CN114154559A - Image recognition model training method and device - Google Patents

Image recognition model training method and device Download PDF

Info

Publication number
CN114154559A
CN114154559A CN202111338332.7A CN202111338332A CN114154559A CN 114154559 A CN114154559 A CN 114154559A CN 202111338332 A CN202111338332 A CN 202111338332A CN 114154559 A CN114154559 A CN 114154559A
Authority
CN
China
Prior art keywords
image recognition
recognition model
network
feature
training
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111338332.7A
Other languages
Chinese (zh)
Inventor
钟艺豪
蔡锐涛
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Youmi Technology Co ltd
Original Assignee
Youmi Technology Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Youmi Technology Co ltd filed Critical Youmi Technology Co ltd
Priority to CN202111338332.7A priority Critical patent/CN114154559A/en
Publication of CN114154559A publication Critical patent/CN114154559A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic 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

Abstract

The invention discloses an image recognition model training method and device, wherein the method comprises the following steps: determining a trained first image recognition model; the first image recognition model comprises a trained first feature coding network; determining a second image recognition model to be trained; the model parameters of the second image recognition model are less than those of the first image recognition model; determining the network parameters of the second feature coding network in the second image recognition model as the network parameters of the first feature coding network; and determining a loss function as the difference between the characteristic output of the first characteristic coding network and the characteristic output of the second characteristic coding network, and performing joint training on the first image recognition model and the second image recognition model to obtain a trained second image recognition model. Therefore, the second image recognition model obtained through the scheme training of the invention can keep smaller scale and achieve better recognition effect.

Description

Image recognition model training method and device
Technical Field
The invention relates to the technical field of image processing, in particular to an image recognition model training method and device.
Background
With the development of image recognition technology, deep learning technology is gradually introduced into this field to improve the accuracy of image recognition, so that the efficiency of character recognition in an image, for example, is improved, but the cost of improving efficiency and accuracy is often high. Therefore, the image recognition model with higher accuracy often requires higher cost, such as consuming higher computing resources and occupying higher memory, so that it is difficult to deploy into the mobile terminal device to realize local real-time recognition.
In order to solve the problem of high cost of a high-precision large-volume model, a model training scheme in the prior art often has to make a decision between precision and cost, so that the precision and the cost are difficult to be effectively considered, a defect exists, and a solution is urgently needed.
Disclosure of Invention
The technical problem to be solved by the present invention is to provide a method and a device for determining training of an image recognition model, which can improve the speed of model training by using a trained network parameter, reduce the training cost, and achieve a better recognition effect while maintaining a smaller scale for a second image recognition model obtained by training.
In order to solve the above technical problem, a first aspect of the present invention discloses an image recognition model training method, including:
determining a trained first image recognition model; the first image recognition model comprises a trained first feature coding network;
determining a second image recognition model to be trained; the model parameters of the second image recognition model are less than the model parameters of the first image recognition model;
determining a network parameter of a second feature coding network in the second image recognition model as a network parameter of the first feature coding network;
and determining a loss function as the difference between the characteristic output of the first characteristic coding network and the characteristic output of the second characteristic coding network, and performing joint training on the first image recognition model and the second image recognition model to obtain the trained second image recognition model.
As an optional implementation manner, in the first aspect of the present invention, the jointly training the first image recognition model and the second image recognition model to obtain the trained second image recognition model includes:
inputting a training data set to the first image recognition model and the second image recognition model simultaneously for training;
fixing all model parameters of the first image recognition model and parameters of the second feature coding network of the second image recognition model in the training to be kept unchanged;
and in the training, optimizing other model parameters in the second image recognition model except the parameters of the second feature coding network until the loss function is converged to obtain the trained second image recognition model.
As an optional implementation manner, in the first aspect of the present invention, the number of network layers of the second signature coding network is the same as the number of network layers of the first signature coding network; and/or the first feature encoding network is a BilSTM network; and/or the second feature encoding network is a BilSTM network.
As an alternative implementation, in the first aspect of the present invention, the model parameters of the second image recognition model are less than those of the first image recognition model and more than 1/10 of the model parameters of the first image recognition model;
and/or the presence of a gas in the gas,
the first image recognition model further comprises a first feature extraction network; the second image recognition model further comprises a second feature extraction network; the network parameters of the second feature extraction network are less than the network parameters of the first feature extraction network and more than 1/10 of the network parameters of the first feature extraction network.
As an optional implementation manner, in the first aspect of the present invention, the determining the loss function as a difference between the feature output of the first feature encoding network and the feature output of the second feature encoding network includes:
determining a vector distance between the feature output of the first feature encoding network and the feature output of the second feature encoding network as a loss function;
and/or the presence of a gas in the gas,
determining a vector distance between a feature output of the first feature encoding network and a feature output of the second feature encoding network;
determining the identification loss corresponding to the second image identification model; the recognition loss is used for measuring the difference between the output recognition result of the second image recognition model and the label of the training data;
and determining a loss function according to the vector distance and the identification loss.
As an optional implementation manner, in the first aspect of the present invention, the determining a loss function according to the vector distance and the identification loss includes:
determining a product between the identification loss and a preset weight;
determining the sum of the vector distance and the product as a loss function.
As an alternative embodiment, in the first aspect of the present invention, the vector distance includes at least one of an L1 distance, an L2 distance, a cosine distance, and a KL divergence; and/or, the identified loss is a CTC loss; and/or the preset weight is 0.1.
The second aspect of the present invention discloses an image recognition model training apparatus, which includes:
the first determining module is used for determining the trained first image recognition model; the first image recognition model comprises a trained first feature coding network;
the second determining module is used for determining a second image recognition model to be trained; the model parameters of the second image recognition model are less than the model parameters of the first image recognition model;
the parameter sharing module is used for determining the network parameters of the second feature coding network in the second image recognition model as the network parameters of the first feature coding network;
a loss determination module to determine a loss function as a difference between the feature output of the first feature encoding network and the feature output of the second feature encoding network;
and the model training module is used for carrying out joint training on the first image recognition model and the second image recognition model to obtain the trained second image recognition model.
As an alternative embodiment, in the second aspect of the present invention, the model training module includes:
a training unit, configured to input a training data set to the first image recognition model and the second image recognition model at the same time for training;
a parameter fixing unit, configured to fix all model parameters of the first image recognition model and parameters of the second feature coding network of the second image recognition model in the training so as to remain unchanged;
and the parameter optimization unit is used for optimizing other model parameters except the parameters of the second feature coding network in the second image recognition model in the training process until the loss function is converged to obtain the trained second image recognition model.
As an optional implementation manner, in the second aspect of the present invention, the number of network layers of the second signature coding network is the same as the number of network layers of the first signature coding network; and/or the first feature encoding network is a BilSTM network; and/or the second feature encoding network is a BilSTM network.
As an alternative embodiment, in the second aspect of the present invention, the model parameters of the second image recognition model are less than those of the first image recognition model and more than 1/10 of the model parameters of the first image recognition model;
and/or the presence of a gas in the gas,
the first image recognition model further comprises a first feature extraction network; the second image recognition model further comprises a second feature extraction network; the network parameters of the second feature extraction network are less than the network parameters of the first feature extraction network and more than 1/10 of the network parameters of the first feature extraction network.
As an optional implementation manner, in the second aspect of the present invention, a specific manner in which the loss determining module determines the loss function as a difference between the feature output of the first feature encoding network and the feature output of the second feature encoding network includes:
determining a vector distance between the feature output of the first feature encoding network and the feature output of the second feature encoding network as a loss function;
and/or the presence of a gas in the gas,
determining a vector distance between a feature output of the first feature encoding network and a feature output of the second feature encoding network;
determining the identification loss corresponding to the second image identification model; the recognition loss is used for measuring the difference between the output recognition result of the second image recognition model and the label of the training data;
and determining a loss function according to the vector distance and the identification loss.
As an optional implementation manner, in the second aspect of the present invention, the specific manner in which the loss determining module determines the loss function according to the vector distance and the identification loss includes:
determining a product between the identification loss and a preset weight;
determining the sum of the vector distance and the product as a loss function.
As an alternative embodiment, in the second aspect of the present invention, the vector distance includes at least one of an L1 distance, an L2 distance, a cosine distance, and a KL divergence; and/or, the identified loss is a CTC loss; and/or the preset weight is 0.1.
The third aspect of the present invention discloses another image recognition model training apparatus, including:
a memory storing executable program code;
a processor coupled with the memory;
the processor calls the executable program code stored in the memory to execute part or all of the steps of the image recognition model training method disclosed by the first aspect of the embodiment of the invention.
A fourth aspect of the embodiments of the present invention discloses a computer storage medium, where the computer storage medium stores computer instructions, and the computer instructions, when called, are used to execute part or all of the steps in the image recognition model training method disclosed in the first aspect of the embodiments of the present invention.
Compared with the prior art, the embodiment of the invention has the following beneficial effects:
in the embodiment of the invention, a trained first image recognition model is determined; the first image recognition model comprises a trained first feature coding network; determining a second image recognition model to be trained; the model parameters of the second image recognition model are less than the model parameters of the first image recognition model; determining a network parameter of a second feature coding network in the second image recognition model as a network parameter of the first feature coding network; and determining a loss function as the difference between the characteristic output of the first characteristic coding network and the characteristic output of the second characteristic coding network, and performing joint training on the first image recognition model and the second image recognition model to obtain the trained second image recognition model. Therefore, the invention can enable the second image recognition model with smaller scale to share the feature coding network parameters in the trained first image recognition model, and perform combined training on the two image models to enable the recognition effect of the second image recognition model to approach the first image recognition model, thereby improving the model training speed by using the trained network parameters, reducing the training cost, and achieving better recognition effect while keeping smaller scale of the trained second image recognition model.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 is a schematic flow chart of a training method for an image recognition model according to an embodiment of the present invention;
FIG. 2 is a schematic flow chart illustrating another image recognition model training method disclosed in the embodiments of the present invention;
FIG. 3 is a schematic structural diagram of an image recognition model training apparatus according to an embodiment of the present invention;
FIG. 4 is a schematic structural diagram of another image recognition model training apparatus disclosed in the embodiment of the present invention;
FIG. 5 is a schematic structural diagram of another image recognition model training apparatus according to an embodiment of the present invention.
Detailed Description
In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The terms "first," "second," and the like in the description and claims of the present invention and in the above-described drawings are used for distinguishing between different objects and not for describing a particular order. Furthermore, the terms "include" and "have," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, article, or article that comprises a list of steps or elements is not limited to only those steps or elements listed, but may alternatively include other steps or elements not listed, or inherent to such process, method, article, or article.
Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the invention. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is explicitly and implicitly understood by one skilled in the art that the embodiments described herein can be combined with other embodiments.
The invention discloses an image recognition model training method and device, which can enable a second image recognition model with smaller scale to share the characteristic coding network parameters in a trained first image recognition model, and perform combined training on two image models to enable the recognition effect of the second image recognition model to approach the first image recognition model, so that the model training speed can be improved by using the trained network parameters, the training cost is reduced, and the second image recognition model obtained by training can achieve better recognition effect while keeping smaller scale. The following are detailed below.
Example one
Referring to fig. 1, fig. 1 is a schematic flow chart of an image recognition model training method according to an embodiment of the present invention. The method described in fig. 1 is applied to an image recognition device, where the recognition device may be a corresponding recognition terminal, a corresponding recognition device, or a corresponding recognition server, and the server may be a local server or a cloud server, which is not limited in the embodiment of the present invention. As shown in fig. 1, the image recognition model training method may include the following operations:
101. and determining the trained first image recognition model.
Optionally, the first image recognition model includes a trained first feature coding network. Alternatively, the first image recognition model may be a large-scale image recognition model that has been sufficiently trained by using a training data set, which may be used for recognizing characters or characters in an image or recognizing other features in an image, and which has a large scale and is difficult to deploy on a small-scale device with low computational power or low computational resources, such as a mobile device, and therefore, the model needs to be migrated by combining the method of the present invention. Optionally, the first image recognition model may be trained by using a CTC (connected semantic Temporal Classification) loss as a loss function in training, where the CTC loss may avoid manual alignment of input and output to calculate a difference between input and output, and may be applied to image recognition, particularly in the field of character recognition in images, to achieve a beneficial effect.
102. And determining a second image recognition model to be trained.
Optionally, the model parameters of the second image recognition model are less than the model parameters of the first image recognition model, that is, the second image recognition model belongs to a smaller scale model, so as to share the parameters of the first image recognition model and be trained as a small scale model with the same or similar recognition effect. Preferably, the model parameters of the second image recognition model are less than those of the first image recognition model but more than 1/10 of the model parameters of the first image recognition model, wherein 1/10 is an empirical value to ensure that the model parameters of the second image recognition model are not too small to cause the training failure.
103. And determining the network parameters of the second feature coding network in the second image recognition model as the network parameters of the first feature coding network.
Optionally, the trained network parameters of the first feature coding network may be directly copied to the second feature coding network in the second image recognition model, so as to implement parameter sharing between the two models. Optionally, the number of network layers of the second signature coding network is the same as the number of network layers of the first signature coding network, that is, the vector dimension numbers output by the two signature coding networks are the same, so as to facilitate the duplication of network parameters and the difference measurement between the outputs of the two subsequent signature coding networks.
Optionally, the first feature coding network and the second feature coding network both include a plurality of hidden layer LSTM networks or BiLSTM networks for decoding image features, and the number of hidden layers of the first feature coding network and the second feature coding network is the same, so as to implement the above, that the vector dimension numbers output by the two feature coding networks are the same, so as to facilitate the duplication of network parameters and the difference measurement between the outputs of the two subsequent feature coding networks.
104. And determining a loss function as the difference between the characteristic output of the first characteristic coding network and the characteristic output of the second characteristic coding network, and performing joint training on the first image recognition model and the second image recognition model according to the loss function to obtain a trained second image recognition model.
Optionally, the training data set labeled with the feature recognition label may be simultaneously input to the first image recognition model and the second image recognition model for joint training until the loss function converges, that is, the feature extraction output of the second image recognition model is close to the feature extraction accuracy of the first image recognition model, so as to obtain the trained second image recognition model, and it is proved in an actual experiment that the thus trained second image recognition model not only can maintain a smaller scale, that is, fewer model parameters, but also can obtain the recognition accuracy close to or even exceeding the first image recognition model.
Therefore, by implementing the method described in the embodiment of the invention, the second image recognition model with smaller scale can share the feature coding network parameters in the trained first image recognition model, and the two image models are jointly trained to make the recognition effect of the second image recognition model approach to that of the first image recognition model, so that the trained network parameters can be utilized to improve the model training speed and reduce the training cost, and the trained second image recognition model can achieve better recognition effect while keeping smaller scale.
As an alternative embodiment, the first image recognition model further includes a first feature extraction network, the second image recognition model further includes a second feature extraction network, and accordingly, the network parameters of the second feature extraction network are less than those of the first feature extraction network and more than 1/10 of the network parameters of the first feature extraction network.
Optionally, the first feature extraction network and the second feature extraction network may both be ResNet networks and have different network depths, and specifically, the network depth of the second feature extraction network is less than the network depth of the first feature extraction network and is greater than 1/10 of the network depth of the first feature extraction network. For example, the first feature extraction network may be a ResNet-101 network with a network depth of 101, and the second feature extraction network may be a ResNet-18 network with a network depth of 18, so as to ensure that the model parameters of the second image recognition model are not too small, which may result in a training failure.
Optionally, the first image recognition model may include a first feature extraction network, a first feature coding network, and a first classification layer, which are connected in sequence, and the second image recognition model may include a second feature extraction network, a second feature coding network, and a second classification layer, which are connected in sequence, where both the first classification layer and the second classification layer may be softmax classification layers for outputting the image recognition result.
Therefore, through the optional implementation mode, the relationship of the network parameters between the feature extraction networks of the two models can be specifically limited, firstly, the image recognition performance of the models can be improved by using the feature extraction networks, and on the other hand, the model parameters of the second image recognition model can be ensured not to be too small, so that the subsequent training effect is improved.
As an alternative implementation, in the step 104, determining the loss function as a difference between the characteristic output of the first feature encoding network and the characteristic output of the second feature encoding network includes:
a vector distance between the feature output of the first feature encoding network and the feature output of the second feature encoding network is determined as a loss function.
Optionally, the vector distance may include at least one of the L1 distance, the L2 distance, the cosine distance, and the KL divergence, which may be any one of the vector distances, or may be a weighted summation result of any multiple distances, for example, the L1 distance and the L2 distance are considered together, and the weighted summation result of the L1 distance and the L2 distance is taken as the vector distance.
It can be seen that by implementing this alternative embodiment, the vector distance between the feature output of the first feature coding network and the feature output of the second feature coding network can be determined as a loss function, so that the feature extraction difference of the two models can be measured more accurately during training, and compared with the method of approximating the outputs of the two models by using KL divergence, which is common in the prior art during model training for image classification tasks, when the L1 distance and/or the L2 distance are/is used as the vector distance, the selection of the vector distance is more suitable for the model training task for the image recognition task, and a more excellent training effect can be achieved.
As an alternative implementation, in the step 104, determining the loss function as a difference between the characteristic output of the first feature encoding network and the characteristic output of the second feature encoding network includes:
determining a vector distance between a feature output of the first feature encoding network and a feature output of the second feature encoding network;
determining the identification loss corresponding to the second image identification model;
and determining a loss function according to the vector distance and the identification loss.
Wherein the recognition penalty is used to measure a difference between the output recognition result of the second image recognition model and the label of the training data. Alternatively, the recognition loss may be a CTC loss function, which can avoid manual alignment of input and output to calculate the difference between input and output, and can be applied to the field of image recognition, especially character recognition in images, to achieve beneficial effects. Accordingly, in an alternative embodiment where the second image recognition model may include a second feature extraction network, a second feature encoding network, and a second classification layer connected in series, the recognition loss may be a CTC loss function used to measure the difference between the recognition result output of the second classification layer and the labeling of the training data.
Therefore, by implementing the optional implementation mode, the vector distance between the feature output of the first feature coding network and the feature output of the second feature coding network and the recognition loss corresponding to the second image recognition model can be comprehensively considered to determine the loss function, so that the loss function can simultaneously measure the feature extraction difference of the two models and the recognition accuracy of the second image recognition model in the training process, the model training convergence speed can be improved, and the trained second image recognition model can achieve a more excellent recognition effect.
As an optional implementation manner, in the above step, determining a loss function according to the vector distance and the recognition loss includes:
determining a product between the recognition loss and a preset weight;
the sum of the vector distance and the product is determined as a loss function.
Optionally, the preset weight is generally set to be less than 1. Optionally, the preset weight is 0.1, which is an empirical value obtained through multiple tests, so that the recognition loss in the loss function occupies a proper proportion, the cost and efficiency of the whole training are not too high and too low due to too high recognition loss proportion, and the trained second image recognition model cannot achieve the due recognition effect due to too low recognition loss proportion.
Therefore, by implementing the optional implementation mode, the sum of the product and the vector distance between the recognition loss and the preset weight can be determined as the loss function, so that the loss function can more reasonably measure the feature extraction difference of the two models in the training and the recognition accuracy of the second image recognition model at the same time, the speed of model training convergence is favorably improved, and the trained second image recognition model can achieve a better recognition effect.
Example two
Referring to fig. 2, fig. 2 is a schematic flow chart of another image recognition model training method according to an embodiment of the present invention. The method described in fig. 2 is applied to an image recognition device, where the recognition device may be a corresponding recognition terminal, a corresponding recognition device, or a corresponding recognition server, and the server may be a local server or a cloud server, which is not limited in the embodiment of the present invention. As shown in fig. 2, the image recognition model training method may include the following operations:
201. and determining the trained first image recognition model.
202. And determining a second image recognition model to be trained.
203. And determining the network parameters of the second feature coding network in the second image recognition model as the network parameters of the first feature coding network.
204. A loss function is determined as a difference between the feature output of the first feature encoding network and the feature output of the second feature encoding network.
The detailed technical details and technical noun explanations of the steps 201-204 can refer to the description of the steps 101-103 in the first embodiment, which will not be repeated herein.
205. And simultaneously inputting the training data set into the first image recognition model and the second image recognition model for training.
206. All model parameters of the first image recognition model and parameters of the second feature encoding network of the second image recognition model are fixed in the training to remain unchanged.
207. And in the training process, other model parameters except the parameters of the second feature coding network in the second image recognition model are optimized until the loss function is converged to obtain the trained second image recognition model.
Optionally, as described in the first embodiment, the first image recognition model may include a first feature extraction network and a first feature coding network, and the second image recognition model includes a second feature extraction network and a second feature coding network, in the training, parameters of the first feature extraction network and the first feature coding network in the first image recognition model and parameters of the second feature coding network in the second image recognition model should be fixed, only parameters of the second feature extraction network in the second image recognition model are optimized until the loss function converges, so as to obtain a trained second feature extraction network, and the trained second feature extraction network and the second feature coding network with copied parameters are determined as the trained second image recognition model.
Therefore, by implementing the method described in the embodiment of the present invention, all the model parameters of the first image recognition model and the parameters of the second feature coding network of the second image recognition model can be fixed to be kept unchanged during training, and other model parameters in the second image recognition model except the parameters of the second feature coding network are optimized until the loss function converges, so as to obtain the trained second image recognition model, thereby reducing the training cost and improving the training efficiency, and on the other hand, ensuring that the output of the second image recognition model effectively approaches the recognition effect of the first image recognition model, and achieving a good training effect.
EXAMPLE III
Referring to fig. 3, fig. 3 is a schematic structural diagram of an image recognition model training apparatus according to an embodiment of the present invention. The apparatus described in fig. 3 may be applied to a corresponding image recognition apparatus, where the recognition apparatus may be a corresponding recognition terminal, a recognition device, or a recognition server, and the server may be a local server or a cloud server, which is not limited in the embodiment of the present invention. As shown in fig. 3, the apparatus may include:
a first determining module 301, configured to determine the trained first image recognition model.
Optionally, the first image recognition model includes a trained first feature coding network. Alternatively, the first image recognition model may be a large-scale image recognition model that has been sufficiently trained by using a training data set, which may be used for recognizing characters or characters in an image or recognizing other features in an image, and which has a large scale and is difficult to deploy on a small-scale device with low computational power or low computational resources, such as a mobile device, and therefore, the model needs to be migrated by combining the method of the present invention. Optionally, the first image recognition model may be trained by using a CTC (connected semantic Temporal Classification) loss as a loss function in training, where the CTC loss may avoid manual alignment of input and output to calculate a difference between input and output, and may be applied to image recognition, particularly in the field of character recognition in images, to achieve a beneficial effect.
A second determining module 302, configured to determine a second image recognition model to be trained.
Optionally, the model parameters of the second image recognition model are less than the model parameters of the first image recognition model, that is, the second image recognition model belongs to a smaller scale model, so as to share the parameters of the first image recognition model and be trained as a small scale model with the same or similar recognition effect. Preferably, the model parameters of the second image recognition model are less than those of the first image recognition model but more than 1/10 of the model parameters of the first image recognition model, wherein 1/10 is an empirical value to ensure that the model parameters of the second image recognition model are not too small to cause the training failure.
The parameter sharing module 303 is configured to determine a network parameter of the second feature coding network in the second image recognition model as a network parameter of the first feature coding network.
Optionally, the trained network parameters of the first feature coding network may be directly copied to the second feature coding network in the second image recognition model, so as to implement parameter sharing between the two models. Optionally, the number of network layers of the second signature coding network is the same as the number of network layers of the first signature coding network, that is, the vector dimension numbers output by the two signature coding networks are the same, so as to facilitate the duplication of network parameters and the difference measurement between the outputs of the two subsequent signature coding networks.
Optionally, the first feature coding network and the second feature coding network both include a plurality of hidden layer LSTM networks or BiLSTM networks for decoding image features, and the number of hidden layers of the first feature coding network and the second feature coding network is the same, so as to implement the above, that the vector dimension numbers output by the two feature coding networks are the same, so as to facilitate the duplication of network parameters and the difference measurement between the outputs of the two subsequent feature coding networks.
A loss determination module 304 for determining a loss function as a difference between the feature output of the first feature encoding network and the feature output of the second feature encoding network.
And the model training module 305 is configured to perform joint training on the first image recognition model and the second image recognition model to obtain a trained second image recognition model.
Therefore, the device described in the embodiment of the invention can enable the second image recognition model with smaller scale to share the feature coding network parameters in the trained first image recognition model, and perform joint training on the two image models to enable the recognition effect of the second image recognition model to approach the first image recognition model, so that the trained network parameters can be used to improve the model training speed and reduce the training cost, and the trained second image recognition model can achieve better recognition effect while keeping smaller scale.
As an alternative embodiment, the first image recognition model further includes a first feature extraction network, the second image recognition model further includes a second feature extraction network, and accordingly, the network parameters of the second feature extraction network are less than those of the first feature extraction network and more than 1/10 of the network parameters of the first feature extraction network.
Optionally, the first feature extraction network and the second feature extraction network may both be ResNet networks and have different network depths, and specifically, the network depth of the second feature extraction network is less than the network depth of the first feature extraction network and is greater than 1/10 of the network depth of the first feature extraction network. For example, the first feature extraction network may be a ResNet-101 network with a network depth of 101, and the second feature extraction network may be a ResNet-18 network with a network depth of 18, so as to ensure that the model parameters of the second image recognition model are not too small, which may result in a training failure.
Optionally, the first image recognition model may include a first feature extraction network, a first feature coding network, and a first classification layer, which are connected in sequence, and the second image recognition model may include a second feature extraction network, a second feature coding network, and a second classification layer, which are connected in sequence, where both the first classification layer and the second classification layer may be softmax classification layers for outputting the image recognition result.
Therefore, through the optional implementation mode, the relationship of the network parameters between the feature extraction networks of the two models can be specifically limited, firstly, the image recognition performance of the models can be improved by using the feature extraction networks, and on the other hand, the model parameters of the second image recognition model can be ensured not to be too small, so that the subsequent training effect is improved.
As an alternative embodiment, as shown in fig. 4, the model training module 305 includes:
a training unit 3051, configured to input a training data set to the first image recognition model and the second image recognition model at the same time for training;
a parameter fixing unit 3052, configured to fix all model parameters of the first image recognition model and parameters of the second feature coding network of the second image recognition model so as to remain unchanged in the training;
and the parameter optimization unit 3053 is configured to optimize, in the training, other model parameters in the second image recognition model except for the parameter of the second feature coding network until the loss function converges, so as to obtain a trained second image recognition model.
Optionally, the first image recognition model may include a first feature extraction network and a first feature coding network, and the second image recognition model includes a second feature extraction network and a second feature coding network, then the parameter fixing unit 3052 should fix parameters of the first feature extraction network and the first feature coding network in the first image recognition model and parameters of the second feature coding network in the second image recognition model in the above training, and the parameter optimizing unit 3053 only optimizes parameters of the second feature extraction network in the second image recognition model until the loss function converges, to obtain the trained second feature extraction network, and determines the trained second feature extraction network and the second feature coding network with copied parameters as the trained second image recognition model.
Therefore, by implementing the method described in the embodiment of the present invention, all the model parameters of the first image recognition model and the parameters of the second feature coding network of the second image recognition model can be fixed to be kept unchanged during training, and other model parameters in the second image recognition model except the parameters of the second feature coding network are optimized until the loss function converges, so as to obtain the trained second image recognition model, thereby reducing the training cost and improving the training efficiency, and on the other hand, ensuring that the output of the second image recognition model effectively approaches the recognition effect of the first image recognition model, and achieving a good training effect.
As an alternative embodiment, the specific way of determining the loss function as the difference between the feature output of the first feature encoding network and the feature output of the second feature encoding network by the loss determination module 304 includes:
a vector distance between the feature output of the first feature encoding network and the feature output of the second feature encoding network is determined as a loss function.
Optionally, the vector distance may include at least one of the L1 distance, the L2 distance, the cosine distance, and the KL divergence, which may be any one of the vector distances, or may be a weighted summation result of any multiple distances, for example, the L1 distance and the L2 distance are considered together, and the weighted summation result of the L1 distance and the L2 distance is taken as the vector distance.
It can be seen that by implementing this alternative embodiment, the vector distance between the feature output of the first feature coding network and the feature output of the second feature coding network can be determined as a loss function, so that the feature extraction difference of the two models can be measured more accurately during training, and compared with the method of approximating the outputs of the two models by using KL divergence, which is common in the prior art during model training for image classification tasks, when the L1 distance and/or the L2 distance are/is used as the vector distance, the selection of the vector distance is more suitable for the model training task for the image recognition task, and a more excellent training effect can be achieved.
As an alternative embodiment, the specific way of determining the loss function as the difference between the feature output of the first feature encoding network and the feature output of the second feature encoding network by the loss determination module 304 includes:
determining a vector distance between a feature output of the first feature encoding network and a feature output of the second feature encoding network;
determining the identification loss corresponding to the second image identification model;
and determining a loss function according to the vector distance and the identification loss.
Wherein the recognition penalty is used to measure a difference between the output recognition result of the second image recognition model and the label of the training data. Alternatively, the recognition loss may be a CTC loss function, which can avoid manual alignment of input and output to calculate the difference between input and output, and can be applied to the field of image recognition, especially character recognition in images, to achieve beneficial effects. Accordingly, in an alternative embodiment where the second image recognition model may include a second feature extraction network, a second feature encoding network, and a second classification layer connected in series, the recognition loss may be a CTC loss function used to measure the difference between the recognition result output of the second classification layer and the labeling of the training data.
Therefore, by implementing the optional implementation mode, the vector distance between the feature output of the first feature coding network and the feature output of the second feature coding network and the recognition loss corresponding to the second image recognition model can be comprehensively considered to determine the loss function, so that the loss function can simultaneously measure the feature extraction difference of the two models and the recognition accuracy of the second image recognition model in the training process, the model training convergence speed can be improved, and the trained second image recognition model can achieve a more excellent recognition effect.
As an alternative implementation, the loss determining module 304 determines a specific manner of the loss function according to the vector distance and the identification loss, including:
determining a product between the recognition loss and a preset weight;
the sum of the vector distance and the product is determined as a loss function.
Optionally, the preset weight is generally set to be less than 1. Optionally, the preset weight is 0.1, which is an empirical value obtained through multiple tests, so that the recognition loss in the loss function occupies a proper proportion, the cost and efficiency of the whole training are not too high and too low due to too high recognition loss proportion, and the trained second image recognition model cannot achieve the due recognition effect due to too low recognition loss proportion.
Therefore, by implementing the optional implementation mode, the sum of the product and the vector distance between the recognition loss and the preset weight can be determined as the loss function, so that the loss function can more reasonably measure the feature extraction difference of the two models in the training and the recognition accuracy of the second image recognition model at the same time, the speed of model training convergence is favorably improved, and the trained second image recognition model can achieve a better recognition effect.
Example four
Referring to fig. 5, fig. 5 is a schematic structural diagram of another image recognition model training apparatus according to an embodiment of the present invention. As shown in fig. 5, the apparatus may include:
a memory 401 storing executable program code;
a processor 402 coupled with the memory 401;
the processor 402 calls the executable program code stored in the memory 401 to execute part or all of the steps of the image recognition model training method disclosed in the first embodiment or the second embodiment of the present invention.
EXAMPLE five
The embodiment of the invention discloses a computer storage medium, which stores computer instructions, and when the computer instructions are called, the computer instructions are used for executing part or all of the steps in the image recognition model training method disclosed in the first embodiment or the second embodiment of the invention.
While certain embodiments of the present disclosure have been described above, other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily have to be in the particular order shown or in sequential order to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.
The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the apparatus, device, and non-volatile computer-readable storage medium embodiments, since they are substantially similar to the method embodiments, the description is relatively simple, and in relation to the description, reference may be made to some portions of the description of the method embodiments.
The apparatus, the device, the nonvolatile computer readable storage medium, and the method provided in the embodiments of the present specification correspond to each other, and therefore, the apparatus, the device, and the nonvolatile computer storage medium also have similar advantageous technical effects to the corresponding method.
In the 90 s of the 20 th century, improvements in a technology could clearly distinguish between improvements in hardware (e.g., improvements in circuit structures such as diodes, transistors, switches, etc.) and improvements in software (improvements in process flow). However, as technology advances, many of today's process flow improvements have been seen as direct improvements in hardware circuit architecture. Designers almost always obtain the corresponding hardware circuit structure by programming an improved method flow into the hardware circuit. Thus, it cannot be said that an improvement in the process flow cannot be realized by hardware physical modules. For example, a Programmable Logic Device (PLD), such as a Field Programmable Gate Array (FPGA), is an integrated circuit whose Logic functions are determined by programming the Device by a user. A digital system is "integrated" on a PLD by the designer's own programming without requiring the chip manufacturer to design and fabricate application-specific integrated circuit chips. Furthermore, nowadays, instead of manually making an integrated Circuit chip, such Programming is often implemented by "logic compiler" software, which is similar to a software compiler used in program development and writing, but the original code before compiling is also written by a specific Programming Language, which is called Hardware Description Language (HDL), and HDL is not only one but many, such as abel (advanced Boolean Expression Language), ahdl (alternate Language Description Language), traffic, pl (core unified Programming Language), HDCal, JHDL (Java Hardware Description Language), langue, Lola, HDL, laspam, hardsradware (Hardware Description Language), vhjhd (Hardware Description Language), and vhigh-Language, which are currently used in most common. It will also be apparent to those skilled in the art that hardware circuitry that implements the logical method flows can be readily obtained by merely slightly programming the method flows into an integrated circuit using the hardware description languages described above.
The controller may be implemented in any suitable manner, for example, the controller may take the form of, for example, a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro) processor, logic gates, switches, an Application Specific Integrated Circuit (ASIC), a programmable logic controller, and an embedded microcontroller, examples of which include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320, the memory controller may also be implemented as part of the control logic for the memory. Those skilled in the art will also appreciate that, in addition to implementing the controller as pure computer readable program code, the same functionality can be implemented by logically programming method steps such that the controller is in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Such a controller may thus be considered a hardware component, and the means included therein for performing the various functions may also be considered as a structure within the hardware component. Or even means for performing the functions may be regarded as being both a software module for performing the method and a structure within a hardware component.
The systems, devices, modules or units illustrated in the above embodiments may be implemented by a computer chip or an entity, or by a product with certain functions. One typical implementation device is a computer. In particular, the computer may be, for example, a personal computer, a laptop computer, a cellular telephone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.
For convenience of description, the above devices are described as being divided into various units by function, and are described separately. Of course, the functions of the various elements may be implemented in the same one or more software and/or hardware implementations of the present description.
As will be appreciated by one skilled in the art, the present specification embodiments may be provided as a method, system, or computer program product. Accordingly, embodiments of the present description may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, embodiments of the present description may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and so forth) having computer-usable program code embodied therein.
The description has been presented with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the description. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
In a typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.
The memory may include forms of volatile memory in a computer readable medium, Random Access Memory (RAM) and/or non-volatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.
Computer-readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), Digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium which can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.
It should also be noted that 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 an … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.
This description may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.
The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the system embodiment, since it is substantially similar to the method embodiment, the description is simple, and for the relevant points, reference may be made to the partial description of the method embodiment.
Finally, it should be noted that: the method and apparatus for training an image recognition model disclosed in the embodiments of the present invention are only preferred embodiments of the present invention, and are only used for illustrating the technical solutions of the present invention, rather than limiting the same; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those skilled in the art; the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. An image recognition model training method, characterized in that the method comprises:
determining a trained first image recognition model; the first image recognition model comprises a trained first feature coding network;
determining a second image recognition model to be trained; the model parameters of the second image recognition model are less than the model parameters of the first image recognition model;
determining a network parameter of a second feature coding network in the second image recognition model as a network parameter of the first feature coding network;
and determining a loss function as the difference between the characteristic output of the first characteristic coding network and the characteristic output of the second characteristic coding network, and performing joint training on the first image recognition model and the second image recognition model to obtain the trained second image recognition model.
2. The method for training the image recognition model according to claim 1, wherein the jointly training the first image recognition model and the second image recognition model to obtain the trained second image recognition model comprises:
inputting a training data set to the first image recognition model and the second image recognition model simultaneously for training;
fixing all model parameters of the first image recognition model and parameters of the second feature coding network of the second image recognition model in the training to be kept unchanged;
and in the training, optimizing other model parameters in the second image recognition model except the parameters of the second feature coding network until the loss function is converged to obtain the trained second image recognition model.
3. The image recognition model training method of claim 1, wherein the number of network layers of the second feature coding network is the same as the number of network layers of the first feature coding network; and/or the first feature encoding network is a BilSTM network; and/or the second feature encoding network is a BilSTM network.
4. The method of claim 1, wherein the model parameters of the second image recognition model are less than the model parameters of the first image recognition model and more than 1/10 of the model parameters of the first image recognition model;
and/or the presence of a gas in the gas,
the first image recognition model further comprises a first feature extraction network; the second image recognition model further comprises a second feature extraction network; the network parameters of the second feature extraction network are less than the network parameters of the first feature extraction network and more than 1/10 of the network parameters of the first feature extraction network.
5. The method of training an image recognition model according to claim 1, wherein the determining a loss function as a difference between the feature output of the first feature encoding network and the feature output of the second feature encoding network comprises:
determining a vector distance between the feature output of the first feature encoding network and the feature output of the second feature encoding network as a loss function;
and/or the presence of a gas in the gas,
determining a vector distance between a feature output of the first feature encoding network and a feature output of the second feature encoding network;
determining the identification loss corresponding to the second image identification model; the recognition loss is used for measuring the difference between the output recognition result of the second image recognition model and the label of the training data;
and determining a loss function according to the vector distance and the identification loss.
6. The method of claim 5, wherein determining a loss function according to the vector distance and the recognition loss comprises:
determining a product between the identification loss and a preset weight;
determining the sum of the vector distance and the product as a loss function.
7. The image recognition model training method of claim 6, wherein the vector distance comprises at least one of an L1 distance, an L2 distance, a cosine distance, and a KL divergence; and/or, the identified loss is a CTC loss; and/or the preset weight is 0.1.
8. An image recognition model training apparatus, characterized in that the apparatus comprises:
the first determining module is used for determining the trained first image recognition model; the first image recognition model comprises a trained first feature coding network;
the second determining module is used for determining a second image recognition model to be trained; the model parameters of the second image recognition model are less than the model parameters of the first image recognition model;
the parameter sharing module is used for determining the network parameters of the second feature coding network in the second image recognition model as the network parameters of the first feature coding network;
a loss determination module to determine a loss function as a difference between the feature output of the first feature encoding network and the feature output of the second feature encoding network;
and the model training module is used for carrying out joint training on the first image recognition model and the second image recognition model to obtain the trained second image recognition model.
9. An image recognition model training apparatus, characterized in that the apparatus comprises:
a memory storing executable program code;
a processor coupled with the memory;
the processor calls the executable program code stored in the memory to execute the image recognition model training method according to any one of claims 1 to 7.
10. A computer storage medium having stored thereon computer instructions which, when invoked, perform the image recognition model training method of any one of claims 1-7.
CN202111338332.7A 2021-11-12 2021-11-12 Image recognition model training method and device Pending CN114154559A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111338332.7A CN114154559A (en) 2021-11-12 2021-11-12 Image recognition model training method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111338332.7A CN114154559A (en) 2021-11-12 2021-11-12 Image recognition model training method and device

Publications (1)

Publication Number Publication Date
CN114154559A true CN114154559A (en) 2022-03-08

Family

ID=80460102

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111338332.7A Pending CN114154559A (en) 2021-11-12 2021-11-12 Image recognition model training method and device

Country Status (1)

Country Link
CN (1) CN114154559A (en)

Similar Documents

Publication Publication Date Title
CN108170667B (en) Word vector processing method, device and equipment
CN108345580B (en) Word vector processing method and device
CN107957989B (en) Cluster-based word vector processing method, device and equipment
CN113221555A (en) Keyword identification method, device and equipment based on multitask model
CN110826894A (en) Hyper-parameter determination method and device and electronic equipment
CN109299276B (en) Method and device for converting text into word embedding and text classification
CN112308113A (en) Target identification method, device and medium based on semi-supervision
CN111222315B (en) Movie scenario prediction method
CN115455166A (en) Method, device, medium and equipment for detecting abnormality of intelligent dialogue system
CN115146601A (en) Method and device for executing language processing task, readable storage medium and equipment
CN111507726B (en) Message generation method, device and equipment
CN116402165B (en) Operator detection method and device, storage medium and electronic equipment
CN107247704B (en) Word vector processing method and device and electronic equipment
CN117216271A (en) Article text processing method, device and equipment
CN111368902A (en) Data labeling method and device
CN110019295B (en) Database retrieval method, device, system and storage medium
CN114154559A (en) Image recognition model training method and device
CN115600090A (en) Ownership verification method and device for model, storage medium and electronic equipment
CN113344197A (en) Training method of recognition model, service execution method and device
CN110704742B (en) Feature extraction method and device
CN109325127B (en) Risk identification method and device
CN108733564B (en) Browser performance testing method, device and equipment
CN111539962A (en) Target image classification method, device and medium
CN113343716B (en) Multilingual translation method, device, storage medium and equipment
CN115017915B (en) Model training and task execution method and device

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination