WO2022036921A1 - Acquisition of target model - Google Patents

Abstract

A target model acquisition method and apparatus, an electronic device and a storage medium. The target model acquisition method comprises: pre-training an original model by using a first training sample set, so as to adjust network parameters of the original model, the original model comprising a first sub-network for feature extraction (S11); obtaining a target model by using a second sub-network and at least some structures of the pre-trained first sub-network (S12); and training the target model by using a second training sample set corresponding to a target task, so as to adjust network parameters of the target model (S13).

Description

Obtaining the target model

CROSS-REFERENCE TO RELATED APPLICATIONS

This disclosure claims the priority of the Chinese patent application filed on August 21, 2020 with the application number of 202010852192.4 and the invention titled "Method and Device for Obtaining Target Model, Electronic Device and Storage Medium", which is disclosed in the Chinese patent application. The entire contents are incorporated herein by reference.

technical field

The present disclosure relates to the field of information technology, and in particular, to a method and device for acquiring a target model, an electronic device, and a storage medium.

Background technique

Transfer learning aims to correlate the original model used to perform a certain task to obtain a target model to apply to the target task. With the rapid development of deep learning, computer vision and other technologies, transfer learning has been applied in many scenarios. Therefore, how to improve the performance of the target model has become a topic of great research value.

SUMMARY OF THE INVENTION

The present disclosure provides a method and apparatus for acquiring a target model, an electronic device and a storage medium.

A first aspect of the present disclosure provides a method for acquiring a target model. The method may include: pre-training an original model by using a first training sample set to adjust network parameters of the original model; wherein the original model includes a a first sub-network; using the second sub-network and at least part of the pre-trained structure of the first sub-network to obtain a target model; wherein the second sub-network is used to perform the target task based on the features extracted by the first sub-network; using and The second training sample set corresponding to the target task trains the target model to adjust the network parameters of the target model.

In some embodiments, the first sub-network may include a plurality of network segments, and each of the network segments may include sequentially connected at least one feature extraction unit for performing feature extraction.

In some embodiments, the first sub-network may include at least one branch network, and each branch network may include at least one of the network segments connected in sequence.

In some embodiments, the obtaining the target model by using the second sub-network and at least part of the pre-trained structure of the first sub-network may include: obtaining by using different partial structures of the first sub-network at least one candidate sub-network, and selecting the candidate sub-network that satisfies a preset condition as a selected sub-network; using the selected sub-network and the second sub-network to obtain the target model.

In some embodiments, the preset condition may include: a candidate model obtained by using the candidate sub-network and the second sub-network satisfies a preset performance condition.

In some embodiments, the preset condition may further include: the number of feature extraction units in the candidate sub-network reaches a preset number.

In some embodiments, the candidate sub-networks may include at least one feature extraction unit in each of the network segments in the same branch network, and the feature extraction units in different candidate sub-networks may at least Parts are different.

In some embodiments, the obtaining at least one candidate sub-network by using different partial structures of the first sub-network, and selecting the candidate sub-network that satisfies a preset condition as the selected sub-network may include: selecting the same sub-network The first feature extraction unit in each of the network segments in the branch network obtains an initial pending sub-network; for the pending sub-network, at least one feature extraction unit is added to obtain a candidate sub-network, wherein the added feature The extraction unit is located after the feature extraction unit selected in the network section; a plurality of candidate sub-networks and the second sub-network are respectively formed into candidate models, and one of the plurality of candidate models with the best performance conditions is selected The candidate sub-network corresponding to the candidate model is used as the pending sub-network; in the case where the number of feature extraction units in the pending sub-network is less than the preset number, obtain a new candidate sub-network based on the pending sub-network, and Select a candidate sub-network corresponding to a candidate model with the best performance condition among the candidate models composed of multiple new candidate sub-networks and the second sub-network as the new undetermined sub-network; When the number of feature extraction units in the network is not less than the preset number, the undetermined sub-network is used as the selected sub-network.

In some embodiments, adding at least one feature extraction unit to the undetermined sub-network to obtain a candidate sub-network may include: taking each of the network segments as a target segment, and using the target segment The number of selected feature extraction units is increased by an integer value, while keeping the number of selected feature extraction units in other network segments unchanged, to obtain a candidate sub-network corresponding to the target segment.

In some embodiments, the obtaining the new candidate sub-network based on the pending sub-network may include: adding at least one feature extraction unit to the pending sub-network to obtain a new candidate sub-network, wherein: The added feature extraction unit is located after the selected feature extraction unit in the network section.

In some embodiments, the candidate model obtained by using the candidate sub-network and the second sub-network satisfies a preset performance condition, may include: using a verification sample corresponding to the target task, comparing the candidate model using the candidate sub-network. The sub-network verifies the candidate model obtained by the second sub-network, and obtains a performance score of the candidate model for performing the target task; based on the performance score, it is determined whether the candidate model satisfies the preset performance condition.

The first sub-network includes a first number of feature extraction units, the first sub-network includes a second number of network segments, and the preset number is less than the first number and greater than or equal to the second number quantity.

In some embodiments, the first sub-network may comprise a one-way branch network, and the branch network may comprise a plurality of network segments connected in sequence, each of the network segments may comprise at least one feature extraction connected in sequence unit; the use of the first training sample set to pre-train the original model to adjust the network parameters of the original model may include: using a preset selection strategy before each training, selecting a The feature extraction unit; using the first training sample set, the part located before the selected feature extraction unit in each of the network segments is trained to adjust the feature extraction located in the selected feature extraction unit in each of the network segments Network parameters for the part before the cell.

In some embodiments, the first sub-network may include multiple branch networks, and each branch network may include at least one network segment connected in sequence, and each of the network segments may include at least one network segment connected in sequence. A feature extraction unit; the use of the first training sample set to pre-train the original model to adjust the network parameters of the original model may include: using a preset selection strategy before each training, selecting from the first sub-network All the way to the branch network and select one of the feature extraction units in each of the network segments included in the selected branch network; using the first training sample set, for each of the network segments The portion preceding the selected feature extraction unit is trained to adjust the network parameters of the portion preceding the selected feature extraction unit in each of said network segments.

In some embodiments, a preset selection strategy is used before each training, one of the branch networks is selected in the first sub-network, and each network segment included in the selected branch network is selected. Selecting one of the feature extraction units in the first sub-network may include: randomly selecting one of the branch networks in the first sub-network and selecting one of the network segments included in the selected branch network Feature extraction unit.

In some embodiments, a preset selection strategy is used before each training, one of the branch networks is selected in the first sub-network, and each network segment included in the selected branch network is selected. Selecting one of the feature extraction units in the first sub-network may include: selecting a branch network that includes the largest number of the feature extraction units in the first sub-network, and selecting each of the network segments included in the selected branch network Select one of the feature extraction units.

In some embodiments, the first sub-network may further include a downsampling layer between adjacent segments of the network.

In some embodiments, the feature extraction unit may include sequentially connected convolutional layers, activation layers, and batching layers.

In some embodiments, after the use of the first training sample set to pre-train the original model to adjust the network parameters of the original model, and after the use of the second sub-network and the pre-trained first sub-network Before obtaining the target model, the method may further include: training the original model by using the second training sample set to adjust the network parameters of the original model.

In some embodiments, after the target model is obtained by using the second sub-network and at least part of the pre-trained structure of the first sub-network, and after the using the first sub-network corresponding to the target task Before training the target model with the second training sample set to adjust the network parameters of the target model, the method may further include: using the first training sample set to train the target model to adjust the network parameters of the target model parameter.

In some embodiments, the original model may further include a third sub-network for performing a preset task based on the extracted features, wherein the preset task may be the same as the target task or different.

In some embodiments, the number of first training samples in the first training sample set may be greater than the number of second training samples in the second training sample set.

A second aspect of the present disclosure provides an apparatus for obtaining a target model, including: a first training module, a model obtaining module, and a second training module, where the first training module is used to pre-train an original model by using the first training sample set , to adjust the network parameters of the original model; wherein, the original model includes a first sub-network for feature extraction; the model acquisition module is used to use the second sub-network and at least part of the structure of the pre-trained first sub-network to obtain the target model; wherein, the second sub-network is used to perform the target task based on the features extracted by the first sub-network; the second training module is used to use the second training sample set corresponding to the target task to train the target model to adjust the network parameters of the target model .

A third aspect of the present disclosure provides an electronic device, including a memory and a processor coupled to each other, where the processor is configured to execute program instructions stored in the memory, so as to implement the method for acquiring the target model in the first aspect.

A fourth aspect of the present disclosure provides a computer-readable storage medium on which program instructions are stored, and when the program instructions are executed by a processor, implement the method for acquiring the target model in the first aspect above.

Description of drawings

1 is a schematic flowchart of a method for acquiring a target model according to an embodiment of the present disclosure;

2 is a schematic diagram of a framework of a first sub-network according to an embodiment of the present disclosure;

3 is a schematic diagram of a framework of a first sub-network according to another embodiment of the present disclosure;

FIG. 4 is a schematic flowchart of step S12 in FIG. 1 according to an embodiment of the present disclosure;

5 is a schematic diagram of a framework of a sub-network to be determined according to an embodiment of the present disclosure;

6 is a schematic flowchart of a method for acquiring a target model according to another embodiment of the present disclosure;

7 is a schematic diagram of a framework of an apparatus for acquiring a target model according to an embodiment of the present disclosure;

8 is a schematic diagram of a framework of an electronic device according to an embodiment of the present disclosure;

FIG. 9 is a schematic diagram of a frame of a computer-readable storage medium according to an embodiment of the present disclosure.

detailed description

The embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

In the following description, for purposes of explanation and not limitation, specific details are set forth, such as specific system structures, interfaces, techniques, etc., in order to provide a thorough understanding of the present disclosure.

The terms "system" and "network" are used interchangeably herein. The term "and/or" in this article is only an association relationship to describe the associated objects, indicating that there can be three kinds of relationships, for example, A and/or B, it can mean that A exists alone, A and B exist at the same time, and A and B exist independently B these three cases. In addition, the character "/" in this document generally indicates that the related objects are an "or" relationship. Also, "multiple" herein means two or more than two.

Please refer to FIG. 1 , which is a schematic flowchart of a method for acquiring a target model according to an embodiment of the present disclosure. Specifically, steps S11 to S13 may be included.

Step S11: Pre-training the original model by using the first training sample set to adjust the network parameters of the original model; wherein, the original model includes a first sub-network for feature extraction.

In one implementation scenario, the first sub-network may include a plurality of network segments (stages), and each network segment may include at least one feature extraction unit (block) connected in sequence. Specifically, the feature extraction unit is used for feature extraction, and may include sequentially connected convolutional layers, activation layers, and batch normalization (BN) layers. A convolutional layer can include several convolution kernels for feature extraction. The activation layer can include activation functions such as sigmoid, tanh, ReLu, etc., to introduce nonlinear factors. Batch layers can be used for normalization operations. The sequential connection of convolutional layers, activation layers and batching layers can help improve the learning effect of the feature extraction unit during the training process. In addition, in the feature extraction unit, a pooling layer can be connected after the convolutional layer to downsample the features extracted by the convolutional layer. In addition, the first sub-network also includes a downsampling layer located between adjacent network segments, which can facilitate feature dimension reduction, compress the number of data and parameters, reduce overfitting, and improve fault tolerance. The number of feature extraction units included in each network segment may be the same, for example, each network segment includes 3, or 4, or 5 feature extraction units. The number of feature extraction units contained in each network segment can also be completely different. For example, the first network segment contains 3 feature extraction units, the second network segment contains 4 feature extraction units, and the third network segment contains 4 feature extraction units. A segment contains 5 feature extraction units. The number of feature extraction units included in each network segment may not be exactly the same. For example, the first network segment includes 3 feature extraction units, the second network segment also includes 3 feature extraction units, and the third network segment also includes 3 feature extraction units. The network segment contains 5 feature extraction units. That is to say, it can be set according to actual application requirements, which is not limited here.

In a specific implementation scenario, several network segments in the above-mentioned multiple network segments may have the same input node. In this case, the first sub-network may have multiple branch networks, and each branch network may include a sequence At least one network segment of the connection.

Please refer to FIG. 2. FIG. 2 is a schematic diagram of a frame of a first sub-network according to an embodiment of the present disclosure. As shown in FIG. 2, a dotted rectangle represents a network segment. Each network segment includes 4 feature extraction units. The sub-network includes three branch networks connected in parallel, the first branch network is the first row of network segments (to simplify the schematic diagram, the first branch network only schematically depicts one network segment), the second branch network is the second row network segment, the third branch network is the third row network segment (to simplify the schematic diagram, the third branch network only schematically depicts one network segment), and the three branch networks have the same input node. In other scenarios, the first sub-network including other number of branch networks may also be set according to actual application requirements. For example, the first sub-network may be set to include a 2-way branch network, a 4-way branch network, etc., which is not limited herein.

In another specific implementation scenario, multiple network segments may also be connected in series. In this case, the first sub-network may have only one branch network, and the branch network includes multiple network segments connected in sequence .

Please refer to FIG. 3 . FIG. 3 is a schematic diagram of a frame of a first sub-network according to another embodiment of the present disclosure. As shown in FIG. 3 , the dotted rectangles represent network segments, and each network segment includes 4 feature extraction units. The first sub-network includes two serially connected network segments, in this case the first sub-network contains only one branch network.

In yet another specific implementation scenario, in addition to the first sub-network, the original model may further include another sub-network, for example, a third sub-network, and the other sub-network is used to perform preset tasks based on the extracted features. The preset task may be a target detection task, an image classification task, a scene segmentation task, etc., which are not limited herein. The target detection task means detecting target objects in the image, for example, detecting vehicles, pedestrians, etc. in the image; the image classification task means classifying the image into a certain category, for example, classifying the image into cats, dogs, turtles, etc. ; The scene segmentation task represents the category to which the pixels in the image are detected. For example, the pixels in the image that belong to lanes, cars, green belts, and the sky are detected. The above examples of preset tasks only indicate that there may exist in practical applications. a use case, and it does not limit its scope of use. The specific structure of the other sub-network can be set according to actual application requirements, which is not limited here. For example, when the preset task is a target detection task or an image classification task, another sub-network may include several (eg, 2, 3, etc.) sequentially connected fully-connected layers, softmax layers, etc., here Not limited. For another example, when the preset task is scene segmentation, another sub-network may include a fully connected layer and a softmax layer, which is not limited herein.

In an implementation scenario, in order to improve the accuracy of pre-training, the first training sample set may be a large-scale data set, that is, the number of first training samples in the first training sample set is greater than a preset value (eg, 1000, 5000, 10000 etc). Therefore, the original model can be fully pre-trained by using the first training sample set, which is beneficial to improve the accuracy of the subsequently acquired target model. In an implementation scenario, in order to improve the efficiency of pre-training, the first sub-network may only include a one-way branch network, and the one-way branch network may include a plurality of sequentially connected network segments. In this case, a preset selection strategy can be used before each training to select a feature extraction unit in each network segment, so that the first training sample set can be used to analyze the selected features in each network segment. The portion preceding the extraction unit (including the selected feature extraction unit) is trained to adjust the network parameters of the portion preceding the selected feature extraction unit in each network segment. The above method can help to improve the efficiency of pre-training. Specifically, the above-mentioned preset selection strategy may include: randomly selecting a feature extraction unit in each network segment.

In a specific implementation scenario, the preset selection strategy may include: randomly selecting an integer value within a preset value range corresponding to each network segment, and the upper limit of the preset value range is the value of the corresponding network segment. The number of included feature extraction units, for convenience of description, can be denoted as N _i , which represents the number of feature extraction units included in the i-th network segment. An integer value is randomly selected within the preset value range of Ni. For the convenience of description, the randomly selected integer value can be denoted as S _i , which represents the randomly selected integer value of the _i -th network segment, and the first training sample can be used. Set, train the part (i.e. the 1st to the S _i ) before the S i th feature extraction unit in each network segment to adjust the S _i th feature extraction unit before the S _i th feature extraction unit in each network segment. The network parameters of the part (ie 1st to _Sith ).

As shown by the dashed arrow in FIG. 3 , in the case where the first training sample set is used to train the part located before the selected feature extraction unit in each network segment, the selected features in each network segment can be The output of the extraction unit is used as the input data for the next network segment. Please refer to Figure 3, as shown in the figure, for example, when the feature extraction unit selected in the first network segment is the second feature extraction unit, the output result of the second feature extraction unit can be used as the next feature extraction unit. Input data for the network segment. For another example, when the feature extraction unit selected in the first network segment is the third feature extraction unit, the output result of the third feature extraction unit can be used as the input result of the next network segment. By analogy, no examples will be given here.

In an implementation scenario, when the first sub-network includes a multi-way branch network, a preset selection strategy can be used before each training to select one branch network in the first sub-network, and the selected branch network includes A feature extraction unit is selected in each network segment of The branch network contains the network parameters of the part of each network segment preceding the selected feature extraction unit. For the specific manner of selecting the feature extraction unit in the network segment, reference may be made to the foregoing related descriptions, which will not be repeated here.

In a specific implementation scenario, for the method of selecting a branch network in the first sub-network, reference may be made to the method of selecting a feature extraction unit in a network segment. Specifically, an integer value can be randomly selected within a preset value range, and the upper limit of the preset value range is the number of branch networks included in the first sub-network. The sub-network includes N branch networks in total, and the lower limit of the preset value range may be 1, and an integer value may be randomly selected from the preset value range of 1 to N. For the convenience of description, the randomly selected integer value may be denoted as S, indicating that the S-th branch network is selected in the first sub-network. Please refer to FIG. 2 , the first sub-network shown in FIG. 2 includes a total of 3 branch networks, then an integer value can be randomly selected from 1 to 3, for example, 2, then the second branch network S ₂ can be used as Selected branch network. Then, a feature extraction unit is selected in each network segment included in the _second branch network S2. For details, please refer to the above description, which will not be repeated here.

In another specific implementation scenario, the branch network that contains the largest number of feature extraction units in the first sub-network may be selected. Then, a feature extraction unit is selected in each network segment included in the branch network. For details, please refer to the foregoing description, which will not be repeated here.

In one implementation scenario, when a preset end condition is satisfied, the pre-training of the original model can be ended. Specifically, the preset end condition may include: the number of times each first training sample participates in training has reached a preset number of times threshold, and the preset number of times threshold may be set according to actual application needs, for example, may be set to 100, 120, 150, etc. etc., which are not limited here.

Step S12: Obtain a target model by using the second sub-network and at least part of the pre-trained structure of the first sub-network.

In the embodiment of the present disclosure, the second sub-network is configured to perform the target task based on the features extracted by the first sub-network. In one implementation scenario, the target task may include any one of the following: target detection task, image classification task, and scene segmentation task. For the specific meanings of the target detection task, the image classification task, and the scene segmentation task, reference may be made to the foregoing descriptions, which will not be repeated here. In addition, as described above, the original model may further include a third sub-network, and the third sub-network is used to perform a preset task based on the extracted features, and the preset task may be the same as or different from the target task. For example, both the preset task and the target task may be image classification tasks. For another example, the preset task and the target task may both be target detection tasks. For another example, the preset task may be a target detection task, and the target task may be an image classification task, which is not limited herein. In addition, the third sub-network may be the same as the second sub-network, or may not be the same. For example, the third sub-network may include a fully connected layer and a softmax layer, and the second sub-network may include two sequentially connected fully connected layers and a softmax layer connected after these two fully connected layers. For another example, the second sub-network, like the third sub-network, may also include a fully connected layer and a softmax layer, which is not limited herein.

In an implementation scenario, different partial structures of the first sub-network can be used to obtain at least one candidate sub-network, and a candidate sub-network that satisfies a preset condition can be selected as the selected sub-network, so that the selected sub-network and the second sub-network can be used. network to get the target model.

In a specific implementation scenario, the candidate sub-network includes at least one feature extraction unit in each network segment in the same branch network, and the feature extraction units in different candidate sub-networks are at least partially different, then each time the first sub-network is selected When there is a partial structure of the network, one branch network can be selected, and a feature extraction unit can be selected in each network section of the selected branch network, so that the part of each network section located before the selected feature extraction unit can be extracted. The combination is used as a partial structure of the first sub-network, so that different partial structures of the first sub-network can be obtained. Taking the first sub-network only includes one branch network as an example, please refer to Fig. 3 in conjunction with Fig. 3. When selecting part of the structure of the first sub-network for the first time, the third feature extraction unit can be randomly selected in the first network section, and the The second network segment randomly selects the second feature extraction unit, then the part of the first network segment before the third feature extraction unit and the part of the second network segment located before the second feature extraction unit can be combined The combination of the parts of the first sub-network is used as the partial structure of the first sub-network; when selecting the partial structure of the first sub-network for the second time, the second feature extraction unit can be randomly selected in the first network section, and the second feature extraction unit can be selected in the second network section. If the segment randomly selects the third feature extraction unit, then the part of the first network segment before the second feature extraction unit and the part of the second network segment before the third feature extraction unit can be combined, As a part of the structure of the first sub-network, and so on, and will not be exemplified one by one here. Specifically, the number of times of selection can be set according to actual application requirements, for example, 10, or 15, or 20, etc. can be used as the number of times of selection according to the computational complexity, which is not limited here.

In another specific implementation scenario, the preset condition may include: the candidate model obtained by using the candidate sub-network and the second sub-network satisfies the preset performance condition. Specifically, the verification samples corresponding to the target task can be used to verify the candidate model obtained by using the candidate sub-network and the second sub-network to obtain the performance score of the candidate model for executing the target task, so as to determine whether the candidate model satisfies the performance score based on the performance score. Preset performance conditions. For example, when the performance score is the highest value of the performance scores of all candidate models, it may be considered that the corresponding candidate model satisfies the preset performance condition.

In yet another specific implementation scenario, the preset condition may further include: the number of feature extraction units in the candidate sub-network is not less than the preset number. The preset number can be set according to actual application requirements, for example, it can be set to 4, 5, 6, etc., which is not limited here. The preset number can constrain the complexity of the target model.

In yet another specific implementation scenario, the selected sub-network and the second sub-network may be sequentially connected to obtain the target model.

In another implementation scenario, the candidate sub-network includes at least one feature extraction unit in each network segment of the same network branch, and the feature extraction units in different candidate sub-networks are at least partially different. In order to obtain the global optimal solution, It is possible to exhaustively enumerate all the different partial structures in the first sub-network, and use each partial structure as the corresponding candidate sub-network, and select the candidate sub-network that satisfies the preset conditions as the selected sub-network, so that the selected sub-network can be used. network and the second sub-network to get the target model. Specifically, for the setting method of the preset condition, reference may be made to the foregoing description, and details are not repeated here.

In a specific implementation scenario, taking the first sub-network including only one network branch as an example, when exhaustively enumerating all the different partial structures in the first sub-network, in order to avoid repeated selection, each sub-network in the first sub-network can be The network segment is assigned a count value count _i , which represents the count value of the i-th network segment, and is used to combine the feature extraction units located before the count value count _i in each network segment as In the partial structure selected this time, the number of network segments included in the first sub-network may be denoted as N _s . The initial value of the count value count _i can be set to 1, and after each partial structure is selected, the count is incremented by 1 until all the different partial structures in the first sub-network are exhaustively exhausted. Take the first sub-network including 4 network segments, and each network segment also includes 4 feature extraction units as an example, then when a partial structure is selected for the first time, the count values of the 4 network segments can be recorded as 1 respectively. , 1, 1, 1, at this time, the combination of the first feature extraction unit in each network segment can be used as the partial structure selected this time; when the partial structure is selected for the second time, the count value of the four network segments can be They are denoted as 1, 1, 1, and 2, respectively. At this time, the combination of the first feature extraction unit of the first three network segments and the first two feature extraction units of the last network segment can be used as part of the structure selected this time. ; When the partial structure is selected for the third time, the count values of the four network segments can be recorded as 1, 1, 1, and 3, respectively. At this time, the first feature extraction unit and the last network segment of the first three network segments can be The combination of the first three feature extraction units of the segment is used as the partial structure selected this time; when the partial structure is selected for the fourth time, the count values of the four network segments can be recorded as 1, 1, 1, and 4, respectively. The first feature extraction unit of the first three network segments and all the feature extraction units of the last network segment are taken as part of the structure selected this time; when the partial structure is selected for the fifth time, the count values of the four network segments can be Denoted as 1, 1, 2, and 1, respectively, at this time, the first feature extraction unit of the first two network segments, the first two feature extraction units of the third network segment, and the first feature extraction unit of the last network segment can be The combination of feature extraction units is used as the partial structure selected this time; when the partial structure is selected for the sixth time, the count values of the four network segments can be recorded as 1, 1, 2, and 2 respectively. The combination of the first feature extraction unit of the segment and the first two feature extraction units of the next two network segments is used as part of the structure selected this time, and so on, which will not be repeated here.

In an implementation scenario, in order to improve the accuracy of the dimension adjustment of the network structure, before obtaining the target model by using the second sub-network and at least part of the pre-trained structure of the first sub-network, the second training sample set may be used for training original model to tune the network parameters of the original model. Therefore, the original model can first adjust the network parameters based on the target task, which can help to improve the accuracy of the network structure dimension adjustment. In this case, at least part of the structure of the first sub-network includes the adjusted network parameters after being trained by the second training sample set, that is, during the transfer learning process, both the network parameters and the network structure are adjusted.

In a specific implementation scenario, the number of first training samples in the first training sample set is greater than the number of second training samples in the second training sample set, then through the embodiments of the present disclosure, the large-scale data samples (ie the first On the basis of the training sample set) and the small-scale data samples corresponding to the target task (ie the second training sample set), a target model suitable for the target task is obtained, which can help reduce the difficulty of collecting small-scale data samples and labeling. The workload of the second training sample set can further improve the efficiency of obtaining the target model. Specifically, the number of first training samples in the first training sample set may be 5000, 10000, 15000, etc. Correspondingly, the number of second training samples in the second training sample set may be 100, 200, 300, etc., which can be determined according to It is set according to the actual usage, which is not limited here.

Step S13 : using the second training sample set corresponding to the target task to train the target model to adjust the network parameters of the target model.

In one implementation scenario, the target model may be trained only by using the second training sample set corresponding to the target task, so as to adjust the network parameters of the target model. Specifically, the number of second training samples in the second training sample set is smaller than the number of first training samples in the first training sample set, for example, the second training sample set is small-scale data samples, and the first training sample set is large-scale data For the sample, please refer to the foregoing description for details, which will not be repeated here.

In another implementation scenario, in order to improve the accuracy of the target model, the target model can also be trained by using the first training sample set to adjust the network parameters of the target model, and then the target model can be trained by using the second training sample set corresponding to the target task. model to adjust the network parameters of the target model again.

In the above embodiment, the original model is first pre-trained by using the first training sample set to adjust the network parameters of the original model, wherein the original model includes a first sub-network for feature extraction. A target model is then obtained using the second sub-network and at least part of the pre-trained structure of the first sub-network, wherein the second sub-network is used to perform the target task based on the features extracted by the first sub-network. Finally, the target model is trained by using the second training sample set corresponding to the target task to adjust the network parameters of the target model. Therefore, not only can the second training sample set corresponding to the target task be used for adjustment in the "network parameter dimension", but also the original model can be adjusted in the "network structure dimension", which can greatly improve the degree of freedom of network adjustment. "Network parameter dimension" and "network structure dimension" fully tap the potential of the pre-trained original model, which is beneficial to improve the performance of the target model.

Please refer to FIG. 4 , which is a schematic flowchart of step S12 in FIG. 1 according to an embodiment of the present disclosure. Specifically, in the embodiment of the present disclosure, the original model is a single-branch network structure, that is, the first sub-network includes a branch network, and the branch network includes a plurality of network segments connected in sequence. Each network segment includes at least one feature extraction unit connected in sequence. The candidate sub-network includes at least one feature extraction unit in each network segment. Feature extraction units in different candidate sub-networks are at least partially different. Step S12 may include steps S41 to S46.

Step S41: Select the first feature extraction unit in each network segment to obtain an initial undetermined sub-network.

The first feature extraction unit in each network segment is selected, and each first feature extraction unit is connected together to obtain the initial undetermined sub-network. Please refer to FIG. 3 , as shown by the dashed arrow in FIG. 3 , the output result of the first feature extraction unit in the first network segment can be used as the input data of the second network segment. The initial pending sub-network is denoted as [1,1].

Step S42, for the undetermined sub-network, add at least one feature extraction unit to obtain a candidate sub-network, wherein the added feature extraction unit is located after the feature extraction unit selected by at least one network segment.

In this embodiment of the present disclosure, the at least one feature extraction unit may be one feature extraction unit, two feature extraction units, etc., which is not limited herein. For example, in the case where the accuracy requirements of network structure adjustment are relatively low, at least one feature extraction unit may be two, three, etc. multiple feature extraction units; while in the case of high accuracy requirements for network structure adjustment, it can be It is a feature extraction unit, which can be specifically set according to actual application needs, which is not limited here.

Each network segment can be used as a target segment, and then the sequence number of the selected feature extraction unit of the target segment in the previous step is increased by 1, that is, a feature extraction unit is added, while maintaining other networks in the first sub-network. In the segment, the number of selected feature extraction units in the previous step remains unchanged, and the candidate sub-network corresponding to the target segment is obtained. Please refer to FIG. 3 , the first network segment and the second network segment can be respectively used as target segments. Utilize the first two feature extraction units in the first target segment (in step S41, the feature extraction unit selected by the first network segment is the first feature extraction unit, so it becomes the first two features in this step extraction unit) and the first feature extraction unit in the second network segment to obtain candidate sub-networks for the first target segment. Specifically, as shown by the dotted arrow in FIG. 3 , the output result of the second feature extraction unit in the first target segment can be used as the input data of the second network segment. Using the first two feature extraction units in the second target section and the first extraction unit in the first network section, the candidate sub-network of the second target section is obtained. Specifically, as shown by the dotted arrow in Figure 3, The output result of the first feature extraction unit in the first network segment can be used as the input data of the second target segment. Please continue to refer to Figure 3. For the convenience of description, the candidate sub-network corresponding to the first target segment can be marked as [2, 1], indicating that the candidate sub-network corresponding to the first target segment is composed of the first network. The first two feature extraction units of the segment and the first feature extraction unit of the second network segment, and the initial candidate sub-network corresponding to the second target segment is recorded as [1, 2], indicating that the second The candidate sub-network corresponding to the target segment is composed of the first feature extraction unit of the first network segment and the first two feature extraction units of the second network segment. Other cases can be deduced by analogy, and we will not give examples here. .

In an implementation scenario, each network segment can be used as a target segment, and then the sequence number of the selected feature extraction unit of the target segment in the previous step is added to an integer such as 2, 3, etc., that is, an increase of two A feature extraction unit, three feature extraction units..., while keeping the number of selected feature extraction units in the previous step in other network segments in the first sub-network unchanged, to obtain a candidate sub-network corresponding to the target segment. The number of added feature extraction units can be set according to actual application requirements, which is not limited here.

Step S43 : forming a candidate model from the plurality of candidate sub-networks and the second sub-network respectively, and selecting a candidate sub-network corresponding to a candidate model with the best performance condition among the plurality of candidate models as the sub-network to be determined.

The verification samples corresponding to the target task can be used to verify the candidate models obtained by multiple candidate sub-networks and the second sub-network, respectively, to obtain the performance scores of the multiple candidate models to perform the target task, and determine the best performance based on the performance scores. A good candidate model, where the validation samples can be the same as or part of the second set of training samples. For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here. It is assumed that, after verification by the validation samples, the candidate sub-network [2, 1] can be taken as the pending sub-network.

Step S44: Determine whether the number of feature extraction units in the sub-network to be determined is less than the preset number, and if so, go to Step S42, otherwise go to Step S45.

In the embodiment of the present disclosure, the preset number may be preset according to actual application requirements. Specifically, it can be set according to the desired complexity of the target model. For example, it can be set to 4, 5, 6, etc., which is not limited here. For example, when the preset number is 4, since the number of feature extraction units in the undetermined sub-network [2, 1] is 3, which is less than the preset number, step S42 may be executed. For another example, when the preset number is 3, since the number of feature extraction units in the undetermined sub-network [2,1] is 3, which is not less than the preset number, step S45 can be executed, that is, the undetermined sub-network [2,1] can be directly ], as the selected subnet. Other situations can be deduced by analogy, and no examples are given here.

In one implementation scenario, the first sub-network may include a first number of feature extraction units, the first sub-network may include a second number of network segments, and the preset number may be set to be smaller than the first number and greater than the second number quantity.

In an implementation scenario, the preset number may be set according to the computational complexity of the target model for completing the target task.

Taking the verification of the verified samples, the candidate sub-network [2, 1] can be used as an example to continue the description. When step S44 determines that the feature extraction units of the sub-network [2, 1] to be determined are less than the preset number, jump to step S42 again. At this time, based on the undetermined sub-network [2, 1], a new candidate sub-network can be obtained.

In an implementation scenario, the first feature extraction unit located after the selected feature extraction unit in the target segment may be added after the selected feature extraction unit of the target segment, thereby obtaining a new candidate sub-network. Please continue to refer to Figure 3, taking the sub-network [2, 1] to be determined as an example, the first network segment can be used as the target segment first, and the segment located after the selected feature extraction unit (the second feature extraction unit) The first feature extraction unit (the third feature extraction unit) is added to the selected feature extraction unit (the second feature extraction unit) of the target segment to obtain a new candidate sub-network, which can be expressed as [ 3,1], that is, the new candidate sub-network is composed of the first three feature extraction units in the first network segment and the first feature extraction unit in the second network segment. Similarly, the second network segment can be used as the target segment, and the first feature extraction unit (the first feature extraction unit) after the selected feature extraction unit (the first feature extraction unit) in the target segment ), after adding to the selected feature extraction unit (the first feature extraction unit) of the target segment, another new candidate sub-network is obtained, which can be expressed as [2, 2] for the convenience of description. Other situations can be deduced by analogy, and no examples are given here.

It can be seen from the above steps that, based on the number of feature extraction units, a sub-network with the best performance conditions is selected from the current candidate sub-networks as the pending sub-network, and then at least one feature extraction unit is added to the previous number of feature extraction units. The unit calculates a new candidate sub-network based on the previous pending sub-network, and selects the sub-network with the best performance condition from the new candidate sub-network as the new pending sub-network. Still taking the example described above as an example, select the pending sub-network from the candidate sub-networks [2, 1] and [1, 2], assuming that [2, 1] has better performance, then [2, 1] is pending Sub-network, and then based on [2, 1], the pending sub-network is selected from [3, 1] and [2, 2]. In this case, [1, 3] will not be used as candidate sub-networks for the second round of performance evaluation. Thus, candidate sub-networks for calculating the performance score in step S43 are saved. When there are many network segments, this method can significantly save computation.

Please refer to FIG. 5. FIG. 5 is a schematic diagram of a frame of a sub-network to be determined according to an embodiment of the present disclosure. As shown in FIG. 5, the feature extraction unit in the solid rectangle represents the selected feature extraction unit, and the feature extraction unit in the dotted rectangle represents the selected feature extraction unit. Indicates an unselected feature extraction unit. The undetermined sub-network shown in Figure 5 is [2, 2], and the undetermined sub-network is [2, 2] and the preset number is 4 as an example. Since the feature extraction in the un-determined sub-network is If the number of units is equal to 4, the following step S45 may be performed, that is, the undetermined sub-network [2, 2] may be used as the selected sub-network.

Step S45: Take the undetermined sub-network as the selected sub-network.

Specifically, when the number of feature extraction units in the sub-network to be determined is not less than the preset number, the sub-network to be determined may be used as the selected sub-network. In the above manner, the selected sub-network at the performance level can be obtained under the model complexity constrained by the preset number.

Step S46: Obtain the target model by using the selected sub-network and the second sub-network.

Specifically, the selected sub-network and the second sub-network can be sequentially connected to obtain the target model.

In the embodiments of the present disclosure, the above method can not only constrain the target model in terms of "model complexity" and "model performance", but also improve the efficiency of acquiring the target model.

In an implementation scenario, when the first sub-network includes a multi-channel branch network, a branch network may be selected from the first sub-network first, and then steps S41 to S46 are performed for the branch network.

Please refer to FIG. 6 , which is a schematic flowchart of a method for acquiring a target model according to another embodiment of the present disclosure. Specifically, in the embodiment of the present disclosure, the original model includes a first sub-network for feature extraction, the first sub-network includes a branch network, and the branch network includes a plurality of network segments connected in sequence, each network segment A segment includes at least one feature extraction unit connected sequentially. Steps S601 to S612 may be included.

Step S601 : Select a feature extraction unit in each network segment by using a preset selection strategy before training the original model each time.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S602: Use the first training sample set to train the part of each network segment before the selected feature extraction unit to adjust the network parameters of the part of each network segment before the selected feature extraction unit.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S603: Use the second training sample set corresponding to the target task to train the original model to adjust the network parameters of the original model.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S604: Select the first feature extraction unit in each network segment to obtain the initial undetermined sub-network.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S605, for the undetermined sub-network, add at least one feature extraction unit to obtain a candidate sub-network, wherein the added feature extraction unit is located after the feature extraction unit selected by at least one network segment.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S606: The multiple candidate sub-networks and the second sub-network are respectively formed into candidate models, and the candidate sub-network corresponding to a candidate sub-model with the best performance condition among the multiple candidate models is selected as the undetermined sub-network.

In the embodiment of the present disclosure, the second sub-network is configured to perform the target task based on the features extracted by the first sub-network. For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S607: Determine whether the number of feature extraction units of the sub-network to be determined is less than the preset number, if so, go to step S605, otherwise go to step S608.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S608: Obtain a new candidate sub-network based on the undetermined sub-network.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S609: Take the undetermined sub-network as the selected sub-network.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S610: Obtain the target model by using the selected sub-network and the second sub-network.

Please refer to FIG. 5 , the sub-network to be determined is the sub-network [2, 2] shown in FIG. 5, and when the preset number is 4, the sub-network to be determined [2, 2] can be used as the selected sub-network, and the selected sub-network can be used. network and the second sub-network to get the target model. Specifically, the selected sub-network and the second sub-network can be sequentially connected to obtain the target model.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S611: Use the first training sample set to train the target model to adjust the network parameters of the target model.

Please continue to refer to Fig. 5. When the neutralized sub-network is the undetermined sub-network [2, 2] shown in Fig. 5, the first training sample set can be used to form the undetermined sub-network [2, 2] and the second sub-network The target model is trained to adjust the network parameters of the target model.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

Step S612: Use the second training sample set corresponding to the target task to train the target model to adjust the network parameters of the target model.

Please continue to refer to Fig. 5. When the neutralized sub-network is the undetermined sub-network [2, 2] shown in Fig. 5, the second training sample set can be further used to pair the undetermined sub-network [2, 2] and the second sub-network The formed target model is trained to adjust the network parameters of the target model.

For details, reference may be made to the relevant descriptions in the foregoing disclosed embodiments, which will not be repeated here.

In the embodiment of the present disclosure, the above method can help to improve the efficiency of acquiring the selected sub-network at the level of "network structure adjustment". Since the original model is adjusted at the level of "network parameter adjustment" and "network structure adjustment", the degree of freedom of network adjustment can be greatly improved, and the original model of pre-training can be fully excavated from the "network parameter dimension" and "network structure dimension". The potential of the model is beneficial to improve the performance of the target model.

In an implementation scenario, when the first sub-network includes a multi-channel branch network, a branch network may be selected from the first sub-network first, and then steps S601 to S612 are performed for the branch network.

Please refer to FIG. 7 , which is a schematic diagram of a framework of an apparatus 70 for acquiring a target model according to an embodiment of the present disclosure. The device 70 for acquiring the target model may include: a first training module 71 , a model acquiring module 72 and a second training module 73 , where the first training module 71 is used to pre-train the original model by using the first training sample set to adjust the network of the original model parameters; wherein, the original model may include a first sub-network for feature extraction; the model acquisition module 72 is configured to obtain a target model by using the second sub-network and at least part of the pre-trained first sub-network structure; wherein, the first sub-network The second sub-network is used to perform the target task based on the features extracted by the first sub-network; the second training module 73 is used to train the target model using the second training sample set corresponding to the target task to adjust the network parameters of the target model.

In some disclosed embodiments, the first sub-network may include a plurality of network segments, and each of the network segments may include sequentially connected at least one feature extraction unit for performing feature extraction.

In some disclosed embodiments, the first sub-network may include at least one branch network, and each branch network includes at least one of the network segments connected in sequence. Therefore, the first sub-network can be set as a "single-chain" single-branch network, or as a "multi-chain" multi-branch network, so that the target model can be obtained in the multi-branch network, and the single branch network can be obtained. The target model is obtained in the network, which can help to expand the scope of use.

In some disclosed embodiments, the model acquisition module 72 may include a structure search sub-module, configured to obtain at least one candidate sub-network by using different partial structures of the first sub-network, and select a candidate sub-network that satisfies a preset condition as the selected sub-network , the model obtaining module 72 may include a model building module for obtaining the target model by using the selected sub-network and the second sub-network.

By using different partial structures of the first sub-network, at least one candidate sub-network is obtained, and the candidate sub-network that satisfies the preset conditions is selected as the selected sub-network, so that the selected sub-network and the second sub-network are used to obtain the target model, which can have It is beneficial to expand the adjustment space of "network structure dimension", which can help to improve the performance of the target model.

In some disclosed embodiments, the preset condition may include: the candidate model obtained by using the candidate sub-network and the second sub-network satisfies the preset performance condition.

In some disclosed embodiments, the preset condition may further include: the number of feature extraction units in the candidate sub-network reaches a preset number.

Since the number of feature extraction units can reflect the complexity of the target model to a certain extent, and the preset performance conditions can reflect the performance of the target model to a certain extent, the target model can be constrained from the levels of "model complexity" and "model performance". .

In some disclosed embodiments, the candidate sub-networks may include at least one feature extraction unit in each network segment in the same branch network, and the feature extraction units in different candidate sub-networks are at least partially different.

In some disclosed embodiments, the structure search sub-module may include an initialization unit for selecting a first feature extraction unit in each of the network segments in the same branch network to obtain an initial undetermined sub-network, the The initialization unit is further configured to add at least one feature extraction unit to the undetermined sub-network to obtain a candidate sub-network, wherein the added feature extraction unit is located after the feature extraction unit selected in the network section, and the structure search sub-module may A performance evaluation unit is included, which is used to form a candidate model with a plurality of the candidate sub-networks and the second sub-network respectively, and select a candidate sub-network corresponding to a candidate model with the best performance condition among the plurality of candidate models, as the candidate sub-network. The sub-network to be determined, the structure search sub-module includes a repeated search unit, used for obtaining a new candidate sub-network based on the sub-network to be determined when the number of feature extraction units in the sub-network to be determined is less than a preset number, and Select a candidate sub-network corresponding to a candidate model with the best performance condition among the candidate models composed of a plurality of new candidate sub-networks and the second sub-network, as the new to-be-determined sub-network, the structure search sub-module may A selection acquisition unit is included, which is used to use the pending sub-network as the selected sub-network under the condition that the number of feature extraction units in the pending sub-network is not less than the preset number.

In some disclosed embodiments, the initialization unit may be configured to use each of the network segments as a target segment, increase the number of feature extraction units selected by the target segment by an integer value, while maintaining the number of feature extraction units in other network segments. The number of selected feature extraction units is unchanged, and the repeated search unit to obtain the candidate sub-network corresponding to the target section can be used to add at least one feature extraction unit to the selected sub-network to obtain a new candidate sub-network, wherein adding The feature extraction unit of is located after the selected feature extraction unit in the network section.

By taking each network segment as a target segment and using the first feature extraction unit in each target segment to obtain the initial undetermined sub-network, it can be beneficial to extract each network segment from the first sub-network The head starts to adjust the network structure. By increasing the number of feature extraction units selected in the target segment by an integer value, while keeping the number of feature extraction units selected in other network segments unchanged, the candidates corresponding to the target segment are obtained. Therefore, in the subsequent adjustment process, the feature extraction units of different network segments can be adjusted one by one, which can help to improve the accuracy of network adjustment.

In some disclosed embodiments, the performance evaluation unit may be configured to use the verification samples corresponding to the target task to verify the candidate model obtained by using the candidate sub-network and the second sub-network to obtain a performance score of the candidate model for executing the target task.

In some disclosed embodiments, the first sub-network includes a first number of feature extraction units, the first sub-network includes a second number of network segments, and the preset number is less than the first number and greater than or equal to the second number.

By using the verification samples corresponding to the target task, the candidate model obtained by using the candidate sub-network and the second sub-network is verified to obtain the performance score of the candidate model for executing the target task, and based on the performance score to determine whether the candidate model meets the preset performance conditions , so the accuracy of selecting the selected sub-network can be improved; in addition, the first sub-network includes a first number of feature extraction units, and the first sub-network includes a second number of network segments, and the preset number is smaller than the first number, Greater than or equal to the second number can help reduce the complexity of the target model.

In some disclosed embodiments, the first sub-network includes a branch network of one way, the branch network includes a plurality of network segments connected in sequence, each network segment includes at least one feature extraction unit connected in sequence, and the first training module 71 It includes a unit selection sub-module for using a preset selection strategy before each training to select a feature extraction unit in each network segment, and the first training module 71 includes a sample training sub-module for using the first training sample set. , train the part of each network segment before the selected feature extraction unit to adjust the network parameters of the part located before the selected feature extraction unit in each network segment.

In some disclosed embodiments, the first sub-network comprises a multi-way branch network, and each of the branch networks comprises at least one network segment connected in sequence, each segment of the network comprising at least one feature connected in sequence The extraction unit, the unit selection sub-module is further configured to use a preset selection strategy before each training, select one of the branch networks in the first sub-network, and select a branch network in each branch network included in the selected branch network. One of the feature extraction units is selected from the network section, and the sample training sub-module is used to use the first training sample set to train the part of each of the network sections before the selected feature extraction unit , to adjust the network parameters of the portion of each of the network segments preceding the selected feature extraction unit.

Through the above configuration, it is beneficial to fully train each part of the first sub-network after multiple trainings, and improve the pre-training efficiency.

In some disclosed embodiments, one of the branch networks may be randomly selected in the first sub-network and one of the feature extraction units may be selected in each of the network segments included in the selected branch network.

In some disclosed embodiments, a branch network including the largest number of feature extraction units may be selected in the first sub-network, and a branch network may be selected in each of the network segments included in the selected branch network Describe the feature extraction unit.

In some disclosed embodiments, the first sub-network may also include a downsampling layer between adjacent network segments.

In some disclosed embodiments, the feature extraction unit may include sequentially connected convolutional layers, activation layers, and batching layers.

By setting the feature extraction unit to include sequentially connected convolutional layers, activation layers and batch processing layers, the learning effect of the feature extraction unit in the training process can be improved; The downsampling layer between them can help to achieve feature dimensionality reduction, compress the number of data and parameters, reduce overfitting, and improve fault tolerance.

In some disclosed embodiments, the apparatus 70 for acquiring the target model may further include a third training module for training the original model by using the second training sample set to adjust the network parameters of the original model.

After pre-training, the original model is first trained with the second training sample set corresponding to the target task to adjust the network parameters of the original model, which can help to improve the accuracy of subsequent network structure dimension adjustment.

In some disclosed embodiments, the apparatus 70 for acquiring the target model may further include a fourth training module, configured to train the target model by using the first training sample set, so as to adjust the network parameters of the target model.

After completing the adjustment of the network structure dimension, the first training sample set is used to train the target model, and then the second training sample set corresponding to the target task is used to train the target model again, which can help improve the performance of the target model.

In some disclosed embodiments, the original model may further include a third sub-network for performing a preset task based on the extracted features, wherein the preset task is the same as or different from the target task.

By setting the original model to include a third sub-network, and the third sub-network is used to perform a preset task based on the extracted features, and the preset task is the same or different from the target task, it can be beneficial to further expand the application for obtaining the target model range.

In some disclosed embodiments, the number of first training samples in the first training sample set may be greater than the number of second training samples in the second training sample set.

By setting the number of the first training samples in the first training sample set to be greater than the number of the second training samples in the second training sample set, it can be beneficial to reduce the workload of sample labeling on the target task.

Please refer to FIG. 8 , which is a schematic diagram of a frame of an electronic device 80 according to an embodiment of the present disclosure. The electronic device 80 includes a memory 81 and a processor 82 coupled to each other, and the processor 82 is configured to execute program instructions stored in the memory 81 to implement the steps of any of the above-mentioned embodiments of the method for acquiring a target model. In a specific implementation scenario, the electronic device 80 may include, but is not limited to, a microcomputer and a server. In addition, the electronic device 80 may also include mobile devices such as a notebook computer and a tablet computer, which are not limited herein.

Specifically, the processor 82 is configured to control itself and the memory 81 to implement the steps of the above-mentioned method for acquiring any target model. The processor 82 may also be referred to as a CPU (Central Processing Unit, central processing unit). The processor 82 may be an integrated circuit chip with signal processing capability. The processor 82 can also be a general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field-Programmable Gate Array, FPGA) or other Programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. In addition, the processor 82 may be jointly implemented by an integrated circuit chip.

In the embodiment of the present disclosure, not only can the second training sample set corresponding to the target task be used for adjustment in the "network parameter dimension", but also the original model can be adjusted in the "network structure dimension", which can greatly improve the freedom of network adjustment It is possible to fully tap the potential of the pre-trained original model from the "network parameter dimension" and "network structure dimension", which is beneficial to improve the performance of the target model.

Please refer to FIG. 9 , which is a schematic diagram of a framework of an embodiment of the disclosed computer-readable storage medium 90 . The computer-readable storage medium 90 stores program instructions 901 that can be executed by the processor, and the program instructions 901 are used to implement the steps of any of the above-mentioned embodiments of the method for acquiring a target model.

In the embodiment of the present disclosure, not only the second training sample set corresponding to the target task can be used for adjustment in the "network parameter dimension", but also the original model can be adjusted in the "network structure dimension", which can greatly improve the freedom of network adjustment. It is possible to fully tap the potential of the pre-trained original model from the "network parameter dimension" and "network structure dimension", which is beneficial to improve the performance of the target model.

In some embodiments, the functions or modules included in the apparatuses provided in the embodiments of the present disclosure may be used to execute the methods described in the above method embodiments. For specific implementation, reference may be made to the descriptions of the above method embodiments. For brevity, here No longer.

The above descriptions of the various embodiments tend to emphasize the differences between the various embodiments, and the similarities or similarities can be referred to each other. For the sake of brevity, details are not repeated herein.

In the several embodiments provided in the present disclosure, it should be understood that the disclosed method and apparatus may be implemented in other manners. For example, the device implementations described above are only illustrative. For example, the division of modules or units is only a logical function division. In actual implementation, there may be other divisions. For example, units or components may be combined or integrated. to another system, or some features can be ignored, or not implemented. On the other hand, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, which may be in electrical, mechanical or other forms.

Units described as separate components may or may not be physically separated, and components shown as units may or may not be physical units, that is, may be located in one place, or may be distributed over network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this implementation manner.

In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware, or may be implemented in the form of software functional units.

The integrated unit, if implemented as a software functional unit and sold or used as a stand-alone product, may be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present disclosure essentially or the parts that contribute to the prior art, or all or part of the technical solutions can be embodied in the form of software products, and the computer software products are stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor (processor) to execute all or part of the steps of the methods of the various embodiments of the present disclosure. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes .

Claims (25)

A method for acquiring a target model, comprising: Using the first training sample set to pre-train the original model to adjust the network parameters of the original model; wherein, the original model includes a first sub-network for feature extraction; Using the second sub-network and at least part of the pre-trained structure of the first sub-network, the target model is obtained; wherein the second sub-network is used to perform the target task based on the features extracted by the first sub-network ; The target model is trained by using the second training sample set corresponding to the target task to adjust the network parameters of the target model. The method of claim 1, wherein the first sub-network comprises a plurality of network segments, each of the network segments comprising at least one feature extraction unit connected in sequence, the feature extraction unit for performing a feature extraction extract. The method of claim 2, wherein the first sub-network comprises at least one branch network, and each branch network comprises at least one of the network segments connected in sequence. The method according to any one of claims 1 to 3, wherein the obtaining the target model by using the second sub-network and at least part of the pre-trained structure of the first sub-network comprises: Obtain at least one candidate sub-network by utilizing different partial structures of the first sub-network, and select the candidate sub-network that satisfies a preset condition as the selected sub-network; Using the selected sub-network and the second sub-network, the target model is obtained. The method according to claim 4, wherein the preset condition comprises: a candidate model obtained by using the candidate sub-network and the second sub-network satisfies a preset performance condition. The method according to claim 5, wherein the preset condition further comprises: the number of feature extraction units in the candidate sub-network reaches a preset number. 6. The method according to any one of claims 4 to 6, wherein the candidate sub-network comprises at least one feature extraction unit in each of the network segments in the same branch network, and is different from the candidate sub-network The feature extraction units in the network are at least partially different. The method according to claim 7, wherein obtaining at least one candidate sub-network by using different partial structures of the first sub-network, and selecting the candidate sub-network that satisfies a preset condition as the selected sub-network, comprising: Selecting the first feature extraction unit in each of the network segments in the same branch network to obtain an initial undetermined sub-network; For the to-be-determined sub-network, at least one feature extraction unit is added to obtain a candidate sub-network, wherein the added feature extraction unit is located after the feature extraction unit selected in the network section; A plurality of candidate sub-networks and the second sub-network are respectively formed into candidate models, and a candidate sub-network corresponding to a candidate model with the best performance condition among the plurality of candidate models is selected as the sub-network to be determined; When the number of feature extraction units in the undetermined sub-network is less than a preset number, obtain a new candidate sub-network based on the undetermined sub-network, and select a plurality of new candidate sub-networks and the The candidate sub-network corresponding to the candidate model with the best performance condition among the candidate models composed of the second sub-network is used as the new undetermined sub-network; In the case that the number of feature extraction units in the to-be-determined sub-network is not less than the preset number, the to-be-determined sub-network is used as the selected sub-network. The method according to claim 8, wherein, for the undetermined sub-network, adding at least one feature extraction unit to obtain a candidate sub-network, comprising: Taking each of the network segments as a target segment, increasing the number of feature extraction units selected by the target segment by an integer value, while keeping the number of feature extraction units selected in other network segments unchanged, we get candidate subnetworks corresponding to the target segment. The method according to claim 8 or 9, wherein the obtaining the new candidate sub-network based on the undetermined sub-network comprises: For the undetermined sub-network, at least one feature extraction unit is added to obtain a new candidate sub-network, wherein the added feature extraction unit is located after the feature extraction unit selected in the network section. The method according to any one of claims 5 to 10, wherein the candidate model obtained by using the candidate sub-network and the second sub-network satisfies a preset performance condition, comprising: Using the verification sample corresponding to the target task, verify the candidate model obtained by using the candidate sub-network and the second sub-network, and obtain a performance score of the candidate model for executing the target task; Whether the candidate model satisfies the preset performance condition is determined based on the performance score. The method of any one of claims 6 to 11, wherein, The first sub-network includes a first number of feature extraction units, the first sub-network includes a second number of network segments, and the preset number is less than the first number and greater than or equal to the second number quantity. 13. The method of any one of claims 1 to 12, wherein the first sub-network comprises a one-way branch network, and the one-way branch network comprises a plurality of network segments connected in sequence, each of the network segments Including at least one feature extraction unit connected in sequence; the use of the first training sample set to pre-train the original model to adjust the network parameters of the original model, including: Before each training, a preset selection strategy is used to select one of the feature extraction units in each of the network segments; Using the first training sample set, the portion of each of the network segments preceding the selected feature extraction unit is trained to adjust the portion of each of the network segments preceding the selected feature extraction unit. Network parameters. 12. The method of any one of claims 1 to 12, wherein the first sub-network comprises a multi-way branch network, and each branch network comprises at least one network segment connected in sequence, each of the branch networks The network section includes at least one feature extraction unit connected in sequence; the pre-training of the original model using the first training sample set to adjust the network parameters of the original model includes: Using a preset selection strategy before each training, select one of the branch networks in the first sub-network and select one of the feature extraction units in each of the network segments included in the selected branch network ; Using the first training sample set, a portion of each of the network segments preceding the selected feature extraction unit is trained to adjust the portion of each of the network segments preceding the selected feature extraction unit. Network parameters. The method according to claim 14, wherein a preset selection strategy is used before each training, one branch network is selected in the first sub-network, and each branch network included in the selected branch network is selected. Selecting one of the feature extraction units in the network segment includes: One of the branch networks is randomly selected in the first sub-network and one of the feature extraction units is selected in each of the network segments included in the selected branch network. The method according to claim 14, wherein a preset selection strategy is used before each training, one branch network is selected in the first sub-network, and each branch network included in the selected branch network is selected. Selecting one of the feature extraction units in the network segment includes: A branch network containing the largest number of the feature extraction units is selected in the first sub-network, and one of the feature extraction units is selected in each of the network segments included in the selected branch network. 17. The method of any one of claims 2 to 16, wherein the first sub-network further comprises a downsampling layer located between adjacent segments of the network. The method according to any one of claims 2 to 17, wherein the feature extraction unit comprises sequentially connected convolutional layers, activation layers and batch processing layers. The method according to any one of claims 1 to 18, wherein after the using the first training sample set to pre-train the original model to adjust the network parameters of the original model, and after the using the second sub-model Before obtaining the target model, the method further includes: The original model is trained using the second training sample set to adjust network parameters of the original model. The method according to any one of claims 1 to 19, wherein after the target model is obtained by using the second sub-network and at least part of the pre-trained structure of the first sub-network, and after the target model is obtained Before the training of the target model by using the second training sample set corresponding to the target task to adjust the network parameters of the target model, the method further includes: The target model is trained by using the first training sample set to adjust network parameters of the target model. The method of any one of claims 1 to 20, wherein, The original model further includes a third sub-network for performing a preset task based on the extracted features, wherein the preset task is the same as or different from the target task. The method according to any one of claims 1 to 21, wherein the number of first training samples in the first training sample set is greater than the number of second training samples in the second training sample set. A device for acquiring a target model, comprising: a first training module for pre-training an original model by using a first training sample set to adjust network parameters of the original model; wherein the original model includes a first sub-network for feature extraction; a model obtaining module, configured to obtain the target model by using the second sub-network and at least part of the pre-trained structure of the first sub-network; wherein the second sub-network is configured to be based on the first sub-network The extracted features perform the target task; The second training module is used for training the target model by using the second training sample set corresponding to the target task, so as to adjust the network parameters of the target model. An electronic device, comprising a memory and a processor coupled to each other, and the processor is used to execute program instructions stored in the memory, so as to realize the method for obtaining a target model according to any one of claims 1 to 22 . A computer-readable storage medium on which program instructions are stored, wherein, when the program instructions are executed by a processor, the method for obtaining a target model according to any one of claims 1 to 22 is implemented.

Images