Perception model training method and device, electronic equipment and storage medium

Abstract

The present disclosure relates to the technical field of artificial intelligence, in particular to the technical field of deep learning, computer vision, image processing, autonomous driving perception and the like. The specific implementation scheme is as follows: obtaining a preliminary training iteration number; constructing N groups of practical training hyperparameters including the preliminary training iteration number, N being an integer greater than or equal to 2; training N initial perception models based on the N groups of practical training hyperparameters to obtain N candidate perception models; continuing to train a to-be-trained model with the best performance in the N candidate perception models to obtain a target perception model, the target perception model being a deep learning model used for identifying environmental perception data to obtain an environmental recognition result. The present disclosure can effectively reduce the training time of the perception model, and at the same time, reduce the occupation of computing resources.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, in particular to the technical field of deep learning, computer vision, image processing, autonomous driving perception, etc. BACKGROUND

[0002] With the development of artificial intelligence technology, the accuracy of deep learning models is also continuously improving. However, the cost of high accuracy is usually longer training time and higher operation resource requirements. SUMMARY

[0003] The present disclosure provides a perception model training method and device, an electronic device, and a storage medium.

[0004] According to an aspect of the present disclosure, a perception model training method is provided, comprising:

[0005] obtaining a preliminary training iteration number;

[0006] constructing N sets of real training hyperparameters including the preliminary training iteration number, N being an integer greater than or equal to 2;

[0007] training N initial perception models based on the N sets of real training hyperparameters to obtain N candidate perception models;

[0008] continuing to train a to-be-trained model with the best performance in the N candidate perception models to obtain a target perception model, the target perception model being a deep learning model for identifying environmental perception data to obtain an environmental recognition result.

[0009] According to a second aspect of the present disclosure, a perception model training device is provided, comprising:

[0010] an iteration number obtaining unit configured to obtain a preliminary training iteration number;

[0011] a hyperparameter constructing unit configured to construct N sets of real training hyperparameters including the preliminary training iteration number, N being an integer greater than or equal to 2;

[0012] a preliminary training unit configured to train N initial perception models based on the N sets of real training hyperparameters to obtain N candidate perception models;

[0013] a continuous training unit configured to continue to train a to-be-trained model with the best performance in the N candidate perception models to obtain a target perception model, the target perception model being a deep learning model for identifying environmental perception data to obtain an environmental recognition result.

[0014] According to a third aspect of the present disclosure, an electronic device is provided, comprising:

[0015] at least one processor;

[0016] a memory in communication with the at least one processor;

[0017] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method provided in the first aspect.

[0018] According to a fourth aspect of the present disclosure, a non-transitory computer readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable a computer to perform the method provided in the first aspect.

[0019] According to a fifth aspect of the present disclosure, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the method provided in the first aspect.

[0020] The present disclosure can effectively reduce the training time of the perception model, and reduce the occupation of operation resources.

[0021] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present disclosure, nor to limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0022] The accompanying drawings are used to better understand the present scheme and do not limit the present disclosure. Among them:

[0023] Figure 1 A determination method of an optimal number of training iterations in the prior art is shown in the auxiliary diagram;

[0024] Figure 2 A flowchart of a perception model training method provided by an embodiment of the present disclosure is shown in the diagram;

[0025] Figure 3 A real-time training result diagram provided by an embodiment of the present disclosure is shown in the diagram;

[0026] Figure 4 A complete flowchart of a perception model training method provided by an embodiment of the present disclosure is shown in the diagram;

[0027] Figure 5 An application scenario diagram of a perception model training method provided by an embodiment of the present disclosure is shown in the diagram;

[0028] Figure 6 A schematic structural block diagram of a perception model training device provided by an embodiment of the present disclosure is shown in the diagram;

[0029] Figure 7 A schematic structural block diagram of an electronic device provided by an embodiment of the present disclosure is shown in the diagram. DETAILED DESCRIPTION

[0030] Exemplary embodiments of the present disclosure are described herein below with reference to the accompanying drawings, which include various details of embodiments of the present disclosure to assist in understanding them. These should be considered in a descriptive sense only and not limiting. As such, various changes and modifications of the embodiments described herein can be made by one of ordinary skill in the art without departing from the scope of the present disclosure. Also, for the purpose of clarity and the brevity of description, the description below omits descriptions of well-known functions and constructions.

[0031] As described in the background, with the development of artificial intelligence technology, the accuracy of deep learning models is also constantly improving. However, the price of high accuracy is usually longer training time and higher operation resource requirements. Specifically, among the multiple hyperparameters required for model training, the number of training iterations (Epoch) is a hyperparameter closely related to the model training time. If the Epoch is too small, the model training time is too short, which will lead to insufficient model training, so that the model cannot converge to the best state, or even cannot converge; if the Epoch is too large, the model training time is too long, which not only wastes operation resources, but also may lead to model overfitting.

[0032] Currently, the optimal Epoch is usually determined according to historical training experience. Specifically, during the model training process, the performance value change of the model under multiple performance evaluation indicators can be observed manually, and then a performance balance point is determined. The Epoch corresponding to the performance balance point is taken as the optimal Epoch. Please refer to Figure 1 As the Epoch increases, the model bias (Loss) continues to decrease at a large amplitude in the initial training stage, and in the later training stage, the Loss tends to be stable, but is also in a decreasing trend. The overall error of the model starts to decrease gradually in the initial stage, but after decreasing to a certain extent, it tends to increase. Since the overall error of the model reaches the minimum value when the corresponding Loss is relatively low, it can meet the usual model application requirements, therefore, the performance balance point can be taken as the overall error of the model reaching the minimum value, and the Epoch corresponding to this time is taken as the optimal Epoch.

[0033] After determining the optimal Epoch, the model can be trained based on the optimal Epoch. In the initial stage of model training, multiple hyperparameters usually need to be determined, and multiple hyperparameter values corresponding to each hyperparameter are set. Then, multiple hyperparameters corresponding to the multiple hyperparameter values are arranged and combined to obtain multiple groups of hyperparameters. For example, the multiple hyperparameters include learning rate, batch size (Batch_Size), network layer number, weight initialization method and activation function, and three hyperparameter values corresponding to each hyperparameter are set. Specifically, it can be shown in Table 1:

[0034] Table 1

[0035]

[0036]

[0037] According to the full permutation combination mode, 3 hyperparameter values corresponding to the above 5 hyperparameter items are arranged in combination, and 3 5 = 243 groups of hyperparameters can be obtained.

[0038] Suppose that the optimal Epoch is 1000 times, and based on each group of hyperparameters, the model is trained according to the optimal Epoch, which needs to run 8 Graphics Processing Unit (GPU) cards for 10 days. Therefore, in the case of obtaining a total of 243 groups of hyperparameters, based on the 243 groups of hyperparameters, 243 models are trained according to the optimal Epoch respectively, and 243 candidate models are obtained, which needs to occupy:

[0039] 8 GPU cards * 10 days * 243 groups * 24 hours = 466560 card hours

[0040] That is, 466560 card hours are needed to complete the training of 243 models and obtain 243 candidate models. Finally, the candidate with the best performance is selected from the 243 candidate models as the target model.

[0041] Obviously, using the existing model training method not only takes a long time to train the model, but also occupies a large amount of computing resources, especially for a perception model with a large training data set. This defect will be more prominent.

[0042] Based on the above background, the embodiment of the present disclosure provides a perception model training method, which can be applied to an electronic device. In the following, the perception model training method provided by the embodiment of the present disclosure will be described in conjunction with the flowchart shown in FIG. 1. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described can also be executed in other orders. Figure 2

[0043] Step S201, obtaining a preliminary Epoch;

[0044] Step S202, constructing N groups of real training hyperparameters including the preliminary Epoch, N being an integer greater than or equal to 2;

[0045] Step S203, based on the N groups of real training hyperparameters, training N initial perception models respectively to obtain N candidate perception models;

[0046] ​Step S204, continue to train the to-be-trained model with the best performance in the N candidate perception models to obtain a target perception model. The target perception model is a deep learning model used to identify environment perception data to obtain an environment recognition result.

[0047] wherein the preliminary Epoch can be any reasonable value less than the target Epoch, and the target Epoch is the optimal Epoch determined according to historical training experience. For example, if the optimal Epoch is 1000 times, the preliminary Epoch can be 200 times, 400 times, 600 times, etc., and the present disclosure does not make specific limitations thereto.

[0048] After obtaining the preliminary Epoch, N sets of real training hyperparameters including the preliminary Epoch can be constructed, and based on the N sets of real training hyperparameters, the N initial perception models are trained respectively to obtain N candidate perception models. Then, the to-be-trained model with the best performance is selected from the N candidate perception models. Finally, only the to-be-trained model is trained to obtain the target perception model. Each set of real training hyperparameters can include other hyperparameters in addition to the preliminary Epoch, such as the hyperparameter values corresponding to the learning rate, Batch_Size, network layer number, weight initialization method, and activation function. The target perception model is a deep learning model used to identify environment perception data to obtain an environment recognition result, and the environment perception data can be real environment data collected by multiple perception modules during the driving process of an autonomous vehicle. For example, it can be real environment data collected by imaging units, lidar, millimeter wave radar, ultrasonic radar, etc. set on the autonomous vehicle.

[0049] In addition, in the present disclosure, when the to-be-trained model is trained to obtain the target perception model, the to-be-trained model can be trained based on the remaining Epoch to obtain the target perception model. The remaining Epoch can be the difference between the target Epoch and the preliminary Epoch, or any reasonable value less than the difference, and the present disclosure does not make specific limitations thereto.

[0050] For example, if the target Epoch is 1000 times, and the model training based on each set of hyperparameters is performed according to the target Epoch, it needs to run for 10 days with 8 GPU cards. Then, if the preliminary Epoch is 200, and the value of N is 243, after constructing 243 sets of real training hyperparameters including the preliminary Epoch, based on the 243 sets of real training hyperparameters, the 243 initial perception models are trained respectively to obtain 243 candidate perception models, which needs to occupy:

[0051] 8 GPU cards * [(200 / 1000) * 10] days * 243 groups * 24 hours = 93312 card hours

[0052] Subsequently, if the remaining Epoch is based on the N candidate perception models, the performance of the best trained model is trained to obtain the target perception model, and the remaining Epoch is the difference between the target Epoch and the initial Epoch, that is, the remaining Epoch is 800 times, and it also needs to occupy:

[0053] 8 GPU cards * [(800 / 1000) * 10] days * 1 group * 24 hours = 1536 card hours

[0054] That is, the perception model training method provided by the embodiment of the present disclosure can obtain the target perception model only by (93312 card hours + 1536 card hours) = 94848 card hours, and compared with the existing model training method, the model training time is shortened by (466560 card hours - 94848 card hours) / 466560 card hours = 80%, and the operation resources are also saved.

[0055] The perception model training method provided by the embodiment of the present disclosure can obtain the initial Epoch, and construct N groups of real training hyperparameters including the initial Epoch, and then train the N initial perception models based on the N groups of real training hyperparameters to obtain N candidate perception models. Subsequently, only the best trained model in the N candidate perception models can be trained to obtain the target perception model, without training the other N-1 candidate perception models. Therefore, the model training time and operation resources consumed by training the other N-1 candidate perception models are saved, especially for the perception model with a large training data set, which can effectively reduce the model training time and reduce the occupation of operation resources.

[0056] In some optional embodiments, "obtaining the initial Epoch" can include the following steps:

[0057] Constructing M groups of experimental hyperparameters, M is an integer greater than or equal to 2;

[0058] Training the M initial experimental models based on the M groups of experimental hyperparameters to obtain real-time training results;

[0059] Based on the real-time training results, determining the experimental Epoch corresponding to the performance evaluation index of the M initial experimental models satisfying the preset index requirement as the initial Epoch.

[0060] Each set of experimental hyperparameters can include hyperparameter values corresponding to hyperparameter items such as a learning rate, a Batch_Size, a number of network layers, a weight initialization method, and an activation function. The real-time training result can be used to represent the change in the performance evaluation indicator with respect to the Epoch. In this case, the preset indicator requirement can be that the change rate of the performance value of the performance evaluation indicator is lower than a preset change rate threshold. The preset change rate threshold can be set according to actual model application requirements, and the present disclosure does not make specific limitations on this. In addition, in the present disclosure, the performance evaluation indicator can be one of a Loss, a model overall error, a model precision, a recall rate, and the like. The present disclosure also does not make limitations on this.

[0061] Based on the above description, it can be understood that, in the present disclosure, when the performance evaluation indicators of the M initial experimental models meet the preset indicator requirements, it indicates that the performance evaluation indicators have become stable and will not fluctuate greatly in the future. Therefore, when the performance evaluation indicators of the M initial experimental models meet the preset indicator requirements, the relative good or bad of the subsequent training effect can be predicted.

[0062] It should be further noted that, in the present disclosure, when it is determined based on the real-time training result that the performance evaluation indicators of the M initial experimental models meet the preset indicator requirements, the training of the M initial experimental models can be stopped, so as to effectively reduce the model training time in the experimental stage, and reduce the occupation of computing resources.

[0063] Through the above steps included in the "obtaining a preliminary Epoch", in the present disclosure, the preliminary Epoch is determined based on actual training experiments. Therefore, after training the N initial perception models based on the N sets of experimental hyperparameters including the preliminary Epoch, the subsequent training effect of the N candidate perception models can be reliably predicted according to the current performance of the N candidate perception models, that is, it can be ensured that the to-be-trained model is the candidate perception model with the optimal plasticity among the N candidate perception models. On the one hand, after training the to-be-trained model to obtain the target perception model, the overall performance of the target perception model can be improved to meet the actual model application requirements. On the other hand, since the target perception model can meet the actual model application requirements, it is not necessary to retrain, thereby reducing the training time of the perception model and reducing the occupation of computing resources.

[0064] In some optional embodiments, "constructing M sets of experimental hyperparameters" can include the following steps:

[0065] determining a plurality of experimental hyperparameter items;

[0066] for each experimental hyperparameter item, setting a plurality of experimental hyperparameter values corresponding to the experimental hyperparameter item;

[0067] constructing multiple sets of experimental hyperparameters related to the dominant hyperparameter, each set of experimental hyperparameters including an experimental hyperparameter value of the dominant hyperparameter and fixed hyperparameter values of multiple other hyperparameters, the dominant hyperparameter being any of the multiple experimental hyperparameters;

[0068] integrating the multiple sets of experimental hyperparameters related to each experimental hyperparameter to obtain M sets of experimental hyperparameters.

[0069] As described before, each set of experimental hyperparameters can include hyperparameter values corresponding to the learning rate, Batch_Size, network layer number, weight initialization method, and activation function, etc. hyperparameters in the embodiments of the present disclosure. Based on this, it can be understood that the multiple experimental hyperparameters in the embodiments of the present disclosure can also include the learning rate, Batch_Size, network layer number, weight initialization method, and activation function, etc. hyperparameters.

[0070] For example, the multiple experimental hyperparameters include the learning rate, Batch_Size, network layer number, weight initialization method, and activation function, and for each experimental hyperparameter, 10 experimental hyperparameter values corresponding to the experimental hyperparameter are set, which can be shown in Table 2 as follows:

[0071] Table 2

[0072]

[0073]

[0074] Taking the learning rate as the dominant hyperparameter, 10 sets of experimental hyperparameters related to the learning rate are constructed, each set of experimental hyperparameters including an experimental hyperparameter value of the learning rate and fixed hyperparameter values of multiple other hyperparameters, that is, the other hyperparameters all take fixed hyperparameter values. For example, the 10 sets of experimental hyperparameters can be:

[0075] Experimental hyperparameter LR01: learning rate (A201) & Batch_Size (B201) & network layer number (C201) & weight initialization method (D201) & activation function (E201);

[0076] Experimental hyperparameter LR02: learning rate (A202) & Batch_Size (B201) & network layer number (C201) & weight initialization method (D201) & activation function (E201);

[0077] Experimental hyperparameter LR03: learning rate (A203) & Batch_Size (B201) & network layer number (C201) & weight initialization method (D201) & activation function (E201);

[0078] Experimental hyperparameters LR04: learning rate (A204) & Batch_Size (B201) & network layers (C201) & weight initialization method (D201) & activation function (E201);

[0079] Experimental hyperparameters LR05: learning rate (A205) & Batch_Size (B201) & network layers (C201) & weight initialization method (D201) & activation function (E201);

[0080] Experimental hyperparameters LR06: learning rate (A206) & Batch_Size (B201) & network layers (C201) & weight initialization method (D201) & activation function (E201);

[0081] Experimental hyperparameters LR07: learning rate (A207) & Batch_Size (B201) & network layers (C201) & weight initialization method (D201) & activation function (E201);

[0082] Experimental hyperparameters LR08: learning rate (A208) & Batch_Size (B201) & network layers (C201) & weight initialization method (D201) & activation function (E201);

[0083] Experimental hyperparameters LR09: learning rate (A209) & Batch_Size (B201) & network layers (C201) & weight initialization method (D201) & activation function (E201);

[0084] Experimental hyperparameters LR10: learning rate (A210) & Batch_Size (B201) & network layers (C201) & weight initialization method (D201) & activation function (E201).

[0085] By analogy, 10 groups of experimental hyperparameters related to Batch_Size, 10 groups of experimental hyperparameters related to network layers, 10 groups of experimental hyperparameters related to weight initialization method, and 10 groups of experimental hyperparameters related to activation function can also be constructed. Finally, 10 groups of experimental hyperparameters related to each experimental hyperparameter are integrated to obtain 50 groups of experimental hyperparameters, that is, the value of M is 50 at this time.

[0086] Through the above steps included in the “constructing the M groups of experimental hyperparameters”, in the embodiments of the present disclosure, after determining the multiple experimental hyperparameters, firstly, for each experimental hyperparameter, multiple experimental hyperparameter values corresponding to the experimental hyperparameter are set, then multiple groups of experimental hyperparameters related to the dominant hyperparameter are constructed, and multiple groups of experimental hyperparameters related to each experimental hyperparameter are integrated to obtain the M groups of experimental hyperparameters. Therefore, the comprehensiveness of the M groups of experimental hyperparameters can be ensured, so as to further improve the reliability of the preliminary Epoch, thereby improving the overall performance of the target perception model.

[0087] In some optional embodiments, “determining, based on the real-time training result, an experimental Epoch at which the performance evaluation indicator of the M initial experimental models meets the preset indicator requirement” can include the following steps:

[0088] Based on the real-time training result, determining a candidate Epoch corresponding to when the performance evaluation indicator of each initial experimental model reaches the preset indicator requirement, to obtain M candidate Epochs;

[0089] Selecting the maximum Epoch from the M candidate Epochs as the experimental Epoch.

[0090] As described before, in the embodiments of the present disclosure, the real-time training result can be used to represent the change of the performance evaluation indicator with respect to the Epoch, and the preset indicator requirement can be that the change rate of the performance value of the performance evaluation indicator is lower than the preset change rate threshold.

[0091] Please refer to Figure 3 For example, the value of M is 10, the performance evaluation indicator is Loss, and the real-time training result includes the change of Loss with respect to Epoch during the training of 10 initial experimental models (for convenience, hereinafter referred to as initial experimental model 301 to initial experimental model 310) based on the 10 groups of experimental hyperparameters. Assume that before the Epoch reaches 200 times, the change rates of the performance values of the performance evaluation indicator Loss of 9 initial experimental models (initial experimental model 301 to initial experimental model 309) have all been lower than the preset change rate threshold, and the change rate of the performance value of the performance evaluation indicator Loss of the initial experimental model 310 is lower than the preset change rate threshold only when the Epoch reaches 200 times. Therefore, 200 times can be taken as the experimental Epoch.

[0092] By the above steps included in "determining, based on the real-time training result, an experiment Epoch corresponding to a situation that the performance evaluation index of the M initial experiment models meets the preset index requirement", in the embodiment of the present disclosure, the candidate Epoch corresponding to the situation that the performance evaluation index of each initial experiment model meets the preset index requirement can be determined based on the real-time training result, so as to obtain M candidate Epochs, and the maximum Epoch is selected from the M candidate Epochs as the experiment Epoch, that is, the preliminary Epoch. In this way, it can be further ensured that after training the N initial perception models based on the N groups of training hyperparameters including the preliminary Epoch, respectively, to obtain N candidate perception models, the subsequent training effect of the N candidate perception models can be reliably predicted according to the current performance of the N candidate perception models, so as to ensure that the model to be trained is the candidate perception model with the optimal plasticity among the N candidate perception models, thereby further improving the overall performance of the target perception model.

[0093] In some optional embodiments, "constructing N groups of training hyperparameters including the preliminary Epoch" can include the following steps:

[0094] determining a plurality of training hyperparameters;

[0095] for each training hyperparameter, setting a plurality of training hyperparameter values corresponding to the training hyperparameter;

[0096] arranging and combining the plurality of training hyperparameter values corresponding to the plurality of training hyperparameters in a full permutation and combination manner to obtain N groups of to-be-processed hyperparameters;

[0097] adding the preliminary Epoch to each group of to-be-processed hyperparameters to obtain N groups of training hyperparameters.

[0098] The plurality of training hyperparameters can include learning rate, Batch_Size, network layer number, weight initialization method, and activation function.

[0099] For example, the plurality of training hyperparameters include learning rate, Batch_Size, network layer number, weight initialization method, and activation function, and for each training hyperparameter, three training hyperparameter values corresponding to the training hyperparameter are set, which can be shown in Table 3 as follows:

[0100] Table 3

[0101]

[0102] Arranging and combining the three training hyperparameter values corresponding to the above five training hyperparameters in a full permutation and combination manner can obtain 3 5 = 243 groups of to-be-processed hyperparameters.

[0103] add the preliminary Epoch to each set of to-be-processed hyperparameters to obtain 243 sets of training hyperparameters. For example, the 243 sets of training hyperparameters can be:

[0104] Experimental hyperparameter HY001: learning rate (A301) & Batch_Size (B301) & network layer number (C301) & weight initialization method (D301) & activation function (E301) & Epoch (preliminary Epoch);

[0105] Experimental hyperparameter HY002: learning rate (A301) & Batch_Size (B301) & network layer number (C301) & weight initialization method (D301) & activation function (E302) & Epoch (preliminary Epoch);

[0106] Experimental hyperparameter HY003: learning rate (A301) & Batch_Size (B301) & network layer number (C301) & weight initialization method (D301) & activation function (E303) & Epoch (preliminary Epoch);

[0107] …

[0108] Experimental hyperparameter HY243: learning rate (A303) & Batch_Size (B303) & network layer number (C303) & weight initialization method (D303) & activation function (E303) & Epoch (preliminary Epoch).

[0109] Through the above steps of "constructing N sets of training hyperparameters including the preliminary Epoch", in the embodiments of the present disclosure, a plurality of training hyperparameter items can be determined, for each training hyperparameter item, a plurality of training hyperparameter values corresponding to the training hyperparameter item are set, and the plurality of training hyperparameter values corresponding to the plurality of training hyperparameter items are arranged and combined in a full permutation combination manner, N sets of to-be-processed hyperparameters are obtained, and the preliminary Epoch is added to each set of to-be-processed hyperparameters to obtain N sets of training hyperparameters, thereby ensuring the comprehensiveness of the N sets of training hyperparameters, so that the to-be-trained model selected from the N candidate perception models has higher performance, thereby further improving the overall performance of the target perception model.

[0110] In some optional embodiments, "setting a plurality of training hyperparameter values corresponding to the training hyperparameter item" can include the following steps:

[0111] In response to determining that the training hyperparameter item belongs to a numerical value type hyperparameter item, the number of data types of the sample perception data in the training data set is obtained;

[0112] According to the number of data types, a hyperparameter value setting range is determined, and the size of the hyperparameter value setting range is in a positive proportional relationship with the size of the number of data types.

[0113] In the range of the hyperparameter value setting, a plurality of real training hyperparameter values corresponding to the real training hyperparameter items are set.

[0114] As described above, in the embodiments of the present disclosure, the plurality of real training hyperparameter items can include learning rate, Batch_Size, network layer number, weight initialization method, and activation function, etc. hyperparameter items. Among them, learning rate, Batch_Size and network layer number belong to numerical value type hyperparameter items.

[0115] In addition, in the embodiments of the present disclosure, the training data set is a sample data set used to train N initial perception models and a to-be-trained model, and the sample perception data in the training data set can be collected by a plurality of perception modules arranged on the autonomous vehicle. For example, it can be collected by imaging units, laser radars, millimeter wave radars, ultrasonic radars, etc. perception modules arranged on the autonomous vehicle.

[0116] The training data set can include at least one of the six categories of sample perception data, i.e. road surface data, static obstacle data, dynamic obstacle data, pedestrian data, vehicle data and weather data, thereby ensuring that the training data set has higher diversity and complexity. Among them, the sample perception data belonging to the road surface data type can at least include 6 small categories of sample data related to road infrastructure such as roads, bridges, tunnels, intersections, lane lines and roads along the road; the sample perception data belonging to the static obstacle data type can at least include 5 small categories of sample data such as road edges, trees, green belts, buildings and traffic indicating props; the sample perception data belonging to the dynamic obstacle data type can at least include 4 small categories of sample data such as electric vehicles, tricycles, bicycles and animals; and the sample perception data belonging to the weather data type can at least include 5 small categories of sample data such as rain, snow, freezing, hail and fog.

[0117] In a specific example, when calculating the number of data types of sample perception data, the number of large categories of sample perception data can be calculated. For example, the data types of sample perception data only include road surface data, static obstacle data, dynamic obstacle data, pedestrian data, vehicle data and weather data, and the number of data types is 6.

[0118] In another specific example, when calculating the number of data types of sample perception data, it can be the number of small classes of sample perception data. For example, the sample perception data belonging to the road surface data type only includes 6 small classes of sample data such as roads, bridges, tunnels, intersections, lane lines, and road along; the sample perception data belonging to the static obstacle data type only includes 5 small classes of sample data such as road along, trees, green belts, buildings, and traffic indicating props; the sample perception data belonging to the dynamic obstacle data type only includes 4 small classes of sample data such as electric vehicles, tricycles, bicycles, and animals; and the sample perception data belonging to the meteorological data type only includes 5 small classes of sample data such as rain, snow, freezing, hail, and fog. Then, adding the pedestrian data and the driving data without smaller classification, the number of data types is 22.

[0119] After obtaining the number of data types, the hyperparameter value setting range can be determined according to the number of data types, so that the size of the hyperparameter value setting range is in a positive proportional relationship with the size of the number of data types. In a specific example, after obtaining the number of data types, the target number interval in which the number of data types is located can be further determined, and then the hyperparameter value setting range corresponding to the target number interval and the training hyperparameter item can be queried from the pre-set range query file. For example, the training hyperparameter item is the learning rate, when the number of data types is 5, the target number interval in which the number of data types is located can be further determined as [1, 5], and then the hyperparameter value setting range corresponding to the target number interval [1, 5] and the learning rate is queried from the range query file as (0~0.005]; when the number of data types is 10, the target number interval in which the number of data types is located can be further determined as [6, 10], and then the hyperparameter value setting range corresponding to the target number interval [6, 10] and the learning rate is queried from the range query file as (0~0.01]. For another example, the training hyperparameter item is Batch_Size, when the number of data types is 5, the target number interval in which the number of data types is located can be further determined as [1, 5], and then the hyperparameter value setting range corresponding to the target number interval [1, 5] and Batch_Size is queried from the range query file as (0~20]; when the number of data types is 10, the target number interval in which the number of data types is located can be further determined as [6, 10], and then the hyperparameter value setting range corresponding to the target number interval [6, 10] and Batch_Size is queried from the range query file as (0~40].

[0120] After determining the hyperparameter value setting range, a plurality of training hyperparameter values corresponding to the training hyperparameter item can be set in the hyperparameter value setting range, and the specific number of training hyperparameters can be in a positive proportional relationship with the size of the hyperparameter value setting range.

[0121] Through the above steps included in the step of setting the plurality of training hyperparameter values corresponding to the training hyperparameter item, in the embodiment of the present disclosure, for the training hyperparameter item belonging to the numerical type, the number of data types of the sample perception data in the training data set can also be obtained, and according to the number of data types, a hyperparameter value setting range in a positive proportional relationship with the number of data types is determined, and then within the hyperparameter value setting range, the plurality of training hyperparameter values corresponding to the training hyperparameter item are set, so as to ensure that in the case that the training data set has higher diversity and complexity, the N groups of training hyperparameters obtained also have higher diversity. In this way, it can be further ensured that the to-be-trained model selected from the N candidate perception models has higher performance, so as to improve the overall performance of the target perception model.

[0122] In addition, as mentioned above, in the embodiment of the present disclosure, the remaining Epoch can be the difference between the target Epoch and the preliminary Epoch, or any reasonable value smaller than the difference. Based on this, in the embodiment of the present disclosure, the step of continuing to train the to-be-trained model with the best performance in the N candidate perception models to obtain the target perception model can include the following steps:

[0123] obtaining a target Epoch, the target Epoch being greater than the preliminary Epoch;

[0124] calculating the difference between the target Epoch and the preliminary Epoch as a remaining Epoch;

[0125] continuing to train the to-be-trained model with the best performance in the N candidate perception models based on the remaining Epoch to obtain the target perception model.

[0126] The target Epoch is the optimal Epoch determined according to historical training experience. For example, the target Epoch is 1000 times, and the preliminary Epoch is 200, so the remaining Epoch is the difference between the target Epoch and the preliminary Epoch, that is, the remaining Epoch is 800 times.

[0127] Through the above steps included in the step of continuing to train the to-be-trained model with the best performance in the N candidate perception models to obtain the target perception model, in the embodiment of the present disclosure, since the target Epoch is the optimal Epoch determined according to historical training experience, calculating the difference between the target Epoch and the preliminary Epoch as the remaining Epoch can improve the rationality of the remaining Epoch, which can not only avoid that the total Epoch is too small, resulting in insufficient model training, so that the target perception model cannot converge to the best state, or even cannot converge, but also avoid that the total Epoch is too large, resulting in waste of computing resources or overfitting of the target perception model.

[0128] In some optional embodiments, the perception model training method further comprises the following steps for selecting the to-be-trained model with the best performance from the N candidate perception models:

[0129] The performance value of each candidate perception model under each performance evaluation index is obtained respectively;

[0130] For each performance evaluation index, the N candidate perception models are ranked in descending order of the performance value;

[0131] The candidate perception model with the highest ranking frequency is selected from the N candidate perception models as the to-be-trained model.

[0132] The performance evaluation indexes can include Loss, model overall error, model accuracy, recall rate, etc. In the case of N=243, i.e., there are 243 candidate perception models, for each performance evaluation index, the 243 candidate perception models are ranked in descending order of the performance value, and the candidate perception model with the highest ranking frequency is selected from the 243 candidate perception models as the to-be-trained model.

[0133] For example, for Loss, the 243 candidate perception models are ranked in descending order of the performance value, and the ranking result is as follows:

[0134] candidate perception model MO1> candidate perception model MO2> candidate perception model MO3>…;

[0135] For model overall error, the 243 candidate perception models are ranked in descending order of the performance value, and the ranking result is as follows:

[0136] candidate perception model MO2> candidate perception model MO1> candidate perception model MO3>…;

[0137] For model accuracy, the 243 candidate perception models are ranked in descending order of the performance value, and the ranking result is as follows:

[0138] candidate perception model MO1> candidate perception model MO3> candidate perception model MO1>…;

[0139] For recall rate, the 243 candidate perception models are ranked in descending order of the performance value, and the ranking result is as follows:

[0140] candidate perception model MO1> candidate perception model MO3> candidate perception model MO2>…;

[0141] Therefore, the candidate perception model MO1 with the highest frequency of performance ranking in the top can be selected from the 243 candidate perception models as the to-be-trained model.

[0142] Through the above steps included in the perception model training method, in the embodiments of the present disclosure, the performance value of each candidate perception model under multiple performance evaluation indicators can be obtained respectively, and for each performance evaluation indicator, the N candidate perception models are ranked in performance from high to low, and then the candidate perception model with the highest frequency of performance ranking in the top is selected from the N candidate perception models as the to-be-trained model, thereby ensuring the comprehensive superiority of the to-be-trained model in multiple performance evaluation dimensions, so that after the to-be-trained model is trained to obtain the target perception model, the overall performance of the target perception model can be further improved.

[0143] In the following, the perception model training method provided by the embodiments of the present disclosure will be described in combination with Figure 4 The complete flow of the perception model training method provided by the embodiments of the present disclosure will be described.

[0144] Firstly, M sets of experimental hyperparameters are constructed, and based on the M sets of experimental hyperparameters, M initial experimental models are trained respectively to obtain real-time training results, and then based on the real-time training results, the experimental Epoch corresponding to the performance evaluation indicators of the M initial experimental models meeting the preset indicator requirements is determined as the preliminary Epoch.

[0145] Thereafter, N sets of real training hyperparameters including the preliminary Epoch are constructed, and based on the N sets of real training hyperparameters, N initial perception models are trained respectively to obtain N candidate perception models, and then the to-be-trained model with the best performance is selected from the N candidate perception models.

[0146] Finally, the target Epoch is obtained, and the iteration number difference between the target Epoch and the preliminary Epoch is calculated as the remaining Epoch, and then based on the remaining Epoch, the to-be-trained model is further trained to obtain the target perception model. The target Epoch is the optimal Epoch determined according to historical training experience, and the target perception model is a deep learning model used for identifying environmental perception data to obtain environmental recognition results.

[0147] Please refer to Figure 5 The application scenario diagram of the perception model training method provided by the embodiments of the present disclosure.

[0148] As described above, the perception model training method provided by the embodiments of the present disclosure is applied to an electronic device. The electronic device is intended to represent various forms of digital computers, such as servers, blade servers, workstations, mainframes, desktop computers, laptop computers, or other suitable computers.

[0149] In the embodiments of the present disclosure, the electronic device can be used to execute the perception model training method.

[0150] an initial epoch is obtained;

[0151] N sets of real training hyperparameters including the initial epoch are constructed, N being an integer greater than or equal to 2;

[0152] Based on the N sets of real training hyperparameters, the N initial perception models are trained respectively to obtain N candidate perception models;

[0153] The to-be-trained model with the best performance in the N candidate perception models is continuously trained to obtain a target perception model, the target perception model being a deep learning model used for identifying environment perception data to obtain an environment recognition result.

[0154] It should be noted that in the embodiments of the present disclosure, the N initial perception models and the to-be-trained model are actually required to be trained by a training data set, and the sample perception data in the training data set can be collected by a plurality of perception modules arranged on an autonomous vehicle. For example, the sample perception data can be collected by imaging units, laser radars, millimeter wave radars, ultrasonic radars, and the like arranged on the autonomous vehicle. Based on this, it can be understood that in the application stage of the target perception model, the environment perception data can be real environment data collected by the plurality of perception modules during the driving process of the autonomous vehicle.

[0155] In addition, it should also be noted that in the embodiments of the present disclosure, Figure 5 The scene schematic diagram shown is only illustrative and not restrictive, and those skilled in the art can make various obvious changes and / or replacements based on the Figure 5 Examples, and the technical solutions obtained still belong to the disclosure range of the embodiments of the present disclosure.

[0156] In order to better implement the above perception model training method, the embodiments of the present disclosure also provide a perception model training device 600, which can be integrated in an electronic device. In the following, a perception model training device 600 provided by the disclosed embodiments will be described in conjunction with the structure schematic diagram shown. Figure 6

[0157] The perception model training device 600 comprises:

[0158] The iteration number obtaining unit 601 is configured to obtain an initial epoch;

[0159] The hyperparameter construction unit 602 is configured to construct N sets of real training hyperparameters including the initial epoch, N being an integer greater than or equal to 2; ​

[0160] The preliminary training unit 603 is configured to train the N initial perception models respectively based on the N sets of real training hyperparameters, to obtain N candidate perception models.

[0161] The continuous training unit 604 is configured to continue training the to-be-trained model with the best performance in the N candidate perception models, to obtain a target perception model, the target perception model being a deep learning model used for identifying the environment perception data to obtain an environment recognition result.

[0162] In some optional embodiments, the iteration number acquisition unit 601 is configured to:

[0163] construct M sets of experimental hyperparameters, M being an integer greater than or equal to 2;

[0164] train the M initial experimental models respectively based on the M sets of experimental hyperparameters, to obtain real-time training results;

[0165] based on the real-time training results, determine an experimental Epoch corresponding to a case where the performance evaluation indicators of the M initial experimental models meet preset indicator requirements, as the preliminary Epoch.

[0166] In some optional embodiments, the iteration number acquisition unit 601 is configured to:

[0167] determine a plurality of experimental hyperparameter items;

[0168] for each experimental hyperparameter item, set a plurality of experimental hyperparameter values corresponding to the experimental hyperparameter item;

[0169] construct a plurality of sets of experimental hyperparameters related to a dominant hyperparameter item, each set of experimental hyperparameters including an experimental hyperparameter value of the dominant hyperparameter item and fixed hyperparameter values of a plurality of other hyperparameter items, the dominant hyperparameter item being any one of the plurality of experimental hyperparameter items;

[0170] integrate the plurality of sets of experimental hyperparameters related to each experimental hyperparameter item, to obtain the M sets of experimental hyperparameters.

[0171] In some optional embodiments, the iteration number acquisition unit 601 is configured to:

[0172] based on the real-time training results, determine a candidate Epoch corresponding to a case where the performance evaluation indicator of each initial experimental model meets preset indicator requirements, to obtain M candidate Epochs;

[0173] select a maximum Epoch from the M candidate Epochs, as the experimental Epoch.

[0174] In some optional embodiments, the hyperparameter construction unit 602 is configured to:

[0175] determining a plurality of training hyperparameters;

[0176] For each training hyperparameter, a plurality of training hyperparameter values corresponding to the training hyperparameter are set;

[0177] According to the full permutation and combination mode, the plurality of training hyperparameter values corresponding to the plurality of training hyperparameters are permuted and combined to obtain N groups of to-be-processed hyperparameters;

[0178] A preliminary epoch is added to each group of to-be-processed hyperparameters to obtain N groups of training hyperparameters.

[0179] In some optional embodiments, the hyperparameter construction unit 602 is configured to:

[0180] In response to determining that the training hyperparameter belongs to a numerical hyperparameter, the number of data types of the sample perception data in the training data set is obtained;

[0181] According to the number of data types, a hyperparameter value setting range is determined, and the size of the hyperparameter value setting range is in a positive proportional relationship with the size of the number of data types;

[0182] Within the hyperparameter value setting range, a plurality of training hyperparameter values corresponding to the training hyperparameter are set.

[0183] In some optional embodiments, the data types of the sample perception data include at least one of road surface data, static obstacle data, dynamic obstacle data, pedestrian data, driving data, and weather data.

[0184] In some optional embodiments, the continuous training unit 604 is configured to:

[0185] A target epoch is obtained, and the target epoch is greater than the preliminary epoch;

[0186] The difference in the number of iterations between the target epoch and the preliminary epoch is calculated as a remaining epoch;

[0187] Based on the remaining epoch, the to-be-trained model with the best performance in the N candidate perception models is continuously trained to obtain a target perception model.

[0188] In some optional embodiments, the perception model training apparatus 600 further includes a model selection unit configured to:

[0189] The performance value of each candidate perception model under each performance evaluation index is obtained respectively;

[0190] For each performance evaluation index, the N candidate perception models are sorted in performance according to the order from high to low of the performance value;

[0191] Select the candidate perception model with the highest frequency of performance ranking from the N candidate perception models as the to-be-trained model.

[0192] In practice, the above modules can be implemented as independent entities, or combined as the same or several entities, and the specific implementation of the above modules can refer to the perception model training method embodiments described above, which will not be repeated here.

[0193] The perception model training device provided in the embodiments of the present disclosure can obtain the preliminary Epoch, and construct N groups of real training hyperparameters including the preliminary Epoch, and then train the N initial perception models based on the N groups of real training hyperparameters to obtain N candidate perception models. Subsequently, only the to-be-trained model with the best performance in the N candidate perception models needs to be trained to obtain the target perception model, and the other N-1 candidate perception models do not need to be trained. Therefore, the model training time and computing resources consumed for training the other N-1 candidate perception models are saved. Especially for a perception model with a large training data set, the model training time can be effectively reduced, and the occupation of computing resources is also reduced.

[0194] In the technical solutions of the present disclosure, the acquisition, storage and application of user personal information comply with relevant laws and regulations and do not violate public order and good customs.

[0195] According to the embodiments of the present disclosure, the present disclosure further provides an electronic device, a storage medium and a computer program product.

[0196] Figure 7 A schematic block diagram of an example electronic device 700 that can be used to implement embodiments of the present disclosure is shown.

[0197] As described previously, the electronic device in the embodiments of the present disclosure is intended to represent a variety of forms of digital computers, such as laptops, desktops, workstations, personal digital processors, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent a variety of forms of mobile devices, such as personal digital processors, cellular telephones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not intended to limit implementations of the present disclosure described and / or claimed in this document.

[0198] As Figure 7As shown, the electronic device 700 includes a computing unit 701 that can perform various appropriate actions and processes in accordance with a computer program stored in a Read-Only Memory (ROM) 702 or a computer program loaded into a Random Access Memory (RAM) 703 from a storage unit 708. Various programs and data required for the operation of the electronic device 700 can also be stored in the RAM 703. The computing unit 701, the ROM 702, and the RAM 703 are connected to each other through a bus 704. An Input / Output (I / O) interface 705 is also connected to the bus 704.

[0199] A plurality of components in the electronic device 700 are connected to the I / O interface 705, including an input unit 706, e.g., a keyboard, a mouse, etc., an output unit 707, e.g., various types of displays, a speaker, etc., a storage unit 708, e.g., a magnetic disk, an optical disk, etc., and a communication unit 709, e.g., a network card, a modem, a wireless communication transceiver, etc. The communication unit 709 allows the electronic device 700 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.

[0200] The computing unit 701 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various special-purpose Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Process (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 701 performs various methods and processes described above, e.g., the perception model training method. For example, the perception model training method can be implemented as a computer software program, respectively, which is tangibly embodied in a non-transitory computer-readable storage medium, e.g., the storage unit 708, in some optional embodiments. In some optional embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device 700 via the ROM 702 and / or the communication unit 709. When the computer program is loaded into the RAM 703 and executed by the computing unit 701, one or more steps of the perception model training method described above can be performed. Alternatively, in other embodiments, the computing unit 701 can be configured to perform the perception model training method by any other appropriate means, e.g., by means of firmware.

[0201] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a complex programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0202] Program code for carrying out methods of the present disclosure can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, enables the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can execute entirely on a machine, partly on the machine, as a stand-alone software package, partly on the machine and partly on a remote machine or entirely on the remote machine or server.

[0203] In the context of this disclosure, a non-transitory computer-readable storage medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A non-transitory computer- readable storage medium can be a machine-readable signal medium or a machine-readable storage medium. A non-transitory computer-readable storage medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a non-transitory computer-readable storage medium can include, but are not limited to, an electrical connection based on one or more wires, a portable computer disk, a hard disk, a RAM, a ROM, an erasable programmable read-only memory (EPROM) or a flash memory, an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0204] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a Cathode Ray Tube (CRT) monitor or a Liquid Crystal Display (LCD)) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0205] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0206] The computer system can include clients and servers. The clients and the servers are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The servers can be cloud servers, servers of a distributed system, or servers combined with a blockchain.

[0207] The embodiments of the present disclosure further provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to execute the aforementioned perception model training method.

[0208] The embodiments of the present disclosure further provide a computer program product comprising a computer program which, when executed by a processor, implements the aforementioned perception model training method.

[0209] The electronic device, the storage medium and the computer program product provided by the embodiments of the present disclosure can obtain a preliminary Epoch, construct N groups of training hyperparameters including the preliminary Epoch, and then train N initial perception models based on the N groups of training hyperparameters to obtain N candidate perception models. Subsequently, only the to-be-trained model with the best performance among the N candidate perception models needs to be trained to obtain the target perception model, and the other N-1 candidate perception models do not need to be trained. Therefore, the model training time and the computing resources consumed for training the other N-1 candidate perception models in the later stage are saved. Especially for a perception model with a large training data set, the model training time can be effectively reduced, and the occupation of computing resources is also reduced.

[0210] It should be understood that the steps shown in the above forms of flowcharts can be reordered, added, or deleted. For example, the steps described in the present disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, which is not limited herein. In addition, in the present disclosure, relationship terms such as "first", "second", "third", etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Furthermore, in the present disclosure, "a plurality of" can be understood as at least two.

[0211] The foregoing detailed description does not constitute a limitation of the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A perception model training method, comprising: obtaining a preliminary training iteration number; constructing N sets of real training hyperparameters including the preliminary training iteration number, N being an integer greater than or equal to 2; training N initial perception models based on the N sets of real training hyperparameters respectively to obtain N candidate perception models; continuing to train a to-be-trained model with the best performance in the N candidate perception models to obtain a target perception model, the target perception model being a deep learning model for identifying environment perception data to obtain an environment recognition result; the environment perception data being real environment data collected by multiple perception modules during driving of an autonomous vehicle; wherein the obtaining of the preliminary training iteration number comprises: constructing M sets of experimental hyperparameters, M being an integer greater than or equal to 2; training M initial experimental models based on the M sets of experimental hyperparameters respectively to obtain real-time training results; determining, based on the real-time training results, an experimental training iteration number corresponding to a case where performance evaluation indexes of the M initial experimental models meet preset index requirements as the preliminary training iteration number; wherein the constructing of the M sets of experimental hyperparameters comprises: determining multiple experimental hyperparameter items; for each experimental hyperparameter item, setting multiple experimental hyperparameter values corresponding to the experimental hyperparameter item; constructing multiple sets of experimental hyperparameters related to a dominant hyperparameter item, each set of experimental hyperparameters including one experimental hyperparameter value of the dominant hyperparameter item and fixed hyperparameter values of multiple other hyperparameter items, the dominant hyperparameter item being any one of the multiple experimental hyperparameter items; integrating multiple sets of experimental hyperparameters related to each experimental hyperparameter item to obtain the M sets of experimental hyperparameters.

2. The method of claim 1, wherein, The determining, based on the real-time training results, of an experimental training iteration number corresponding to a case where performance evaluation indexes of the M initial experimental models meet preset index requirements comprises: determining, based on the real-time training results, a candidate training iteration number corresponding to a case where a performance evaluation index of each initial experimental model meets the preset index requirements to obtain M candidate training iteration numbers; selecting a maximum training iteration number from the M candidate training iteration numbers as the experimental training iteration number.

3. The method of claim 1, wherein, The constructing of N sets of real training hyperparameters including the preliminary training iteration number comprises: determining multiple real training hyperparameter items; for each real training hyperparameter item, setting multiple real training hyperparameter values corresponding to the real training hyperparameter item; arranging and combining the multiple real training hyperparameter values corresponding to the multiple real training hyperparameter items in a full permutation and combination manner to obtain N sets of to-be-processed hyperparameters; adding the preliminary training iteration number to each set of to-be-processed hyperparameters to obtain the N sets of real training hyperparameters.

4. The method of claim 3, wherein, The setting of multiple real training hyperparameter values corresponding to the real training hyperparameter item comprises: in response to determining that the real training hyperparameter item belongs to a numerical value type hyperparameter item, obtaining a data type number of sample perception data in a training data set; determining a hyperparameter value setting range according to the data type number, the size of the hyperparameter value setting range being in a positive proportional relationship with the size of the data type number; setting the multiple real training hyperparameter values corresponding to the real training hyperparameter item within the hyperparameter value setting range.

5. The method of claim 4, wherein, The data type of the sample perception data includes at least one of road surface data, static obstacle data, dynamic obstacle data, pedestrian data, driving data and weather data.

6. The method of claim 1, wherein, The continuing training of the to-be-trained model with the best performance in the N candidate perception models obtains a target perception model, including: Obtaining a target training iteration number, which is greater than the preliminary training iteration number; Calculating the iteration number difference between the target training iteration number and the preliminary training iteration number as a remaining training iteration number; Based on the remaining training iteration number, the to-be-trained model with the best performance in the N candidate perception models is continuously trained to obtain a target perception model.

7. The method of claim 1 or 6, further comprising: respectively obtaining the performance value of each candidate perception model under each performance evaluation index; for each performance evaluation index, the N candidate perception models are sorted in descending order of the performance value; selecting the candidate perception model with the most times of high performance from the N candidate perception models as the to-be-trained model.

8. A perception model training device, comprising: an iteration number acquisition unit configured to acquire a preliminary training iteration number; a hyperparameter construction unit configured to construct N sets of real training hyperparameters including the preliminary training iteration number, N being an integer greater than or equal to 2; a preliminary training unit configured to train N initial perception models based on the N sets of real training hyperparameters to obtain N candidate perception models; a continuous training unit configured to continuously train a to-be-trained model with the best performance in the N candidate perception models to obtain a target perception model, the target perception model being a deep learning model for identifying environmental perception data to obtain an environmental recognition result; the environmental perception data being real environmental data collected by multiple perception modules during driving of an autonomous vehicle; wherein the iteration number acquisition unit is configured to: construct M sets of experimental hyperparameters, M being an integer greater than or equal to 2; train M initial experimental models based on the M sets of experimental hyperparameters to obtain real-time training results; determine the experimental training iteration number corresponding to the performance evaluation index of the M initial experimental models satisfying the preset index requirement as the preliminary training iteration number based on the real-time training results; wherein the iteration number acquisition unit is configured to: determine a plurality of experimental hyperparameter items; for each experimental hyperparameter item, set a plurality of experimental hyperparameter values corresponding to the experimental hyperparameter item; construct a plurality of sets of experimental hyperparameters related to the dominant hyperparameter item, each set of experimental hyperparameters including an experimental hyperparameter value of the dominant hyperparameter item and fixed hyperparameter values of a plurality of other hyperparameter items, the dominant hyperparameter item being any one of the plurality of experimental hyperparameter items; integrate the plurality of sets of experimental hyperparameters related to each experimental hyperparameter item to obtain the M sets of experimental hyperparameters.

9. The apparatus of claim 8, wherein, the iteration number acquisition unit is configured to: determine, based on the real-time training result, a candidate training iteration number corresponding to a performance evaluation index of each initial experimental model reaching the preset index requirement, to obtain M candidate training iteration numbers; select a maximum training iteration number from the M candidate training iteration numbers as the experimental training iteration number.

10. The apparatus of claim 8, wherein, The super parameter construction unit is configured to: determine a plurality of real training super parameter items; for each real training super parameter item, set a plurality of real training super parameter values corresponding to the real training super parameter item; arrange and combine the plurality of real training super parameter values corresponding to the plurality of real training super parameter items in a full permutation and combination manner to obtain N groups of to-be-processed super parameters; add the preliminary training iteration number to each group of to-be-processed super parameters to obtain the N groups of real training super parameters.

11. The apparatus of claim 10, wherein, The super parameter construction unit is configured to: obtain a data type number of sample perception data in a training data set; determine a super parameter value setting range according to the data type number, the size of the super parameter value setting range being in a positive proportional relationship with the size of the data type number; set the plurality of real training super parameter values corresponding to the real training super parameter item within the super parameter value setting range.

12. The apparatus of claim 11, wherein, The data types of the sample perception data include at least one of road surface data, static obstacle data, dynamic obstacle data, pedestrian data, driving data, and weather data.

13. The apparatus of claim 8, wherein, The continuous training unit is configured to: obtain a target training iteration number, the target training iteration number being greater than the preliminary training iteration number; calculate an iteration number difference between the target training iteration number and the preliminary training iteration number as a remaining training iteration number; based on the remaining training iteration number, continue to train the to-be-trained model with the best performance among the N candidate perception models to obtain a target perception model.

14. The apparatus of claim 8 or 13, further comprising a model selection unit configured to: obtain a performance value of each candidate perception model under a plurality of performance evaluation indexes, respectively; for each performance evaluation index, perform performance sorting on the N candidate perception models in a descending order of the performance value; select a candidate perception model with a highest number of performance sorting from the N candidate perception models as the to-be-trained model.

15. An electronic device, comprising: at least one processor; a memory in communication connection with the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

16. A non-transitory computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to enable the computer to perform the method of any one of claims 1-7.

17. A computer program product comprising a computer program which, when executed by a processor, implements the method of any one of claims 1-7.

17. A computer program product comprising a computer program which, when executed by a processor, implements the method of any one of claims 1-7.

Images