CN114359767A

CN114359767A - Product data processing method and device, storage medium and processor

Info

Publication number: CN114359767A
Application number: CN202011065307.1A
Authority: CN
Inventors: 黄梁华; 潘攀; 王彬; 刘宇
Original assignee: Alibaba Group Holding Ltd
Current assignee: Alibaba Group Holding Ltd
Priority date: 2020-09-30
Filing date: 2020-09-30
Publication date: 2022-04-15

Abstract

The invention discloses a method and a device for processing product data, a storage medium and a processor. The method comprises the following steps: acquiring an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos; distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process; polymerizing the obtained at least one local gradient to obtain a polymerization gradient; the data characterization model is trained using the aggregate gradient. The invention solves the technical problem that the structured modeling result of the product data is inaccurate.

Description

Product data processing method and device, storage medium and processor

Technical Field

The invention relates to the field of data processing, in particular to a method and a device for processing product data, a storage medium and a processor.

Background

Currently, in the process of structured modeling of product data, a supervised learning method is generally used for modeling.

The supervised learning method is to label the product information of the commodity by manpower and then complete the structural modeling by learning the mapping relation between the product information and the artificial label. However, the supervised learning method cannot fully utilize larger-scale unlabelled data to improve the model accuracy, and is difficult to cover complicated and variable scenes, commodity types and shooting conditions, so that the technical problem that the structured modeling result of product data is inaccurate exists.

Aiming at the technical problem that the structured modeling result of the product data is inaccurate, an effective solution is not provided at present.

Disclosure of Invention

The embodiment of the invention provides a method and a device for processing product data, a storage medium and a processor, which are used for at least solving the technical problem that the structured modeling result of the product data is inaccurate.

According to an aspect of an embodiment of the present invention, a method for processing product data is provided. The method can comprise the following steps: acquiring an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process; polymerizing the obtained at least one local gradient to obtain a polymerization gradient; training data characterization models using aggregate gradients

According to another aspect of the embodiment of the invention, another product data processing method is also provided. The method can comprise the following steps: an entry page on the operation interface enters an image data set of the product object, wherein the image data set comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; inducing a model generation instruction in an operation interface, distributing a plurality of images in an image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregation gradient; and displaying prompt information of the training data representation model on the operation interface, wherein the training data representation model is trained by using the aggregation gradient.

According to another aspect of the embodiment of the invention, another product data processing method is also provided. The method can comprise the following steps: displaying an image dataset of a product object on an interactive interface, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; sensing a model generation instruction in the interactive interface; responding to a model generation instruction, selecting a plurality of images from an image data set, distributing the plurality of images to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregation gradient; and displaying prompt information of the training data representation model on the interactive interface, wherein the training data representation model is trained by using the aggregation gradient.

According to another aspect of the embodiment of the invention, another product data processing method is also provided. The method can comprise the following steps: the front-end client uploads an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; the front-end client transmits the image data set to a background server; the front-end client receives a trained data representation model returned by the background server, wherein a plurality of images in the image data set are distributed to a plurality of processes by the background server for distribution calculation, at least one process acquires a local gradient in a back propagation process, the acquired local gradients are aggregated to obtain an aggregation gradient, and the aggregation gradient is used for training the data representation model.

According to another aspect of the embodiment of the invention, a device for processing product data is also provided. The apparatus may include: an acquisition unit for acquiring an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; the distribution unit is used for distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, and at least one process acquires a local gradient in a back propagation process; the polymerization unit is used for polymerizing the acquired at least one local gradient to obtain a polymerization gradient; a training unit for training the data characterization model using the aggregate gradient.

According to another aspect of the embodiment of the invention, another processing device for product data is also provided. The apparatus may include: the entry unit is used for entering an image data set of the product object on an entry page on the operation interface, wherein the image data set comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; the first sensing unit is used for sensing a model generation instruction in an operation interface, distributing a plurality of images in an image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating the acquired at least one local gradient to obtain an aggregate gradient; and the first display unit is used for displaying prompt information of the training data representation model on the operation interface, wherein the training data representation model is trained by using the aggregation gradient.

According to another aspect of the embodiment of the invention, another processing device for product data is also provided. The apparatus may include: a second display unit for displaying an image dataset of the product object on the interactive interface, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; the second sensing unit is used for sensing a model generation instruction in the interactive interface; the response unit is used for responding to the model generation instruction, selecting a plurality of images from the image data set, distributing the plurality of images to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregate gradient; and the third display unit is used for displaying prompt information of the training data representation model on the interactive interface, wherein the training data representation model is trained by using the aggregation gradient.

According to another aspect of the embodiment of the invention, another processing device for product data is also provided. The apparatus may include: an upload unit configured to enable a front-end client to upload an image dataset of a product object, wherein the image dataset includes: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; the transmission unit is used for enabling the front-end client to transmit the image data set to the background server; the receiving unit is used for enabling the front-end client to receive the trained data representation model returned by the background server, wherein a plurality of images in the image data set are distributed to a plurality of processes by the background server for distribution calculation, at least one process acquires a local gradient in a back propagation process, the acquired local gradients are aggregated to obtain an aggregation gradient, and the aggregation gradient is used for training the data representation model.

According to another aspect of the embodiments of the present invention, there is also provided a computer-readable storage medium. The computer readable storage medium includes a stored program, wherein the program, when executed by a processor, controls an apparatus in which the computer readable storage medium is located to perform a method of processing product data of an embodiment of the present invention.

According to another aspect of the embodiments of the present invention, there is also provided a processor. The processor is configured to execute a program, wherein the program executes a method of processing product data according to an embodiment of the present invention.

According to another aspect of the embodiment of the invention, a system for processing product data is also provided. The system may include: a processor; a memory coupled to the processor for providing instructions to the processor for processing the following processing steps: acquiring an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process; polymerizing the obtained at least one local gradient to obtain a polymerization gradient; the data characterization model is trained using the aggregate gradient.

In an embodiment of the invention, an image dataset of a product object is acquired, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process; polymerizing the obtained at least one local gradient to obtain a polymerization gradient; the data characterization model is trained using the aggregate gradient. That is to say, the supervised machine learning method is used, distributed training can be performed on the unmarked product object according to the image data set of the unmarked product object, so that a data representation model is effectively constructed, the data representation model is not interfered by other large amount of noise at all, the problem that the supervised learning method cannot fully utilize larger-scale unmarked data to improve the model precision is avoided, the problem that a small amount of marked data is difficult to cover complicated and changeable scenes, commodity types and shooting conditions, and the model constructed on the data possibly has certain data deviation is also avoided, the technical problem that the structured modeling result of the product data is inaccurate is solved, and the technical effect of improving the accuracy of the structured modeling result of the product data is achieved.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the invention without limiting the invention. In the drawings:

fig. 1 is a block diagram of a hardware configuration of a computer terminal (or mobile device) for implementing processing of product data according to an embodiment of the present invention;

FIG. 2 is a flow chart of a method of processing product data according to an embodiment of the present invention;

FIG. 3 is a flow chart of another method of processing product data according to an embodiment of the present invention;

FIG. 4 is a flow chart of another method of processing product data according to an embodiment of the present invention;

FIG. 5 is a flow chart of another method of processing product data according to an embodiment of the present invention;

FIG. 6 is a schematic diagram of a noise contrast loss function according to an embodiment of the present invention;

FIG. 7A is a schematic diagram of a distributed training process according to an embodiment of the present invention;

FIG. 7B is a schematic illustration of an interactive interface of a method of processing product data according to an embodiment of the present invention;

FIG. 7C is a schematic diagram of a scenario illustrating a method for processing product data according to an embodiment of the present invention;

FIG. 8 is a schematic diagram of a product data processing apparatus according to an embodiment of the present invention;

FIG. 9 is a schematic view of another product data processing apparatus according to an embodiment of the present invention;

FIG. 10 is a schematic view of another product data processing apparatus according to an embodiment of the present invention;

FIG. 11 is a schematic view of another product data processing apparatus according to an embodiment of the present invention; and

fig. 12 is a block diagram of a computer terminal according to an embodiment of the present invention.

Detailed Description

In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the invention described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

First, some terms or terms appearing in the description of the embodiments of the present application are applicable to the following explanations:

the deep learning is an artificial neural network structure with a high layer number and is used for realizing the functions of intelligent image detection, classification and the like.

Self-supervision, a machine learning method which can directly build a model from data without manually providing a data label;

the data characterization model is a model obtained by carrying out structural modeling on data of a product object;

a Standardized Product Unit (SPU) is a set of reusable and easily retrievable standardized information, which describes the characteristics of a Product;

a Noise contrast loss function (NCE) for calculating local loss according to the local similarity matrix;

transfer Learning (Transfer Learning), which is a machine Learning method, is a method for transferring knowledge in one field (source field) to another field (target field) so that the target field can obtain a better Learning effect.

Example 1

There is also provided, in accordance with an embodiment of the present invention, an embodiment of a method for processing product data, it being noted that the steps illustrated in the flowchart of the figure may be performed in a computer system such as a set of computer-executable instructions and that, although a logical order is illustrated in the flowchart, in some cases the steps illustrated or described may be performed in an order different than that presented herein.

The method provided by the first embodiment of the present application may be executed in a mobile terminal, a computer terminal, or a similar computing device. Fig. 1 is a block diagram of a hardware configuration of a computer terminal (or mobile device) for implementing processing of product data according to an embodiment of the present invention. As shown in fig. 1, the computer terminal 10 (or mobile device 10) may include one or more (shown as 102a, 102b, … …, 102 n) processors 102 (the processors 102 may include, but are not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions. Besides, the method can also comprise the following steps: a display, an input/output interface (I/O interface), a Universal Serial Bus (USB) port (which may be included as one of the ports of the I/O interface), a network interface, a power source, and/or a camera. It will be understood by those skilled in the art that the structure shown in fig. 1 is only an illustration and is not intended to limit the structure of the electronic device. For example, the computer terminal 10 may also include more or fewer components than shown in FIG. 1, or have a different configuration than shown in FIG. 1.

It should be noted that the one or more processors 102 and/or other data processing circuitry described above may be referred to generally herein as "data processing circuitry". The data processing circuitry may be embodied in whole or in part in software, hardware, firmware, or any combination thereof. Further, the data processing circuit may be a single stand-alone processing module, or incorporated in whole or in part into any of the other elements in the computer terminal 10 (or mobile device). As referred to in the embodiments of the application, the data processing circuit acts as a processor control (e.g. selection of a variable resistance termination path connected to the interface).

The memory 104 may be used to store software programs and modules of application software, such as program instructions/data storage devices corresponding to the processing method of the product data in the embodiment of the present invention, and the processor 102 executes various functional applications and data processing, i.e., implements the processing method of the product data of the application program, by running the software programs and modules stored in the memory 104. The memory 104 may include high speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include memory located remotely from the processor 102, which may be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The transmission device 106 is used for receiving or transmitting data via a network. Specific examples of the network described above may include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network adapter (NIC) that can be connected to other Network devices through a base station to communicate with the internet. In one example, the transmission device 106 can be a Radio Frequency (RF) module, which is used to communicate with the internet in a wireless manner.

The display may be, for example, a touch screen type Liquid Crystal Display (LCD) that may enable a user to interact with a user interface of the computer terminal 10 (or mobile device).

It should be noted here that in some alternative embodiments, the computer device (or mobile device) shown in fig. 1 described above may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that fig. 1 is only one example of a particular specific example and is intended to illustrate the types of components that may be present in the computer device (or mobile device) described above.

In the operating environment shown in fig. 1, the present application provides a method of processing product data as shown in fig. 2. It should be noted that the processing method of the product data of this embodiment may be executed by the mobile terminal of the embodiment shown in fig. 1.

Fig. 2 is a flowchart of a method for processing product data according to an embodiment of the present invention. As shown in fig. 2, the method may include the steps of:

step S202, acquiring an image data set of the product object, wherein the image data set includes: a plurality of images of the plurality of pictures and the plurality of image frames of the video, and none of the images in the image dataset are labeled.

In the technical solution provided by the above step S202, the product object may be a commodity object, for example, a new commodity to be issued by a seller. The method includes acquiring an image data set of a product object, where the image data set may be data of a plurality of images obtained by shooting the product object and may include a plurality of images of the product object and a plurality of image frames of a video, and in the case that the product object is a commodity, the image of the product object may be referred to as commodity image data. Optionally, the embodiment obtains the image data set of the product object from the e-commerce platform, and the e-commerce platform is not particularly limited herein.

In this embodiment, none of the images in the image dataset are labeled, that is, the image dataset may be large-scale (massive) label-free data, so that the method for processing product data in this embodiment is for a label-free image dataset.

Step S204, distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process.

In the technical solution provided by step S204 of the present invention, after the image dataset of the product object is acquired, a plurality of images in the image dataset are distributed to a plurality of processes for distribution calculation, and at least one process acquires a local gradient in a back propagation process.

The embodiment may be a distributed training scheme of parallel processes, and distribute a plurality of images in an image data set to a plurality of processes for distributed computation, where the plurality of processes may be all processes currently, for example, process 1, process 2 … …, process N. At least one of the processes may be configured to compute a local similarity matrix and a local loss, and optionally, the at least one process may perform distributed computation on a plurality of images in the image dataset on a corresponding Graphics Processing Unit (GPU). Optionally, in the process of performing back propagation, the embodiment returns the local gradient, that is, the gradient information, to each process, thereby implementing efficient training of the model.

And step S206, polymerizing the acquired at least one local gradient to obtain a polymerization gradient.

In the technical solution provided in step S206 of the present invention, after at least one process acquires a local gradient in a back propagation process, the acquired at least one local gradient may be aggregated in the at least one process, so as to obtain an aggregated gradient under the at least one process, so as to complete optimization of a model on each process, that is, the aggregated gradient in this embodiment is for the at least one process, and this training process improves training efficiency of a data characterization model.

In step S208, the data characterization model is trained using the aggregate gradient.

In the technical solution provided in step S208 of the present invention, after aggregating at least one obtained local gradient to obtain an aggregate gradient, the aggregate gradient is used to train a data characterization model, for example, the aggregate gradient is used to learn (optimize) a data characterization model with strong discriminability, and the data characterization model may be referred to as a pre-training model. Optionally, the data characterization model of this embodiment may serve as a feature extractor and a classifier, where the feature extractor is configured to extract features, the classifier is configured to classify the features, and the data characterization model may be migrated to a product identification task to improve accuracy and generalization performance of product object identification.

It should be noted that the above-mentioned method of this embodiment can directly construct the data representation model from the image data set itself without manually providing a data tag of the image, and thus, the above-mentioned method of this embodiment is a self-supervised machine learning method.

The embodiment builds the data representation model in a self-supervision learning mode only according to the information of a large amount of image data sets, and fine tuning can be carried out by transferring the data representation model to a small amount of manually marked data, so that higher performance of the model is achieved.

In this embodiment, the above steps S202 to S208 may be executed in a loop, so as to continuously achieve the purpose of optimizing the data characterization model, so as to improve the performance of the data characterization model.

The method for processing product data in this embodiment is a data structured modeling process of a product object, and an image dataset of the product object is obtained through the above steps S202 to S208 in this application, where the image dataset includes: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process; polymerizing the obtained at least one local gradient to obtain a polymerization gradient; the data characterization model is trained using the aggregate gradient. That is to say, the embodiment uses the supervised machine learning method, and the data representation model can be effectively constructed only according to the image data set of the unmarked product object, which is not interfered by other large amount of noise at all, so that the problem that the model constructed on the data may have certain data deviation due to the fact that a small amount of marked data is difficult to cover complicated and changeable scenes, commodity types and shooting conditions and the problem that the structured modeling result of the product data is inaccurate is solved, and the technical effect of improving the accuracy of the structured modeling result of the product data is achieved.

The above-described method of this embodiment is further described below.

As an optional implementation manner, in step S204, distributing the multiple images in the image dataset to multiple processes for distribution calculation, where at least one process acquires a local gradient in a back propagation process, and the method includes: acquiring the image characteristics of each image in the image data set, and distributing the image characteristics of each image to different processes; calculating a similarity matrix between image features, and acquiring a local similarity matrix generated under at least one process; performing local loss calculation on the local similarity matrix generated under the at least one process to generate local loss under the at least one process; and performing gradient calculation based on the local loss acquired by at least one process to acquire a local gradient in the back propagation process.

In this embodiment, the image dataset includes a plurality of images, and in the process of training the data representation model, the algorithm needs to aggregate and store all image features in different machines and different processes, so that in the embodiment, when the distribution of the plurality of images in the image dataset to the plurality of processes is implemented to perform distribution calculation, and at least one process obtains a local gradient in the back propagation process, the image features may be extracted from each image in the image dataset, each image may be input to the representation model to obtain the image features of each image, and then the image features of each image are distributed to different processes, for example, the image features of each image are uniformly distributed to different processes, at least one process is only used to calculate a local similarity matrix and a local loss, and the similarity matrix calculation may be performed, And the loss calculation is uniformly distributed to each process so as to complete the calculation of local loss, thereby realizing a distributed training scheme based on parallel processes.

Optionally, in this embodiment, a similarity matrix between image features of each image is calculated through at least one process, a local similarity matrix generated under the at least one process is obtained according to the similarity matrix, and then local loss calculation is performed on the local similarity matrix generated under the at least one process, so as to generate local loss (loss value) under the at least one process. It should be noted that the above-mentioned procedure of this embodiment has a large amount of data transmission, calculation, and storage.

After the local loss under at least one process is generated, gradient calculation can be performed based on the local loss acquired by at least one process, that is, local gradients are distributed and calculated on each GPU, so that a plurality of local gradients in a back propagation process are acquired, the local gradients on the GPUs can be averaged, and thus data are processed in a large scale, a purpose of aggregating the local gradients and training a data representation model according to the local gradients is achieved, and training efficiency of the data representation model is improved.

As an alternative embodiment, the image characteristics of each image in the image dataset are acquired: the neural network model is used to identify image features of each image in the image dataset and extract image features of each image.

In this embodiment, in implementing the acquiring of the image features of each image in the image dataset, it may be to read a pre-trained neural network model, which may include, but is not limited to, a convolutional neural network model, and then use the neural network model to identify the image features of each image in the image dataset. After the image features of each image in the image dataset are identified, the image features can be extracted from each image to distribute the image features of each image to different processes, so that a distributed training scheme of the data characterization model is realized.

As an optional implementation manner, calculating a similarity matrix between image features, and obtaining a local similarity matrix generated in at least one process includes: in the process of carrying out local similarity matrix calculation on any process, image features generated in other processes are pulled; and performing similarity calculation by using the pulled image features and the image features in the current process to generate a local similarity matrix in the current process.

In this embodiment, when the similarity matrix between the image features is calculated and the local similarity matrix generated in at least one process is obtained, in the process of performing the local similarity matrix calculation in any one process, the image features generated in the processes except the process may also be simultaneously pulled, that is, all the image features are pulled from other GPUs and collected on one GPU to calculate the local similarity matrix, where any one process may also be referred to as a current process, and the pulled image features and the image features in the current process may be used to perform the similarity calculation, so as to generate the local similarity matrix in the current process.

It should be noted that, the local similarity matrix under at least one of the processes may be obtained according to the method.

As an optional implementation manner, a noise contrast loss function is used to perform local loss calculation on a local similarity matrix generated in at least one process, so as to generate local loss in the at least one process, where the noise contrast loss function is used to reduce image feature distances between different transformations in the same image, so that image feature distances between different images are expanded.

In this embodiment, when calculating the local loss according to the local similarity matrix, a noise contrast loss function may be read, and then the local loss calculation is performed on the local similarity matrix generated in at least one process by using the noise contrast loss function, for example, the noise contrast loss function may be represented by the following formula:

where i, j are used for two transformations representing the image of the same product object, which may include, but are not limited to, scaling, matting, color transformation, etc., where in the various transformations the meaning of the image is not changed and k is used to represent the image of other product objects.

By the noise contrast loss function of the embodiment, the image feature distances between different transformations in the same image can be gradually reduced, that is, the distances between the image features of the image between different transformations are gradually reduced; the noise contrast loss function of this embodiment may also gradually enlarge the image feature distance between different images, i.e. gradually push away the image feature distance between different transformations of the image. Taking the image a and the image B of the product object as an example, the noise contrast loss function may require that the distances between the image features of the same image between different transformations are as close as possible, and the distances between the image features between different transformations are as far as possible, so as to effectively learn the data characterization model with strong discriminability.

As an optional implementation manner, in step S206, aggregating the obtained at least one local gradient to obtain an aggregated gradient, including: distributing the local gradient under any process to other processes; and at least one process uses the local gradient in the current process and the local gradients distributed by other processes for aggregation to generate an aggregation gradient under at least one process.

In this embodiment, when the obtained at least one local gradient is aggregated to obtain an aggregated gradient, the local gradient in any one process may be distributed to other processes except any one process in the multiple processes, and at least one process is determined as a current process, so that the local gradient in the current process and the local gradients distributed by other processes are aggregated by the at least one process, thereby generating the aggregated gradient in the at least one process.

Optionally, in the process of performing aggregation gradient in any one process, local gradients in other processes may be pulled; and aggregating the pulled local gradient and the local gradient in the current process to generate an aggregate gradient under the current process.

As an alternative embodiment, the frequency of gradient aggregation is determined based on the amount of communication, the amount of computation, and the number of accesses between at least one process.

In this embodiment, the obtained at least one local gradient is aggregated, and the obtained aggregated gradient has a certain frequency, which may be determined based on a communication amount, a calculation amount, and an access frequency between at least one process, where the communication amount is a data amount for data communication between at least one process, the calculation amount is a data amount for data calculation between at least one process, and the access frequency is a number of times for data storage and reading between at least one process.

As an alternative embodiment, after training the data characterization model using the gradient of aggregation in step S208, the method further includes: acquiring an identification task of product information; and identifying the identification task by using the trained data characterization model to obtain product information.

In this embodiment, the trained data representation model may be migrated to a recognition task of Product information through migration learning, where the Product information may be a text description for describing a Product object, for example, commodity information, which may include but is not limited to information such as a Product title and a Product selling point of the Product object, and may also be structured information such as a category, an attribute, a class, and a Standardized Product Unit (SPU) of the Product object, which may be generated by acquiring Product data of the Product object, analyzing the Product data, generating multi-modal information of the Product object, and processing the multi-modal information by using a multi-modal network model, where the Product data includes at least one of: picture information, video information and character information of the product; analyzing the product data to generate multi-modal information for the product object, wherein the multi-modal information comprises: a feature sequence between different modality information, wherein the multi-modal network model may be a multi-modal transformer network model.

In this embodiment, the identification task is identified by using the trained data representation model, which may be copying the model parameters of the trained data representation model into the model of the identification task of the product information as initialization, and then continuing to train the model of the identification task of the product information, where the identification task may be a task for distinguishing different product objects, for example, the identification task is a task for distinguishing a shirt from a body-shirt. Optionally, when the product object is a commodity, the model of the identification task of the product information may be referred to as a commodity identification model, and the identification task is identified through the model, so that the product information for describing the product object is identified, the accuracy of identification of the category, the attribute, the category and the like of the product object is higher, the performance is more stable, and the accuracy, the generalization performance and the cross-domain capability of the product information identification are improved.

As an optional embodiment, after identifying the product information using the trained data characterization model, the method further comprises: generating various types of product materials based on the product information; and issuing a plurality of product materials.

In this embodiment, after the product information describing the product object is generated, for example, after the product title and the product selling point of the product object are generated, a plurality of types of product materials may be generated therethrough, the product materials are materials required for releasing the product object, and may be picture product materials, video product materials, text product materials, and the like, and each type of product material may include the above-mentioned product information, and then a plurality of product materials are released.

As an optional embodiment, after generating the plurality of types of product materials, the method further comprises: uploading a product material to be issued, and extracting a plurality of product contents to be verified in the product material to be issued; judging whether at least one product content to be verified meets an entry standard; if yes, successfully inputting the product materials into the release template; otherwise, preprocessing the product content failed in verification, and recording the product material into the release template under the condition that the preprocessed product content meets the recording standard.

In this embodiment, after the product material to be issued is generated, the product material to be issued may be uploaded to a product issuing platform, and a plurality of product contents to be verified are extracted from the product material to be issued, where the plurality of product contents are contents that need to be entered into an issuing template in the product material, at least one product content to be verified may be determined from the plurality of product contents, and then it is determined whether the at least one product content to be verified satisfies an entry standard, where the entry standard is also used to determine whether the at least one product content meets a standard, and if it is determined that the at least one product content to be verified satisfies the entry standard, it is determined that the product material is successfully authenticated, and then the product material may be successfully entered into the issuing template; if the at least one product content to be verified is judged not to meet the standard, the at least one product content is determined to fail in verification, the product content failed in verification is preprocessed, for example, the product content failed in verification is modified and adjusted, whether the preprocessed product content meets the entry standard or not is judged, if the preprocessed product content meets the entry standard, the product material can be entered into the publishing template, and the product object is published to the product publishing platform through the publishing template.

The embodiment of the invention also provides another product data processing method from the aspect of human-computer interaction.

Fig. 3 is a flowchart of another product data processing method according to an embodiment of the present invention. As shown in fig. 3, the method may include the steps of:

step S302, inputting an image data set of a product object on an input page on an operation interface, wherein the image data set comprises: a plurality of images of the plurality of pictures and the plurality of image frames of the video, and none of the images in the image dataset are labeled.

In the technical solution provided in step S302 of the present invention, an entry page is displayed on the operation interface, where the entry page is used to enter an image data set of a product object, the product object may be a commodity object, for example, a new commodity to be released by a seller, and the image data set may be data of multiple images obtained by shooting the product object, and may include multiple images of the multiple product objects and multiple image frames of a video.

It should be noted that none of the images in the image dataset are labeled, that is, the image dataset may be large-scale (massive) label-free data, and thus the method for processing product data according to this embodiment is for a label-free image dataset.

And S304, inducing a model generation instruction in the operation interface, distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregate gradient.

In the technical solution provided by step S304 of the present invention, after the image data set of the product object is entered on the entry page on the operation interface, a model generation instruction for instructing generation of the data representation model may be received and sensed in the operation interface. Optionally, the operation interface of this embodiment may distribute, in response to the model generation instruction, the plurality of images in the image data set to a plurality of processes for distribution calculation, where the plurality of processes may be all current processes. At least one process may be used to compute local similarity matrices and local penalties, optionally, at least one process may perform a distribution computation on multiple images in the image dataset on a corresponding GPU. Optionally, in the process of performing back propagation, the embodiment returns the local gradient to each process, thereby implementing efficient training of the model.

And S306, displaying prompt information of the training data representation model on the operation interface, wherein the training data representation model is trained by using the aggregation gradient.

In the technical solution provided in step S306 of the present invention, after aggregating at least one obtained local gradient to obtain an aggregated gradient, the aggregated gradient training data characterization model may be used, and then prompt information is displayed on the operation interface, where the prompt information may be used to indicate that the data characterization model is being trained, or to complete training of the data characterization model.

This embodiment uses the aggregate gradient to learn (optimize) discriminative data characterization models. Optionally, the data characterization model of this embodiment may be used as a feature extractor and a classifier, and may be migrated to a product identification task to improve the accuracy and generalization performance of product object identification.

The embodiment builds the data representation model in a self-supervision learning mode only according to the information of a large amount of image data sets, and fine tuning can be carried out by transferring the data representation model to a small amount of manually marked data, so that higher performance of the model is realized.

As an optional implementation manner, after displaying prompt information of the training data characterization model on the operation interface in step S306, the method further includes: displaying the identification task of the product information on an operation interface; and displaying product information on the operation interface, wherein the product information is obtained by using the trained data representation model to identify the identification task.

In this embodiment, the trained data representation model may be migrated to a task of identifying product information, where the product information may be a text description for describing a product object, for example, commodity information, which may include, but is not limited to, a product title, a product selling point, and other information of the product object. After the prompt information of the training data representation model is displayed on the operation interface, the identification task of the product information can be displayed on the operation interface, and then the identification operation instruction of the task can be received and responded on the operation interface, and the identification task is identified by using the trained data representation model.

Optionally, in this embodiment, the trained model parameters of the data representation model are copied to the model of the identification task of the product information as initialization, and then the model of the identification task of the product information is trained continuously, where, when the product object is a commodity, the model of the identification task of the product information may be called a commodity identification model, and the identification task is identified through the model, so as to identify the product information used for describing the product object, and further display the product information on the operation interface, so that the accuracy of identification of the category, attribute, class, and the like of the product object is higher, and the performance is more stable, thereby achieving the purposes of improving the accuracy, generalization performance, and cross-domain capability of product information identification.

As an optional implementation, after the product information is displayed on the operation interface, the method further includes: popping up guide information on an operation interface, wherein the guide information comprises defect information existing in the product information; displaying an authoring material generated based on the guidance information on an operation interface, wherein the authoring material is basic information constituting a product material; generating various types of product materials based on the authoring materials; and issuing a plurality of product materials.

In this embodiment, after the product information describing the product object is displayed on the operation interface, guidance information may be popped up on the operation interface, where the guidance information may include defect information existing in the product information, the defect information indicating a problem existing in the product information when generating the product material, and may be used to guide generation of an authoring material, where the authoring material is basic information constituting the product material. The embodiment may generate the authoring material based on the guidance information, for example, by omitting the omission based on the guidance information to generate the authoring material, and further display the authoring material on the operation interface.

After the creation material generated based on the guidance information is displayed on the operation interface, a plurality of types of product materials can be generated based on the creation material, the product materials are materials required when the product object is released, and can be picture product materials, video product materials, character product materials and the like, and each type of product material can comprise the product information, so that a plurality of product materials are released.

Fig. 4 is a flowchart of another product data processing method according to an embodiment of the present invention. As shown in fig. 4, the method may include the steps of:

step S402, displaying an image data set of the product object on the interactive interface, wherein the image data set comprises: a plurality of images of the plurality of pictures and the plurality of image frames of the video, and none of the images in the image dataset are labeled.

In the technical solution provided by step S402 of the present invention, an image data set obtained by identifying the product object may be directly displayed on the interactive interface, where the image data set may be data of a plurality of images obtained by shooting the product object, and may include a plurality of image frames of pictures and videos of the product object.

It should be noted that none of the images in the image dataset of the embodiment are labeled, that is, the image dataset may be large-scale (massive) label-free data, and thus the method for processing product data of the embodiment is for the label-free image dataset.

And S404, sensing a model generation instruction in the interactive interface.

In the technical solution provided by step S404 of the present invention, after the image data set of the product object is displayed on the interactive interface, a model generation instruction may be sensed in the interactive interface, where the model generation instruction is used to instruct to generate a data representation model, and may be an instruction triggered on the operation interface by the user to generate the data representation model.

And S406, responding to a model generation instruction, selecting a plurality of images from the image data set, distributing the plurality of images to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregated gradient.

In the technical solution provided in step S406 of the present invention, after the model generation instruction is sensed in the interactive interface, a plurality of images are selected from the image data set in response to the model generation instruction, for example, 1000 images are randomly extracted from the image data set for training. After selecting a plurality of images from the image data set, distributing the selected plurality of images to a plurality of processes for distribution calculation, wherein the plurality of processes can be all the processes at present. At least one process may be used to compute local similarity matrices and local penalties, optionally, at least one process may perform a distribution computation on multiple images in the image dataset on a corresponding GPU. Optionally, in the process of performing back propagation, the embodiment returns the local gradient to each process, thereby implementing efficient training of the model.

And step S408, displaying prompt information of the training data representation model on the interactive interface, wherein the training data representation model is trained by using the aggregation gradient.

In the technical solution provided in step S408 of the present invention, after at least one process acquires a local gradient in a back propagation process, and aggregates the acquired at least one local gradient to obtain an aggregated gradient, prompt information of a training data characterization model may be displayed on an interactive interface, where the prompt information may be used to indicate that the training data characterization model is being trained, or to complete the training data characterization model.

Fig. 5 is a flowchart of another product data processing method according to an embodiment of the present invention. As shown in fig. 5, the method may include the steps of:

step S502, the front-end client uploads an image data set of the product object, wherein the image data set comprises: a plurality of images of the plurality of pictures and the plurality of image frames of the video, and none of the images in the image dataset are labeled.

In the technical solution provided by step S502 of the present invention, the front-end client may be a merchant issuing end, and may receive an upload operation instruction applied to the operation interface, and start uploading an image data set of a product object in response to the upload operation instruction, where the product object may be a commodity object, and the image data set may be data of a plurality of images obtained by shooting the product object, and may include a plurality of image frames of images and videos of the plurality of product objects.

It should be noted that none of the images in the image dataset of the product object uploaded by the front-end client is labeled, that is, the image dataset may be large-scale (massive) label-free data, so that the method for processing the product data in this embodiment is for the label-free image dataset.

Step S504, the front-end client transmits the image data set to the background server.

In the technical solution provided by step S504 of the present invention, after the front-end client uploads the image data set of the product object, the front-end client may transmit the image data set of the product object to the background server.

In this embodiment, a communication connection is established between the front-end client and the backend server, and the image dataset of the product object may be transmitted to the backend server, so that the backend server generates the data representation model based on the image dataset.

Step S506, the front-end client receives the trained data representation model returned by the background server.

In the technical scheme provided by step S506 of the present invention, a plurality of images in an image data set are distributed to a plurality of processes by a background server for distributed computation, at least one process obtains a local gradient in a back propagation process, at least one obtained local gradient is aggregated to obtain an aggregated gradient, and a data characterization model is trained by using the aggregated gradient.

In this embodiment, after the background server receives the image dataset of the product object, the background server distributes a plurality of images in the image dataset to a plurality of processes for distribution calculation, and at least one process acquires a local gradient in a back propagation process.

Optionally, the background server of this embodiment may distribute, for a distributed training scheme of parallel processes, a plurality of images in the image data set to a plurality of processes for distributed computation. At least one process may be used to compute local similarity matrices and local penalties, optionally, at least one process may perform a distribution computation on multiple images in the image dataset on a corresponding GPU. Optionally, in the process of performing back propagation, the background server returns the local gradient to each process, so as to implement efficient training of the model.

After at least one process acquires the local gradient in the back propagation process, the background server may aggregate the acquired at least one local gradient in the at least one process, so as to obtain an aggregate gradient under the at least one process, so as to complete optimization of the model on each process.

After the background server aggregates the acquired at least one local gradient to obtain an aggregate gradient, the aggregate gradient is used to train the data characterization model, for example, the aggregate gradient is used to learn (optimize) the data characterization model with strong discriminability. Optionally, the data characterization model of this embodiment may serve as a feature extractor and classifier. After the data representation model is trained by the front-end server, the data representation model can be returned to the front-end client.

In the related art, a supervised learning method and a weakly supervised learning method are used in the structured modeling process for product data to perform modeling. The supervised learning method is used for marking the product information of the commodity manually and then completing the structured modeling by learning the mapping relation between the product information and the manual label, but the supervised learning method cannot fully utilize larger-scale unmarked data to improve the model precision and is difficult to cover complex and variable scenes, commodity types and shooting conditions; the weak supervised learning method extracts effective information from product description as a label, and constructs the label by means of the mapping relation between product data and externally provided description information, but the constructed result is greatly influenced by noise information in a description text, and the technical problem that the structured modeling result of the product data is inaccurate exists.

However, the embodiment is a complete recognition pre-training scheme of the product object based on the self-supervision learning algorithm, the distributed training and the transfer learning, the data representation model can be constructed by effectively utilizing the image data sets of the mass unmarked product objects, and the solution is transferred to the recognition task of the product information, and the method aims to fully utilize the image data sets of the mass product objects of the e-commerce platform to complete the structural modeling, alleviate the cross-domain problem (reduce the recognition rate of the commodities with large difference with the labeled data), and improve the accuracy of the product object recognition so as to improve the user experience of the seller and the buyer.

Example 2

The following further describes a preferred implementation of the above method of this embodiment, specifically, the product object is taken as a commercial product for example.

In an e-commerce platform, structured modeling needs to be performed on commodity data, and an algorithm is usually required to be designed to intelligently identify structured information such as categories, attributes, categories, SPUs and the like of commodities. The existing algorithm usually depends on manual labeling of commodity information, and structured modeling is completed by learning the mapping relation between commodity data and manual labels. Most algorithms are supervised learning methods, and are based on a data-driven modeling mode represented by deep learning, which requires intensive manual labeling and data auditing, and the data cost is high. Because the manual labeling is high in cost and difficult to scale, the labeled data only accounts for a very small part of the total commodity data. Therefore, on one hand, the supervised learning method cannot fully utilize larger-scale unmarked data to improve the model precision; on the other hand, the labeled data labeled manually usually only occupies a very small part of the mass commodity data, and is difficult to cover complicated and variable scenes, commodity types, shooting angles, shooting environments and the like, and a model constructed on the data may have certain data deviation, so that the performance may be reduced when the labeled data is applied to unusual scenes, namely, the labeled data faces the cross-domain problem in machine learning. Therefore, the above algorithm cannot sufficiently use large-scale product data, and has a certain cross-domain problem (a decrease in the rate of identifying a product having a large difference from the labeled data).

One solution to solve the above problem is to use the commodity description information provided by the merchant as weak supervision information to construct a commodity model. For example, the product name and introduction characters of the product may be subjected to word segmentation and selected to be used as a label of product data. The method can greatly reduce manual intervention, save cost and is beneficial to scale production. However, a large amount of text noise exists in the commodity description information, which causes significant interference in the structured modeling process and greatly reduces the model effect. Thus, the existing commodity data structured modeling process has the following limitations:

(1) the manual marking and auditing are excessively depended, the data cost is high, and the large-scale production is difficult;

(2) the utilization rate of commodity data is low, and a scheme for effectively utilizing unmarked data is lacked;

(3) there is a cross-domain problem, and the recognition accuracy may be reduced in the face of scenes, commodity categories, and the like that do not appear in the annotation data.

The related technology can extract effective information from the commodity description as a label, and a supervision model is constructed by means of the mapping relation between the commodity data and the description information provided by the seller. The method does not need additional artificial labels and belongs to a weak supervision learning method. However, such methods are greatly affected by noise information in the description text, and the effect is not good. At present, a solution capable of effectively utilizing massive unmarked commodity data to construct a data representation model and transferring the data representation model to a commodity identification task does not exist. Therefore, there is a need to research a pre-training method for self-supervision commodity identification that effectively utilizes massive non-labeled data.

However, the embodiment provides a complete commodity identification pre-training scheme based on a self-supervision learning algorithm, distributed training and transfer learning, and mainly aims to complete structured modeling by fully utilizing mass data of the e-commerce platform, alleviate the cross-domain problem, and improve the accuracy of commodity identification so as to improve the user experience of sellers and buyers. The method of this embodiment is further described below.

According to the embodiment, on the large-scale unmarked commodity picture data, the distance between different transformations of the same commodity picture is minimized and the distance between different commodity pictures is maximized on the basis of the noise contrast estimation loss function, so that the data representation model with strong discriminability is learned, and the representation model is migrated to a commodity identification task, so that the precision and generalization performance of commodity identification are improved. The data characterization model may be a feature extractor and a classifier, among others.

In the process of training the data characterization model, the algorithm needs to aggregate and store all image features in different machines and different processes, calculate the similarity matrix between the image features and the similarity matrix, and further calculate the loss value. The data transmission amount, the calculation amount and the storage amount of the process are large.

Optionally, in this embodiment, 1000 pictures are randomly extracted from an image data set of a commodity to perform training, iteration, and calculation of a similarity matrix of the 1000 pictures, and may be dispersed to three GPUs to obtain image features, and then summarized to one GPU to calculate a similarity matrix between the image features, to obtain a local similarity matrix generated in at least one process, to perform local loss calculation on the local similarity matrix generated in the at least one process, to generate local loss in the at least one process, and to perform gradient calculation based on the local loss obtained in the at least one process, to obtain a local gradient in a back propagation process. Wherein the frequency of gradient aggregation is determined based on at least one inter-process (inter-GPU) traffic, computation, and number of accesses.

Optionally, the similarity and loss calculation task is uniformly distributed to all processes in the embodiment, at least one process is only used for calculating the local similarity matrix and the local loss, and gradient information is distributed back to each process in the back propagation process, so that efficient training is realized.

The training method of the data characterization model of this embodiment may include the steps of:

and S1, acquiring the image characteristics of each image in the multiple images of the commodity.

S2, calculating a similarity matrix between the image features (here, the image features of each image need to be obtained and distributed to different processes, that is, all the image features need to be pulled from other GPUs).

And S3, obtaining a local similarity matrix generated under at least one process according to the similarity matrix, and performing local loss calculation on the local similarity matrix generated under at least one process by adopting a noise contrast loss function to generate local loss under at least one process.

S4, based on the local loss obtained by at least one process, perform gradient calculation to obtain a local gradient in the back propagation process (calculating local gradients distributed on each GPU).

And S5, distributing the local gradient under any process to other processes (to each GPU), and aggregating the local gradient in the current process and the local gradients distributed by other processes by at least one process to generate an aggregate gradient under at least one process. Where local gradients across multiple GPUs can be averaged to scale the processing data.

S6, training the data characterization model by using the aggregation gradient.

The above steps S1 to S6 are repeated.

In this embodiment, the noise contrast loss function can be expressed by the following formula:

wherein, i, j are used to represent two transformations (such as scaling, matting, color transformation, etc., wherein the various transformations do not change the picture meaning) of the same commodity picture, and k is used to represent pictures of other commodities.

FIG. 6 is a schematic diagram of a noise contrast loss function according to an embodiment of the present invention. As shown in fig. 6, taking a product image a and a product image B as an example, the product image a is transformed 1, the product image a is transformed 2, an image feature is extracted from the product image a after transformation 1 by a convolutional neural network, an image feature is extracted from the product image a after transformation 2 by a convolutional neural network, an image feature is extracted from the product image B after transformation 1 by a convolutional neural network, and an image feature is extracted from the product image B after transformation 2 by a convolutional neural network. The noise contrast loss function of this embodiment requires that the distance between image features between different transformations of the same image be as close as possible, while the features between image features of different images be as far apart as possible. In fig. 6, the solid line is used to indicate the distance between the two groups of image features, and the dotted line is used to indicate the distance between the two groups of image features.

Fig. 7A is a schematic diagram of a distributed training process according to an embodiment of the present invention. As shown in fig. 7A, there are multiple processes, including process 1 and process 2 … …, where multiple images in an image data set are distributed to multiple processes for distributed computation, and a neural network model may be used in at least one process to identify image features of each image in the image data set, and a similarity matrix between the image features is computed, so as to obtain a local similarity matrix generated in at least one process, where in the process of performing local similarity matrix computation in any one process, image features generated in other processes are pulled, and the pulled image features and image features in the current process are used to perform similarity computation, so as to generate a local similarity matrix in the current process.

The embodiment performs local loss calculation on the local similarity matrix generated under at least one process to generate local loss under at least one process, and can distribute local gradients under any one process to other processes; and at least one process uses the local gradient in the current process and the local gradients distributed by other processes to carry out aggregation to generate an aggregation gradient under at least one process, thereby completing the optimization purpose of the data characterization model on at least one process.

Fig. 7B is a schematic view of an interactive interface of a method for processing product data according to an embodiment of the present invention. As shown in fig. 7B, a user may enter an image dataset of a product object on an operator interface, wherein the image dataset includes at least one of: the method comprises the steps that a plurality of images B and a plurality of image frames P of a video are obtained, the images in an image data set are not marked, the images in the image data set are distributed to a plurality of processes for distribution calculation by clicking a 'model training' key, at least one process obtains a local gradient in a back propagation process, the obtained local gradients are aggregated to obtain an aggregation gradient, a data representation model is trained by using the aggregation gradient, and prompt information of the training data representation model is displayed on an operation interface.

Fig. 7C is a schematic view of a scenario of a processing method of product data according to an embodiment of the present invention. As shown in fig. 7C, the computing device obtains an image dataset of the product object, wherein the image dataset includes: the product data of the product object can be displayed on an interactive interface of the computing device without labeling the images in the image data set. And then inducing a model generation instruction in an interactive interface of the computing equipment, responding to the model generation instruction, selecting a plurality of images from the image data set, distributing the plurality of images to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, aggregating at least one acquired local gradient to obtain an aggregate gradient, and further displaying prompt information of a training data representation model on the interactive interface of the computing equipment, wherein the training data representation model is trained by using the aggregate gradient.

Compared with the scheme of the related art of building a supervised model by means of the mapping relationship between commodity data and the description information provided by the seller, the method for commodity recognition and pre-training based on self-supervision is completely free from interference of a large amount of noise contained in the description information of commodities, and the pre-training model is built in a self-supervision learning mode only according to the information of a large number of commodity pictures, and is fine-tuned by migrating the pre-training model to a small amount of manually marked data to achieve higher performance. In terms of technical effects, the model after migration in the embodiment has higher identification precision for the commodity category, the attribute, the product category and the like, and is more stable in performance.

The embodiment provides a complete commodity identification pre-training scheme based on an auto-supervision learning algorithm, distributed training and transfer learning, model pre-training can be completed by utilizing a large amount of unlabeled data, and the model pre-training scheme is transferred to a commodity identification task to improve identification precision and cross-domain capability. The method specifically comprises the following steps:

(1) a whole set of commodity identification pre-training scheme comprises self-supervision learning, distributed training, transfer learning and the like.

(2) The image self-supervision method based on noise contrast estimation loss gradually draws the distance of image features between different transformations of the same image in the training process, and gradually advances the distance of the image features between different images, so that a data characterization model with strong discrimination capability can be effectively learned.

(3) The distributed training scheme based on the parallel processes uniformly distributes similarity matrix calculation and loss calculation to each process, and then gathers gradients after local loss calculation is completed in each process, so that optimization of a data representation model on at least one process is completed, and the training efficiency of the data representation model is improved.

(4) And in the transfer learning scheme, after the self-supervision training on the unlabeled data is completed, the parameters of the trained data representation model can be copied into the model of the commodity identification task as initialization, and the commodity identification model is continuously trained. The accuracy and cross-domain capability of commodity identification are improved through the transfer learning.

It should be noted that, for simplicity of description, the above-mentioned method embodiments are described as a series of acts or combination of acts, but those skilled in the art will recognize that the present invention is not limited by the order of acts, as some steps may occur in other orders or concurrently in accordance with the invention. Further, those skilled in the art should also appreciate that the embodiments described in the specification are preferred embodiments and that the acts and modules referred to are not necessarily required by the invention.

Through the above description of the embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware, but the former is a better implementation mode in many cases. Based on such understanding, the technical solutions of the present invention may be embodied in the form of a software product, which is stored in a storage medium (e.g., ROM/RAM, magnetic disk, optical disk) and includes instructions for enabling a terminal device (e.g., a mobile phone, a computer, a server, or a network device) to execute the method according to the embodiments of the present invention.

Example 3

According to an embodiment of the present invention, there is also provided a product data processing apparatus for implementing the product data processing method shown in fig. 2.

Fig. 8 is a schematic diagram of a product data processing apparatus according to an embodiment of the present invention. As shown in fig. 8, the processing device 80 of the product data may include: an acquisition unit 81, a distribution unit 82, an aggregation unit 83, and a training unit 84.

An acquisition unit 81 for acquiring an image dataset of the product object, wherein the image dataset comprises: a plurality of images of the plurality of pictures and the plurality of image frames of the video, and none of the images in the image dataset are labeled.

The distribution unit 82 is configured to distribute a plurality of images in the image data set to a plurality of processes for distribution calculation, where at least one process acquires a local gradient in a back propagation process.

And the aggregation unit 83 is configured to aggregate the obtained at least one local gradient to obtain an aggregation gradient.

A training unit 84 for training the data characterization model using the aggregate gradient.

It should be noted here that the acquiring unit 81, the distributing unit 82, the aggregating unit 83, and the training unit 84 correspond to steps 202 to 208 in embodiment 1, and the four units are the same as the corresponding steps in the implementation example and application scenario, but are not limited to the disclosure in the first embodiment. It should be noted that the above units as a part of the apparatus may operate in the computer terminal 10 provided in the first embodiment.

According to an embodiment of the present invention, there is provided another product data processing apparatus for implementing the product data processing method shown in fig. 3.

Fig. 9 is a schematic diagram of another product data processing apparatus according to an embodiment of the present invention. As shown in fig. 9, the product data processing device 90 may include: an entry unit 91, a first sensing unit 92, and a first display unit 93.

An entry unit 91, configured to enter an image data set of the product object on an entry page on the operation interface, where the image data set includes: a plurality of images of the plurality of pictures and the plurality of image frames of the video, and none of the images in the image dataset are labeled.

The first sensing unit 92 is configured to sense a model generation instruction in the operation interface, distribute a plurality of images in the image data set to a plurality of processes for distribution calculation, where at least one process acquires a local gradient in a back propagation process, and aggregate at least one acquired local gradient to obtain an aggregate gradient.

And a first display unit 93, configured to display prompt information of a training data characterization model on the operation interface, where the training data characterization model is trained using the aggregation gradient.

It should be noted here that the entry unit 91, the first sensing unit 92 and the first display unit 93 correspond to steps 302 to S306 in embodiment 1, and the three units are the same as the corresponding steps in the implementation example and application scenario, but are not limited to the disclosure in the first embodiment. It should be noted that the above units as a part of the apparatus may operate in the computer terminal 10 provided in the first embodiment.

According to an embodiment of the present invention, there is provided another product data processing apparatus for implementing the product data processing method shown in fig. 4.

Fig. 10 is a schematic diagram of another product data processing apparatus according to an embodiment of the present invention. As shown in fig. 10, the processing apparatus 100 of the product data may include: a second display unit 101, a second sensing unit 102, a response unit 103, and a third display unit 104.

A second display unit 101 for displaying an image data set of the product object on the interactive interface, wherein the image data set comprises: a plurality of images of the plurality of pictures and the plurality of image frames of the video, and none of the images in the image dataset are labeled.

And the second sensing unit 102 is used for sensing a model generation instruction in the interactive interface.

The response unit 103 is configured to respond to the model generation instruction, select a plurality of images from the image data set, distribute the plurality of images to a plurality of processes for distribution calculation, where at least one process acquires a local gradient in a back propagation process, and aggregate at least one acquired local gradient to obtain an aggregate gradient.

And a third display unit 104, configured to display prompt information of the training data characterization model on the interactive interface, where the training data characterization model is trained using the aggregation gradient.

It should be noted that the second display unit 101, the second sensing unit 102, the responding unit 103, and the third display unit 104 correspond to steps 402 to S408 in embodiment 1, and the four units are the same as the corresponding steps in the implementation example and application scenario, but are not limited to the disclosure in the first embodiment. It should be noted that the above units as a part of the apparatus may operate in the computer terminal 10 provided in the first embodiment.

According to an embodiment of the present invention, there is provided another product data processing apparatus for implementing the product data processing method shown in fig. 5.

Fig. 11 is a schematic diagram of another product data processing apparatus according to an embodiment of the present invention. As shown in fig. 11, the processing device 110 for the product data may include: an upload unit 111, a transmission unit 112, and a reception unit 113.

An upload unit 111, configured to enable a front-end client to upload an image data set of a product object, where the image data set includes: a plurality of images of the plurality of pictures and the plurality of image frames of the video, and none of the images in the image dataset are labeled.

A transmission unit 112, configured to enable the front-end client to transmit the image data set to the backend server.

The receiving unit 113 is configured to enable the front-end client to receive the trained data characterization model returned by the background server, where multiple images in the image data set are distributed by the background server to multiple processes for distributed computation, at least one process obtains a local gradient in a back propagation process, aggregates the obtained local gradients to obtain an aggregate gradient, and trains the data characterization model using the aggregate gradient.

In the product data processing apparatus of this embodiment, a supervised machine learning method is used, distributed training may be performed on an unmarked product object only according to an image data set of the unmarked product object, so as to effectively construct a data characterization model, which is not interfered by other large amounts of noise at all, thereby avoiding that a supervised learning method cannot fully utilize larger-scale unmarked data to improve model accuracy, and also avoiding that a model constructed on the data may have a certain data deviation due to a small amount of marked data being difficult to cover complex and changeable scenes, commodity types, and shooting conditions, thereby solving a technical problem that a structured modeling result of product data is inaccurate, and achieving a technical effect of improving accuracy of the structured modeling result of product data.

Example 4

Embodiments of the present invention may provide a product data processing system, which may include a computer terminal, which may be any one of computer terminal devices in a computer terminal group. Optionally, in this embodiment, the computer terminal may also be replaced with a terminal device such as a mobile terminal.

Optionally, in this embodiment, the computer terminal may be located in at least one network device of a plurality of network devices of a computer network.

In this embodiment, the computer terminal may execute the program code of the following steps in the processing method of the product data of the application program: acquiring an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process; polymerizing the obtained at least one local gradient to obtain a polymerization gradient; the data characterization model is trained using the aggregate gradient.

Alternatively, fig. 12 is a block diagram of a computer terminal according to an embodiment of the present invention. As shown in fig. 12, the computer terminal a may include: one or more processors 1202 (only one of which is shown), a memory 1204, and a transmitting device 1206.

The memory may be used to store software programs and modules, such as program instructions/modules corresponding to the method and apparatus for processing product data in the embodiments of the present invention, and the processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, implements the method for processing product data described above. The memory may include high speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory may further include memory located remotely from the processor, which may be connected to the computer terminal a via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The processor can call the information and application program stored in the memory through the transmission device to execute the following steps: acquiring an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process; polymerizing the obtained at least one local gradient to obtain a polymerization gradient; the data characterization model is trained using the aggregate gradient.

Optionally, the processor may further execute the program code of the following steps: acquiring the image characteristics of each image in the image data set, and distributing the image characteristics of each image to different processes; calculating a similarity matrix between image features, and acquiring a local similarity matrix generated under at least one process; performing local loss calculation on the local similarity matrix generated under the at least one process to generate local loss under the at least one process; and performing gradient calculation based on the local loss acquired by at least one process to acquire a local gradient in the back propagation process.

Optionally, the processor may further execute the program code of the following steps: the neural network model is used to identify image features of each image in the image dataset and extract image features of each image.

Optionally, the processor may further execute the program code of the following steps: in the process of carrying out local similarity matrix calculation on any process, image features generated in other processes are pulled; and performing similarity calculation by using the pulled image features and the image features in the current process to generate a local similarity matrix in the current process.

Optionally, the processor may further execute the program code of the following steps: and performing local loss calculation on the local similarity matrix generated under the at least one process by adopting a noise contrast loss function to generate local loss under the at least one process, wherein the noise contrast loss function is used for reducing the image characteristic distance between different transformations in the same image so as to expand the image characteristic distance between different images.

Optionally, the processor may further execute the program code of the following steps: distributing the local gradient under any process to other processes; and at least one process uses the local gradient in the current process and the local gradients distributed by other processes for aggregation to generate an aggregation gradient under at least one process.

Optionally, the processor may further execute the program code of the following steps: the frequency of gradient aggregation is determined based on the amount of traffic, the amount of computation, and the number of accesses between the at least one process.

Optionally, the processor may further execute the program code of the following steps: after the aggregation gradient is used for training the data characterization model, acquiring an identification task of product information; and identifying the identification task by using the trained data characterization model to obtain product information.

Optionally, the processor may further execute the program code of the following steps: after the trained data representation model is used for identifying product information, generating various types of product materials based on the product information; and issuing a plurality of product materials.

Optionally, the processor may further execute the program code of the following steps: after various types of product materials are generated, uploading the product materials to be issued, and extracting a plurality of product contents to be verified in the product materials to be issued; judging whether at least one product content to be verified meets an entry standard; if yes, successfully inputting the product materials into the release template; otherwise, preprocessing the product content failed in verification, and recording the product material into the release template under the condition that the preprocessed product content meets the recording standard.

As an alternative, the processor may call the information and application stored in the memory through the transmission device to execute the following steps: an entry page on the operation interface enters an image data set of the product object, wherein the image data set comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; inducing a model generation instruction in an operation interface, distributing a plurality of images in an image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregation gradient; and displaying prompt information of the training data representation model on the operation interface, wherein the training data representation model is trained by using the aggregation gradient.

Optionally, the processor may further execute the program code of the following steps: after the prompt information of the training data representation model is displayed on the operation interface, displaying the identification task of the product information on the operation interface; and displaying product information on the operation interface, wherein the product information is obtained by using the trained data representation model to identify the identification task.

Optionally, the processor may further execute the program code of the following steps: after the product information is displayed on the operation interface, popping up guide information on the operation interface, wherein the guide information comprises defect information existing in the product information; displaying an authoring material generated based on the guidance information on an operation interface, wherein the authoring material is basic information constituting a product material; generating various types of product materials based on the authoring materials; and issuing a plurality of product materials.

As an alternative, the processor may call the information and application stored in the memory through the transmission device to execute the following steps: displaying an image dataset of a product object on an interactive interface, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; sensing a model generation instruction in the interactive interface; responding to a model generation instruction, selecting a plurality of images from an image data set, distributing the plurality of images to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregation gradient; and displaying prompt information of the training data representation model on the interactive interface, wherein the training data representation model is trained by using the aggregation gradient.

As an alternative, the processor may call the information and application stored in the memory through the transmission device to execute the following steps: the front-end client uploads an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; the front-end client transmits the image data set to a background server; the front-end client receives a trained data representation model returned by the background server, wherein a plurality of images in the image data set are distributed to a plurality of processes by the background server for distribution calculation, at least one process acquires a local gradient in a back propagation process, the acquired local gradients are aggregated to obtain an aggregation gradient, and the aggregation gradient is used for training the data representation model.

The embodiment of the invention provides a method for processing product data. By acquiring an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process; polymerizing the obtained at least one local gradient to obtain a polymerization gradient; the data characterization model is trained using the aggregate gradient. That is to say, the supervised machine learning method is used, distributed training can be performed on the unmarked product object according to the image data set of the unmarked product object, so that a data representation model is effectively constructed, the interference of a large amount of other noise is avoided completely, the problem that the supervised learning method cannot fully utilize larger-scale unmarked data to improve the model precision is avoided, the problem that a small amount of marked data is difficult to cover complicated and changeable scenes, commodity types and shooting conditions, and the model constructed on the data possibly has certain data deviation is also avoided, the technical problem that the structured modeling result of the product data is inaccurate is solved, and the technical effect of improving the accuracy of the structured modeling result of the product data is achieved.

It can be understood by those skilled in the art that the structure shown in fig. 12 is only an illustration, and the computer terminal a may also be a terminal device such as a smart phone (e.g., an Android phone, an iOS phone, etc.), a tablet computer, a palmtop computer, a Mobile Internet Device (MID), a PAD, and the like. Fig. 12 is not intended to limit the structure of the computer terminal a. For example, the computer terminal a may also include more or fewer components (e.g., network interfaces, display devices, etc.) than shown in fig. 12, or have a different configuration than shown in fig. 12.

Those skilled in the art will appreciate that all or part of the steps in the methods of the above embodiments may be implemented by a program instructing hardware associated with the terminal device, where the program may be stored in a computer-readable storage medium, and the storage medium may include: flash disks, Read-Only memories (ROMs), Random Access Memories (RAMs), magnetic or optical disks, and the like.

Example 5

Embodiments of the present invention also provide a computer-readable storage medium. Optionally, in this embodiment, the computer-readable storage medium may be used to store the program code executed by the processing method of the product data provided in the first embodiment.

Optionally, in this embodiment, the computer-readable storage medium may be located in any one of a group of computer terminals in a computer network, or in any one of a group of mobile terminals.

Optionally, in this embodiment, the computer readable storage medium is configured to store program code for performing the following steps: acquiring an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process; polymerizing the obtained at least one local gradient to obtain a polymerization gradient; the data characterization model is trained using the aggregate gradient.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: acquiring the image characteristics of each image in the image data set, and distributing the image characteristics of each image to different processes; calculating a similarity matrix between image features, and acquiring a local similarity matrix generated under at least one process; performing local loss calculation on the local similarity matrix generated under the at least one process to generate local loss under the at least one process; and performing gradient calculation based on the local loss acquired by at least one process to acquire a local gradient in the back propagation process.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: the neural network model is used to identify image features of each image in the image dataset and extract image features of each image.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: in the process of carrying out local similarity matrix calculation on any process, image features generated in other processes are pulled; and performing similarity calculation by using the pulled image features and the image features in the current process to generate a local similarity matrix in the current process.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: and performing local loss calculation on the local similarity matrix generated under the at least one process by adopting a noise contrast loss function to generate local loss under the at least one process, wherein the noise contrast loss function is used for reducing the image characteristic distance between different transformations in the same image so as to expand the image characteristic distance between different images.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: distributing the local gradient under any process to other processes; and at least one process uses the local gradient in the current process and the local gradients distributed by other processes for aggregation to generate an aggregation gradient under at least one process.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: the frequency of gradient aggregation is determined based on the amount of traffic, the amount of computation, and the number of accesses between the at least one process.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: after the aggregation gradient is used for training the data characterization model, acquiring an identification task of product information; and identifying the identification task by using the trained data characterization model to obtain product information.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: after identifying the product information using the trained data characterization model, the method further comprises: generating various types of product materials based on the product information; and issuing a plurality of product materials.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: after various types of product materials are generated, uploading the product materials to be issued, and extracting a plurality of product contents to be verified in the product materials to be issued; judging whether at least one product content to be verified meets an entry standard; if yes, successfully inputting the product materials into the release template; otherwise, preprocessing the product content failed in verification, and recording the product material into the release template under the condition that the preprocessed product content meets the recording standard.

As an alternative example, the computer readable storage medium is arranged to store program code for performing the steps of: an entry page on the operation interface enters an image data set of the product object, wherein the image data set comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; inducing a model generation instruction in an operation interface, distributing a plurality of images in an image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregation gradient; and displaying prompt information of the training data representation model on the operation interface, wherein the training data representation model is trained by using the aggregation gradient.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: after the prompt information of the training data representation model is displayed on the operation interface, displaying the identification task of the product information on the operation interface; and displaying product information on the operation interface, wherein the product information is obtained by using the trained data representation model to identify the identification task.

Optionally, the computer readable storage medium is further arranged to store program code for performing the steps of: after the product information is displayed on the operation interface, popping up guide information on the operation interface, wherein the guide information comprises defect information existing in the product information; displaying an authoring material generated based on the guidance information on an operation interface, wherein the authoring material is basic information constituting a product material; generating various types of product materials based on the authoring materials; and issuing a plurality of product materials.

As an alternative example, the computer readable storage medium is arranged to store program code for performing the steps of: displaying an image dataset of a product object on an interactive interface, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; sensing a model generation instruction in the interactive interface; responding to a model generation instruction, selecting a plurality of images from an image data set, distributing the plurality of images to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregation gradient; and displaying prompt information of the training data representation model on the interactive interface, wherein the training data representation model is trained by using the aggregation gradient.

As an alternative example, the computer readable storage medium is arranged to store program code for performing the steps of: the front-end client uploads an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos, and the images in the image data set are not labeled; the front-end client transmits the image data set to a background server; the front-end client receives a trained data representation model returned by the background server, wherein a plurality of images in the image data set are distributed to a plurality of processes by the background server for distribution calculation, at least one process acquires a local gradient in a back propagation process, the acquired local gradients are aggregated to obtain an aggregation gradient, and the aggregation gradient is used for training the data representation model.

The above-mentioned serial numbers of the embodiments of the present invention are merely for description and do not represent the merits of the embodiments.

In the above embodiments of the present invention, the descriptions of the respective embodiments have respective emphasis, and for parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.

In the embodiments provided in the present application, it should be understood that the disclosed technology can be implemented in other ways. The above-described embodiments of the apparatus are merely illustrative, and for example, the division of the units is only one type of division of logical functions, and there may be other divisions when actually implemented, for example, a plurality of units or components may be combined or may be integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, units or modules, and may be in an electrical or other form.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit.

The integrated unit, if implemented in the form of a software functional unit and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a removable hard disk, a magnetic or optical disk, and other various media capable of storing program codes.

The foregoing is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, various modifications and decorations can be made without departing from the principle of the present invention, and these modifications and decorations should also be regarded as the protection scope of the present invention.

Claims

1. A method for processing product data, comprising:

acquiring an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos;

distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process;

polymerizing the obtained at least one local gradient to obtain a polymerization gradient;

training a data characterization model using the aggregate gradient.

2. The method of claim 1, wherein distributing the plurality of images in the image dataset to a plurality of processes for distribution computation, at least one process obtaining a local gradient in a back propagation process, comprises:

acquiring the image characteristics of each image in the image data set, and distributing the image characteristics of each image to different processes;

calculating a similarity matrix between image features, and acquiring a local similarity matrix generated under the at least one process;

performing local loss calculation on the local similarity matrix generated under the at least one process to generate local loss under the at least one process;

and performing gradient calculation based on the local loss acquired by the at least one process to acquire the local gradient in the back propagation process.

3. The method of claim 2, wherein obtaining image features for each image in the image dataset comprises:

image features of each image in the image dataset are identified using a neural network model and extracted.

4. The method of claim 2, wherein computing a similarity matrix between image features to obtain a local similarity matrix generated under the at least one process comprises:

in the process of carrying out local similarity matrix calculation on any process, image features generated in other processes are pulled;

and performing similarity calculation by using the pulled image features and the image features in the current process to generate a local similarity matrix in the current process.

5. The method according to claim 2, wherein a local loss calculation is performed on the local similarity matrix generated in the at least one process by using a noise contrast loss function, so as to generate the local loss in the at least one process, wherein the noise contrast loss function is used for reducing image feature distances between different transformations in the same image, so that the image feature distances between different images are expanded.

6. The method according to any one of claims 1 to 5, wherein aggregating the obtained at least one local gradient to obtain an aggregated gradient comprises:

distributing the local gradient under any process to other processes;

the at least one process uses the local gradient in the current process and the local gradients distributed by other processes for aggregation, and the aggregation gradient under the at least one process is generated.

7. The method of claim 6, wherein the frequency of gradient aggregation is determined based on traffic, computational load, and number of accesses between the at least one process.

8. The method of any one of claims 1 to 5, wherein after training a data characterization model using the aggregate gradient, the method further comprises:

acquiring an identification task of product information;

and identifying the identification task by using the trained data characterization model to obtain the product information.

9. The method of claim 8, wherein after identifying the product information using the trained data characterization model, the method further comprises:

generating various types of product materials based on the product information;

and issuing a plurality of the product materials.

10. The method of claim 9, wherein after generating a plurality of types of product material, the method further comprises:

uploading the product material to be issued, and extracting a plurality of product contents to be verified in the product material to be issued;

judging whether the content of the at least one product to be verified meets an entry standard;

if yes, successfully inputting the product materials into the release template;

otherwise, preprocessing the product content failed in verification, and recording the product material into the release template under the condition that the preprocessed product content meets the recording standard.

11. A method for processing product data, comprising:

an entry page on an operator interface enters an image dataset for a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos;

inducing a model generation instruction in the operation interface, distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregation gradient;

and displaying prompt information of a training data representation model on the operation interface, wherein the training data representation model is trained by using the aggregation gradient.

12. The method of claim 11, wherein after displaying prompt information for training a data characterization model on the operator interface, the method further comprises:

displaying an identification task of the product information on the operation interface;

and displaying the product information on the operation interface, wherein the product information is obtained by identifying the identification task by using the trained characterization model.

13. The method of claim 12, wherein after displaying the product information on the operator interface, the method further comprises:

popping up guide information on the operation interface, wherein the guide information comprises defect information existing in the product information;

displaying an authoring material generated based on the guidance information on the operation interface, wherein the authoring material is basic information constituting a product material;

generating various types of product materials based on the authoring materials;

and issuing a plurality of the product materials.

14. A method for processing product data, comprising:

displaying an image dataset of a product object on an interactive interface, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos;

sensing a model generation instruction in the interactive interface;

responding to the model generation instruction, selecting a plurality of images from the image data set, distributing the images to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregate gradient;

and displaying prompt information of a training data representation model on the interactive interface, wherein the training data representation model is trained by using the aggregation gradient.

15. A method for processing product data, comprising:

the front-end client uploads an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos;

the front-end client transmits the image data set to a background server;

the front-end client receives the trained data representation model returned by the background server, wherein a plurality of images in the image data set are distributed to a plurality of processes by the background server for distributed calculation, at least one process acquires a local gradient in a back propagation process, the acquired local gradients are aggregated to obtain an aggregation gradient, and the aggregation gradient is used for training the data representation model.

16. An apparatus for processing product data, comprising:

an acquisition unit for acquiring an image dataset of a product object, wherein the image dataset comprises:

a plurality of image frames of a plurality of pictures and videos;

the distribution unit is used for distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, and at least one process acquires a local gradient in a back propagation process;

the polymerization unit is used for polymerizing the acquired at least one local gradient to obtain a polymerization gradient;

a training unit to train a data characterization model using the aggregate gradient.

17. An apparatus for processing product data, comprising:

the input unit is used for inputting an image data set of the product object on an input page on the operation interface, wherein the image data set comprises: a plurality of image frames of a plurality of pictures and videos;

the first sensing unit is used for sensing a model generation instruction in the operation interface, distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating the acquired at least one local gradient to obtain an aggregation gradient;

and the first display unit is used for displaying prompt information of a training data representation model on the operation interface, wherein the training data representation model is trained by using the aggregation gradient.

18. An apparatus for processing product data, comprising:

a second display unit for displaying an image dataset of a product object on an interactive interface, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos;

the second sensing unit is used for sensing a model generation instruction in the interactive interface;

the response unit is used for responding to the model generation instruction, selecting a plurality of images from the image data set, distributing the images to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process, and aggregating at least one acquired local gradient to obtain an aggregate gradient;

and the third display unit is used for displaying prompt information of a training data representation model on the interactive interface, wherein the training data representation model is trained by using the aggregation gradient.

19. An apparatus for processing product data, comprising:

an upload unit configured to enable a front-end client to upload an image dataset of a product object, wherein the image dataset includes: a plurality of image frames of a plurality of pictures and videos;

a transmission unit, configured to enable the front-end client to transmit the image data set to a background server;

and the receiving unit is used for enabling the front-end client to receive the trained data representation model returned by the background server, wherein a plurality of images in the image data set are distributed to a plurality of processes by the background server for distribution calculation, at least one process acquires a local gradient in a back propagation process, the acquired local gradients are aggregated to obtain an aggregation gradient, and the aggregation gradient is used for training the data representation model.

20. A computer-readable storage medium, comprising a stored program, wherein the program, when executed by a processor, controls an apparatus in which the computer-readable storage medium is located to perform the method of any of claims 1-15.

21. A processor, characterized in that the processor is configured to run a program, wherein the program when running performs the method of any of claims 1 to 15.

22. A system for processing product data, comprising:

a processor;

a memory coupled to the processor for providing instructions to the processor for processing the following processing steps: acquiring an image dataset of a product object, wherein the image dataset comprises: a plurality of image frames of a plurality of pictures and videos; distributing a plurality of images in the image data set to a plurality of processes for distribution calculation, wherein at least one process acquires a local gradient in a back propagation process; polymerizing the obtained at least one local gradient to obtain a polymerization gradient; training a data characterization model using the aggregate gradient.