WO2022205835A1 - 深度学习模型生产系统、电子设备和存储介质 - Google Patents

深度学习模型生产系统、电子设备和存储介质 Download PDF

Info

Publication number
WO2022205835A1
WO2022205835A1 PCT/CN2021/124453 CN2021124453W WO2022205835A1 WO 2022205835 A1 WO2022205835 A1 WO 2022205835A1 CN 2021124453 W CN2021124453 W CN 2021124453W WO 2022205835 A1 WO2022205835 A1 WO 2022205835A1
Authority
WO
WIPO (PCT)
Prior art keywords
production line
deep learning
learning model
setting
model
Prior art date
Application number
PCT/CN2021/124453
Other languages
English (en)
French (fr)
Inventor
林达华
曹阳
李兆松
张行程
陈恺
杨冠姝
Original Assignee
上海商汤科技开发有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 上海商汤科技开发有限公司 filed Critical 上海商汤科技开发有限公司
Publication of WO2022205835A1 publication Critical patent/WO2022205835A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Definitions

  • the present disclosure relates to the field of computer technology, and in particular, to a deep learning model production system, an electronic device and a storage medium.
  • Deep learning technology has a wide range of applications in computer vision, natural language processing, speech recognition, recommendation systems, etc. When applying deep learning technology, it usually relies on professional technicians to customize the required deep learning models for different application scenarios, such as customizing data collection methods, network type selection, network parameter configuration, etc.
  • the present disclosure proposes a technical solution for producing a deep learning model.
  • a deep learning model production system comprising: a production line selection module for displaying a setting list of the selected model production line in response to a selection operation for a model production line in the production line list, wherein the The selected model production line is used to generate the target deep learning model to be deployed, the setting list includes the first setting item of the selected model production line, and the first setting item is used to set the target deep learning model, the target The training method of the deep learning model and the data set used for training the target deep learning model, the production line list is used to display the production line information of the model production line for selection; the setting module is used for responding to the first in the setting list.
  • a setting operation of a setting item determining a target deep learning model for realizing the target task, a training method for the target deep learning model, and a data set for training the target deep learning model; the training module is used to respond to the target deep learning model.
  • the training trigger operation of the target deep learning model is to train the target deep learning model according to the data set and the training method to obtain the target deep learning model to be deployed.
  • the setting operation of the first setting item includes: a data set setting operation, a network type setting operation, and a training mode setting operation; the response is directed to the first setting item in the setting list.
  • the setting operation determining the target deep learning model for realizing the target task, the training method of the target deep learning model, and the data set for training the target deep learning model, including: in response to the target depth
  • the network type setting operation of the learning model determining the target deep learning model for realizing the target task; in response to the data set setting operation for the data set of the target task, determining the data used for training the target deep learning model set; in response to a training mode setting operation for the target deep learning model, determining the training mode of the target deep learning model.
  • the training method includes a training device for performing training, a training end indicator, and a specified size of sample data; the target depth is determined according to the data set and the training method.
  • the learning model is trained to obtain the target deep learning model to be deployed, including: adjusting the size of the sample data in the data set according to the specified size to obtain an adjusted data set; according to the adjusted data set and the training End index, train the target deep learning model in the training device to obtain the target deep learning model to be deployed.
  • the setting list further includes a second setting item, where the second setting item is used to set packaging parameters of the target deep learning model to be deployed, and the packaging parameters include packaging methods and/or deployment device
  • the user development platform further includes: a deployment module, configured to encapsulate the target deep learning model to be deployed in response to the setting operation for the second setting item in the setting list, and obtain the The encapsulation file of the target deep learning model to be deployed
  • the setting operation of the second setting item includes: the setting operation of the encapsulation mode and/or the deployment device, and the encapsulation file is used to deploy in the deployment device The target deep learning model to be deployed.
  • the setting list further includes a third setting item for importing a dataset, a fourth setting item for labeling a dataset, and evaluating the target deep learning model to be deployed the fifth setting item
  • the user development platform further includes: an import module, configured to obtain the original data set of the target task in response to the import operation for the third setting item in the setting list, the original data set
  • the sample data is the data that meets the preset data collection standard, and the preset data collection standard is used to indicate the collection of the sample data in the original data set
  • the labeling module is used for responding to the first in the setting list.
  • the evaluation module is used to respond to the target deep learning model.
  • the setting operation of the fifth setting item in the setting list shows the performance evaluation result of the target deep learning model to be deployed according to the set network evaluation index, and the network evaluation index is used for the deep learning of the to-be-deployed performance evaluation of the model.
  • the user development platform further includes a production line purchase item for entering the production line store platform in response to a click operation on the production line purchase item.
  • the system further includes a production line store platform, the system further includes the production line store platform, and information transmission is performed between the production line store platform and the user development platform through a communication interface, so The production line store platform described above is used to sell model production lines.
  • the production line information of the successfully purchased model production line is added to the production line list.
  • the system further includes an expert development platform, and information transmission is performed between the expert development platform and the production line store platform through a communication interface
  • the expert development platform includes: a production line building module for responding to In the construction operation for the model production line, a pre-built model production line is obtained, and the pre-built model production line is used to generate the target deep learning model to be deployed; the production line publishing module is used to respond to the pre-built model production line.
  • the publishing operation is to publish the pre-built model production line to the production line store platform, so as to display and purchase the pre-built model production line on the production line store platform.
  • the building operation includes a network configuration operation for the model production line;
  • the obtaining a pre-built model production line in response to the building operation for the model production line includes: responding to the model production line A network configuration operation to obtain at least one deep learning model, where the network configuration operation includes an operation of configuring at least one of a network structure, a network algorithm, and an algorithm parameter; pre-training the at least one deep learning model, The pre-trained deep learning model in the pre-built model production line is obtained, and the pre-trained deep learning model corresponds to the network type to be set in the setting module of the user development platform.
  • the construction operation further includes: an information editing operation, a training configuration operation, a standard editing operation, and an index configuration operation for the model production line;
  • the obtaining of the pre-built model production line in response to the construction operation for the model production line further includes: in response to the information editing operation for the model production line, obtaining the production line information of the pre-built model production line, and the production line information is used for Identifying the pre-built model production line in the production line store platform and the user development platform; in response to the training configuration operation for the model production line, obtaining the training mode to be set in the pre-built model production line, The training mode to be set is used for setting in the setting module of the user development platform; in response to the standard configuration operation for the model production line, the preset data collection standard and the preset data collection standard of the pre-built model production line are obtained.
  • a data labeling standard is set, wherein the preset data collection standard is used to be displayed in the interface of the import module of the user development platform, and the preset data labeling standard is used to be displayed in the labeling module of the user development platform.
  • the interface in response to the index configuration operation for the model production line, the network evaluation index to be set in the pre-built model production line is obtained, and the network evaluation index to be set is used in the evaluation module of the user development platform Make settings.
  • the target task includes an image processing task
  • the image processing task includes at least one of image recognition, image segmentation, image classification, and key point detection.
  • an electronic device comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to invoke the instructions stored in the memory to execute the above system.
  • a computer-readable storage medium having computer program instructions stored thereon, the computer program instructions implementing the above-described system when executed by a processor.
  • a computer program comprising computer readable code, when the computer readable code is executed in an electronic device, a processor in the electronic device executes the above-described system.
  • the user development platform can implement a streamlined setting operation for generating a target deep learning model to be deployed based on a pre-built model production line, so that when an operation is triggered in response to the training of the target deep learning model, It can efficiently generate the target deep learning model to be deployed based on the process-based setting operation. Meet the industry's needs for efficient, streamlined, and automated generation of neural networks.
  • FIG. 1 shows a structural diagram of a deep learning model production system according to an embodiment of the present disclosure.
  • FIG. 2 illustrates a block diagram of a user development platform according to an embodiment of the present disclosure.
  • FIG. 3 shows a schematic diagram of a production line management interface according to an embodiment of the present disclosure.
  • FIG. 4 shows a schematic diagram of a production line list implemented in accordance with the present disclosure.
  • FIG. 5 shows a schematic diagram of a setting interface of a first setting item according to an embodiment of the present disclosure.
  • FIG. 6 shows a schematic diagram of a setting interface of a second setting item according to an embodiment of the present disclosure.
  • FIG. 7 shows a schematic diagram of a setting interface of a third setting item according to an embodiment of the present disclosure.
  • FIG. 8 shows a schematic diagram of a setting interface of a fourth setting item according to an embodiment of the present disclosure.
  • FIG. 9 shows a schematic diagram of a setting interface of a fifth setting item according to an embodiment of the present disclosure.
  • FIG. 10 illustrates a block diagram of a production line store platform in accordance with an embodiment of the present disclosure.
  • FIG. 11 shows a schematic diagram of a production line display interface according to an embodiment of the present disclosure.
  • FIG. 12 illustrates a block diagram of an expert development platform according to an embodiment of the present disclosure.
  • FIG. 13 shows a schematic diagram of a management interface of a model production line according to an embodiment of the present disclosure.
  • FIG. 14 shows a schematic diagram of a network configuration interface according to an embodiment of the present disclosure.
  • FIG. 15 shows a schematic diagram of an information editing interface according to an embodiment of the present disclosure.
  • FIG. 16 shows a schematic diagram of a training configuration interface according to an embodiment of the present disclosure.
  • FIG. 17 shows a schematic diagram of a standard configuration interface according to an embodiment of the present disclosure.
  • FIG. 18 shows a block diagram of an electronic device according to an embodiment of the present disclosure.
  • FIG. 19 shows a block diagram of an electronic device according to an embodiment of the present disclosure.
  • FIG. 1 shows a structural diagram of a deep learning model production system according to an embodiment of the present disclosure.
  • the deep learning model production system includes:
  • the expert development platform 11 is used to build a model production line, and publish the pre-built or built model production line to the production line store platform 12, and the pre-built or built model production line is used to generate the target deep learning model to be deployed;
  • the production line store platform 12 is used to display and sell pre-built or built model production lines, so as to add the purchased model production lines to the user development platform 13;
  • the user development platform 13 is used to generate the target deep learning model to be deployed based on the purchased model production line.
  • the deep learning model production system may execute the method steps in the system on an electronic device such as a terminal device or a server, and the terminal device may be a user equipment (User Equipment, UE), a mobile device, a user terminal , terminal, cellular phone, cordless phone, personal digital assistant (Personal Digital Assistant, PDA), handheld device, computing device, vehicle-mounted device, wearable device, etc.
  • the method steps in the system can be called by the processor in the terminal device
  • the method steps in the system may be implemented by means of computer readable instructions stored in a memory, or by a server.
  • information is transmitted between the production line store platform 12 and the user development platform 13 through a communication interface, and information is transmitted between the expert development platform 11 and the production line store platform 12 through a communication interface. That is, the expert development platform 11 , the production line store platform 12 , and the user development platform 13 can implement information exchange by invoking the communication interface.
  • the expert development platform 11 , the production line store platform 12 , and the user development platform 13 may be independently developed application programs, or may be integrated application programs, which are not limited by the embodiments of the present disclosure.
  • Those skilled in the art can use any known technology to realize the development of the expert development platform 11 , the production line store platform 12 , and the user development platform 13 , which are not limited to the embodiments of the present disclosure.
  • an expert development platform for professional technicians to build a model production line a store user platform for purchasing pre-built or built model production lines, and an ordinary user to facilitate the generation of pre-built or built model production lines.
  • a user development platform for the target deep learning model to be deployed so that the model production line built by professional technicians can be purchased and used by ordinary users, and a deep learning model that can be deployed on the device can be generated.
  • the efficiency and accuracy of the to-be-deployed deep learning model generated by the production line can approach the level of professional technicians and have high practicability and versatility.
  • the deep learning model production system includes a user development platform
  • FIG. 2 shows a block diagram of the user development platform according to an embodiment of the present disclosure.
  • the user development platform includes:
  • the production line selection module 131 is configured to display a setting list of the selected model production line in response to the selection operation for the model production line in the production line list.
  • the selected model production line is used to generate the target deep learning model to be deployed.
  • the model production line may be a vehicle detection model production line for generating vehicle detection models, etc., which is not limited by the embodiment of the present disclosure.
  • the setting list includes the first setting item of the selected model production line.
  • the first setting item is used to set the target deep learning model, the training method of the target deep learning model, and the data set used to train the target deep learning model. Line information to display model lines for selection.
  • the production line selection module may be provided with a production line list, and the production line display list may display the production line information of the model production line purchased by the user in the production line store, so that in response to the selection operation for the model production line, display A list of settings for the selected model line. It should be understood that, the setting list of the selected model production line may be displayed on the display interface of the terminal device.
  • a production line list such as production line name, production line identification, etc.
  • the production line list may include a trigger control in response to the selection operation, for example, a trigger control in response to a click operation or a touch operation, so as to determine the model production line selected by the user.
  • the implementation manner of the trigger control is not limited in this embodiment of the present disclosure.
  • FIG. 3 shows a schematic diagram of a production line management interface according to an embodiment of the present disclosure.
  • Figure 4 shows a schematic diagram of a production line list implemented in accordance with the present disclosure.
  • the production line list shown in FIG. 4 can be a list expanded by clicking the production line information “AAA” of the model production line shown in FIG. 3 , so that it is convenient for the user to select any model production line and respond to any model production line.
  • the selection operation displays a list of settings for the selected model line.
  • the production line list may include trigger controls in response to a click operation on the model production line, and controls (eg, sliders) for pulling up and down the list.
  • the display under “AAA” in FIG. 3 can be the setting list of the selected model production line, and the name or logo of each setting item of the model production line can be displayed in the setting list, in response to the user clicking on any setting item operation, the setting interface of the selected setting item is displayed, so that the user can realize the setting operation of generating the target deep learning model to be deployed in the displayed setting interface.
  • the setting list may include trigger controls in response to a click operation on a setting item.
  • first, second, third, etc. in the present disclosure can be used to distinguish different setting items, and do not limit the position of each setting item in the setting list. It should be understood that those skilled in the art can set the position of the first setting item in the setting list according to actual requirements.
  • setting item c in FIG. 3 may be the first setting item.
  • the setting module 132 is used to determine the target deep learning model for realizing the target task, the training method of the target deep learning model, and the data set for training the target deep learning model in response to the setting operation for the first setting item in the setting list .
  • the target tasks may include image processing tasks.
  • the image processing tasks include at least one of image recognition, image segmentation, image classification, and keypoint detection. It should be understood that the target tasks may also include speech processing tasks, such as speech recognition tasks, semantic recognition tasks, voiceprint recognition tasks, natural language processing tasks, and the like.
  • speech processing tasks such as speech recognition tasks, semantic recognition tasks, voiceprint recognition tasks, natural language processing tasks, and the like. The embodiment of the present disclosure does not limit the task type of the target task.
  • FIG. 5 shows a schematic diagram of a setting interface of a first setting item according to an embodiment of the present disclosure.
  • the setting interface of the first setting item can provide a drop-down box for setting the dataset, a drop-down box for setting the network type, a drop-down box for setting the training end indicator, and a drop-down box for setting the execution Drop-down box to set the training equipment for training.
  • the network type can be used to indicate the deep learning model under different network structures, network algorithms, and algorithm parameters, and the deep learning model indicated by the network type can be a pre-trained deep learning model.
  • the data set set in the first setting item may be the data set used to train the target deep learning model, and the data set may be a labeled data set. It is understood that a plurality of labeled data sets may be provided for set up.
  • the training method may include a training end indicator and a training device (for example, an image processing unit GPU, an X86 central processing unit CPU, etc.) for performing the training.
  • the training end indicator is used to indicate the condition for ending the training of the target deep learning model, for example, the number of iteration rounds reaches the set threshold, the iteration duration reaches the set threshold, and the like.
  • the training method of the target deep learning model can be determined by setting the training end indicator and training equipment.
  • the user can click each drop-down box in the setting interface to display the corresponding drop-down list, so that the user can set the target deep learning model, training method and data set according to actual needs.
  • the target deep learning model, training method, and data set set by the user are the determined target deep learning model, training method, and data set.
  • a related drop-down box control may be provided in the setting interface of the first setting item, so as to realize the setting operation for the first setting item, which is not limited by the embodiment of the present disclosure.
  • the setting operation for the first setting item may also be implemented through other types of controls, such as a multi-select box, which is not limited in this embodiment of the present disclosure.
  • the setting content displayed in the setting interface of the first setting item in FIG. 5 is an implementation manner provided by an embodiment of the present disclosure. It should be understood that the present disclosure should not be limited to this, and those skilled in the art can refine the settings of the target deep learning model, training method, and data set according to actual needs.
  • the data set can be divided into training set and test set, and the proportion of training set and test set can be set; when the training end indicator is set to stop by iteration rounds, the iteration round threshold can also be set to determine according to the The set iteration round threshold ends the training of the target deep learning model.
  • the setting content in the first setting item for generating the target deep learning model to be deployed may be determined according to actual requirements, which is not limited in this embodiment of the present disclosure.
  • the training module 133 is configured to train the target deep learning model according to the data set and the training method in response to the training trigger operation for the target deep learning model, and obtain the target deep learning model to be deployed.
  • the training trigger operation for the target deep learning model can be implemented, for example, by clicking the trigger control at "Start training" in the setting interface of the first setting item shown in FIG. 5 , by responding to According to the trigger operation of the trigger control, the training of the target deep learning model is started.
  • the style, location, and implementation manner of the trigger control can be determined according to actual requirements, which are not limited by the embodiments of the present disclosure.
  • the training modality may include an end-of-training indicator and a training device used to perform the training.
  • the target deep learning model is trained according to the data set and the training method to obtain the target deep learning model to be deployed, which may include: according to the data set and the training end indicator, in the set training equipment Execute the training of the target deep learning model to obtain the target deep learning model to be deployed.
  • the embodiment of the present disclosure does not limit the training process of the target deep learning model.
  • a streamlined setting operation for generating a target deep learning model to be deployed can be implemented, so that when an operation is triggered in response to the training of the target deep learning model, it can be efficiently based on The streamlined setting operation realizes the automatic generation of the target deep learning model to be deployed.
  • the setting operation for the target deep learning model, training method and training set can be performed in the setting interface of the first setting item.
  • the setting operation of the first setting item includes: data A set setting operation, a network type setting operation, and a training mode setting operation, in response to the setting operation for the first setting item in the setting list, determining the target deep learning model for realizing the target task, the training method of the target deep learning model, and The dataset used to train the target deep learning model, including:
  • the target deep learning model In response to the network type setting operation for the target deep learning model, determine the target deep learning model for realizing the target task; in response to the data set setting operation for the data set of the target task, determine the data set used for training the target deep learning model ; In response to the setting operation for the training mode of the target deep learning model, determine the training mode of the target deep learning model.
  • the network type can be used to indicate deep learning models under different network structures, network algorithms, and algorithm parameters, and the deep learning model indicated by the network type can be a pre-trained deep learning model.
  • the training method may include the training equipment used to perform the training and the training end indicator.
  • the training mode setting operation may include the setting operation for training equipment and training end indicators.
  • the training method may further include a specified size of the sample data.
  • the operation of setting the training mode may further include the operation of setting the specified size of the sample data.
  • the specified size may include the specified image resolution.
  • the sample data in the dataset may be adjusted by using the default specified size, in which case, the specified size of the sample data may not be set; or if the sample data is non-image data (such as In the case of voice), the specified size of the sample data is not set. It should be understood that, in these cases, the setting content of the specified size of the sample data may not be displayed in the setting interface of the first setting item.
  • the network type setting for the target deep learning model can be realized by clicking each drop-down box control provided in the setting interface of the first setting item, and triggering the drop-down list displayed by each drop-down box control. operations, dataset setting operations, and training mode setting operations.
  • the network type setting operation can be, for example, clicking the network type identifier displayed in the corresponding drop-down list
  • the data set setting operation can be, for example, clicking the data set identifier displayed in the corresponding drop-down list
  • the training method setting operation can be, for example, clicking the corresponding Training method identification in the drop-down list, etc.
  • trigger controls for each setting operation can be provided in the drop-down list, so as to determine the target deep learning model corresponding to the data set, training method, and network type selected by the user.
  • the target deep learning model, the data set for training the target deep learning model, and the training of the target deep learning model can be conveniently determined through the network setting operation, the data set setting operation, and the training mode setting operation. method, so as to facilitate and efficiently realize the automatic generation of the target deep learning model to be deployed.
  • the training method may include the training equipment used to perform the training, the training end index and the specified size of the sample data.
  • Train to get the target deep learning model to be deployed including:
  • the target deep learning model is trained in the training device, and the target deep learning model to be deployed is obtained.
  • the sample data may include image data
  • the specified size may include a specified image resolution.
  • Adjusting the size of the sample data in the dataset according to the specified size to obtain the adjusted dataset may include: adjusting the image resolution of the image data in the dataset to the specified image resolution to obtain the adjusted dataset.
  • any known image processing technology may be used to realize the adjustment of the image resolution of the image data, for example, normalization processing, scaling processing, etc., which is not limited by the embodiment of the present disclosure.
  • the training device used to perform the training may be the training device set for the training mode in the first setting item; it is understandable that the default training device may also be used, that is, the The training equipment is not set, and the implementation of the present disclosure is not limited.
  • the training of the target deep learning model in the training device means that the training of the target deep learning model is performed in the training device.
  • the training end indicator may be the training end indicator set for the training mode in the first setting item; it is understandable that the default training end indicator may also be used, that is, the training may not be terminated. Indicators are used for equipment, and the implementation of this disclosure is not limited.
  • the setting operation for the training mode in the first setting item may include the setting operation for at least one of the training equipment for performing training, the training end indicator and the specified size of the sample data.
  • the adjusted dataset, the set training end indicator, and the target deep learning model may be transmitted to a training device for training the target deep learning model, so as to perform the target depth analysis in the training device. Learning model training.
  • the target deep learning model is trained in the training device to obtain the target deep learning model to be deployed, which may include: combining samples in the adjusted data set The data is input into the target deep learning model, and an output result is obtained; according to the loss of the output result, the network parameters of the target deep learning model are adjusted; when the adjusted target deep learning model meets the training end index, the to-be-deployed target deep learning model is obtained.
  • Target deep learning model any known deep learning model training system may be used to train the target deep learning model, which is not limited to this embodiment of the present disclosure.
  • the training of the target deep learning model can be automatically realized according to the set training set and training method efficiently, and the target deep learning model to be deployed that meets the actual needs of the user is obtained.
  • the deep learning model obtained by training is usually packaged, that is, packaged, so as to realize the deployment of the deep learning model in related devices.
  • the setting list further includes a second setting item, and the second setting item is used to encapsulate the target deep learning model to be deployed, and the system further includes:
  • the target deep learning model to be deployed is encapsulated to obtain an encapsulation file of the target deep learning model to be deployed, wherein the setting operation of the second setting item includes: encapsulating the method And/or the setting operation of the deployment device, the package file is used to deploy the target deep learning model to be deployed in the deployment device.
  • the encapsulation method may indicate an inference framework required to encapsulate the target deep learning model to be deployed, and the inference framework can adapt the target deep learning model to be deployed to various deployment devices.
  • Inference frameworks such as TensorRT (a high-performance deep learning model inference engine (Inference) launched by NVIDIA, used to deploy applications of deep learning models), OpenVINO (a tool launched by Intel for deploying deep learning models) set).
  • the deployment device may indicate the processor type of the device to be deployed by the target deep learning model to be deployed, for example, it may include: X86 processor, Arm processor, CUDA (Compute Unified Device Architecture) processor, GPU, etc.
  • the encapsulation file may be, for example, an SDK (Software Development Kit, software development kit), wherein any known encapsulation technology can be used to encapsulate the target deep learning model to be deployed, which is not limited in this embodiment of the present disclosure.
  • SDK Software Development Kit, software development kit
  • FIG. 6 shows a schematic diagram of a setting interface of a second setting item according to an embodiment of the present disclosure.
  • the setting interface of the second setting item can provide a drop-down box for setting the encapsulation method, a drop-down box for setting the deployment device, a drop-down box for setting the network version, and whether to quantify the deployment.
  • Set radio buttons can be provided.
  • the network version is used to select different versions of the target deep learning model to be deployed. It should be understood that multiple setting operations can be performed based on the above-mentioned first setting item, that is, the target deep learning model is trained multiple times to obtain Multiple versions of the target deep learning model to be deployed; whether the deployment is quantized is used to indicate whether the target deep learning model to be deployed is quantified, and the quantized target deep learning model is encapsulated. Wherein, any known quantification technology can be used to realize the quantification of the target deep learning model to be deployed, which is not limited in this embodiment of the present disclosure.
  • the user can display the corresponding drop-down list by clicking each drop-down box in the setting interface, so that the user can set the encapsulation mode and the deployment device according to actual needs.
  • the encapsulation mode, deployment device, etc. set by the user that is, the determined encapsulation mode, deployment device, and the like.
  • a related drop-down box control may be provided in the setting interface of the second setting item, so as to realize the setting operation for the second setting item, which is not limited by the embodiment of the present disclosure.
  • the setting operation for the second setting item may also be implemented through other types of controls, such as a multi-select box, which is not limited to this embodiment of the present disclosure.
  • a trigger operation for triggering the start of encapsulation of the target deep learning model to be deployed may also be included, for example, by clicking the setting shown in FIG. 6 .
  • the trigger control at "Confirm" in the interface triggers the encapsulation of the target deep learning model to be deployed. It should be understood that the style, location, and implementation manner of the trigger control can be determined according to actual requirements, which are not limited by the embodiments of the present disclosure.
  • the setting interface of the second setting item shown in FIG. 6 above is an implementation manner provided by an embodiment of the present disclosure.
  • Those skilled in the art can adjust the content set in the setting interface of the second setting item according to actual needs, which is not limited in this embodiment of the present disclosure.
  • the target deep learning model to be deployed may be quantified by default, and the radio button of whether to quantify the deployment may be removed in the setting interface of the second setting item.
  • various encapsulation requirements of the user for the target deep learning model to be deployed can be fulfilled, so that the encapsulated files obtained can be effectively deployed in various deployment devices.
  • the setting list further includes a third setting item for importing a dataset, a fourth setting item for labeling a dataset, and evaluating the target deep learning model to be deployed
  • the fifth setting item the user development platform also includes:
  • the import module is used to obtain the original data set of the target task in response to the import operation for the third setting item in the setting list.
  • the sample data in the original data set meets the preset data collection standard, and the preset data collection standard is used to indicate Collection of sample data in the original data set;
  • the labeling module is used for labeling the sample data in the original data set according to the preset data labeling standard in response to the labeling operation for the fourth setting item in the setting list, to obtain a data set for training the target deep learning model;
  • the evaluation module is used to display the performance evaluation result of the target deep learning model to be deployed according to the set network evaluation index in response to the setting operation for the fifth setting item in the setting list, and the network evaluation index is used for the to-be-deployed target deep learning model. Performance evaluation of deep learning models.
  • a preset data collection standard may be displayed in the setting interface of the third setting item to guide the user to collect an original sample set that meets the preset data collection standard.
  • the data collection standard of vehicle images may be Including: the shooting height is 3.5-7 meters, the front of the car and the whole body are shot, the resolution is 1080p; at least 100 images, each image does not exceed 100 vehicles, etc.
  • the setting interface of the third setting item may include a file upload control for importing datasets, for example, supporting importing datasets by dragging and dropping files, or supporting searching for datasets in The data set is imported by means of a local storage path, etc.
  • the embodiment of the present disclosure does not limit the import method of the data set.
  • Imported raw datasets can be transferred to other electronic devices and stored for automated labeling of raw datasets.
  • FIG. 7 shows a schematic diagram of a setting interface of a third setting item according to an embodiment of the present disclosure.
  • the local storage path of the original data set can be determined by clicking the control at "Select”; by clicking the "Import” button, it is triggered to obtain the original data set according to the storage path of the original data set, so as to realize the original data set set import operation.
  • the original data sets of the target task may include multiple ones.
  • a drop-down box for selecting the original data sets may be displayed in the setting interface of the fourth setting item, so as to facilitate the user Select any original dataset for labeling to obtain a dataset for training the target deep learning model.
  • the labeling operation for the fourth setting item may include a selection operation for the original data set.
  • the preset data labeling standard may be a preset data labeling standard, for example, a labeling box with a size smaller than 30*30 is filtered.
  • the preset data annotation standard can be displayed in the setting interface of the fourth setting item to inform the user of the preset data annotation standard.
  • a plurality of preset data annotation standards can also be provided for the user to select, and this disclosure is implemented in this disclosure. Examples are not limited.
  • any known data labeling technology can be used to label the sample data in the original data set according to preset data labeling standards.
  • the original technology set can be labelled through a data labeling algorithm. Automatic labeling.
  • the labeling operation for the fourth setting item may further include a triggering operation that triggers the start of labeling the sample data in the original data set.
  • a labeling tool for manual labeling (eg LabelMe: a Javascript labeling tool for online image labeling) may also be provided in the setting interface of the fourth setting item, so as to manually label the original data set Label.
  • the labeling operation for the fourth setting item may further include a labeling operation of manually labeling the original data set.
  • the labeling method of the original data set may be set according to actual requirements, which is not limited in this embodiment of the present disclosure.
  • FIG. 8 shows a schematic diagram of a setting interface of a fourth setting item according to an embodiment of the present disclosure.
  • the setting interface of the fourth setting item displays a drop-down box control for selecting the original data set.
  • the original data set displayed in the drop-down box may be the selected original data set; the user can Click the button at "Automatic Labeling" to automatically label the selected original data set; you can also click the button at "Manual Labeling" to trigger the display of the labeling tool for manual labeling.
  • the network evaluation index is used to evaluate the performance of the target deep learning model obtained by training.
  • the network evaluation index may include, for example, at least one of: accuracy rate, precision rate, recall rate, and F1-score (F1-Score).
  • the setting operation for the fifth setting item may include a setting operation for the network evaluation index, so as to display the performance evaluation result in response to the set network evaluation index.
  • FIG. 9 shows a schematic diagram of a setting interface of a fifth setting item according to an embodiment of the present disclosure.
  • the setting of network evaluation indicators can be implemented in the form of a multi-select box. Among them, the performance evaluation result can be triggered to be displayed by clicking the “Confirm” button in FIG. 9 .
  • the setting operation for the fifth setting item may include a triggering operation for triggering the display of the performance evaluation result.
  • the network evaluation indicator setting interface may further include an option of a confidence threshold, so that the user can view the performance evaluation results under different confidence thresholds.
  • the confidence level can be the confidence level of the output result of the deep learning model, for example, the confidence level of vehicle detection, which can represent the possibility that the detection result is indicated as a vehicle, and the confidence level threshold can be used to indicate that the display target deep learning model reaches the confidence level threshold performance evaluation results.
  • the performance evaluation results may be displayed in the form of lists, graphs (such as graphs), and the like.
  • the display form of the performance evaluation result may be set according to actual requirements, which is not limited in this embodiment of the present disclosure.
  • the setting interfaces of the third setting item, the fourth setting item, and the fifth setting item respectively shown in FIG. 7 , FIG. 8 , and FIG. 9 are an implementation manner provided by an embodiment of the present disclosure. It should be understood that the present disclosure should not be limited thereto, and those skilled in the art can design the setting interfaces of the third setting item, the fourth setting item and the fifth setting item according to actual needs, which is not limited by the embodiment of the present disclosure.
  • the user development platform further includes a production line purchase item for entering the production line store platform in response to a click operation on the production line purchase item. In this way, it is convenient for users to enter the production line store platform at any time to purchase pre-built or built model production lines.
  • the production line purchase item can be realized in the form of the entry button of the production line store platform; the production line purchase item can be set corresponding to the jump address of the production line store platform, so that when the user clicks the production line purchase item, he can enter the production line store based on the jump address. platform.
  • the embodiment of the present disclosure does not limit the implementation form of the production line purchase item.
  • the production line purchase item can be set at any position on any interface in the user development platform, for example, it can be set in the production line list, the setting list of the model production line, the setting of each setting item In the interface, this embodiment of the present disclosure is not limited.
  • the deep learning model production system also includes a production line store platform, the production line store platform and the user development platform perform information transmission through a communication interface, and the production line store platform is used to sell model production lines.
  • FIG. 10 shows a block diagram of a production line store platform according to an embodiment of the present disclosure. As shown in FIG. 10 , the production line store platform includes:
  • the production line display module 121 is used to display a pre-built or built model production line, and the pre-built or built model production line is a model production line built through an expert development platform.
  • FIG. 11 shows a schematic diagram of a production line display interface according to an embodiment of the present disclosure.
  • the production line display interface can provide a button “Buy” for purchasing a model production line, and the user can click the “Buy” button to enter payment page; users can also click the "Details” button in the production line display interface to view the detailed information of the selected model production line; they can also click the "Search" case to search for the model production line to be purchased.
  • the production line display interface shown in FIG. 11 is an implementation provided by the embodiment of the present disclosure. It should be understood that the present disclosure should not be limited to this, and those skilled in the art can design the production line display interface according to actual needs. , the embodiments of the present disclosure are not limited.
  • the production line purchase module 122 is configured to determine a purchased model production line in response to a purchase operation for the displayed model production line.
  • the purchase operation for the displayed model production line may include: a click operation for the "purchase” button displayed in the production line display interface as shown in FIG. 11 . After clicking the "Buy” button, you can enter the payment interface.
  • a payment control for online payment can be provided in the payment interface, and those skilled in the art can use any known technology to realize online payment for the model production line. The disclosed embodiments are not limiting.
  • a model production line that has completed payment is a model production line that has been successfully purchased.
  • the production line information of the successfully purchased model production line is added to the production line list of the user development platform, so that the user can use the successfully purchased model production line to generate The target deep learning model to be deployed.
  • the production line store platform may further include a production line sending module for sending the successfully purchased model production line to the user development platform, so as to display the successfully purchased model production line in the production line display list of the user development platform, Among them, the production line display list is used to display and select the successfully purchased model production line.
  • the user can purchase different model production lines through the production line store platform to generate target deep learning models to be deployed that achieve different target tasks.
  • the deep learning model production system may include an expert development platform.
  • FIG. 12 shows a block diagram of an expert development platform according to an embodiment of the present disclosure. As shown in FIG. 12 , the expert development platform includes:
  • the production line building module 111 is used for obtaining a pre-built or built model production line in response to a building operation for a model production line, and the pre-built or built model production line is used to generate a target deep learning model to be deployed.
  • FIG. 13 shows a schematic diagram of a management interface of a model production line according to an embodiment of the present disclosure.
  • the information of the built model production line can also be displayed in the management interface as shown in FIG. 13 .
  • the person in charge can be the name of the person who built the model production line.
  • the management interface may provide a control (such as a slider) for pulling up and down the list to display more model production lines.
  • the management interface of the model production line can also provide a search box for searching the model production line (the box on the left of "Search” in Figure 13) and a search trigger control ("Search" in Figure 13), to quickly search for the model production line; click the trigger control at the "Details” corresponding to any model production line in Figure 13 to display more detailed model production line information, click the trigger control at "Edit” in Figure 13, you can Edit the model production line, click the trigger control at "Publish” in Figure 13, and the pre-built or built model production line can be released to the production line store platform.
  • the way of entering the building interface of the model production line can be determined according to actual needs, which is not limited by the embodiment of the present disclosure. It should be understood that various types of controls on the interface may be provided in the construction interface, for example, an edit box, a drop-down box, a multi-select box, a radio box, etc., so as to realize the construction operation for the model production line.
  • the building operations for the model production line may include: network configuration operations, information editing operations, training configuration operations, standard editing operations, and index configuration operations for the model production line.
  • the network configuration operation is used to configure at least one pre-trained deep learning model in the model production line;
  • the information editing operation is used to edit the production line information of the model production line;
  • the training configuration operation is used to configure the model production line to be set.
  • the standard configuration operation is used to configure the preset data collection standard and the preset data labeling standard of the model production line;
  • the indicator configuration operation is used to configure the network evaluation indicators to be set in the model production line.
  • a pre-built or built model production line can be obtained.
  • the configuration results corresponding to each construction operation can be connected in series by using a known component series tool to obtain a pre-built or built model production line, which is not limited by the embodiment of the present disclosure.
  • the production line publishing module 112 is used for publishing the pre-built or built model production line to the production line store platform in response to the release operation for the pre-built or built model production line, so as to display and purchase the pre-built or built model production line on the production line store platform model production line.
  • the pre-built or built model production line can be displayed in the management interface as shown in FIG. 13 , and the pre-built or built model production line can be triggered by clicking the “Publish” button shown in FIG. 13 . Publish the model production line to the production line store platform. It should be understood that a person skilled in the art can set a release manner of a pre-built or built model production line according to actual requirements, which is not limited by the embodiment of the present disclosure.
  • information is transmitted between the expert development platform and the production line store platform through the communication interface, and the pre-built or built model production line is published to the production line store platform.
  • the configuration content is packaged and compressed, and the packaged and compressed files are sent to the production line store platform.
  • the embodiment of the present disclosure does not limit how to publish the pre-built or built model production line to the production line store platform.
  • an expert development platform can be used to build a model production line, so that ordinary users can use the model production line built by professional technicians to generate a target deep learning model to be deployed, and improve the performance of the target deep learning model to be deployed. productivity and precision.
  • the building operation includes a network configuration operation for the model production line.
  • a pre-built or built model production line is obtained, including:
  • Pre-train at least one deep learning model to obtain a pre-built or pre-built deep learning model in the model production line, where the pre-trained deep learning model corresponds to the network type to be set in the setting module of the user development platform.
  • the network structure may include the network structure of the backbone network backbone of the deep learning model, the fusion network neck and the branch network head.
  • the backbone network is used for feature extraction
  • the fusion network is used to fuse the features extracted by the backbone network
  • the branch network is used to perform inference on different tasks for the features extracted by the backbone network or the features of fusion processing.
  • the network structures of the backbone network backbone, the fusion network neck, and the branch network head may, for example, adopt at least: ResNet series, SENet series, MobileNet series, ShuffleNet series, and the like.
  • the network algorithm can at least include: fast convolution Faster R-CNN algorithm, keypoint convolution Keypoint R-CNN algorithm, stepped convolution Cascade R-CNN algorithm, normalization algorithm, etc.
  • the algorithm parameters include at least: batch size, sampling method, data enhancement method, network parameters, etc.
  • batch size can be understood as the number of sample data for one training; sampling methods can at least include: positive and negative sample sampling methods, hard sample (Hard Sample) sampling methods, under-sampling methods, over-sampling methods, etc., and data enhancement methods can be Including: random cropping, distortion, amplification, mirroring, deformation and other data enhancement methods; network parameters can include network hyperparameters, since there are many types of hyperparameters, they can be determined by selecting an existing parameter configuration file.
  • sampling methods can at least include: positive and negative sample sampling methods, hard sample (Hard Sample) sampling methods, under-sampling methods, over-sampling methods, etc.
  • data enhancement methods can be Including: random cropping, distortion, amplification, mirroring, deformation and other data enhancement methods
  • network parameters can include network hyperparameters, since there are many types of hyperparameters, they can be determined by selecting an existing parameter configuration file.
  • the network configuration operation for the model production line may not only include the configuration contents disclosed in the above-mentioned embodiments of the present disclosure;
  • the interactive implementation can also be implemented by invoking a related interface and invoking a related configuration file in the background, as long as a deep learning model can be implemented and configured, which is not limited in this embodiment of the present disclosure.
  • different deep learning models can be obtained in different configuration manners, and the network size, precision, and computing performance of different deep learning models can be different.
  • the network algorithm can be different; for the same network structure and network algorithm, the algorithm parameters can be different; a deep learning model with high precision is large in size; a deep learning model with small volume has low precision and good computing performance.
  • deep learning models with different precisions and different computing performances that is, deep learning models with different network types, can be configured, so that users can select the target deep learning model according to the precision requirements and performance requirements.
  • FIG. 14 shows a schematic diagram of a network configuration interface according to an embodiment of the present disclosure.
  • the user can realize the network configuration operation for the model production line by triggering the corresponding drop-down box, edit box, multi-select box and other controls.
  • the network configuration interface shown in FIG. 14 is a possible implementation, and those skilled in the art can set the configurable content in the network configuration interface, the layout, style, etc. of the network configuration interface according to actual needs, This embodiment of the present disclosure is not limited.
  • any known pre-training manner may be used to pre-train at least one deep learning model obtained by configuration to obtain a pre-trained deep learning model.
  • the training set used in the pre-training can be related to the target tasks and application scenarios of the model production line. For example, the training sets corresponding to the human detection task and the vehicle detection task are different, and the deep learning model obtained by training can also be used. different.
  • the pre-trained deep learning model corresponds to the network type to be set in the setting module of the user development platform, so that the user can determine the corresponding target deep learning model by setting the network type.
  • the target deep learning model set on the user development platform is a pre-trained deep learning model. In this way, a deep learning model with a certain accuracy can be obtained, thereby improving the training efficiency of the above-mentioned training target deep learning model, so as to efficiently generate the target deep learning model to be deployed.
  • multiple processing tasks such as vehicle detection, license plate detection, license plate recognition, etc.
  • vehicle detection, license plate detection, license plate recognition, etc. may be included in the same model production line.
  • corresponding at least one deep learning model can be configured and pre-trained.
  • At least one deep learning model can be established in advance for the model production line, and at least one deep learning model can be pre-trained to obtain a deep learning model for the user to choose, so that the user can set the Operation, that is, selecting the network type, can obtain the target deep learning model that can realize the target task, and improve the generation efficiency of the target deep learning model.
  • a model production line for generating the target deep learning model to be deployed is obtained, further comprising:
  • the production line information of the pre-built or built model production line is obtained, and the production line information is used to identify the pre-built or built model production line in the production line store platform and the user development platform;
  • the training mode to be set in the pre-built or built model production line is obtained, and the training mode to be set is used for setting in the setting module of the user development platform;
  • the preset data collection standard and the preset data labeling standard of the pre-built or built model production line are obtained, wherein the preset data collection standard is used to display the import module of the user development platform.
  • the preset data annotation standard is used to display in the interface of the annotation module of the user development platform;
  • the network evaluation index to be set in the pre-built or built model production line is obtained, and the network evaluation index to be set is used for setting in the evaluation module of the user development platform.
  • FIG. 15 shows a schematic diagram of an information editing interface according to an embodiment of the present disclosure.
  • the information editing interface can display information editing prompt information.
  • You can enter the task name, processing object, application scenario, data acquisition equipment and other information through the editing box, that is, to realize the information editing operation for the model production line.
  • the item information may include "task name: vehicle detection, processing object: vehicle, application scene: street, data collection device: camera".
  • the information input by the user in the edit box that is, the obtained production line information of the model production line, can be determined by clicking the “Confirm” button shown in FIG. 15 .
  • FIG. 16 shows a schematic diagram of a training configuration interface according to an embodiment of the present disclosure.
  • the training end indicator can be configured through the multi-select box. It should be understood that the training end indicator configured here is also the training end indicator for the user to set in the setting module of the user platform; The training equipment is configured through the multi-check box, and the training equipment configured here is also the training equipment for the user to set in the setting module of the user platform; in this way, multiple training end indicators and multiple training equipment can be configured. For users to choose any training end indicator and training equipment for training the target deep learning model according to their needs.
  • the training configuration interface may also provide an edit box for editing performance estimation information for the training device, where the performance estimation information may be, for example, the execution target depth of the training device.
  • the edited performance evaluation information of the training device can be displayed in the setting interface of the first setting item, so that the user can select the required training device to perform the training of the target deep learning model based on the displayed running time.
  • the training configuration operation for the model production line may include a multi-selection operation for training end indicators, a multi-selection operation for training equipment, and an editing operation for performance estimation information.
  • the user may click the “Confirm” button shown in FIG. 16 to determine the selected training end indicator and training equipment, as well as the edited performance estimation information, etc., which is not limited to this embodiment of the present disclosure.
  • the preset data collection standard and the preset data labeling standard can be respectively input through the edit box on the display interface, that is, the standard editing operation for the model production line can be realized; the relevant interface can also be called through the background, Realize the configuration of preset data collection standards and preset data labeling standards.
  • the configured preset data collection standard and preset data labeling standard can be displayed on the setting interface of the third setting item for importing the dataset and the setting interface for the fourth setting item for labeling the dataset, respectively. .
  • FIG. 17 shows a schematic diagram of a standard configuration interface according to an embodiment of the present disclosure.
  • the preset data collection standard and the preset data labeling standard can be respectively input through the edit box, and the data collection can also be configured through the drop-down box. standards and data labelling standards.
  • the user can determine the edited preset data collection standard and preset data labeling standard by clicking the “Confirm” button shown in FIG. 17 , which is not limited in this embodiment of the present disclosure.
  • the index configuration operation of the network evaluation index can be realized in the form of a multi-check box in the index configuration interface, wherein, for the index configuration interface, refer to the fifth setting item shown in FIG. 9 above.
  • Set interface the network evaluation index selected in the index configuration interface is the network evaluation index displayed in the setting interface of the fifth setting item, and the network evaluation index to be set in the setting interface of the fifth setting item can be. It is a part or all of the network evaluation indicators displayed in the indicator configuration interface, which is not limited in this embodiment of the present disclosure.
  • the construction interface of the model production line includes: network configuration interface, information editing interface, training configuration interface, standard editing interface and indicator configuration interface.
  • the existing component series tools can be used to Each configuration result is connected in series to build a model production line.
  • professional technicians can configure the content of each setting item in the model production line to build a model production line suitable for different scene requirements and different task requirements.
  • the target deep learning model to be deployed that meets the actual needs can be efficiently obtained.
  • At least one model production line and a recommended configuration contained therein can be established, wherein the recommended configuration may include, for example, at least a pre-trained deep learning model, data collection standards, data labeling standards, deep learning model structure, super Parameters, evaluation metrics, training configuration, inference configuration, etc. Therefore, it is convenient for the user to perform a streamlined setting operation based on the built model production line and the recommended configuration contained therein, so as to efficiently generate the target deep learning model to be deployed.
  • the generation process of deep learning models requires professional and technical personnel to have professional knowledge and experience, such as data collection methods, deep learning model selection, hyperparameters, pre-training models, data enhancement methods, number of iterations, iterations Expertise in the logic of the way etc.
  • knowledge and experience vary with the application scenarios and processing tasks of the deep learning model. Therefore, in practical application scenarios, the deep learning model generation process usually needs to be customized by professional technicians.
  • the migration and adaptability of deep learning models are weak. For example, a vehicle recognition network with very good accuracy is trained in one city, but in another city, the performance is poor due to changes in the distribution of vehicle models.
  • a configuration environment for a model production line for professional technicians and a setting environment for an ordinary user to generate a required deep learning model using the model production line.
  • professional technicians can define and generate a model production line for deep learning models, and each model production line can support the generation of deep learning models for at least one application scenario.
  • the model production line built by professional and technical personnel can be selected and used by ordinary users.
  • the setting environment of the model production line the generation process of the deep learning model is completed, and the deep learning model that can be deployed on the device is obtained. In this way, the efficiency and accuracy of the deep learning model generated by ordinary users based on the model production line can approach the level of professional technicians.
  • the generation process of deep learning models cannot meet the accuracy requirements of deep learning model training in various scenarios; and the migration ability of deep learning models in the field of deep learning is not strong enough, so users need to target different application scenarios. , re-collect data to train and validate the deep learning model. And the deep learning models generated by the same generation process in related technologies cannot automatically adapt to the generation needs of deep learning models in various scenarios.
  • the embodiments of the present disclosure by defining the application scenarios of the model production line, customizing the model production line in different application scenarios can be realized, and the generation of automated deep learning models in various application scenarios can be supported.
  • the single generation process of the deep learning model makes the provided deep learning model after training less accurate in similar scenarios.
  • different setting operations performed on the same model production line can make the generated deep learning model suitable for similar application scenarios, and obtain a higher precision deep learning model through the optimized setting process and training configuration.
  • different training sets can be used to train the same target deep learning model to efficiently obtain different target deep learning models to be deployed, which can be applied to similar vehicle detection in different cities and streets Scenes.
  • professional technicians can establish model production lines for different application scenarios, and ordinary users can use professional technologies to provide configuration generation processes and related configurations, which can satisfy various application scenarios and deep learning models of similar application scenarios.
  • To generate demand you can also obtain a high-precision, high-performance deep learning model that is close to the level of professional and technical personnel and suitable for scene requirements.
  • the embodiments of the present disclosure also provide a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the above-mentioned system is implemented.
  • the computer-readable storage medium may be a non-volatile computer-readable storage medium.
  • An embodiment of the present disclosure further provides an electronic device, including: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to call the instructions stored in the memory to execute the above-mentioned system. method steps.
  • Embodiments of the present disclosure also provide a computer program product, including computer-readable codes.
  • a processor in the device executes the production for implementing the deep learning model provided by any of the above embodiments. system commands.
  • Embodiments of the present disclosure also provide another computer program product for storing computer-readable instructions, which, when executed, cause the computer to perform the operations of the deep learning model production system provided by any of the foregoing embodiments.
  • Embodiments of the present disclosure also provide a computer program, including computer-readable codes, when the computer-readable codes are executed in an electronic device, a processor in the electronic device executes the above-mentioned system.
  • the electronic device may be provided as a terminal, server or other form of device.
  • FIG. 18 shows a block diagram of an electronic device 800 according to an embodiment of the present disclosure.
  • electronic device 800 may be a mobile phone, computer, digital broadcast terminal, messaging device, game console, tablet device, medical device, fitness device, personal digital assistant, etc. terminal.
  • an electronic device 800 may include one or more of the following components: a processing component 802, a memory 804, a power supply component 806, a multimedia component 808, an audio component 810, an input/output (I/O) interface 812, a sensor component 814 , and the communication component 816 .
  • the processing component 802 generally controls the overall operation of the electronic device 800, such as operations associated with display, phone calls, data communications, camera operations, and recording operations.
  • the processing component 802 may include one or more processors 820 to execute instructions to perform all or part of the steps of the system described above. Additionally, processing component 802 may include one or more modules that facilitate interaction between processing component 802 and other components. For example, processing component 802 may include a multimedia module to facilitate interaction between multimedia component 808 and processing component 802.
  • Memory 804 is configured to store various types of data to support operation at electronic device 800 . Examples of such data include instructions for any application or system operating on electronic device 800, contact data, phonebook data, messages, pictures, videos, and the like. Memory 804 may be implemented by any type of volatile or nonvolatile storage device or combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), Magnetic Memory, Flash Memory, Magnetic or Optical Disk.
  • SRAM static random access memory
  • EEPROM electrically erasable programmable read only memory
  • EPROM erasable Programmable Read Only Memory
  • PROM Programmable Read Only Memory
  • ROM Read Only Memory
  • Magnetic Memory Flash Memory
  • Magnetic or Optical Disk Magnetic Disk
  • Power supply assembly 806 provides power to various components of electronic device 800 .
  • Power supply components 806 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to electronic device 800 .
  • Multimedia component 808 includes a screen that provides an output interface between the electronic device 800 and the user.
  • the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user.
  • the touch panel includes one or more touch sensors to sense touch, swipe, and gestures on the touch panel. The touch sensor may not only sense the boundaries of a touch or swipe action, but also detect the duration and pressure associated with the touch or swipe action.
  • the multimedia component 808 includes a front-facing camera and/or a rear-facing camera. When the electronic device 800 is in an operation mode, such as a shooting mode or a video mode, the front camera and/or the rear camera may receive external multimedia data. Each of the front and rear cameras can be a fixed optical lens system or have focal length and optical zoom capability.
  • Audio component 810 is configured to output and/or input audio signals.
  • audio component 810 includes a microphone (MIC) that is configured to receive external audio signals when electronic device 800 is in operating modes, such as calling mode, recording mode, and voice recognition mode.
  • the received audio signal may be further stored in memory 804 or transmitted via communication component 816 .
  • audio component 810 also includes a speaker for outputting audio signals.
  • the I/O interface 812 provides an interface between the processing component 802 and a peripheral interface module, which may be a keyboard, a click wheel, a button, or the like. These buttons may include, but are not limited to: home button, volume buttons, start button, and lock button.
  • Sensor assembly 814 includes one or more sensors for providing status assessment of various aspects of electronic device 800 .
  • the sensor assembly 814 can detect the on/off state of the electronic device 800, the relative positioning of the components, such as the display and the keypad of the electronic device 800, the sensor assembly 814 can also detect the electronic device 800 or one of the electronic device 800 Changes in the position of components, presence or absence of user contact with the electronic device 800 , orientation or acceleration/deceleration of the electronic device 800 and changes in the temperature of the electronic device 800 .
  • Sensor assembly 814 may include a proximity sensor configured to detect the presence of nearby objects in the absence of any physical contact.
  • Sensor assembly 814 may also include a light sensor, such as a complementary metal oxide semiconductor (CMOS) or charge coupled device (CCD) image sensor, for use in imaging applications.
  • CMOS complementary metal oxide semiconductor
  • CCD charge coupled device
  • the sensor assembly 814 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.
  • Communication component 816 is configured to facilitate wired or wireless communication between electronic device 800 and other devices.
  • the electronic device 800 may access a wireless network based on a communication standard, such as wireless network (WiFi), second generation mobile communication technology (2G) or third generation mobile communication technology (3G), or a combination thereof.
  • the communication component 816 receives broadcast signals or broadcast related information from an external broadcast management system via a broadcast channel.
  • the communication component 816 also includes a near field communication (NFC) module to facilitate short-range communication.
  • the NFC module may be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.
  • RFID radio frequency identification
  • IrDA infrared data association
  • UWB ultra-wideband
  • Bluetooth Bluetooth
  • electronic device 800 may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable A programmed gate array (FPGA), controller, microcontroller, microprocessor or other electronic component implementation is used to perform the method steps in the above system.
  • ASICs application specific integrated circuits
  • DSPs digital signal processors
  • DSPDs digital signal processing devices
  • PLDs programmable logic devices
  • FPGA field programmable A programmed gate array
  • controller microcontroller, microprocessor or other electronic component implementation is used to perform the method steps in the above system.
  • a non-volatile computer-readable storage medium is also provided, such as a memory 804 including computer program instructions that can be executed by the processor 820 of the electronic device 800 to complete the above-mentioned system. method steps.
  • FIG. 19 shows a block diagram of an electronic device 1900 according to an embodiment of the present disclosure.
  • the electronic device 1900 may be provided as a server.
  • the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource, represented by memory 1932, for storing instructions executable by the processing component 1922, such as an application program.
  • An application program stored in memory 1932 may include one or more modules, each corresponding to a set of instructions.
  • the processing component 1922 is configured to execute instructions to perform method steps in the systems described above.
  • the electronic device 1900 may also include a power supply assembly 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input output (I/O) interface 1958 .
  • the electronic device 1900 can operate based on an operating system stored in the memory 1932, such as a Microsoft server operating system (Windows Server TM ), a graphical user interface based operating system (Mac OS X TM ) introduced by Apple, a multi-user multi-process computer operating system (Unix TM ), Free and Open Source Unix-like Operating System (Linux TM ), Open Source Unix-like Operating System (FreeBSD TM ) or the like.
  • Microsoft server operating system Windows Server TM
  • Mac OS X TM graphical user interface based operating system
  • Uniix TM multi-user multi-process computer operating system
  • Free and Open Source Unix-like Operating System Linux TM
  • FreeBSD TM Open Source Unix-like Operating System
  • a non-volatile computer-readable storage medium such as a memory 1932 comprising computer program instructions executable by the processing component 1922 of the electronic device 1900 to implement the method of the system described above step.
  • the present disclosure may be a system and/or a computer program product.
  • the computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the present disclosure.
  • a computer-readable storage medium may be a tangible device that can hold and store instructions for use by the instruction execution device.
  • the computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.
  • Non-exhaustive list of computer readable storage media include: portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM) or flash memory), static random access memory (SRAM), portable compact disk read only memory (CD-ROM), digital versatile disk (DVD), memory sticks, floppy disks, mechanically coded devices, such as printers with instructions stored thereon Hole cards or raised structures in grooves, and any suitable combination of the above.
  • RAM random access memory
  • ROM read only memory
  • EPROM erasable programmable read only memory
  • flash memory static random access memory
  • SRAM static random access memory
  • CD-ROM compact disk read only memory
  • DVD digital versatile disk
  • memory sticks floppy disks
  • mechanically coded devices such as printers with instructions stored thereon Hole cards or raised structures in grooves, and any suitable combination of the above.
  • Computer-readable storage media are not to be construed as transient signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (eg, light pulses through fiber optic cables), or through electrical wires transmitted electrical signals.
  • the computer readable program instructions described herein may be downloaded to various computing/processing devices from a computer readable storage medium, or to an external computer or external storage device over a network such as the Internet, a local area network, a wide area network, and/or a wireless network.
  • the network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and/or edge servers.
  • a network adapter card or network interface in each computing/processing device receives computer-readable program instructions from a network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing/processing device .
  • Computer program instructions for carrying out operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or instructions in one or more programming languages.
  • Source or object code written in any combination, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as the "C" language or similar programming languages.
  • the computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server implement.
  • the remote computer may be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (eg, using an Internet service provider through the Internet connect).
  • LAN local area network
  • WAN wide area network
  • custom electronic circuits such as programmable logic circuits, field programmable gate arrays (FPGAs), or programmable logic arrays (PLAs) can be personalized by utilizing state information of computer readable program instructions.
  • Computer readable program instructions are executed to implement various aspects of the present disclosure.
  • These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer or other programmable data processing apparatus to create a machine that causes the instructions when executed by the processor of the computer or other programmable data processing apparatus , resulting in means for implementing the functions/acts specified in one or more blocks of the flowchart and/or block diagrams.
  • These computer readable program instructions can also be stored in a computer readable storage medium, these instructions cause a computer, programmable data processing apparatus and/or other equipment to operate in a specific manner, so that the computer readable medium on which the instructions are stored includes An article of manufacture comprising instructions for implementing various aspects of the functions/acts specified in one or more blocks of the flowchart and/or block diagrams.
  • Computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other equipment to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other equipment to produce a computer-implemented process , thereby causing instructions executing on a computer, other programmable data processing apparatus, or other device to implement the functions/acts specified in one or more blocks of the flowcharts and/or block diagrams.
  • each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more functions for implementing the specified logical function(s) executable instructions.
  • the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved.
  • each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations can be implemented in dedicated hardware-based systems that perform the specified functions or actions , or can be implemented in a combination of dedicated hardware and computer instructions.
  • the computer program product can be specifically implemented by hardware, software or a combination thereof.
  • the computer program product is embodied as a computer storage medium, and in another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (Software Development Kit, SDK), etc. Wait.
  • a software development kit Software Development Kit, SDK

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Feedback Control In General (AREA)

Abstract

本公开涉及一种深度学习模型生产系统、电子设备和存储介质，所述系统包括用户开发平台，所述用户开发平台包括：生产线选择模块，用于响应于针对生产线列表中模型生产线的选择操作，展示选中的模型生产线的设置列表；设置模块，用于响应于针对设置列表中第一设置项的设置操作，确定用于实现目标任务的目标深度学习模型、目标深度学习模型的训练方式以及用于训练目标深度学习模型的数据集；训练模块，用于响应于针对目标深度学习模型的训练触发操作，根据数据集以及训练方式，对目标深度学习模型进行训练，得到待部署的目标深度学习模型。

Description

深度学习模型生产系统、电子设备和存储介质
本公开要求在2021年4月2日提交中国专利局、申请号为202110363166.X、申请名称为“深度学习模型生产系统、电子设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本公开中。
技术领域
本公开涉及计算机技术领域，尤其涉及一种深度学习模型生产系统、电子设备和存储介质。
背景技术
深度学习技术，在计算机视觉、自然语言处理、语音识别、推荐系统等方向上，具有广泛的应用。在应用深度学习技术时，通常依赖专业的技术人员针对不同的应用场景定制所需的深度学习模型，例如，定制数据采集方式、网络类型选择、网络参数配置等。
发明内容
本公开提出了一种深度学习模型生产技术方案。
根据本公开的一方面，提供了一种深度学习模型生产系统，包括：生产线选择模块，用于响应于针对生产线列表中模型生产线的选择操作，展示选中的模型生产线的设置列表，其中，所述选中的模型生产线用于生成待部署的目标深度学习模型，所述设置列表中包括所述选中的模型生产线的第一设置项，所述第一设置项用于设置目标深度学习模型、所述目标深度学习模型的训练方式以及用于训练所述目标深度学习模型的数据集，所述生产线列表用于展示供选择的模型生产线的生产线信息；设置模块，用于响应于针对所述设置列表中第一设置项的设置操作，确定用于实现目标任务的目标深度学习模型、所述目标深度学习模型的训练方式以及用于训练所述目标深度学习模型的数据集；训练模块，用于响应于针对所述目标深度学习模型的训练触发操作，根据所述数据集以及所述训练方式，对所述目标深度学习模型进行训练，得到待部署的目标深度学习模型。
在一种可能的实现方式中，所述第一设置项的设置操作，包括：数据集设置操作、网络类型设置操作以及训练方式设置操作；所述响应于针对所述设置列表中第一设置项的设置操作，确定用于实现所述目标任务的目标深度学习模型、所述目标深度学习模型的训练方式以及用于训练所述目标深度学习模型的数据集，包括：响应于针对所述目标深度学习模型的网络类型设置操作，确定用于实现所述目标任务的目标深度学习模型；响应于针对所述目标任务的数据集的数据集设置操作，确定用于训练所述目标深度学习模型的数据集；响应于针对所述目标深度学习模型的训练方式设置操作，确定所述目标深度学习模型的训练方式。
在一种可能的实现方式中，所述训练方式包括用于执行训练的训练设备、训练结束指标及样本数据的指定尺寸；所述根据所述数据集以及所述训练方式，对所述目标深度学习模型进行训练，得到待部署的目标深度学习模型，包括：根据所述指定尺寸调整所述数据集中样本数据的尺寸，得到调整后的数据集；根据所述调整后的数据集以及所述训练结束指标，在所述训练设备中训练所述目标深度学习模型，得到待部署的目标深度学习模型。
在一种可能的实现方式中，所述设置列表中还包括第二设置项，所述第二设置项用于设置所述待部署的目标深度学习模型的封装参数，所述封装参数包括封装方式和/或部署设备，所述用户开发平台还包括：部署模块，用于响应于针对所述设置列表中第二设置项的设置操作，对所述待部署的目标深度学习模型进行封装，得到所述待部署的目标深度学习模型的封装文件，其中，所述第二设置项的设置操作包括：对封装方式和/或部署设备的设置操作，所述封装文件用于在所述部署设备中部署所述待部署的目标深度学习模型。
在一种可能的实现方式中，所述设置列表中还包括用于导入数据集的第三设置项、用于标注数据集的第四设置项，以及用于对待部署的目标深度学习模型进行评估的第五设置项，所述用户开发平台 还包括：导入模块，用于响应于针对所述设置列表中第三设置项的导入操作，获得所述目标任务的原始数据集，所述原始数据集中的样本数据是满足预设数据采集标准的数据，所述预设数据采集标准用于指示对所述原始数据集内的样本数据的采集；标注模块，用于响应于针对所述设置列表中第四设置项的标注操作，根据预设数据标注标准，对所述原始数据集中的样本数据进行标注，得到所述用于训练所述目标深度学习模型的数据集；评估模块，用于响应于针对所述设置列表中第五设置项的设置操作，根据设置的网络评估指标，展示所述待部署的目标深度学习模型的性能评估结果，所述网络评估指标用于对所述待部署的深度学习模型的进行性能评估。
在一种可能的实现方式中，所述用户开发平台还包括生产线购买项，用于响应于针对所述生产线购买项的点选操作，进入生产线商店平台。
在一种可能的实现方式中，所述系统还包括生产线商店平台，所述系统还包括所述生产线商店平台，所述生产线商店平台与所述用户开发平台之间通过通信接口进行信息传输，所述生产线商店平台用于售卖模型生产线。
在一种可能的实现方式中，在通过所述生产线商店平台成功购买模型生产线的情况下，在所述生产线列表中增加成功购买的模型生产线的生产线信息。
在一种可能的实现方式中，所述系统还包括专家开发平台，所述专家开发平台与生产线商店平台之间通过通信接口进行信息传输，所述专家开发平台包括：生产线搭建模块，用于响应于针对模型生产线的搭建操作，得到预搭建的模型生产线，所述预搭建的模型生产线用于生成待部署的目标深度学习模型；生产线发布模块，用于响应于针对所述预搭建的模型生产线的发布操作，将所述预搭建的模型生产线发布至所述生产线商店平台，以在所述生产线商店平台展示并购买所述预搭建的模型生产线。
在一种可能的实现方式中，所述搭建操作包括针对模型生产线的网络配置操作；所述响应于针对模型生产线的搭建操作，得到预搭建的模型生产线，包括：响应于所述针对模型生产线的网络配置操作，得到至少一种深度学习模型，所述网络配置操作包括对网络结构、网络算法、算法参数中的至少一种进行配置的操作；对所述至少一种深度学习模型进行预训练，得到所述预搭建的模型生产线中预训练的深度学习模型，所述预训练的深度学习模型与所述用户开发平台的设置模块中待设置的网络类型对应。
在一种可能的实现方式中，所述搭建操作还包括：针对模型生产线的信息编辑操作、训练配置操作、标准编辑操作以及指标配置操作；
所述响应于针对模型生产线的搭建操作，得到预搭建的模型生产线，还包括：响应于所述针对模型生产线的信息编辑操作，得到所述预搭建的模型生产线的生产线信息，所述生产线信息用于在所述生产线商店平台以及所述用户开发平台中标识所述预搭建的模型生产线；响应于所述针对模型生产线的训练配置操作，得到所述预搭建的模型生产线中待设置的训练方式，所述待设置的训练方式用于在所述用户开发平台的设置模块中进行设置；响应于所述针对模型生产线的标准配置操作，得到所述预搭建的模型生产线的预设数据采集标准以及预设数据标注标准，其中，所述预设数据采集标准用于展示在所述用户开发平台的导入模块的界面中，所述预设数据标注标准用于展示在所述用户开发平台的标注模块的界面中；响应于所述针对模型生产线的指标配置操作，得到所述预搭建的模型生产线中待设置的网络评估指标，所述待设置的网络评估指标用于在所述用户开发平台的评估模块进行设置。
在一种可能的实现方式中，所述目标任务包括图像处理任务，所述图像处理任务包括：图像识别、图像分割、图像分类、关键点检测中的至少一种。
根据本公开的一方面，提供了一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为调用所述存储器存储的指令，以执行上述系统。
根据本公开的一方面，提供了一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述系统。
根据本公开的一方面，提供了一种计算机程序，包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行上述系统。
在本公开实施例中，能够在用户开发平台基于预先构建的模型生产线，实现生成待部署的目标深度学习模型的流程化的设置操作，以在响应于对目标深度学习模型的训练触发操作时，能够高效地基于流程化的设置操作，实现自动化生成待部署的目标深度学习模型。满足工业界对于神经网络的高效、流程化、自动化的生成需求。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，而非限制本公开。根据下面参考附图对示例性实施例的详细说明，本公开的其它特征及方面将变得清楚。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，这些附图示出了符合本公开的实施例，并与说明书一起用于说明本公开的技术方案。
图1示出根据本公开实施例的深度学习模型生产系统的结构图。
图2示出根据本公开实施例的用户开发平台的框图。
图3示出根据本公开实施例的一种生产线管理界面的示意图。
图4示出根据本公开实施的一种生产线列表的示意图。
图5示出根据本公开实施例的一种第一设置项的设置界面的示意图。
图6示出根据本公开实施例的一种第二设置项的设置界面的示意图。
图7示出根据本公开实施例的一种第三设置项的设置界面的示意图。
图8示出根据本公开实施例的一种第四设置项的设置界面的示意图。
图9示出根据本公开实施例的一种第五设置项的设置界面的示意图。
图10示出根据本公开实施例的生产线商店平台的框图。
图11示出根据本公开实施例的一种生产线展示界面的示意图。
图12示出根据本公开实施例的专家开发平台的框图。
图13示出根据本公开实施例的一种模型生产线的管理界面的示意图。
图14示出根据本公开实施例的一种网络配置界面的示意图。
图15示出根据本公开实施例的一种信息编辑界面的示意图。
图16示出根据本公开实施例的一种训练配置界面的示意图。
图17示出根据本公开实施例的一种标准配置界面的示意图。
图18示出根据本公开实施例的一种电子设备的框图。
图19示出根据本公开实施例的一种电子设备的框图。
具体实施方式
以下将参考附图详细说明本公开的各种示例性实施例、特征和方面。附图中相同的附图标记表示功能相同或相似的元件。尽管在附图中示出了实施例的各种方面，但是除非特别指出，不必按比例绘制附图。
在这里专用的词“示例性”意为“用作例子、实施例或说明性”。这里作为“示例性”所说明的任何实施例不必解释为优于或好于其它实施例。
本文中术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，本文中术语“至少一种”表示多种中的任意一种或多种中的至少两种的任意组合，例如，包括A、B、C中的至少一种，可以表示包括从A、B和C构成的集合中选择的任意一个或多个元素。
应当理解，本公开的权利要求、说明书及附图中的术语“第一”、“第二”及“第三”等是用于区别不同对象，而不是用于描述特定顺序。本公开的说明书和权利要求书中使用的术语“包括”和“包含”指示所描述特征、整体、步骤、操作、元素和/或组件的存在，但并不排除一个或多个其它特征、整体、步骤、操作、元素、组件和/或其集合的存在或添加。
另外，为了更好地说明本公开，在下文的具体实施方式中给出了众多的具体细节。本领域技术人员应当理解，没有某些具体细节，本公开同样可以实施。在一些实例中，对于本领域技术人员熟知的方法、手段、元件和电路未作详细描述，以便于凸显本公开的主旨。
图1示出根据本公开实施例的深度学习模型生产系统的结构图。如图1所示，所述深度学习模型生产系统，包括：
专家开发平台11，用于搭建模型生产线，并将预搭建或已搭建的模型生产线发布至生产线商店平台12，所述预搭建或已搭建的模型生产线用于生成待部署的目标深度学习模型；
生产线商店平台12，用于展示并售卖预搭建或已搭建的模型生产线，以将购买的模型生产线添加至用户开发平台13；
用户开发平台13，用于基于购买的模型生产线，生成待部署的目标深度学习模型。
在一种可能的实现方式中，所述深度学习模型生产系统可以在终端设备或服务器等电子设备执行系统中的方法步骤，终端设备可以为用户设备(User Equipment，UE)、移动设备、用户终端、终端、蜂窝电话、无绳电话、个人数字助理(Personal Digital Assistant，PDA)、手持设备、计算设备、车载设备、可穿戴设备等，所述系统中的方法步骤可以通过终端设备中的处理器调用存储器中存储的计算机可读指令的方式来实现，或，可通过服务器执行所述系统中的方法步骤。
在一种可能的实现方式中，生产线商店平台12与用户开发平台13之间通过通信接口进行信息传输，专家开发平台11与生产线商店平台12之间通过通信接口进行信息传输。也即，专家开发平台11、生产线商店平台12、用户开发平台13之间可通过调用通信接口实现信息交互。
应理解的是，专家开发平台11、生产线商店平台12、用户开发平台13可是独立开发的应用程序，也可是集成一体的应用程序，对此本公开实施例不作限制。本领域技术人员可采用任何已知技术，实现专家开发平台11、生产线商店平台12、用户开发平台13的开发，对此本公开实施例不作限制。
根据本公开的实施例，能够提供专业技术人员的搭建模型生产线的专家开发平台，用于购买预搭建或已搭建的模型生产线的商店用户平台，以及普通用户利于预搭建或已搭建的模型生产线生成待部署的目标深度学习模型的用户开发平台，使得专业技术人员搭建的模型生产线，可供普通用户购买使用，生成可部署在设备上的深度学习模型，这样普通用户基于预搭建或已搭建的模型生产线所生成的待部署的深度学习模型的效率和精度，就能够接近专业技术人员的水平，具有较高的实用性以及通用性。
如上所述，深度学习模型生产系统包括用户开发平台，图2示出根据本公开实施例的用户开发平台的框图，如图2所示，所述用户开发平台包括：
生产线选择模块131，用于响应于针对生产线列表中模型生产线的选择操作，展示选中的模型生产线的设置列表。
其中，选中的模型生产线用于生成待部署的目标深度学习模型，例如，模型生产线可是车辆检测模型生产线，用于生成车辆检测模型等，对此本公开实施例不作限制。设置列表中包括选中的模型生产线的第一设置项，第一设置项用于设置目标深度学习模型、目标深度学习模型的训练方式以及用于训练所述目标深度学习模型的数据集，生产线列表用于展示供选择的模型生产线的生产线信息。
在一种可能的实现方式中，生产线选择模块中可设有生产线列表，生产线展示列表中可展示用户在生产线商店中购买的模型生产线的生产线信息，以便于响应于针对模型生产线的选择操作，展示选中的模型生产线的设置列表。应理解的是，可以是在终端设备的显示界面中，展示选中的模型生产线的设置列表。
在一种可能的实现方式，可在终端设备的显示界面中展示生产线列表，例如生产线名称、生产线标识等，以便于用户实现对模型生产线的选择操作。应理解的是，生产线列表中可包括响应于该选择操作的触发控件，例如，响应于点选操作、触控操作的触发控件，从而确定出用户选择的模型生产线。对于触发控件的实现方式，本公开实施例不作限制。
图3示出根据本公开实施例的一种生产线管理界面的示意图。图4示出根据本公开实施的一种生产 线列表的示意图。图4中示出的生产线列表，可以是通过点击图3中示出的模型生产线的生产线信息“AAA”处展开的列表，这样可以便于用户选择任一模型生产线，并响应于针对任一模型生产线的选择操作，展示被选中的模型生产线的设置列表。应理解的是，生产线列表中可包含响应于针对模型生产线的点选操作的触发控件，以及用于上拉下滑列表的控件(例如滑动条)。
图3中“AAA”下的展示可以是选中的模型生产线的设置列表，设置列表中可展示模型生产线的各个设置项的名称，或者说标识等，以响应于用户针对任一设置项的点选操作，展示被选中的设置项的设置界面，从而便于用户在展示的设置界面中实现对生成待部署的目标深度学习模型的设置操作。应理解的是，设置列表中可包含响应于针对设置项的点选操作的触发控件。
需要说明的是，本公开中的第一、第二、第三等可以用于区分不同的设置项，并不限制各设置项在设置列表中的位置。应理解的是，本领域技术人员可根据实际需求，设置第一设置项在设置列表中的位置，例如，图3中的设置项c，可以是第一设置项。
设置模块132，用于响应于针对设置列表中第一设置项的设置操作，确定用于实现目标任务的目标深度学习模型、目标深度学习模型的训练方式以及用于训练目标深度学习模型的数据集。
在一种可能的实现方式中，目标任务可以包括图像处理任务。图像处理任务包括：图像识别、图像分割、图像分类、关键点检测中的至少一种。应理解的是，目标任务还可以包括语音处理任务，例如语音识别任务、语义识别任务、声纹识别任务、自然语言处理任务等。对于目标任务的任务类型，本公开实施例不作限制。
在一种可能的实现方式，可以在第一设置项的设置界面中对目标深度学习模型、训练方式以及数据集进行设置。图5示出根据本公开实施例的一种第一设置项的设置界面的示意图。如图5所示，第一设置项的设置界面中可提供用于对数据集进行设置的下拉框、对网络类型进行设置的下拉框、对训练结束指标进行设置的下拉框以及对用于执行训练的训练设备进行设置的下拉框。
其中，网络类型可用于指示不同网络结构、网络算法、算法参数下的深度学习模型，网络类型指示的深度学习模型可以是预训练的深度学习模型，通过对网络类型进行设置可以确定出用于实现目标任务的目标深度学习模型。第一设置项中设置的数据集可是用于训练所述目标深度学习模型的数据集，该数据集可以是已标注的数据集，可理解的是，可提供多个已标注的数据集以供设置。
其中，训练方式可包括训练结束指标以及用于执行训练的训练设备(例如，图像处理器GPU、X86的中央处理器CPU等)。其中，训练结束指标用于指示结束目标深度学习模型训练结束的条件，例如，迭代轮数达到设置阈值、迭代时长达到设置阈值等。目标深度学习模型的训练方式，可通过对训练结束指标及训练设备进行设置确定。
应理解的是，用户可通过点击设置界面中各下拉框来展示对应的下拉列表，以供用户根据实际需求对目标深度学习模型、训练方式及数据集进行设置操作。用户设置的目标深度学习模型、训练方式以及数据集，也即为确定出的目标深度学习模型、训练方式以及数据集。
其中，第一设置项的设置界面中可提供相关的下拉框控件，以实现针对第一设置项的设置操作，对此本公开实施例不作限制。当然，也可通过其他类型的控件，如，多选框等，实现针对第一设置项的设置操作，对此本公开实施例不作限制。
需要说明的是，上述图5第一设置项的设置界面中展示的设置内容，是本公开实施例提供的一种实现方式。应理解的是，本公开应不限于此，本领域技术人员可根据实际需求，细化对于目标深度学习模型、训练方式以及数据集的设置。
例如，可将数据集划分为训练集和测试集，并设置训练集与测试集的占比；在训练结束指标设置为按迭代轮次停止的情况下，还可设置迭代轮次阈值，以根据设置的迭代轮次阈值结束目标深度学习模型的训练。对于用于生成待部署的目标深度学习模型的第一设置项内的设置内容，可根据实际需求确定，对此本公开实施例不作限制。
训练模块133，用于响应于针对目标深度学习模型的训练触发操作，根据数据集以及训练方式，对目标深度学习模型进行训练，得到待部署的目标深度学习模型。
在一种可能的实现方式中，针对目标深度学习模型的训练触发操作，例如可通过点击图5示出的第一设置项的设置界面中的“开始训练”处的触发控件实现，通过响应于针对该触发控件的触发操作，开始对目标深度学习模型进行训练。应理解的是，对于触发控件的样式、位置、实现方式等，可根据实际需求确定，对此本公开实施例不作限制。
如上文所述，训练方式可包括训练结束指标以及用于执行训练的训练设备。在一种可能的实现方式中，根据数据集以及训练方式，对目标深度学习模型进行训练，得到待部署的目标深度学习模型，可以包括：根据数据集以及训练结束指标，在设置的训练设备中执行目标深度学习模型的训练，得到待部署的目标深度学习模型。其中，对于目标深度学习模型的训练过程，本公开实施例不作限制。
在本公开实施例中，能够基于预先构建的模型生产线，实现生成待部署的目标深度学习模型的流程化的设置操作，以在响应于对目标深度学习模型的训练触发操作时，能够高效地基于流程化的设置操作，实现自动化生成待部署的目标深度学习模型。
如上文所述，可在第一设置项的设置界面中进行针对目标深度学习模型、训练方式以及训练集的设置操作，在一种可能的实现方式，第一设置项的设置操作，包括：数据集设置操作、网络类型设置操作以及训练方式设置操作，所述响应于针对设置列表中第一设置项的设置操作，确定用于实现目标任务的目标深度学习模型、目标深度学习模型的训练方式以及用于训练目标深度学习模型的数据集，包括：
响应于针对目标深度学习模型的网络类型设置操作，确定用于实现目标任务的目标深度学习模型；响应于针对目标任务的数据集的数据集设置操作，确定用于训练目标深度学习模型的数据集；响应于针对目标深度学习模型的训练方式设置操作，确定目标深度学习模型的训练方式。
如上所述，网络类型可用于指示不同网络结构、网络算法、算法参数下的深度学习模型，网络类型指示的深度学习模型可以是预训练的深度学习模型，通过对网络类型进行设置可以确定出用于实现目标任务的目标深度学习模型。
如上所述，训练方式可包括用于执行训练的训练设备及训练结束指标。训练方式设置操作，可包括针对训练设备及训练结束指标的设置操作。
在一种可能的实现方式中，针对样本数据为图像数据的情况下，训练方式还可包括样本数据的指定尺寸。相应的是，训练方式设置操作，还可包括对样本数据的指定尺寸的设置操作。其中，指定尺寸可包括指定的图像分辨率。通过该方式，可将数据集中的样本数据调整成该指定尺寸下的样本数据，这样能够使数据集中的样本数据，满足训练过程中目标深度学习模型对输入数据的尺寸要求，提升目标深度学习模型的训练效果。
在一种可能的实现方式，还可以是采用默认的指定尺寸对数据集中的样本数据进行调整，在该情况下，可以不对样本数据的指定尺寸进行设置；或者在样本数据为非图像数据(如语音)的情况下，不对样本数据的指定尺寸进行设置。应理解的是，在该些情况下，可以不在第一设置项的设置界面中展示对于该样本数据的指定尺寸的设置内容。
在一种可能的实现方式，可以通过点击第一设置项的设置界面中提供的各下拉框控件，并在触发各下拉框控件所展示的下拉列表中，实现针对目标深度学习模型的网络类型设置操作、数据集设置操作、以及训练方式设置操作。
其中，网络类型设置操作，例如可以是点击对应下拉列表中展示的网络类型标识；数据集设置操作，例如可以是点击对应下拉列表中展示的数据集标识；训练方式设置操作，例如可以是点击对应下拉列表中的训练方式标识等。应理解的是，下拉列表中可提供针对各设置操作的触发控件，以确定出用户选择的数据集、训练方式、网络类型对应的目标深度学习模型。
在本公开实施例中，能够通过网络设置操作、数据集设置操作以及训练方式设置操作，便捷地确定出目标深度学习模型、用于训练目标深度学习模型的数据集、以及目标深度学习模型的训练方式，从而便于高效地实现自动化生成待部署的目标深度学习模型。
如上所述，训练方式可包括用于执行训练的训练设备、训练结束指标及样本数据的指定尺寸，在 一种可能的实现方式中，所述根据数据集以及训练方式，对目标深度学习模型进行训练，得到待部署的目标深度学习模型，包括：
根据指定尺寸调整数据集中样本数据的尺寸，得到调整后的数据集；
根据调整后的数据集以及训练结束指标，在训练设备中训练目标深度学习模型，得到待部署的目标深度学习模型。
如上所述，样本数据可包括图像数据，指定尺寸可包括指定的图像分辨率。根据指定尺寸调整数据集中样本数据的尺寸，得到调整后的数据集，可以包括：将数据集中图像数据的图像分辨率调整至指定的图像分辨率，得到调整后的数据集。其中，可采用任何已知的图像处理技术，实现调整图像数据的图像分辨率，例如，归一化处理、缩放处理等，对此本公开实施例不作限制。
在一种可能的实现方式中，用于执行训练的训练设备，可以是在第一设置项中针对训练方式设置的训练设备；可理解的是，还可以是采用默认的训练设备，也即可以不对训练设备进行设置，对此本公开实施不作限制。其中，在训练设备中训练目标深度学习模型，也即指在训练设备中执行目标深度学习模型的训练。
在一种可能的实现方式中，训练结束指标可以是在第一设置项中针对训练方式设置的训练结束指标；可理解的是，还可以是采用默认的训练结束指标，也即可以不对训练结束指标进行设备，对此本公开实施不作限制。
应理解的是，在第一设置项中针对训练方式的设置操作，可以包括对用于执行训练的训练设备、训练结束指标及样本数据的指定尺寸中的至少一种的设置操作。
在一种可能的实现方式中，可以将调整后的数据集、设置的训练结束指标以及目标深度学习模型传输至用于训练目标深度学习模型的训练设备中，以在训练设备中执行对目标深度学习模型的训练。
在一种可能的实现方式中，根据调整后的数据集以及训练结束指标，在训练设备中训练目标深度学习模型，得到待部署的目标深度学习模型，可以包括：将调整后的数据集中的样本数据输入至目标深度学习模型中，得到输出结果；根据所述输出结果的损失，调整目标深度学习模型的网络参数；在调整后的目标深度学习模型满足训练结束指标的情况下，得到待部署的目标深度学习模型。应理解的是，可采用任何已知的深度学习模型训练系统，训练目标深度学习模型，对此本公开实施例不作限制。
在本公开实施例中，能够高效地根据设置的训练集以及训练方式，自动实现对目标深度学习模型的训练，得到满足用户实际需求的待部署的目标深度学习模型。
考虑到，通常要对训练得到的深度学习模型进行封装，也即打包，以实现在相关设备中部署深度学习模型。在一种可能的实现方式中，所述设置列表中还包括第二设置项，第二设置项用于对待部署的目标深度学习模型进行封装，所述系统还包括：
响应于针对设置列表中第二设置项的设置操作，对待部署的目标深度学习模型进行封装，得到待部署的目标深度学习模型的封装文件，其中，第二设置项的设置操作包括：对封装方式和/或部署设备的设置操作，封装文件用于在部署设备中部署待部署的目标深度学习模型。
其中，封装方式可指示封装待部署的目标深度学习模型所需的推理框架，推理框架可使待部署的目标深度学习模型适配各种部署设备。推理框架，例如可包括TensorRT(英伟达推出的一种高性能深度学习模型推理(Inference)引擎，用于部署深度学习模型的应用程序)、OpenVINO(英特尔推出的一款用于部署深度学习模型的工具集)。
部署设备可指示待部署的目标深度学习模型所要部署的设备的处理器类型，例如可包括：X86处理器、Arm处理器、CUDA(Compute Unified Device Architecture，统一计算设备架构)处理器、GPU等。
封装文件例如可为SDK(Software Development Kit，软件开发工具包)，其中，可采用任何已知的封装技术，对待部署的目标深度学习模型进行封装，对此本公开实施例不作限制。
图6示出根据本公开实施例的一种第二设置项的设置界面的示意图。如图6所示，第二设置项的设置界面中可提供用于对封装方式进行设备的下拉框、对部署设备进行设置的下拉框、对网络版本进行 设置的下拉框、对部署是否量化进行设置的单选按钮。
其中，网络版本用于选择不同版本的待部署的目标深度学习模型，应理解的是，可基于上述第一设置项进行多次设置操作，也即对目标深度学习模型的进行多次训练，得到多个版本的待部署的目标深度学习模型；部署是否量化，用于指示是否对待部署的目标深度学习模型进行量化，将量化后的目标深度学习模型进行封装。其中，可采用任何已知的量化技术，实现对待部署的目标深度学习模型的量化，对此本公开实施例不作限制。
应理解的是，用户可通过点击设置界面中各下拉框来展示对应的下拉列表，以供用户根据实际需求对封装方式、部署设备进行设置操作。用户设置的封装方式、部署设备等，也即为确定出的封装方式、部署设备等。
其中，第二设置项的设置界面中可提供相关的下拉框控件，以实现针对第二设置项的设置操作，对此本公开实施例不作限制。当然，也可通过其他类型的控件，如，多选框等，实现针对第二设置项的设置操作，对此本公开实施例不作限制。
在一种可能的实现方式中，响应于针对设置列表中第二设置项的设置操作，还可包括触发对待部署的目标深度学习模型开始封装的触发操作，例如可通过点击图6示出的设置界面中的“确认”处的触发控件，触发对待部署的目标深度学习模型进行封装。应理解的是，对于触发控件的样式、位置、实现方式等，可根据实际需求确定，对此本公开实施例不作限制。
应理解的是，以上图6示出的第二设置项的设置界面，是本公开实施例提供的一种实现方式。本领域技术人员可根据实际需求，对第二设置项的设置界面中设置的内容进行调整，对此本公开实施例不作限制。例如，在一些情况下，可默认对待部署的目标深度学习模型进行量化，则可去掉第二设置项的设置界面中，对部署是否量化的单选按钮。
在本公开实施例中，能够实现满足用户对于待部署的目标深度学习模型的各种封装需求，使得到的封装文件有效地部署在各种部署设备中。
在一种可能的实现方式中，所述设置列表中还包括用于导入数据集的第三设置项、用于标注数据集的第四设置项，以及用于对待部署的目标深度学习模型进行评估的第五设置项，所述用户开发平台还包括：
导入模块，用于响应于针对设置列表中第三设置项的导入操作，获得目标任务的原始数据集，原始数据集中的样本数据是满足预设数据采集标准的，预设数据采集标准用于指示对原始数据集内的样本数据的采集；
标注模块，用于响应于针对设置列表中第四设置项的标注操作，根据预设数据标注标准，对原始数据集中的样本数据进行标注，得到用于训练目标深度学习模型的数据集；
评估模块，用于响应于针对设置列表中第五设置项的设置操作，根据设置的网络评估指标，展示待部署的目标深度学习模型的性能评估结果，网络评估指标用于对所述待部署的深度学习模型的进行性能评估。
在一种可能的实现方式中，第三设置项的设置界面中可展示预设数据采集标准，以指引用户采集满足该预设数据采集标准的原始样本集，例如，车辆图像的数据采集标准可包括：拍摄高度3.5-7米，拍摄到车头及整体车身，分辨率1080p；最少100张图像，每张图像不超过100辆车等。
在一种可能的实现方式中，第三设置项的设置界面中，可包括用于导入数据集的文件上传控件，例如，支持通过拖拽文件的方式导入数据集、或支持通过查找数据集在本地的存储路径的方式导入数据集等，对于数据集的导入方式，本公开实施例不作限制。
应理解的是，目标任务的原始数据集可导入不止一个。可将导入的原始数据集传输至其他电子设备并存储，以便于对原始数据集进行自动化标注。
图7示出根据本公开实施例的一种第三设置项的设置界面的示意图。如图7所示，可通过点击“选择”处的控件，确定原始数据集在本地的存储路径；通过点击“导入”按钮，触发按照原始数据集的存储路径获取原始数据集，从而实现原始数据集的导入操作。
如上文所述，目标任务的原始数据集可包括多个，在一种可能的实现方式中，第四设置项的设置界面中可展示用于对原始数据集进行选择的下拉框，以便于用户选择任一原始数据集进行标注，得到用于训练目标深度学习模型的数据集。应理解的是，针对第四设置项的标注操作，可包括针对原始数据集的选择操作。
在一种可能的实现方式中，预设数据标注标准可以是预先设置的数据标注标准，例如，过滤尺寸小于30*30的标注框。该预设数据标注标准，可展示在第四设置项的设置界面中，以告知用户该预设数据标注标准，当然也可提供多个预设数据标注标准以供用户选择，对此本公开实施例不作限制。
在一种可能的实现方式中，可采用任何已知的数据标注技术，实现根据预设数据标注标准，对原始数据集中的样本数据进行标注，例如，可通过数据标注算法，对原始技术集进行自动标注。在该情况下，针对第四设置项的标注操作，还可包括触发开始对原始数据集中的样本数据进行标注的触发操作。
在一种可能的实现方式中，第四设置项的设置界面中还可提供用于进行手动标注的标注工具(例如LabelMe：一个用于在线图像标注的Javascript标注工具)，以手动对原始数据集进行标注。在该情况下，针对第四设置项的标注操作，还可包括手动标注原始数据集的标注操作。对于原始数据集的标注方式，可根据实际需求设定，对此本公开实施例不作限制。
图8示出根据本公开实施例中一种第四设置项的设置界面的示意图。如图8所示，第四设置项的设置界面中展示用于对原始数据集进行选择的下拉框控件，应理解的是，下拉框中展示的原始数据集可是选中的原始数据集；用户可点击“自动标注”处的按钮，实现针对选择的原始数据集进行自动标注；还可点击“手动标注”处的按钮，触发展示用于进行手动标注的标注工具。
在一种可能的实现方式中，网络评估指标用于对训练得到的目标深度学习模型的性能进行评估。其中，网络评估指标例如可包括：准确率、精确率、召回率、F1分数(F1-Score)中的至少一种。
应理解的是，针对第五设置项的设置操作，可包括针对网络评估指标的设置操作，以响应于设置的网络评估指标，展示性能评估结果。图9示出根据本公开实施例的一种第五设置项的设置界面的示意图，如图9所示，可通过多选框的形式，实现对网络评估指标的设置。其中，可通过点击图9中“确认”按键，触发展示性能评估结果。应理解的是，针对第五设置项的设置操作可包括触发展示性能评估结果的触发操作。
在一种可能的实现方式中，网络评估指标设置界面中还可包括置信度阈值的选项，以便于用户查看不同置信度阈值下的性能评估结果。其中，置信度可以是深度学习模型输出结果的置信度，例如，车辆检测的置信度，可表征检测结果指示为车辆的可能性，置信度阈值可用于指示展示目标深度学习模型达到该置信度阈值时的性能评估结果。
在一种可能的实现方式中，可通过列表、图表(如曲线图)等形式展示性能评估结果。对于性能评估结果的展示形式，可根据实际需求设置，对此本公开实施例不作限制。
需要说明的是，上述图7、图8、图9分别示出的第三设置项、第四设置项以及第五设置项的设置界面，是本公开实施例提供的一种实现方式。应理解的是，本公开应不限于此，本领域技术人员可根据实际需求，设计第三设置项、第四设置项以及第五设置项的设置界面，对此本公开实施例不作限制。
在本公开实施例中，能够实现对原始数据集的导入、标注，从而便于在训练目标深度学习模型时对训练集的设置；以及，通过展示待部署的目标深度学习模型的性能评估结果，可以便于用户知晓待部署的目标深度学习模型的训练效果是否满足性能需求。
在一种可能的实现方式中，所述用户开发平台还包括生产线购买项，用于响应于针对生产线购买项的点选操作，进入生产线商店平台。通过该方式，可便于用户随时进入生产线商店平台购买预搭建或已搭建的模型生产线。
其中，生产线购买项，可用生产线商店平台的入口按钮的形式实现；生产线购买项可对应设置生产线商店平台的跳转地址，以便于用户在点击生产线购买项时，可基于该跳转地址进入生产线商店平台。对于生产线购买项的实现形式，本公开实施例不作限制。
在一种可能的实现方式中，生产线购买项可设置在用户开发平台中的任一界面的任一位置处，例如，可设置在生产线列表中、模型生产线的设置列表中、各设置项的设置界面中，对此本公开实施例不作限制。
如上所述，深度学习模型生产系统还包括生产线商店平台，生产线商店平台与用户开发平台之间通过通信接口进行信息传输，生产线商店平台用于售卖模型生产线。图10示出根据本公开实施例的生产线商店平台的框图，如图10所示，所述生产线商店平台包括：
生产线展示模块121，用于展示预搭建或已搭建的模型生产线，预搭建或已搭建的模型生产线是通过专家开发平台搭建的模型生产线。
在一种可能的实现方式中，可在生产线商店平台的生产线展示界面中，展示预搭建或已搭建的模型生产线。图11示出根据本公开实施例的一种生产线展示界面的示意图，如图11所示，生产线展示界面中可提供用于购买模型生产线的按键“购买”，用户点击“购买”按键可进入支付页面；用户还可点击生产线展示界面中的“详情”按键，以查看选中的模型生产线的详细信息；还可点击“搜索”案件，搜索要购买的模型生产线。
应理解的是，图11示出的生产线展示界面是本公开实施例提供的一种实现方式，应理解的是，本公开应不限于此，本领域技术人员可根据实际需求中设计生产线展示界面，对此本公开实施例不作限制。
其中，预搭建或已搭建的模型生产线的搭建过程，将在下文中进行阐述，为行文简洁，在此不做赘述。
生产线购买模块122，用于响应于针对展示的模型生产线的购买操作，确定购买的模型生产线。
在一种可能的实现方式中，针对展示的模型生产线的购买操作可包括：针对如图11所示的生产线展示界面中展示的“购买”按键的点选操作。在点击“购买”按键后，可进入支付界面。应理解的是，支付界面中可提供用于进行线上支付的支付控件，本领域技术人员可采用任何已知的技术，实现针对模型生产线的线上支付，对于线上支付的实现方式，本公开实施例不作限制。
应理解的是，已完成支付的模型生产线，可是成功购买的模型生产线。在一种可能的实现方式中，在通过生产线商店平台成功购买模型生产线的情况下，在用户开发平台的生产线列表中增加成功购买的模型生产线的生产线信息，以便于用户使用成功购买的模型生产线生成待部署的目标深度学习模型。其中，生产线列表的实现方式，可参照上述本公开实施例公开的方式，在此不做赘述。
在一种可能的实现方式中，生产线商店平台还可包括生产线发送模块，用于将成功购买的模型生产线发送至用户开发平台，以在用户开发平台的生产线展示列表中展示成功购买的模型生产线，其中，生产线展示列表用于展示并选择成功购买的模型生产线。
在本公开实施例中，用户可通过生产线商店平台购买不同的模型生产线，以生成实现不同目标任务的待部署的目标深度学习模型。
如上文所述，深度学习模型生产系统可包括专家开发平台。图12示出根据本公开实施例的专家开发平台的框图，如图12所示，所述专家开发平台包括：
生产线搭建模块111，用于响应于针对模型生产线的搭建操作，得到预搭建或已搭建的模型生产线，所述预搭建或已搭建的模型生产线用于生成待部署的目标深度学习模型。
图13示出根据本公开实施例的一种模型生产线的管理界面的示意图。在一种可能的实现方式中，还可以在如图13所示的管理界面中展示已搭建的模型生产线的信息，如图13所示，负责人可为搭建该模型生产线的人员名称。应理解的是，管理界面中可提供用于上拉下滑列表的控件(例如滑动条)，来展示较多的模型生产线。
其中，如图13所示，模型生产线的管理界面还可提供用于搜索模型生产线的搜索框(图13中“搜索”左侧的框)以及搜索触发控件(图13中的“搜索”)，来快速的搜索出模型生产线；点击图13中任意模型生产线对应的“详情”处的触发控件，以展示更为详细的模型生产线的信息，点击图13中“编辑”处的触发控件，可对模型生产线进行编辑，点击图13中的“发布”处的触发控件，可将预搭建或 已搭建的模型生产线发布至生产线商店平台。
在一种可能的实现方式中，可通过点击图13中“创建生产线”处的触发控件，进入模型生产线的搭建界面，以针对模型生产线进行搭建操作。对于模型生产线的搭建界面的进入方式，可根据实际需求确定，对此本公开实施例不作限制。应理解的是，搭建界面中可提供各类界面上的控件，例如，编辑框、下拉框、多选框、单选框等，以实现针对模型生产线的搭建操作。
在一种可能的实现方式中，针对模型生产线的搭建操作可包括：针对模型生产线的网络配置操作、信息编辑操作、训练配置操作、标准编辑操作以及指标配置操作。
其中，网络配置操作用于配置出模型生产线中至少一种预训练的深度学习模型；信息编辑操作，用于编辑出模型生产线的生产线信息；训练配置操作，用于配置出模型生产线中待设置的训练方式；标准配置操作用于配置出模型生产线的预设数据采集标准以及预设数据标注标准；指标配置操作用于配置出模型生产线中待设置的网络评估指标。
应理解的是，在完成上述针对模型生产线的搭建操作后，可得到预搭建或已搭建的模型生产线。其中，可采用已知的组件串联工具将与各搭建操作对应的配置结果串联起来，得到预搭建或已搭建的模型生产线，对此本公开实施例不作限制。
生产线发布模块112，用于响应于针对预搭建或已搭建的模型生产线的发布操作，将预搭建或已搭建的模型生产线发布至生产线商店平台，以在生产线商店平台展示并购买预搭建或已搭建的模型生产线。
在一种可能的实现方式中，预搭建或已搭建的模型生产线可展示如图13示出的管理界面中，可通过点击图13中示出的“发布”按钮，触发将预搭建或已搭建的模型生产线发布至生产线商店平台。应理解的是，本领域技术人员可根据实际需求设置预搭建或已搭建的模型生产线的发布方式，对此本公开实施例不作限制。
如上所述，专家开发平台与生产线商店平台之间通过通信接口进行信息传输，将预搭建或已搭建的模型生产线发布至生产线商店平台，可以是将预搭建或已搭建的模型生产线中包含的全部配置内容进行打包压缩，并将打包压缩后的文件发送至生产线商店平台。对于如何将预搭建或已搭建的模型生产线发布至生产线商店平台，本公开实施例不作限制。
其中，在生产线商店平台展示并购买预搭建或已搭建的模型生产模型，可参照上述本公开实施例中的内容，在此不做赘述。
在本公开实施例中，能够利用专家开发平台搭建出模型生产线，从而使普通用户可以使用专业技术人员搭建出的模型生产线，生成待部署的目标深度学习模型，提高待部署的目标深度学习模型的生产效率以及精度。
如上所述，所述搭建操作包括针对模型生产线的网络配置操作，在一种可能的实现方式中，响应于针对模型生产线的搭建操作，得到预搭建或已搭建的模型生产线，包括：
响应于针对模型生产线的网络配置操作，得到至少一种深度学习模型，网络配置操作包括对网络结构、网络算法、算法参数中的至少一种进行配置的操作；
对至少一种深度学习模型进行预训练，得到预搭建或已搭建的模型生产线中预训练的深度学习模型，预训练的深度学习模型与用户开发平台的设置模块中待设置的网络类型对应。
其中，网络结构可包括深度学习模型的主干网络backbone、融合网络neck以及分支网络head的网络结构。其中，主干网络用于进行特征提取，融合网络用于对主干网络提取的特征进行融合处理，分支网络用于针对主干网络提取的特征或融合处理的特征进行不同任务的推理。
在一种可能的实现方式中，主干网络backbone、融合网络neck以及分支网络head的网络结构例如可至少采用：ResNet系列，SENet系列，MobileNet系列，ShuffleNet系列等。网络算法至少可以包括：快速卷积Faster R-CNN算法，关键点卷积Keypoint R-CNN算法，阶梯卷积Cascade R-CNN算法，归一化算法等。算法参数至少包括：批大小、采样方式、数据增强方式、网络参数等。
其中，批大小，可理解为一次训练的样本数据的数目；采样方式可至少包括：正负样本采样方式、 困难样本(Hard Sample)采样方式、欠采样方式、过采样方式等，数据增强方式可包括：随机裁剪，扭曲，扩增，镜像，形变等数据增强方式；网络参数可包括网络的超参数，由于超参数的种类较多，可通过选取已有的参数配置文件确定。
应理解的是，针对模型生产线的网络配置操作，也即针对深度学习模型的网络配置操作，可不止包括上述本公开实施例公开的配置内容；以及，网络配置操作，可通过显示界面进行人机交互实现，也可通过后台调用相关接口、调用相关配置文件实现，只要可实现配置出深度学习模型即可，对此本公开实施例不作限制。
应理解的是，不同配置方式可得到不同的深度学习模型，不同的深度学习模型的网络大小、精度、运算性能可不同。例如，针对同一网络结构，网络算法可不同；针对同一网络结构及网络算法，算法参数可不同；精度高的深度学习模型，体积大；体积小的深度学习模型精度低而运算性能好。通过该方式，可以配置出不同精度、不同运算性能的深度学习模型，也即不同网络类型的深度学习模型，以供用户根据精度需求、性能需求选择目标深度学习模型。
图14示出根据本公开实施例的一种网络配置界面的示意图。如图14所示，用户可通过触发各项对应的下拉框、编辑框、多选框等控件，实现针对模型生产线的网络配置操作。应理解的是，图14示出的网络配置界面是一种可能的实现方式，本领域技术人员可根据实际需求，设定网络配置界面中可配置的内容、网络配置界面的布局、样式等，对此本公开实施例不作限制。
在一种可能的实现方式中，可采用任何已知的预训练方式，对配置得到的至少一种深度学习模型进行预训练，得到预训练的深度学习模型。应理解的是，预训练中采用的训练集可与模型生产线的目标任务、应用场景等相关，例如，人体检测任务、车辆检测任务各自对应的训练集是不同的，训练得到深度学习模型也可不同。
其中，预训练的深度学习模型与用户开发平台的设置模块中待设置的网络类型对应，这样用户可通过设置网络类型来确定出对应的目标深度学习模型。其中，在用户开发平台设置的目标深度学习模型，是预训练的深度学习模型。通过该方式，可得到有一定精度的深度学习模型，从而提高上述训练目标深度学习模型的训练效率，以高效地生成待部署的目标深度学习模型。
应理解的是，同一模型生产线中可包含多个处理任务，例如，车辆检测、车牌检测、车牌识别等。那么针对同一模型生产线中的多个处理任务可分别配置对应的至少一种深度学习模型，并进行预训练。
在本公开实施例中，能够针对模型生产线，预先建立至少一种深度学习模型，并对至少一种深度学习模型进行预训练，得到供用户选择的深度学习模型，这样用户可以通过简单便捷的设置操作，也即选择网络类型，便得到可实现目标任务的目标深度学习模型，提高目标深度学习模型的生成效率。
在一种可能的实现方式中，响应于针对模型生产线的搭建操作，得到用于生成待部署的目标深度学习模型的模型生产线，还包括：
响应于针对模型生产线的信息编辑操作，得到预搭建或已搭建的模型生产线的生产线信息，生产线信息用于在生产线商店平台以及用户开发平台中标识预搭建或已搭建的模型生产线；
响应于针对模型生产线的训练配置操作，得到预搭建或已搭建的模型生产线中待设置的训练方式，待设置的训练方式用于在用户开发平台的设置模块中进行设置；
响应于针对模型生产线的标准配置操作，得到预搭建或已搭建的模型生产线的预设数据采集标准以及预设数据标注标准，其中，预设数据采集标准用于展示在用户开发平台的导入模块的界面中，预设数据标注标准用于展示在用户开发平台的标注模块的界面中；
响应于针对模型生产线的指标配置操作，得到预搭建或已搭建的模型生产线中待设置的网络评估指标，待设置的网络评估指标用于在用户开发平台的评估模块进行设置。
图15示出根据本公开实施例的一种信息编辑界面的示意图。如图15所示，信息编辑界面中可展示信息编辑的提示信息，可通过编辑框输入任务名称，处理对象，应用场景，数据采集设备等信息，也即实现针对模型生产线的信息编辑操作。举例来说，对于车辆检测项目，项目信息可包括“任务名称：车辆检测，处理对象：车辆，应用场景：街道，数据采集设备：摄像头”。应理解的是，用户在编辑 框中输入的信息，也即得到的模型生产线的生产线信息，可通过点击图15中示出的“确认”按钮，确定输入的生产线信息。
图16示出根据本公开实施例的一种训练配置界面的示意图。如图16所示，可通过多选框对训练结束指标进行配置，应理解的是，此处配置的训练结束指标也即为用户平台的设置模块中以供用户设置的训练结束指标；也可通过多选框对训练设备进行配置，此处配置的训练设备也即为用户平台的设置模块中以供用户设置的训练设备；通过该方式，可配置多个训练结束指标以及多个训练设备，以供用户根据需求选择任一训练结束指标以及用于训练目标深度学习模型的训练设备。
在一种可能的实现方式中，如图16示，训练配置界面中还可提供对针对训练设备的性能预估信息进行编辑的编辑框，其中，性能预估信息例如可为训练设备执行目标深度学习模型训练所需的运行时长。可将编辑的训练设备的性能评估信息展示的第一设置项的设置界面中，以便用户基于展示的运行时长，选择所需的训练设备来执行目标深度学习模型的训练。
应理解的是，针对模型生产线的训练配置操作，可包括针对训练结束指标的多选操作、对训练设备的多选操作以及对性能预估信息的编辑操作。其中，用户可通过点击图16中示出的“确认”按钮，确定选择的训练结束指标及训练设备，以及编辑的性能预估信息等，对此本公开实施例不作限制。
在一种可能的实现方式中，可通过显示界面上的编辑框分别输入预设数据采集标准以及预设数据标注标准，也即实现针对模型生产线的标准编辑操作；也可通过后台调用相关接口，实现预设数据采集标准以及预设数据标注标准的配置。其中，可将配置的预设数据采集标准以及预设数据标注标准，分别展示在用于导入数据集的第三设置项的设置界面，及用于标注数据集的第四设置项的设界面中。
图17示出根据本公开实施例的一种标准配置界面的示意图，如图17所示，可通过编辑框分别输入预设数据采集标准以及预设数据标注标准，也可通过下拉框配置数据采集标准以及数据标注标准。其中，用户可通过点击图17中示出的“确认”按钮，确定编辑的预设数据采集标准以及预设数据标注标准等，对此本公开实施例不作限制。
在一种可能的实现方式中，可在指标配置界面中通过多选框的形式，实现对网络评估指标的指标配置操作，其中，指标配置界面可参照上述图9示出的第五设置项的设置界面。应理解的是，在指标配置界面中选择的网络评估指标，也即为上述第五设置项的设置界面中展示的网络评估指标，第五设置项的设置界面中待设置的网络评估指标，可以是指标配置界面中展示的网络评估指标的部分或全部指标，对此本公开实施例不作限制。
需要说明的是，以上网络配置、信息编辑、训练配置、标准配置、指标配置，可以是同一显示界面中展示，也可在不同的显示界面中展示，也即，网络配置界面、信息编辑界面、训练配置界面、标准配置界面、指标配置界面可以是同一界面，也可是不同界面，对此本公开实施例不作限制。模型生产线的搭建界面包括：网络配置界面、信息编辑界面、训练配置界面、标准编辑界面以及指标配置界面。
在一种可能的实现方式中，在针对深度学习模型、生产线信息、训练方式、预设数据采集标准、预设数据标注标准、网络评估指标等配置后，可通过现有的组件串联工具，对各个配置结果进行串联，从而搭建出模型生产线。
在本公开实施例中，能够使专业的技术人员，对模型生产线中各个设置项的内容进行配置，以搭建出适用于不同场景需求、不同任务需求的模型生产线，这样可以使普通用户基于搭建好的模型生产线，通过简单便捷地流程化设置操作，高效地得到满足实际需求的待部署的目标深度学习模型。
在本公开实施例中，能够建立至少一种模型生产线及其中包含的推荐配置，其中推荐配置例如可至少包括至预训练的深度学习模型、数据采集标准、数据标注标准、深度学习模型结构、超参数、评估指标、训练配置、推理配置等。从而可以便于用户基于搭建的模型生产线及其中包含的推荐配置，进行流程化的设置操作，以高效地生成待部署的目标深度学习模型。
可知晓的是，深度学习模型的生成过程，通过需要专业技术人员具备专业知识及经验，例如针对数据采集方式、深度学习模型选择、超参数、预训练模型，数据增强的方式、迭代次数、迭代方式的 逻辑等的专业知识。这类知识及经验，随着深度学习模型的应用场景、处理任务的不同而不同。因此，在实际应用的场景中，深度学习模型生成过程，通常需要专业的技术人员定制完成。另外，深度学习模型的迁移、适应能力薄弱，例如在一个城市中训练精度非常好的车辆识别网络，在另外一个城市里，由于车型分布的变化，表现较差。
根据本公开的实施例，能够提供专业技术人员的针对模型生产线的配置环境，和普通用户使用模型生产线生成所需的深度学习模型的设置环境。这样可以由专业技术人员定义、生成深度学习模型的模型生产线，每一模型生产线，能够支持至少一个应用场景的深度学习模型的生成。专业技术人员搭建的模型生产线，可供普通用户选择使用，使用模型生产线的设置环境，完成深度学习模型的生成过程，得到可部署在设备的深度学习模型。通过该方式，普通用户基于模型生产线生成深度学习模型的效率和精度，就能够接近专业技术人员的水平。
相关技术中，深度学习模型的生成流程，不能够满足多种场景下深度学习模型训练的精度要求；以及深度学习领域的深度学习模型的迁移能力不够强，因此用户需要在针对不同的应用场景下，重新采集数据训练、验证深度学习模型。以及相关技术中用相同生成流程生成的深度学习模型，无法自动化适应多种场景下的深度学习模型生成需求。根据本公开的实施例，通过定义模型生产线的应用场景，实现不同应用场景下定制模型生产线，能够支持多种应用场景下自动化的深度学习模型的生成。
相关技术中，深度学习模型单一的生成流程，使得提供的训练后的深度学习模型，在类似场景中的精度偏低。根据本公开的实施例，针对同一模型生产线进行的不同设置操作，可以使生成的深度学习模型，适用于类似应用场景，通过优化的设置流程及训练配置得到更高精度的深度学习模型。例如，车辆检测场景下的模型生产线，可采用不同训练集对同一目标深度学习模型训练，高效地得到不同的待部署的目标深度学习模型，从而能够适用于不同城市、不同街道的类似的车辆检测场景。
根据本公开的实施例，能够通过专业技术人员建立不同应用场景的模型生产线，普通用户通过使用专业技术提供配置的生成流程及相关配置，可以满足各类不同应用场景、相似应用场景的深度学习模型生成需求，也可得到接近专业技术人员水平的高精度、高性能并适合场景需求的深度学习模型。
可以理解，本公开提及的上述各个系统实施例，在不违背原理逻辑的情况下，均可以彼此相互结合形成结合后的实施例，限于篇幅，本公开不再赘述。本领域技术人员可以理解，在具体实施方式的上述系统中，各步骤的具体执行顺序应当以其功能和可能的内在逻辑确定。
本公开实施例还提出一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述系统。计算机可读存储介质可以是非易失性计算机可读存储介质。
本公开实施例还提出一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为调用所述存储器存储的指令，以执行上述系统中的方法步骤。
本公开实施例还提供了一种计算机程序产品，包括计算机可读代码，当计算机可读代码在设备上运行时，设备中的处理器执行用于实现如上任一实施例提供的深度学习模型生产系统的指令。
本公开实施例还提供了另一种计算机程序产品，用于存储计算机可读指令，指令被执行时使得计算机执行上述任一实施例提供的深度学习模型生产系统的操作。
本公开实施例还提供了一种计算机程序，包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行上述系统。
电子设备可以被提供为终端、服务器或其它形态的设备。
图18示出根据本公开实施例的一种电子设备800的框图。例如，电子设备800可以是移动电话，计算机，数字广播终端，消息收发设备，游戏控制台，平板设备，医疗设备，健身设备，个人数字助理等终端。
参照图18，电子设备800可以包括以下一个或多个组件：处理组件802，存储器804，电源组件806，多媒体组件808，音频组件810，输入/输出(I/O)的接口812，传感器组件814，以及通信组件816。
处理组件802通常控制电子设备800的整体操作，诸如与显示，电话呼叫，数据通信，相机操作和记录操作相关联的操作。处理组件802可以包括一个或多个处理器820来执行指令，以完成上述的系统 的全部或部分步骤。此外，处理组件802可以包括一个或多个模块，便于处理组件802和其他组件之间的交互。例如，处理组件802可以包括多媒体模块，以方便多媒体组件808和处理组件802之间的交互。
存储器804被配置为存储各种类型的数据以支持在电子设备800的操作。这些数据的示例包括用于在电子设备800上操作的任何应用程序或系统的指令，联系人数据，电话簿数据，消息，图片，视频等。存储器804可以由任何类型的易失性或非易失性存储设备或者它们的组合实现，如静态随机存取存储器(SRAM)，电可擦除可编程只读存储器(EEPROM)，可擦除可编程只读存储器(EPROM)，可编程只读存储器(PROM)，只读存储器(ROM)，磁存储器，快闪存储器，磁盘或光盘。
电源组件806为电子设备800的各种组件提供电力。电源组件806可以包括电源管理系统，一个或多个电源，及其他与为电子设备800生成、管理和分配电力相关联的组件。
多媒体组件808包括在所述电子设备800和用户之间的提供一个输出接口的屏幕。在一些实施例中，屏幕可以包括液晶显示器(LCD)和触摸面板(TP)。如果屏幕包括触摸面板，屏幕可以被实现为触摸屏，以接收来自用户的输入信号。触摸面板包括一个或多个触摸传感器以感测触摸、滑动和触摸面板上的手势。所述触摸传感器可以不仅感测触摸或滑动动作的边界，而且还检测与所述触摸或滑动操作相关的持续时间和压力。在一些实施例中，多媒体组件808包括一个前置摄像头和/或后置摄像头。当电子设备800处于操作模式，如拍摄模式或视频模式时，前置摄像头和/或后置摄像头可以接收外部的多媒体数据。每个前置摄像头和后置摄像头可以是一个固定的光学透镜系统或具有焦距和光学变焦能力。
音频组件810被配置为输出和/或输入音频信号。例如，音频组件810包括一个麦克风(MIC)，当电子设备800处于操作模式，如呼叫模式、记录模式和语音识别模式时，麦克风被配置为接收外部音频信号。所接收的音频信号可以被进一步存储在存储器804或经由通信组件816发送。在一些实施例中，音频组件810还包括一个扬声器，用于输出音频信号。
I/O接口812为处理组件802和外围接口模块之间提供接口，上述外围接口模块可以是键盘，点击轮，按钮等。这些按钮可包括但不限于：主页按钮、音量按钮、启动按钮和锁定按钮。
传感器组件814包括一个或多个传感器，用于为电子设备800提供各个方面的状态评估。例如，传感器组件814可以检测到电子设备800的打开/关闭状态，组件的相对定位，例如所述组件为电子设备800的显示器和小键盘，传感器组件814还可以检测电子设备800或电子设备800一个组件的位置改变，用户与电子设备800接触的存在或不存在，电子设备800方位或加速/减速和电子设备800的温度变化。传感器组件814可以包括接近传感器，被配置用来在没有任何的物理接触时检测附近物体的存在。传感器组件814还可以包括光传感器，如互补金属氧化物半导体(CMOS)或电荷耦合装置(CCD)图像传感器，用于在成像应用中使用。在一些实施例中，该传感器组件814还可以包括加速度传感器，陀螺仪传感器，磁传感器，压力传感器或温度传感器。
通信组件816被配置为便于电子设备800和其他设备之间有线或无线方式的通信。电子设备800可以接入基于通信标准的无线网络，如无线网络(WiFi)，第二代移动通信技术(2G)或第三代移动通信技术(3G)，或它们的组合。在一个示例性实施例中，通信组件816经由广播信道接收来自外部广播管理系统的广播信号或广播相关信息。在一个示例性实施例中，所述通信组件816还包括近场通信(NFC)模块，以促进短程通信。例如，在NFC模块可基于射频识别(RFID)技术，红外数据协会(IrDA)技术，超宽带(UWB)技术，蓝牙(BT)技术和其他技术来实现。
在示例性实施例中，电子设备800可以被一个或多个应用专用集成电路(ASIC)、数字信号处理器(DSP)、数字信号处理设备(DSPD)、可编程逻辑器件(PLD)、现场可编程门阵列(FPGA)、控制器、微控制器、微处理器或其他电子元件实现，用于执行上述系统中的方法步骤。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器804，上述计算机程序指令可由电子设备800的处理器820执行以完成上述系统中的方法步骤。
图19示出根据本公开实施例的一种电子设备1900的框图。例如，电子设备1900可以被提供为一服务器。参照图19，电子设备1900包括处理组件1922，其进一步包括一个或多个处理器，以及由存储器 1932所代表的存储器资源，用于存储可由处理组件1922的执行的指令，例如应用程序。存储器1932中存储的应用程序可以包括一个或一个以上的每一个对应于一组指令的模块。此外，处理组件1922被配置为执行指令，以执行上述系统中的方法步骤。
电子设备1900还可以包括一个电源组件1926被配置为执行电子设备1900的电源管理，一个有线或无线网络接口1950被配置为将电子设备1900连接到网络，和一个输入输出(I/O)接口1958。电子设备1900可以操作基于存储在存储器1932的操作系统，例如微软服务器操作系统(Windows Server TM)，苹果公司推出的基于图形用户界面操作系统(Mac OS X TM)，多用户多进程的计算机操作系统(Unix TM),自由和开放原代码的类Unix操作系统(Linux TM)，开放原代码的类Unix操作系统(FreeBSD TM)或类似。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器1932，上述计算机程序指令可由电子设备1900的处理组件1922执行以完成上述系统的方法步骤。
本公开可以是系统和/或计算机程序产品。计算机程序产品可以包括计算机可读存储介质，其上载有用于使处理器实现本公开的各个方面的计算机可读程序指令。
计算机可读存储介质可以是可以保持和存储由指令执行设备使用的指令的有形设备。计算机可读存储介质例如可以是(但不限于)电存储设备、磁存储设备、光存储设备、电磁存储设备、半导体存储设备或者上述的任意合适的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、静态随机存取存储器(SRAM)、便携式压缩盘只读存储器(CD-ROM)、数字多功能盘(DVD)、记忆棒、软盘、机械编码设备、例如其上存储有指令的打孔卡或凹槽内凸起结构、以及上述的任意合适的组合。这里所使用的计算机可读存储介质不被解释为瞬时信号本身，诸如无线电波或者其他自由传播的电磁波、通过波导或其他传输媒介传播的电磁波(例如，通过光纤电缆的光脉冲)、或者通过电线传输的电信号。
这里所描述的计算机可读程序指令可以从计算机可读存储介质下载到各个计算/处理设备，或者通过网络、例如因特网、局域网、广域网和/或无线网下载到外部计算机或外部存储设备。网络可以包括铜传输电缆、光纤传输、无线传输、路由器、防火墙、交换机、网关计算机和/或边缘服务器。每个计算/处理设备中的网络适配卡或者网络接口从网络接收计算机可读程序指令，并转发该计算机可读程序指令，以供存储在各个计算/处理设备中的计算机可读存储介质中。
用于执行本公开操作的计算机程序指令可以是汇编指令、指令集架构(ISA)指令、机器指令、机器相关指令、微代码、固件指令、状态设置数据、或者以一种或多种编程语言的任意组合编写的源代码或目标代码，所述编程语言包括面向对象的编程语言—诸如Smalltalk、C++等，以及常规的过程式编程语言—诸如“C”语言或类似的编程语言。计算机可读程序指令可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络—包括局域网(LAN)或广域网(WAN)—连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。在一些实施例中，通过利用计算机可读程序指令的状态信息来个性化定制电子电路，例如可编程逻辑电路、现场可编程门阵列(FPGA)或可编程逻辑阵列(PLA)，该电子电路可以执行计算机可读程序指令，从而实现本公开的各个方面。
这里参照根据本公开实施例的系统和计算机程序产品的流程图和/或框图描述了本公开的各个方面。应当理解，流程图和/或框图的每个方框以及流程图和/或框图中各方框的组合，都可以由计算机可读程序指令实现。
这些计算机可读程序指令可以提供给通用计算机、专用计算机或其它可编程数据处理装置的处理器，从而生成出一种机器，使得这些指令在通过计算机或其它可编程数据处理装置的处理器执行时，产生了实现流程图和/或框图中的一个或多个方框中规定的功能/动作的装置。也可以把这些计算机可读程序指令存储在计算机可读存储介质中，这些指令使得计算机、可编程数据处理装置和/或其他设 备以特定方式工作，从而，存储有指令的计算机可读介质则包括一个制造品，其包括实现流程图和/或框图中的一个或多个方框中规定的功能/动作的各个方面的指令。
也可以把计算机可读程序指令加载到计算机、其它可编程数据处理装置、或其它设备上，使得在计算机、其它可编程数据处理装置或其它设备上执行一系列操作步骤，以产生计算机实现的过程，从而使得在计算机、其它可编程数据处理装置、或其它设备上执行的指令实现流程图和/或框图中的一个或多个方框中规定的功能/动作。
附图中的流程图和框图显示了根据本公开的多个实施例的系统和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段或指令的一部分，所述模块、程序段或指令的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个连续的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或动作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
该计算机程序产品可以具体通过硬件、软件或其结合的方式实现。在一个可选实施例中，所述计算机程序产品具体体现为计算机存储介质，在另一个可选实施例中，计算机程序产品具体体现为软件产品，例如软件开发包(Software Development Kit，SDK)等等。
以上已经描述了本公开的各实施例，上述说明是示例性的，并非穷尽性的，并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择，旨在最好地解释各实施例的原理、实际应用或对市场中的技术的改进，或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。

Claims (15)

  1. 一种深度学习模型生产系统，其特征在于，包括用户开发平台，所述用户开发平台包括：
    生产线选择模块，用于响应于针对生产线列表中模型生产线的选择操作，展示选中的模型生产线的设置列表，其中，所述选中的模型生产线用于生成待部署的目标深度学习模型，所述设置列表中包括所述选中的模型生产线的第一设置项，所述第一设置项用于设置目标深度学习模型、所述目标深度学习模型的训练方式以及用于训练所述目标深度学习模型的数据集，所述生产线列表用于展示供选择的模型生产线的生产线信息；
    设置模块，用于响应于针对所述设置列表中第一设置项的设置操作，确定用于实现目标任务的目标深度学习模型、所述目标深度学习模型的训练方式以及用于训练所述目标深度学习模型的数据集；
    训练模块，用于响应于针对所述目标深度学习模型的训练触发操作，根据所述数据集以及所述训练方式，对所述目标深度学习模型进行训练，得到待部署的目标深度学习模型。
  2. 根据权利要求1所述的系统，其特征在于，所述第一设置项的设置操作，包括：数据集设置操作、网络类型设置操作以及训练方式设置操作；
    所述响应于针对所述设置列表中第一设置项的设置操作，确定用于实现所述目标任务的目标深度学习模型、所述目标深度学习模型的训练方式以及用于训练所述目标深度学习模型的数据集，包括：
    响应于针对所述目标深度学习模型的网络类型设置操作，确定用于实现所述目标任务的目标深度学习模型；
    响应于针对所述目标任务的数据集的数据集设置操作，确定用于训练所述目标深度学习模型的数据集；
    响应于针对所述目标深度学习模型的训练方式设置操作，确定所述目标深度学习模型的训练方式。
  3. 根据权利要求1或2所述的系统，其特征在于，所述训练方式包括用于执行训练的训练设备、训练结束指标及样本数据的指定尺寸；
    所述根据所述数据集以及所述训练方式，对所述目标深度学习模型进行训练，得到待部署的目标深度学习模型，包括：
    根据所述指定尺寸调整所述数据集中样本数据的尺寸，得到调整后的数据集；
    根据所述调整后的数据集以及所述训练结束指标，在所述训练设备中训练所述目标深度学习模型，得到待部署的目标深度学习模型。
  4. 根据权利要求1-3任一项所述的系统，其特征在于，所述设置列表中还包括第二设置项，所述第二设置项用于设置所述待部署的目标深度学习模型的封装参数，所述封装参数包括封装方式和/或部署设备，所述用户开发平台还包括：
    部署模块，用于响应于针对所述设置列表中第二设置项的设置操作，对所述待部署的目标深度学习模型进行封装，得到所述待部署的目标深度学习模型的封装文件，
    其中，所述第二设置项的设置操作包括：对封装方式和/或部署设备的设置操作，所述封装文件用于在所述部署设备中部署所述待部署的目标深度学习模型。
  5. 根据权利要求1-4任一项所述的系统，其特征在于，所述设置列表中还包括用于导入数据集的第三设置项、用于标注数据集的第四设置项，以及用于对待部署的目标深度学习模型进行评估的第五设置项，所述用户开发平台还包括：
    导入模块，用于响应于针对所述设置列表中第三设置项的导入操作，获得所述目标任务的原始数据集，所述原始数据集中的样本数据是满足预设数据采集标准的数据，所述预设数据采集标准用于指示对所述原始数据集内的样本数据的采集；
    标注模块，用于响应于针对所述设置列表中第四设置项的标注操作，根据预设数据标注标准，对所述原始数据集中的样本数据进行标注，得到所述用于训练所述目标深度学习模型的数据集；
    评估模块，用于响应于针对所述设置列表中第五设置项的设置操作，根据设置的网络评估指标，展示所述待部署的目标深度学习模型的性能评估结果，所述网络评估指标用于对所述待部署的深度学习模型的进行性能评估。
  6. 根据权利要求1所述的系统，其特征在于，所述用户开发平台还包括生产线购买项，用于响应于针对所述生产线购买项的点选操作，进入生产线商店平台。
  7. 根据权利要求6所述的系统，其特征在于，所述系统还包括所述生产线商店平台，所述生产线商店平台与所述用户开发平台之间通过通信接口进行信息传输，所述生产线商店平台用于售卖模型生产线。
  8. 根据权利要求6或7所述的系统，其特征在于，在通过所述生产线商店平台成功购买模型生产线的情况下，在所述生产线列表中增加成功购买的模型生产线的生产线信息。
  9. 根据权利要求1所述的系统，其特征在于，所述系统还包括专家开发平台，所述专家开发平台与生产线商店平台之间通过通信接口进行信息传输，所述专家开发平台包括：
    生产线搭建模块，用于响应于针对模型生产线的搭建操作，得到预搭建的模型生产线，所述预搭建的模型生产线用于生成待部署的目标深度学习模型；
    生产线发布模块，用于响应于针对所述预搭建的模型生产线的发布操作，将所述预搭建的模型生产线发布至所述生产线商店平台，以在所述生产线商店平台展示并购买所述预搭建的模型生产线。
  10. 根据权利要求7所述的系统，其特征在于，所述搭建操作包括针对模型生产线的网络配置操作；
    所述响应于针对模型生产线的搭建操作，得到预搭建的模型生产线，包括：
    响应于所述针对模型生产线的网络配置操作，得到至少一种深度学习模型，所述网络配置操作包括对网络结构、网络算法、算法参数中的至少一种进行配置的操作；
    对所述至少一种深度学习模型进行预训练，得到所述预搭建的模型生产线中预训练的深度学习模型，所述预训练的深度学习模型与所述用户开发平台的设置模块中待设置的网络类型对应。
  11. 根据权利要求7或8所述的系统，其特征在于，所述搭建操作还包括：针对模型生产线的信息编辑操作、训练配置操作、标准编辑操作以及指标配置操作；
    所述响应于针对模型生产线的搭建操作，得到预搭建的模型生产线，还包括：
    响应于所述针对模型生产线的信息编辑操作，得到所述预搭建的模型生产线的生产线信息，所述生产线信息用于在所述生产线商店平台以及所述用户开发平台中标识所述预搭建的模型生产线；
    响应于所述针对模型生产线的训练配置操作，得到所述预搭建的模型生产线中待设置的训练方式，所述待设置的训练方式用于在所述用户开发平台的设置模块中进行设置；
    响应于所述针对模型生产线的标准配置操作，得到所述预搭建的模型生产线的预设数据采集标准以及预设数据标注标准，其中，所述预设数据采集标准用于展示在所述用户开发平台的导入模块的界面中，所述预设数据标注标准用于展示在所述用户开发平台的标注模块的界面中；
    响应于所述针对模型生产线的指标配置操作，得到所述预搭建的模型生产线中待设置的网络评估指标，所述待设置的网络评估指标用于在所述用户开发平台的评估模块进行设置。
  12. 根据权利要求1所述的系统，其特征在于，所述目标任务包括图像处理任务，所述图像处理任务包括：图像识别、图像分割、图像分类、关键点检测中的至少一种。
  13. 一种电子设备，其特征在于，包括：
    处理器；
    用于存储处理器可执行指令的存储器；
    其中，所述处理器被配置为调用所述存储器存储的指令，以执行权利要求1至12中任意一项所述的系统。
  14. 一种计算机可读存储介质，其上存储有计算机程序指令，其特征在于，所述计算机程序指令被处理器执行时实现权利要求1至12中任意一项所述的系统。
  15. 一种计算机程序，包括计算机可读代码，当所述计算机可读代码在电子设备中运行时，所述电子设备中的处理器执行用于实现权利要求1至12中的任一权利要求所述的系统。
PCT/CN2021/124453 2021-04-02 2021-10-18 深度学习模型生产系统、电子设备和存储介质 WO2022205835A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110363166.XA CN113052328B (zh) 2021-04-02 2021-04-02 深度学习模型生产系统、电子设备和存储介质
CN202110363166.X 2021-04-02

Publications (1)

Publication Number Publication Date
WO2022205835A1 true WO2022205835A1 (zh) 2022-10-06

Family

ID=76517289

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/124453 WO2022205835A1 (zh) 2021-04-02 2021-10-18 深度学习模型生产系统、电子设备和存储介质

Country Status (2)

Country Link
CN (1) CN113052328B (zh)
WO (1) WO2022205835A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116994609A (zh) * 2023-09-28 2023-11-03 苏州芯合半导体材料有限公司 应用于智能生产线的数据分析方法及系统

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113052328B (zh) * 2021-04-02 2023-05-12 上海商汤科技开发有限公司 深度学习模型生产系统、电子设备和存储介质
CN113505895B (zh) * 2021-08-05 2023-05-05 上海高德威智能交通系统有限公司 一种机器学习引擎服务系统及模型训练方法和配置方法
CN114153540A (zh) * 2021-11-30 2022-03-08 上海商汤科技开发有限公司 预训练模型发布方法及装置、电子设备和存储介质
CN114186697B (zh) * 2021-12-10 2023-03-14 北京百度网讯科技有限公司 基于深度学习框架生成和应用深度学习模型的方法及装置
CN114329201B (zh) * 2021-12-27 2023-08-11 北京百度网讯科技有限公司 深度学习模型的训练方法、内容推荐方法和装置
CN115796272B (zh) * 2022-11-24 2024-03-12 北京百度网讯科技有限公司 基于深度学习平台的模型训练方法、数据处理方法和装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110991649A (zh) * 2019-10-28 2020-04-10 中国电子产品可靠性与环境试验研究所（（工业和信息化部电子第五研究所）（中国赛宝实验室）） 深度学习模型搭建方法、装置、设备和存储介质
CN111310934A (zh) * 2020-02-14 2020-06-19 北京百度网讯科技有限公司 一种模型生成方法、装置、电子设备和存储介质
US20200210848A1 (en) * 2018-12-29 2020-07-02 International Business Machines Corporation Deep learning testing
CN112416301A (zh) * 2020-10-19 2021-02-26 山东产研鲲云人工智能研究院有限公司 深度学习模型开发方法及装置、计算机可读存储介质
CN113052328A (zh) * 2021-04-02 2021-06-29 上海商汤科技开发有限公司 深度学习模型生产系统、电子设备和存储介质

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110598868B (zh) * 2018-05-25 2023-04-18 腾讯科技（深圳）有限公司 一种机器学习模型搭建方法、装置及相关设备
CN109934512B (zh) * 2019-03-28 2024-02-09 努比亚技术有限公司 一种预测模型的训练方法及系统
CN110532445A (zh) * 2019-04-26 2019-12-03 长佳智能股份有限公司 提供类神经网络训练模型的云端交易系统及其方法
US20210019665A1 (en) * 2019-07-18 2021-01-21 International Business Machines Corporation Machine Learning Model Repository Management and Search Engine
CN117893845A (zh) * 2019-09-17 2024-04-16 华为云计算技术有限公司 提供ai模型的方法、ai平台、计算设备及存储介质
CN111611239A (zh) * 2020-04-17 2020-09-01 第四范式（北京）技术有限公司 实现自动机器学习的方法、装置、设备及存储介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200210848A1 (en) * 2018-12-29 2020-07-02 International Business Machines Corporation Deep learning testing
CN110991649A (zh) * 2019-10-28 2020-04-10 中国电子产品可靠性与环境试验研究所（（工业和信息化部电子第五研究所）（中国赛宝实验室）） 深度学习模型搭建方法、装置、设备和存储介质
CN111310934A (zh) * 2020-02-14 2020-06-19 北京百度网讯科技有限公司 一种模型生成方法、装置、电子设备和存储介质
CN112416301A (zh) * 2020-10-19 2021-02-26 山东产研鲲云人工智能研究院有限公司 深度学习模型开发方法及装置、计算机可读存储介质
CN113052328A (zh) * 2021-04-02 2021-06-29 上海商汤科技开发有限公司 深度学习模型生产系统、电子设备和存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116994609A (zh) * 2023-09-28 2023-11-03 苏州芯合半导体材料有限公司 应用于智能生产线的数据分析方法及系统
CN116994609B (zh) * 2023-09-28 2023-12-01 苏州芯合半导体材料有限公司 应用于智能生产线的数据分析方法及系统

Also Published As

Publication number Publication date
CN113052328A (zh) 2021-06-29
CN113052328B (zh) 2023-05-12

Similar Documents

Publication Publication Date Title
WO2022205835A1 (zh) 深度学习模型生产系统、电子设备和存储介质
CN107491541B (zh) 文本分类方法及装置
CN108197327B (zh) 歌曲推荐方法、装置及存储介质
US11430427B2 (en) Method and electronic device for separating mixed sound signal
CN105320428B (zh) 用于提供图像的方法和设备
TWI773481B (zh) 圖像處理方法及裝置、電子設備和電腦可讀儲存介質
WO2022011892A1 (zh) 网络训练方法及装置、目标检测方法及装置和电子设备
US20210248718A1 (en) Image processing method and apparatus, electronic device and storage medium
CN111445493B (zh) 图像处理方法及装置、电子设备和存储介质
CN110909815B (zh) 神经网络训练、图像处理方法、装置及电子设备
TWI735112B (zh) 圖像生成方法、電子設備和儲存介質
CN114240882A (zh) 缺陷检测方法及装置、电子设备和存储介质
JP2022522551A (ja) 画像処理方法及び装置、電子機器並びに記憶媒体
US11012388B2 (en) Media enhancement system
CN113792207A (zh) 一种基于多层次特征表示对齐的跨模态检索方法
CN111611490A (zh) 资源搜索方法、装置、设备及存储介质
CN112598063A (zh) 神经网络生成方法及装置、电子设备和存储介质
CN115641518B (zh) 一种无人机用视图感知网络模型及目标检测方法
CN109671051A (zh) 图像质量检测模型训练方法及装置、电子设备和存储介质
CN111435422B (zh) 动作识别方法、控制方法及装置、电子设备和存储介质
WO2022141969A1 (zh) 图像分割方法及装置、电子设备、存储介质和程序
CN115100492B (zh) Yolov3网络训练、pcb表面缺陷检测方法及装置
WO2023097952A1 (zh) 预训练模型发布方法及装置、电子设备、存储介质和计算机程序产品
CN113377976B (zh) 资源搜索方法、装置、计算机设备及存储介质
WO2022110801A1 (zh) 数据处理方法及装置、电子设备和存储介质

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2022531620

Country of ref document: JP

Kind code of ref document: A

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21934471

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21934471

Country of ref document: EP

Kind code of ref document: A1