, such as audio and video, the processing is performed hierarchically, as shown in fig. 1, which is a schematic diagram of a data processing procedure of a convolutional neural network model in the prior art, the neural network model performs 5 layers (L1-L5) processing on a picture, wherein the picture is input in L0 layers, the pixels are 512×512, the L1 layers are 256×256, the L2 layers are 128×128, the L3 layers are 64×64, the L4 layers are 32×32, and then the L5 layers are processed, and the L6 layers are output (output), so as to finally obtain a processing result, wherein a processing manner between L0-L4 is convolution calculation (convolution), and a processing manner between L4-L6 is full connection (full connection). If the number of layers of the neural network model is reduced, the accuracy of the neural network model is reduced. Therefore, according to the scheme, the number of layers is not reduced, but the partial weights in the layers with high energy consumption are trimmed, so that the energy consumption of the neural network model is reduced, and the calculation accuracy of the neural network model is ensured due to trimming only the partial weights in the layers.
As shown in fig. 2, an architecture schematic diagram of a neural network model applied to the control method and the electronic device provided in the present application includes: a signal layer 201, a logo layer 202, an environment awareness layer 203, and a device application layer 204.
Specifically, the signals collected by the sensors are directly input to the signal layer 201, and a scene recognition algorithm (AI, etc.) is applied to the mark layer 202 and the environment sensing layer 203; and the application directly invokes the device application layer 204 to obtain current context information.
In the architecture, the layers are mutually independent, each layer can be reconstructed, and other layers are not influenced; the reconstruction of each layer is determined by the function attribute of the layer; the aggregation algorithm of each layer can be independently designed and implemented;
the architecture may support a variety of algorithms, such as machine learning, neural networks, hidden Markov machines, etc., which are not limited in this application.
Specifically, the signal layer 201 is directly connected to the sensor for controlling data collection and storage, and the signal layer output is raw data with a time tag, which indicates that the raw data will have a certain built-in time characteristic.
In particular, the signature layer 202 is primarily a layer that extracts signature from individual sensor data, converts the time-series signal to a time-independent special space, including but not limited to the frequency domain, complex frequency domain, and so forth.
In particular, the perceived environment layer 203 uses a flag or a combination of flags to determine an environment scene, the output of which is the probability of a user (device) being in a particular environment.
As an example, the current environment is 90% in a moving car (non-highway); in the perception environment layer, the environment can be deeply customized, and an AI (artificial intelligence) tool such as a neural network is applied to learn the mark, so that the whole environment perception has a time tag, namely, the scene on the time axis is switched.
Specifically, the device application layer 204 develops in the device operating system, and applies the results of the entire sensor convergence scene perception system to provide to the required programs.
As shown in fig. 3, the architecture diagram of sensor aggregate environmental scene perception provided in the present application includes a signal layer 301, a flag layer 302, an environmental perception layer 303 and a device application layer 304, where the signal layer 301 is connected to a sensor 1 and a sensor 2 … …, and data acquired by the sensor is transmitted to the flag layer 302 through a signal interface in a plurality of channels (channel 1 and channel 2 … …); the identification layer performs identification extraction on the obtained data to obtain an identification 1 and an identification 2 …, and then transmits the identifications to the sensing environment layer 303 through a mark interface; the perceived environment layer 303 analyzes according to the obtained identifier, and the obtained analysis result characterizes the probability of the user (device) in a certain environment, and transmits the analysis result to the device application layer through the environment interface, and the device application layer outputs the result.
As shown in fig. 4, a schematic diagram of an architecture provided for the present application in an application scenario, where the architecture includes: a signal layer 401, a logo layer 402, an environment awareness layer 403 and a device application layer 404.
The signal layer 401 obtains the following information from an external device such as a sensor: ambient light fundamental frequency, average brightness, brightness standard deviation, maximum temperature, minimum temperature, average temperature, GPS/position data, date data, etc.; the identification layer 402 performs identification extraction on the information transmitted by the signal layer, performs extraction analysis on spectral features and time domain features according to the fundamental frequency, average brightness and standard deviation of ambient light to obtain an artificial light source, namely, performs temperature time domain analysis according to the highest temperature, the lowest temperature and the average temperature to obtain 24H (hour) temperature, performs analysis on the data according to the GPS/position data, date data and the like to obtain the data which is the ground gas image data at the moment, namely, extracts the data which is the ground gas image data at the moment, gathers the identification extracted by the identification layer, performs analysis on the gathered data and the identification by combining an AI model, and can obtain the results of 95% of indoor probability and 5% of outdoor probability. Ultimately, the results may be output through the device application layer 404.
As shown in fig. 5, a flowchart of an embodiment 1 of a control method provided in the present application is applied to an electronic device, where the electronic device has a neural network model, and the method includes the following steps:
step S501: acquiring each layer of a neural network model;
the neural network model adopts a layered mutually independent architecture, and each layer can be reconstructed without affecting other layers; and the reconstruction of each layer is determined by the functional attribute of the layer, and the aggregation algorithm of each layer can be independently designed and implemented.
The neural network model comprises at least two layers.
Specifically, each layer in the neural network model is determined so as to obtain each layer of the neural network model.
Step S502: according to a first calculation rule, calculating the energy consumption of each layer;
in a specific implementation, in the process of processing input data, each layer of the neural network model needs to execute processing operation, and accordingly, each layer has energy consumption.
The method comprises the steps of presetting a first calculation rule, and sequentially calculating energy consumption of each layer in the neural network model to obtain the energy consumption of each layer.
Step S503: a first layer having energy consumption meeting a first condition is selected, and a first set of weights within the first layer is pruned.
Wherein the first set of weights comprises at least one weight;
the first condition is: the energy consumption of the first layer is higher than the energy consumption of the layer with the lowest energy consumption.
Wherein at least one layer within the neural network model is trimmed in order to reduce energy consumption of the neural network model.
Specifically, at least one layer with higher/highest energy consumption is selected for trimming.
Wherein the higher/highest energy consumption is determined in such a way that the energy consumption of the layer is higher than the layer with the lowest energy consumption.
Specifically, the first weight set in the selected first layer is trimmed, so that the processing operation executed in the data processing process of the layer is reduced, and the energy consumption of the layer is reduced.
In the implementation, the layer with the highest energy consumption may be trimmed first, and after trimming, the energy consumption calculation of each layer is performed again in the neural network model, and then the layer with the highest energy consumption is trimmed, and so on.
In summary, the information processing method provided in this embodiment includes: acquiring each layer of a neural network model; according to a first calculation rule, calculating the energy consumption of each layer; selecting a first layer with energy consumption meeting a first condition, and pruning a first weight set in the first layer, wherein the first weight set comprises at least one weight; the first condition is: the first energy consumption of the first layer is not lower than the energy consumption of the layer with the lowest energy consumption. In the scheme, partial weights in the layers with high energy consumption are trimmed through the layered structure of the neural network model, so that the energy consumption of the neural network model is reduced, and the calculation accuracy of the neural network model is ensured because only partial weights in the layers are trimmed.
As shown in fig. 6, a flowchart of an embodiment 2 of a control method provided in the present application includes the following steps:
step S601: acquiring each layer of a neural network model;
step S601 is identical to step S501 in embodiment 1, and is not described in detail in this embodiment.
Step S602: according to the calculation times in any layer, calculating to obtain the calculation energy consumption of the layer;
the present embodiment is mainly explained with respect to a specific process of calculating energy consumption in a layer.
It should be noted that, in the process of processing data by the neural network model, each layer has energy consumption in the process of calculating and storing the data, and in this embodiment, the energy consumption is calculated for different processing modes respectively.
Wherein, according to the number of times of calculation in any layer, the calculation energy consumption of the layer is calculated.
Specifically, the energy consumption of one calculation is increased every time the calculation processing such as adding, subtracting, multiplying and dividing is performed, wherein each calculation mode can correspond to different energy consumption values, for example, one addition/subtraction calculation is performed for 1 joule (J), and one multiplication/division calculation is performed for 1.1 joule; the same energy consumption values, e.g., 1 joule, may also be corresponded. Currently, the value of the energy consumption value corresponding to the calculation is not limited to this, and may be set according to actual situations.
Generally speaking, the computational energy consumption in a layer is related to the complexity of its computational algorithm, the greater the computational energy consumption, the greater the complexity of its computational algorithm; the smaller the computational power consumption, the lower the complexity of its computational algorithm.
Step S603: according to the access times of the internal memory in the layer, calculating to obtain the storage energy consumption of the layer;
and according to the access times of the internal memory in any layer, calculating to obtain the access energy consumption of the layer.
Specifically, the energy consumption is increased once for every access processing, wherein the storage or reading mode can correspond to different energy consumption values, for example, the storage is 0.5 joule (J), and the reading is 0.6 joule; the same energy consumption values, e.g., 0.5 joules, may also be corresponded. Currently, the value of the energy consumption value corresponding to the calculation is not limited to this, and may be set according to actual situations.
Generally speaking, the storage energy consumption within a layer is related to the number of nodes it needs to store, the greater the storage energy consumption, the greater the number of nodes it needs to store; the smaller the energy consumption, the fewer the number of nodes it requires for its storage.
Step S604: obtaining the energy consumption of the layer according to the calculated energy consumption and the stored energy consumption;
and calculating according to the calculated energy consumption and the stored energy consumption of the layer, so as to obtain the energy consumption of the layer.
Specifically, the calculated energy consumption and the stored energy consumption can be directly summed, and the obtained value is the energy consumption of the layer; when the overall consumption is more prone to be determined by a certain energy, different weighting values may be assigned to the calculated energy consumption and the stored energy consumption, for example, the weighting value of the calculated energy consumption is 0.6 and the weighting value of the stored energy consumption is 0.4, and the calculation rule is that the energy consumption of the layer=calculated energy consumption×0.6+stored energy consumption×0.4.
Step S605: a first layer having energy consumption meeting a first condition is selected, and a first set of weights within the first layer is pruned.
Step S605 corresponds to step 503 in embodiment 1.
In summary, in the information processing method provided in the present embodiment, the calculating the energy consumption of each layer according to the first calculation rule includes: according to the calculation times in any layer, calculating to obtain the calculation energy consumption of the layer; according to the access times of the internal memory in the layer, calculating to obtain the storage energy consumption of the layer; and obtaining the energy consumption of the layer according to the calculated energy consumption and the stored energy consumption. In the scheme, the energy consumption of the layer is calculated by calculating the energy consumption of the layer and the energy consumption of the layer is further calculated, and the calculation process is simple.
As shown in fig. 7, a flowchart of an embodiment 3 of a control method provided in the present application includes the following steps:
step S701: acquiring each layer of a neural network model;
step S702: according to a first calculation rule, calculating the energy consumption of each layer;
steps S701 to 702 correspond to steps S501 to 502 in embodiment 1.
Step S703: and selecting a first layer with energy consumption meeting a first condition, and deleting a first weight set with the weight value in the first layer smaller than a first threshold value.
Wherein the pruning is performed on the weights in the first layer, specifically on part of the weights therein.
Specifically, a weight with a weight smaller than a first threshold value is selected in the first layer and is deleted as a first weight set.
Specifically, the first threshold may be a preset value, and may be an adaptively adjusted value.
In general, to improve the accuracy of data processing of the neural network model, the first threshold value adopts an adaptively adjusted value.
The self-adaptive adjustment process is as follows: presetting a threshold value; deleting the weight with the weight value in the first layer smaller than the preset threshold according to the preset threshold; processing preset input information based on the deleted neural network model to obtain an output result; if the output evaluation error and/or the output accuracy do not meet the specific conditions, the deletion operation of the weight smaller than the preset threshold value is canceled, and the threshold value is reduced to a first value from the preset value; deleting the weight with the weight value smaller than the first value in the first layer, and processing preset input information based on the deleted neural network model to obtain an output result; if the output evaluation error and/or the output accuracy do not meet the specific conditions, canceling the deletion operation of the weight smaller than the first value, reducing the threshold value from the first value to the second value until the threshold value reduces the Nth value (the value of N is a natural number), deleting the weight of which the weight value in the first layer is smaller than the Nth value, and processing preset input information based on the deleted neural network model to obtain an output result; and ending if the output evaluation error and the output accuracy meet the specific conditions.
Generally, the threshold is adjusted in a small direction and the values are random.
In a specific implementation, the weights in the first weight set of the pruning may also have a targeted choice, i.e. with respect to the way the energy is consumed.
Specifically, if the calculation energy consumption is high, the calculation algorithm is possibly complicated, so that the pruning needs to be adjusted for the calculation algorithm, and the parameters/the weights corresponding to the steps of the pruning algorithm are adjusted, so that the steps of the algorithm are reduced, and the calculation energy consumption is reduced. If the storage consumes more energy, the storage needs more nodes, so the pruning needs to prune the weight corresponding to the nodes, thereby reducing the stored nodes and the energy consumption.
In summary, in the information processing method provided in this embodiment, the pruning is performed on the first weight set in the first layer, which specifically includes: and deleting the first weight set with the weight value in the first layer smaller than a first threshold value. In the scheme, the partial weight with smaller weight in the first layer is deleted, and the energy consumption in the data processing process of the layer is reduced on the premise that the accuracy of the data processing of the first layer is ensured.
As shown in fig. 8, a flowchart of an embodiment 4 of a control method provided in the present application includes the following steps:
step S801: acquiring each layer of a neural network model;
step S802: according to a first calculation rule, calculating the energy consumption of each layer;
step S803: selecting a first layer with energy consumption meeting a first condition, and deleting a first weight set with the weight value in the first layer smaller than a first threshold value;
steps S801 to 804 correspond to steps S701 to 704 in embodiment 3.
Step S804: processing preset input information according to the trimmed first neural network model to obtain a first output result;
it should be noted that, the weights in the layers and between the layers in the neural network model may have an association relationship, the more the weights have an association relationship with a certain weight, the heavier the correlation between the weight and other weights, and when the weight is deleted, the accuracy/error of the neural network model will be affected more.
Specifically, when one or even several weights in the first weight set have an association relationship with other weights in the non-first weight set, the calculation/storage related to the other weights is also affected after deleting the first weight set.
For this reason, after deleting the first weight set, a return test needs to be performed on the first neural network model obtained by the pruning.
Specifically, preset input information is input into the first neural network model, so that the first neural network model processes the input information to obtain a first output result.
Step S805: according to the first output result analysis, obtaining an output evaluation error and/or output accuracy which does not meet a specific condition, and canceling deletion of the second weight set;
step S806: modifying the weights in the second set of weights from a first value to a second value.
Wherein the first weight set at least comprises the second weight set;
wherein the first value is greater than the second value.
And analyzing according to the first output result to obtain an output evaluation error and/or output accuracy which does not meet a specific condition, namely, the output evaluation error is large and/or the output accuracy is lower.
Then an adjustment is required to the deleted first set of weights.
Specifically, the partial weights in the second weight set may be modified, that is, after the deletion operation on the second weight set is cancelled, the numerical value of each weight in the second weight set is reduced.
In particular implementations, whether to modify the value of a weight in the first set of weights may be determined based on a correlation between the weight and other weights (weights within a layer or weights of other layers).
In a specific implementation, if the correlation between a certain weight in the first weight set and other layers is larger, the value of the weight can be reduced, and the other weights related to the weight are less influenced.
In summary, the information processing method provided in this embodiment further includes: processing preset input information according to the trimmed first neural network model to obtain a first output result; according to the first output result analysis, obtaining output evaluation errors and/or output accuracy which do not meet specific conditions, cancelling deletion of a second weight set, wherein the first weight set at least comprises the second weight set; modifying the weights in the second set of weights from a first value to a second value, the first value being greater than the second value. In the scheme, when the output evaluation error and/or the output accuracy of the output result of the information processing of the pruned neural network model do not meet the specific conditions, the partial weights in the pruned first weight set are cancelled and deleted, and the numerical value of the cancelled weights is reduced, so that the influence of other weights related to the weights in the second weight set is less, and the output evaluation error and/or the output accuracy of the output result of the information processing of the pruned neural network model is also ensured.
As shown in fig. 9, a flowchart of an embodiment 5 of a control method provided in the present application includes the following steps:
step S901: acquiring each layer of a neural network model;
step S902: according to a first calculation rule, calculating the energy consumption of each layer;
step S903: selecting a first layer with energy consumption meeting a first condition, and deleting a first weight set with the weight value in the first layer smaller than a first threshold value;
wherein steps S901-903 are identical to steps S701-703 in embodiment 3.
Step S904: processing preset input information according to the trimmed first neural network model to obtain a first output result;
step S905: and according to the first output result analysis, the output evaluation error and/or the output accuracy do not meet the specific conditions, selecting a second weight set meeting the important conditions from the first weight set, canceling the deletion of the second weight set, and maintaining the deletion of other weights except the second weight set from the first weight set.
Wherein, the weight meeting the important condition includes that the correlation between the weight and other weights is high, if the weight is deleted, the result of the neural network model is greatly affected.
Then, a second weight set specific procedure is selected:
analyzing the correlation between any weight in the first weight set and other weights, and cancelling the deletion of the weights based on the condition that the correlation meets the condition; otherwise, maintaining the deletion of the weight.
Wherein the correlation analysis of any weight with other weights includes correlation analysis of weights within layers and correlation analysis of weights between layers.
In summary, in the information processing method provided in this embodiment, deletion of the second weight set by the revocation includes: selecting a second weight set meeting important conditions from the first weight set, and deleting the second weight set; maintaining the deletion of other weights in the first set of weights than the second set of weights. In the scheme, the cancel deletion operation is carried out on part of weights in the deleted first weight set, and the output evaluation error and/or the output accuracy of the output result of the pruned neural network model in the information processing are ensured.
As shown in fig. 10, a flowchart of an embodiment 6 of a control method provided in the present application includes the following steps:
Step S1001: acquiring each layer of a neural network model;
step S1002: according to a first calculation rule, calculating the energy consumption of each layer;
step S1003: selecting a first layer with energy consumption meeting a first condition, and deleting a first weight set with the weight value in the first layer smaller than a first threshold value;
step S1004: processing preset input information according to the trimmed first neural network model to obtain a first output result;
step S1005: according to the first output result analysis, obtaining an output evaluation error and/or output accuracy which does not meet a specific condition, and canceling deletion of the second weight set;
step S1006: modifying the weights in the second set of weights from a first value to a second value;
steps S1001 to 1006 correspond to steps S801 to 806 in embodiment 4.
Step S1007: acquiring a first weight in a second weight set in the first layer, and acquiring a first parameter corresponding to the first weight;
after the weights in the second set of weights in the layer are modified, in order to improve the overall accuracy of the neural network model and reduce the output evaluation error, the neural network model is further locally and finely adjusted, and the weights in the layer with relevance are adjusted.
Specifically, one weight in a second weight set in the layer is firstly obtained and defined as a first weight, and a first parameter corresponding to the first weight is obtained.
Wherein the first parameter is a parameter of a generally linear or non-linear polynomial, the plurality of parameters calculate a weight, and the first parameter is a first parameter of the polynomial.
Step S1008: acquiring at least one second parameter associated with the first parameter in the first layer;
wherein, there is at least one weight in the first layer having a correlation with the first weight, and the at least one weight also corresponds to a parameter, i.e. a second parameter.
And correspondingly, acquiring a second parameter associated with the first parameter in the layer, wherein the second parameter is at least one.
Specifically, after the architecture of the neural network model is determined, the relevance between the parameters can be determined in the process of training the neural network model.
Step S1009: and adjusting the weight values of the first parameter and the at least one second parameter based on the correlation of the at least one second parameter and the first parameter so that the output evaluation error and the output accuracy of the neural network model based on the adjusted weight values meet preset requirements.
Specifically, based on the association relation between the second parameter and the first parameter, the weight values of the first parameter and the second parameter are respectively adjusted, after each adjustment, the neural network model is detected to obtain an output result, and an output evaluation error and output accuracy of the output result are determined to meet preset requirements.
In practice, the accuracy of the preset request is higher than the specific condition in step S1005, and/or the output evaluation error of the preset request is smaller than the specific condition in step S1005.
In summary, the information processing method provided in this embodiment further includes: acquiring a first weight in a second weight set in the first layer, and acquiring a first parameter corresponding to the first weight; acquiring at least one second parameter associated with the first parameter in the first layer; and adjusting the weight values of the first parameter and the at least one second parameter based on the correlation between the at least one second parameter and the first parameter, so that the output evaluation error and the output accuracy of the neural network model based on the adjusted weight values meet preset requirements. In the scheme, parameters of the weight with correlation in the first layer are adjusted, so that the neural network model is further optimized, and the output accuracy of the neural network model is improved.
As shown in fig. 11, a flowchart of a control method embodiment 7 provided in the present application includes the following steps:
step S1101: acquiring each layer of a neural network model;
step S1102: according to a first calculation rule, calculating the energy consumption of each layer;
step S1103: selecting a first layer with energy consumption meeting a first condition, and deleting a first weight set with the weight value in the first layer smaller than a first threshold value;
step S1104: processing preset input information according to the trimmed first neural network model to obtain a first output result;
step S1105: according to the first output result analysis, obtaining an output evaluation error and/or output accuracy which does not meet a specific condition, and canceling deletion of the second weight set;
step S1106: modifying the weights in the second set of weights from a first value to a second value;
steps S1101 to 1106 correspond to steps S801 to 806 in embodiment 4.
Step S1107: acquiring second weights in all second weight sets of the trimmed first neural network model, and acquiring third parameters corresponding to the second weights;
after modifying the weights in the second set of weights in the layer, the neural network model may be globally fine-tuned to improve the overall accuracy of the neural network model and reduce the output evaluation error, by tuning the weights having relevance between different layers and within the layer.
Specifically, one weight in a second weight set in the body is firstly obtained and defined as a second weight, and a third parameter corresponding to the second weight is obtained.
Wherein the third weight is generally a parameter of a linear or nonlinear polynomial, a plurality of parameters calculate a weight, and the third parameter is a first parameter of the polynomial.
In specific implementation, the step can be performed after no untrimmed layer exists in the neural network model, so that global adjustment of the neural network model can be realized.
Step S1108: acquiring at least one fourth parameter associated with the third parameter in the trimmed first neural network model;
at least one weight (which may or may not be the same layer as the second weight) associated with the second weight in the neural network model also corresponds to a parameter, i.e., a fourth parameter.
And correspondingly, acquiring a fourth parameter which is associated with the third parameter in the layer, wherein the fourth parameter is at least one.
Specifically, after the architecture of the neural network model is determined, the relevance between the parameters can be determined in the process of training the neural network model.
Step S1109: and adjusting the weight values of the third parameter and the at least one fourth parameter based on the correlation of the at least one fourth parameter and the third parameter so that the output evaluation error and the output accuracy of the neural network model based on the adjusted weight values meet preset requirements.
Specifically, based on the association relation between the third parameter and the fourth parameter, the weight values of the third parameter and the fourth parameter are respectively adjusted, after each adjustment, the neural network model is detected to obtain an output result, and an output evaluation error and output accuracy of the output result are determined to meet preset requirements.
In practice, the accuracy of the preset request is higher than the specific condition in step S1105, and/or the output evaluation error of the preset request is smaller than the specific condition in step S1105.
In summary, the information processing method provided in this embodiment further includes: acquiring second weights in all second weight sets of the trimmed first neural network model, and acquiring third parameters corresponding to the second weights; acquiring at least one fourth parameter associated with the third parameter in the trimmed first neural network model; and adjusting the weight values of the third parameter and the at least one fourth parameter based on the correlation of the at least one fourth parameter and the third parameter so that the output evaluation error and the output accuracy of the neural network model based on the adjusted weight values meet preset requirements. In the scheme, parameters of the weight with correlation in the trimmed neural network model are adjusted, so that the neural network model is further optimized, and the output accuracy of the neural network model is improved.
As shown in fig. 12, a flowchart of an embodiment 8 of a control method provided in the present application includes the following steps:
step S1201: acquiring each layer of a neural network model;
step S1202: according to a first calculation rule, calculating the energy consumption of each layer;
step S1203: selecting a first layer with energy consumption meeting a first condition, and pruning a first weight set in the first layer;
steps S1201-1203 are identical to steps S501-503 in embodiment 1, and are not described in detail in this embodiment.
Step S1204: judging whether an untrimmed layer exists according to a preset judging rule;
after trimming the weight set in the primary layer, it is further required to determine whether the neural network model has an untrimmed layer.
Specifically, the step of determining whether an untrimmed layer exists includes:
according to a first calculation rule, calculating the energy consumption of the trimmed first neural network model, wherein the energy consumption is the sum of the energy consumption of each layer in the trimmed first neural network model;
based on the energy consumption of the pruned first neural network model being below an energy consumption threshold, no untrimmed layers are present;
And based on the energy consumption of the trimmed first neural network model not being lower than an energy consumption threshold, an untrimmed layer exists.
Based on the presence of the untrimmed layer, the loop returns to step S1202 of calculating the energy consumption of the layers according to the first calculation rule.
Step S1205: ending based on the absence of untrimmed layers.
Specifically, if there is an untrimmed layer, the energy consumption of each layer of the trimmed neural network model is continuously calculated, a layer meeting the first energy consumption condition is determined, and step S1203 is circularly executed until there is no untrimmed layer, and the process is ended.
In summary, the information processing method provided in this embodiment further includes: judging whether an untrimmed layer exists according to a preset judging rule; and based on the existence of the untrimmed layer, performing the energy consumption step of calculating each layer according to the first calculation rule in a return loop. By adopting the method, whether the untrimmed layers exist or not is judged, so that the energy consumption of each layer in the finally obtained neural network model meets the conditions, the overall energy consumption also meets the conditions, and the output accuracy and/or the output evaluation error meet the conditions.
As shown in fig. 13, an application scenario flowchart of a control method provided in the present application includes the following steps:
Step S1301: inputting a model;
the input model is a neural network model, and the neural network model comprises a plurality of layers.
Step S1302: determining a layer trimming order by energy consumption;
in this step, the order of pruning the layers is determined by calculating the energy consumption of each layer in the neural network model.
Typically, the layer with the greatest energy consumption is trimmed first.
Step S1303: removing weights in the layer that are less than a certain threshold;
in this step, weights smaller than a certain threshold in the layer determined to be pruned are removed, so that the layer becomes sparse, and the energy consumption of the neural network model is reduced.
Step S1304: restoring certain weights to reduce output errors;
in the last step, if the accuracy and bias of the neural network model is affected to a worse direction after the weights are removed, some weights (which may be part of the removed weights) need to be recovered to reduce the output error.
Specifically, the recovered weight may be subjected to a numerical reduction process to further reduce the output error.
Step S1305: locally fine-adjusting the weight;
specifically, through the association relation between the recovered weight and other weights in the layer, the values of the recovered weight and the corresponding parameters of the associated weight are modified, so that the aims of reducing output errors and improving output accuracy are fulfilled.
Step S1306: judging whether other layers which are not trimmed exist or not;
determining by determining whether the overall energy consumption of the trimmed neural network model is smaller than a threshold, if so, determining that no other layers without trimming exist, and executing the next step S1307; otherwise, there are other layers that are not trimmed, returning to execute step S1303 to start trimming the next layer and;
step S1307: globally fine-adjusting the weights;
specifically, through the association relation between the restored weight in each layer and other weights (the weights in the same layer or the weights between layers of different layers) in the whole neural network model, the values of the restored weight and the corresponding parameters of the associated weight are modified, so that the aims of reducing output errors and improving output accuracy are fulfilled.
Step S1308: judging whether the accuracy meets the requirement or not;
if so, the next step S1309 is executed; if not, go back to step S1302 to begin the next iteration.
Specifically, the output accuracy of the adjusted neural network model is judged, if the accuracy requirement is met, the trimming adjustment is judged to be successful, the neural network model has high accuracy and low energy consumption, and the processing can be ended; otherwise, even if the neural network model achieves the purpose of low energy consumption, the initial setting purpose of the neural network model is not achieved due to low accuracy, the neural network model still needs to be processed, and the next iteration is started until the neural network model meets the accuracy requirement.
Step S1309: and outputting the model.
Corresponding to the embodiment of the control method provided by the application, the application also provides an embodiment of the electronic equipment applying the control method.
Fig. 14 is a schematic structural diagram of an embodiment of an electronic device provided in the present application, where the electronic device has a neural network model, and the electronic device includes the following structures: a body 1401 and a processor 1402;
wherein the processor is arranged in the body;
the processor is used for acquiring each layer of the neural network model; according to a first calculation rule, calculating the energy consumption of each layer; selecting a first layer with energy consumption meeting a first condition, and pruning a first weight set in the first layer, wherein the first weight set comprises at least one weight; the first condition is: the energy consumption of the first layer is higher than the energy consumption of the layer with the lowest energy consumption.
Preferably, the processor calculates the energy consumption of each layer according to a first calculation rule, including:
according to the calculation times in any layer, calculating to obtain the calculation energy consumption of the layer;
according to the access times of the internal memory in the layer, calculating to obtain the storage energy consumption of the layer;
And obtaining the energy consumption of the layer according to the calculated energy consumption and the stored energy consumption.
Preferably, the processor prunes the first set of weights in the first layer, including:
and deleting the first weight set with the weight value in the first layer smaller than a first threshold value.
Preferably, after deleting the first set of weights having the weight value in the first layer less than the first threshold, the processor further includes:
processing preset input information according to the trimmed first neural network model to obtain a first output result;
according to the first output result analysis, obtaining output evaluation errors and/or output accuracy which do not meet specific conditions, cancelling deletion of a second weight set, wherein the first weight set at least comprises the second weight set;
modifying the weights in the second set of weights from a first value to a second value, the first value being greater than the second value.
Preferably, the processor cancels the deletion of the second set of weights, including:
selecting a second weight set meeting important conditions from the first weight set, and deleting the second weight set;
maintaining the deletion of other weights in the first set of weights than the second set of weights.
Preferably, after modifying the weights in the second set of weights from the first value to the second value, the processor further includes:
acquiring a first weight in a second weight set in the first layer, and acquiring a first parameter corresponding to the first weight;
acquiring at least one second parameter associated with the first parameter in the first layer;
and adjusting the weight values of the first parameter and the at least one second parameter based on the correlation between the at least one second parameter and the first parameter, so that the output evaluation error and the output accuracy of the neural network model based on the adjusted weight values meet preset requirements.
Preferably, after modifying the weights in the second set of weights from the first value to the second value, the processor further includes:
acquiring second weights in all second weight sets of the trimmed first neural network model, and acquiring third parameters corresponding to the second weights;
acquiring at least one fourth parameter associated with the third parameter in the trimmed first neural network model;
and adjusting the weight values of the third parameter and the at least one fourth parameter based on the correlation of the at least one fourth parameter and the third parameter so that the output evaluation error and the output accuracy of the neural network model based on the adjusted weight values meet preset requirements.
Preferably, after pruning the first set of weights in the first layer, the processor further includes:
judging whether an untrimmed layer exists according to a preset judging rule;
and based on the existence of the untrimmed layer, performing the energy consumption step of calculating each layer according to the first calculation rule in a return loop.
Preferably, the processor determines whether there is an untrimmed layer, comprising:
according to a first calculation rule, calculating the energy consumption of the trimmed first neural network model, wherein the energy consumption is the sum of the energy consumption of each layer in the trimmed first neural network model;
based on the energy consumption of the pruned first neural network model being below an energy consumption threshold, no untrimmed layers are present;
and based on the energy consumption of the trimmed first neural network model not being lower than an energy consumption threshold, an untrimmed layer exists.
In summary, the present embodiment provides an electronic device, by using a hierarchical structure of a neural network model, to prune part of weights in a layer with high energy consumption, so as to reduce the energy consumption of the neural network model, and because only part of weights in the layer are pruned, the calculation accuracy of the neural network model is ensured.
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. The device provided in the embodiment corresponds to the method provided in the embodiment, so that the description is simpler, and the relevant points refer to the description of the method.
The previous description of the provided embodiments is provided to enable any person skilled in the art to make or use the present disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the disclosure. Thus, the present disclosure is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features provided herein.
Claims (9)
1. A control method applied to an electronic device having a neural network model, comprising:
acquiring each layer of a neural network model; the neural network acquires physical data input by a sensor, and extracts characteristics of the data input by the sensor so as to determine an environment scene;
According to a first calculation rule, calculating the energy consumption of each layer;
selecting a first layer with energy consumption meeting a first condition, pruning a first weight set in the first layer to reduce the power consumption of the neural network model, wherein the first weight set comprises at least one weight;
the first condition is:
the energy consumption of the first layer is higher than the energy consumption of the layer with the lowest energy consumption;
the calculating the energy consumption of each layer according to the first calculation rule comprises the following steps:
according to the calculation times in any layer, calculating to obtain the calculation energy consumption of the layer;
according to the access times of the internal memory in the layer, calculating to obtain the storage energy consumption of the layer;
and obtaining the energy consumption of the layer according to the calculated energy consumption and the stored energy consumption.
2. The method of claim 1, the pruning of the first set of weights within the first layer comprising:
and deleting the first weight set with the weight value in the first layer smaller than a first threshold value.
3. The method of claim 2, after the deleting the first set of weights having the first intra-layer weight value less than a first threshold, further comprising:
Processing preset input information according to the trimmed first neural network model to obtain a first output result;
according to the first output result analysis, obtaining output evaluation errors and/or output accuracy which do not meet specific conditions, cancelling deletion of a second weight set, wherein the first weight set at least comprises the second weight set;
modifying the weights in the second set of weights from a first value to a second value, the first value being greater than the second value.
4. A method according to claim 3, the cancelling the deletion of the second set of weights comprising:
selecting a second weight set meeting important conditions from the first weight set, and deleting the second weight set;
maintaining the deletion of other weights in the first set of weights than the second set of weights.
5. The method of claim 3, further comprising, after modifying the weights in the second set of weights from the first value to the second value:
acquiring a first weight in a second weight set in the first layer, and acquiring a first parameter corresponding to the first weight;
acquiring at least one second parameter associated with the first parameter in the first layer;
And adjusting the weight values of the first parameter and the at least one second parameter based on the correlation between the at least one second parameter and the first parameter, so that the output evaluation error and the output accuracy of the neural network model based on the adjusted weight values meet preset requirements.
6. The method of claim 3, further comprising, after modifying the weights in the second set of weights from the first value to the second value:
acquiring second weights in all second weight sets of the trimmed first neural network model, and acquiring third parameters corresponding to the second weights;
acquiring at least one fourth parameter associated with the third parameter in the trimmed first neural network model;
and adjusting the weight values of the third parameter and the at least one fourth parameter based on the correlation of the at least one fourth parameter and the third parameter so that the output evaluation error and the output accuracy of the neural network model based on the adjusted weight values meet preset requirements.
7. The method of claim 1, the pruning of the first set of weights within the first layer further comprising:
judging whether an untrimmed layer exists according to a preset judging rule;
And based on the existence of the untrimmed layer, performing the energy consumption step of calculating each layer according to the first calculation rule in a return loop.
8. The method of claim 7, determining whether an untrimmed layer is present, comprising:
according to a first calculation rule, calculating the energy consumption of the trimmed first neural network model, wherein the energy consumption is the sum of the energy consumption of each layer in the trimmed first neural network model;
based on the energy consumption of the pruned first neural network model being below an energy consumption threshold, no untrimmed layers are present;
and based on the energy consumption of the trimmed first neural network model not being lower than an energy consumption threshold, an untrimmed layer exists.
9. An electronic device having a neural network model, comprising:
a housing;
the processor is used for acquiring each layer of the neural network model; the neural network acquires physical data input by a sensor, and extracts characteristics of the data input by the sensor so as to determine an environment scene; according to a first calculation rule, calculating the energy consumption of each layer; selecting a first layer with energy consumption meeting a first condition, pruning a first weight set in the first layer to reduce the power consumption of the neural network model, wherein the first weight set comprises at least one weight; the first condition is: the energy consumption of the first layer is higher than the energy consumption of the layer with the lowest energy consumption;
The calculating the energy consumption of each layer according to the first calculation rule comprises the following steps:
according to the calculation times in any layer, calculating to obtain the calculation energy consumption of the layer;
according to the access times of the internal memory in the layer, calculating to obtain the storage energy consumption of the layer;
and obtaining the energy consumption of the layer according to the calculated energy consumption and the stored energy consumption.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811639238.3A CN109634401B (en) | 2018-12-29 | 2018-12-29 | Control method and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811639238.3A CN109634401B (en) | 2018-12-29 | 2018-12-29 | Control method and electronic equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109634401A CN109634401A (en) | 2019-04-16 |
| CN109634401B true CN109634401B (en) | 2023-05-02 |
Family
ID=66055120
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811639238.3A Active CN109634401B (en) | 2018-12-29 | 2018-12-29 | Control method and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109634401B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106779075A (en) * | 2017-02-16 | 2017-05-31 | 南京大学 | The improved neutral net of pruning method is used in a kind of computer |
| CN107368885A (en) * | 2017-07-13 | 2017-11-21 | 北京智芯原动科技有限公司 | Network model compression method and device based on more granularity beta prunings |
| CN108229681A (en) * | 2017-12-28 | 2018-06-29 | 郑州云海信息技术有限公司 | A kind of neural network model compression method, system, device and readable storage medium storing program for executing |
| CN109063835A (en) * | 2018-07-11 | 2018-12-21 | 中国科学技术大学 | The compression set and method of neural network |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP4546157B2 (en) * | 2004-06-03 | 2010-09-15 | キヤノン株式会社 | Information processing method, information processing apparatus, and imaging apparatus |
| US11315018B2 (en) * | 2016-10-21 | 2022-04-26 | Nvidia Corporation | Systems and methods for pruning neural networks for resource efficient inference |
-
2018
- 2018-12-29 CN CN201811639238.3A patent/CN109634401B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106779075A (en) * | 2017-02-16 | 2017-05-31 | 南京大学 | The improved neutral net of pruning method is used in a kind of computer |
| CN107368885A (en) * | 2017-07-13 | 2017-11-21 | 北京智芯原动科技有限公司 | Network model compression method and device based on more granularity beta prunings |
| CN108229681A (en) * | 2017-12-28 | 2018-06-29 | 郑州云海信息技术有限公司 | A kind of neural network model compression method, system, device and readable storage medium storing program for executing |
| CN109063835A (en) * | 2018-07-11 | 2018-12-21 | 中国科学技术大学 | The compression set and method of neural network |
Non-Patent Citations (3)
| Title |
|---|
| PRUNING CONVOLUTIONAL NEURAL NETWORKS FOR RESOURCE EFFICIENT INFERENCE;Pavlo Molchanov et al.;《arXiv》;20170608;全文 * |
| 基于显著性分析的神经网络混合修剪算法;蒲兴成等;《智能***学报》;20141231;第09卷(第06期);全文 * |
| 面向嵌入式应用的深度神经网络压缩方法研究;段秉环等;《航空计算技术》;20180925(第05期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109634401A (en) | 2019-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11468262B2 (en) | Deep network embedding with adversarial regularization | |
| CN110837811B (en) | Method, device and equipment for generating semantic segmentation network structure and storage medium | |
| CN115129848B (en) | Method, device, equipment and medium for processing visual question-answering task | |
| KR20180055708A (en) | Device and method for image processing | |
| CN111539941B (en) | Parkinson's disease leg flexibility task evaluation method and system, storage medium and terminal | |
| KR20130000160A (en) | User adaptive augmented reality mobile device and server and method thereof | |
| EP3671441B1 (en) | Application management method and apparatus, storage medium, and electronic device | |
| KR20190056940A (en) | Method and device for learning multimodal data | |
| CN112861518B (en) | Text error correction method and device, storage medium and electronic device | |
| JP2020087103A (en) | Learning method, computer program, classifier, and generator | |
| CN116822382B (en) | Sea surface temperature prediction method and network based on space-time multiple characteristic diagram convolution | |
| CN109740057A (en) | A kind of strength neural network and information recommendation method of knowledge based extraction | |
| US9436912B1 (en) | Symmetric schema instantiation method for use in a case-based reasoning system | |
| JP2020052520A (en) | Determination device, determination method, and program | |
| KR101646926B1 (en) | Method and system of deep concept hioerarchy for reconstruction in multi-modality data | |
| CN114974292A (en) | Audio enhancement method and device, electronic equipment and readable storage medium | |
| CN112084864A (en) | Model optimization method and device, electronic equipment and storage medium | |
| EP3738305B1 (en) | Electronic device and control method thereof | |
| CN109634401B (en) | Control method and electronic equipment | |
| JP2024035192A (en) | System and method for universal purification of input perturbation with denoised diffusion model | |
| Li et al. | Towards communication-efficient digital twin via AI-powered transmission and reconstruction | |
| CN115168722A (en) | Content interaction prediction method and related equipment | |
| CN114358102A (en) | Data classification method, device, equipment and storage medium | |
| CN117975484B (en) | Training method of change detection model, change detection method, device and equipment | |
| CN117852624B (en) | Training method, prediction method, device and equipment of time sequence signal prediction model |
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 |