JP7408042B2 - Distributed heterogeneous data processing method, device, and equipment based on deep learning - Google Patents

Description

TECHNICAL FIELD Embodiments of the present disclosure relate to the field of computer technology, and specifically to a method, apparatus, and equipment for processing distributed heterogeneous data based on deep learning.

Model training means training an initial model so that the model obtained through training has predictive or classification ability. Currently, the commonly adopted method for repopulating model classification is to train the model using a fixed distributed training scheme.
However, when adopting the above method, the following technical problems often exist.
First, the node configuration of each node included in the distributed formula is often different, and the model structures of models trained multiple times are often different. Therefore, when training a model using a fixed distributed training method, Node usage efficiency of nodes in a distributed system often decreases.
Second, if the model includes multiple sub-models and the model structure is complex, the whole model will be trained directly, the model training cycle will be long, and the model training efficiency will be reduced.

The content section of this disclosure is used to introduce ideas in a concise form that are explained in detail in the specific embodiments section below.
Some embodiments of the present disclosure provide a deep learning-based distributed heterogeneous data processing method, apparatus, and method to solve one or more of the technical problems mentioned in the background section above. Suggest equipment.

In a first aspect, some embodiments of the present disclosure provide the steps of: determining a node information group sequence in which one submodel of each submodel included in the initial model is configured; A method for processing distributed heterogeneous data based on deep learning is provided. A set of training samples corresponding to the node information group is determined based on the node configuration information and model structure information corresponding to each node information group in the sequence of node information groups, and the model structure information is added to the node information group. for each node information group in the node information group sequence, characterizing the model structure of the corresponding sub-model, a training sample set corresponding to the node information group, based on model structure information corresponding to the node information group; sample-reconstructing the training samples in the model to generate a target training sample set; Perform training and generate sub-training results. In response to all of the determined sub-training results converging, training is performed based on the current model parameter information corresponding to the sub-model in each of the above-mentioned sub-models. Generate a completed initial model.

In a second aspect, some embodiments of the present disclosure provide an information determining unit configured to determine a sequence of node information groups and a node configuration corresponding to each node information group in the sequence of node information groups. sample determining means for determining a model structure information group corresponding training sample set characterizing a model structure of a sub-model corresponding to the node information group based on information and model structure information; and each node information in the node information group sequence. Sample reconstruction for a group, based on model structure information corresponding to the node information group, training samples in a training sample set corresponding to the node information group to generate a target training sample set. means; a model training unit for model training the sub-model to generate a sub-training result for each sub-model of each of the sub-models via a target training sample set corresponding to the sub-model; is configured to generate a trained initial model based on current model parameter information corresponding to the submodel in each of the submodels in response to convergence of all of the plurality of subtraining results obtained. Model generation means.

In a third aspect, some embodiments of the present disclosure provide for one or more processors and one or more programs executed by the one or more processors, the one or more processors according to the first aspect above. and a storage device storing one or more programs for implementing the method described in any of the above.

In a fourth aspect, some embodiments of the present disclosure provide a computer readable medium storing a computer program that, when executed by a processor, implements a method according to any of the first aspects above. .

Each of the above-described embodiments of the present disclosure includes the deep learning-based distributed heterogeneous data processing method of some embodiments of the present disclosure to improve the node utilization efficiency of distributed nodes and improve model training efficiency. It has the beneficial effect of being able to

Specifically, the causes of low node usage efficiency and model training efficiency are, firstly, that the node configuration of each node included in the distributed formula is often different, and the model structure of the model that is trained multiple times. Because they are often different, training a model using a fixed distributed training scheme often results in less efficient node usage of the nodes in the distributed scheme. Second, if the model includes multiple sub-models and the model structure is complex, the whole model will be trained directly, the model training cycle will be long, and the model training efficiency will be reduced. Based on this, the deep learning-based distributed heterogeneous data processing method of some embodiments of the present disclosure firstly includes a node where one submodel of each submodel included in the initial model is configured. including determining an information group sequence. In reality, the node configurations of the nodes are often different, so the model structures of the submodels are often different. If you adopt a fixed distributed training method and configure submodels on preset fixed nodes, the model structure and node configuration may not match (for example, a model with a simple model structure may be configured on a highly configured node). ), the node utilization of the node may decrease. Therefore, by determining the sequence of node information groups, sub-models are configured into corresponding node groups (for example, a model with a simple model structure is configured into a node with a low configuration), and the node usage efficiency of the nodes is increased. Next, a set of training samples corresponding to the node information group is determined based on the node configuration information and model structure information corresponding to each node information group in the sequence of node information groups, and the model structure information is the node information Characterize the model structure of the submodels corresponding to the groups. In practice, the configurations of node groups corresponding to different node information groups are often different, and the structures of different sub-models are also different, so the model training speeds of different node groups are also different. Therefore, in order to approximate the training time of submodels in each node group and reduce the training time, the number of training samples required for a node information group is determined based on the node configuration information of the node information group and the structure information of the submodels. need to be determined. Then, for each node information group in the node information group sequence, the training samples in the training sample set corresponding to the node information group are sample-reconstructed based on the model structure information corresponding to the node information group. Then, a target training sample set is generated. In reality, model inputs may differ depending on the submodel. Therefore, it is necessary to reconstruct the training samples based on the model input of the submodel corresponding to the model structure information to obtain a target training sample set. Next, for each submodel in each of the submodels, model training is performed on the submodel using a target training sample set corresponding to the submodel to generate a subtraining result. Thereby, by training each sub-model individually, the model training period is shortened and the model training efficiency is improved. Finally, in response to all of the determined sub-training results converging, an initial model for which training has been completed is generated based on the current model parameter information corresponding to the sub-model in each of the sub-models. As a result, a trained initial model is generated using the model parameter information obtained by completing the training of each sub-model in the distributed model, thereby improving the node usage efficiency of the nodes in the distributed model and improving the model training efficiency.

2 is a flowchart of several embodiments of a distributed heterogeneous data processing method based on deep learning according to the present disclosure. 1 is a configuration diagram of some embodiments of a distributed heterogeneous data processing apparatus based on deep learning according to the present disclosure. FIG. FIG. 3 is a structural schematic diagram of an electronic device suitable for implementing some embodiments of the present disclosure.

Hereinafter, the present disclosure will be described in detail in connection with embodiments with reference to the drawings.
FIG. 1 shows a flow 100 of some embodiments of a distributed heterogeneous data processing method based on deep learning according to the present disclosure. This deep learning-based distributed heterogeneous data processing method includes the following steps:

In step 101, a node information group sequence is determined.
In some embodiments, an entity (eg, a computing device) performing a deep learning-based distributed heterogeneous data processing method may determine the sequence of node information groups in various ways. The node group corresponding to each node information group in the node information group sequence constitutes one submodel among the submodels included in the initial model. Node information can characterize a node. The nodes are distributed computational nodes. The node information may include a node number, node configuration information, submodel number, and model structure information of a model deployed at the node. Node configuration information may characterize the configuration of a node. For example, node configuration information can be characterized by a grade. When the node configuration information is characterized by a grade, the higher the grade, the higher the computing power for characterizing the node corresponding to the node configuration information. The model structure information characterizes the model structure of the submodel corresponding to the node information. Model structure information can characterize the model complexity of the submodel. For example, model structure information can be quantized by the model rank of the corresponding submodel. The configuration of nodes corresponding to each piece of node information in a node information group is the same.
In some alternatives of some embodiments, the above-described execution entity may include determining a sequence of node information groups.

The first step is to obtain an initial node information sequence.
Here, the initial node information may include a node number and node configuration information. Node configuration information may characterize the configuration of a node. The node configuration information can be characterized by the maximum amount of data processing per unit time due to the hardware configuration corresponding to the node.
In the second step, calculation capacity requirement information corresponding to the submodel in each of the submodels is determined.
Among them, the computing power demand information characterizes the computing power demand of the submodel. The computational power demand can be obtained by quantizing the maximum amount of data processing for each layer included in the submodel. For example, the computational power requirement may be the amount of floating point computation for the submodel. Note that the number of floating point operations refers to the amount of floating point operations performed by a floating point arithmetic unit per unit time.
In the third step, for each sub-model in each sub-model, an initial node information subset corresponding to the sub-model is extracted from the initial node information sequence based on computing power requirement information corresponding to the sub-model. It is selected as a node information group and made into a node information group.
Here, the amount of floating point operations characterized by the computing power demand information corresponding to the submodel matches the maximum amount of data processing of the node corresponding to the node information in the node information group.

In step 102,
A set of training samples corresponding to the node information group is determined based on node configuration information and model structure information corresponding to each node information group in the node information group sequence.
In some embodiments, the execution entity generates a set of training samples corresponding to the node information group based on node configuration information and model configuration information corresponding to each node information group in the sequence of node information groups. can be determined. here,
The training samples in the training sample set may be samples required during submodel training corresponding to the node information group. Here, the model structure information characterizes the model structure of the submodel corresponding to the node information group.
Optionally, node configuration information includes task scheduler configuration information, memory configuration information,
and data processor configuration information. The memory configuration information may include memory size information and memory read/write speed information. Here, the task scheduler configuration information may include the clock frequency of the node's central processor. The data processor configuration information may include the core frequency of the node's graphics card. The memory size information may include the memory size of the node's memory. The memory read/write speed information may include the read/write frequency of the node's memory.
Optionally, the execution entity determines a set of training samples corresponding to the node information group based on node configuration information and model structure information corresponding to each node information group in the sequence of node information groups, and performs the following: The steps may include:

First, task scheduling capability information of a node group corresponding to the node information group is determined based on task scheduler configuration information included in node configuration information corresponding to the node information group.
Here, the task scheduling capability information may characterize the size of the scheduling capability of the central processors of the nodes in the corresponding node group. The magnitude of the central processor's scheduling capability can be obtained by grade quantization. The task scheduling ability information may include a task scheduling ability grade. for example,
The value of the task scheduling ability grade may be an integer between [1,100]. As an example, if the clock frequency of the central processor of the node included in the task scheduler configuration information is between 3.8 GHz and 4 GHz, the task scheduling capability grade included in the corresponding task scheduling capability information is 10.
Second, the memory caching capacity of the node group corresponding to the node information group is determined based on memory size information and memory read/write speed information included in memory configuration information included in the node configuration information corresponding to the node information group. Confirm information.
Here, the memory cache capability information can characterize the size of the memory cache capability of the nodes in the corresponding node group. The size of memory cache capacity is
It can be obtained by magnitude quantization. The memory cache capability information may include a memory cache capability grade. For example, the value of memory cache capacity grade is [1,100]
It may be an integer between.
As an example, if the memory of the node included in the memory configuration information is 16 GB or more,
When the read/write frequency of the memory is 4000 MHz or higher, the memory cache capability grade included in the corresponding memory cache capability information is 5.
Third, data processing capability information of a node group corresponding to the node information group is determined based on data processor configuration information included in node configuration information corresponding to the node information group.
Here, the data processing capacity information can characterize the data processing capacity of the graphics card of the node in the corresponding node group. The amount of data processing power of a graphics card can be obtained through grade quantization. The data processing capability information may include a data processing capability grade. For example, the value of the data processing capacity grade is [1,100
] may be an integer between.
As an example, if the core frequency of the graphics card of the node included in the data processor configuration information is between 350 Mhz and 400 Mhz, the data processing capability grade included in the corresponding data processing capability information is 5.
Fourth, determining training sample request information for a node group corresponding to the node information group based on the task scheduling capability information, the memory caching capability information, the data processing capability information, and the model structure information.
Here, the training sample demand information may characterize the demand ratio of training samples used to train the submodel corresponding to the node information group. The demand ratio is the proportion of training samples required by a submodel to all training samples.
As an example, first, the execution entity sums up the task scheduling ability grade, memory cache ability grade, data processing ability grade, and the number of model layers included in the model structure information, and calculates the total grade corresponding to the node information group. can be determined. Then, the execution entity can normalize the total grade corresponding to each node information group in the node information group sequence to obtain a weight corresponding to each node information group. Finally, the execution entity specifies the weight corresponding to each node information group as a demand ratio of training samples for training the submodel corresponding to each information group, and Demand information can be obtained.
In step 5, a training sample set corresponding to the node information group is determined based on the training sample request information.
For example, the number of all training samples is 100. The demand ratio of training samples for the submodel is 10%, and the number of training samples used for training the submodel is 10. Here, the execution entity can randomly extract 10 training samples from all the training samples as a set of training samples corresponding to the node information group corresponding to the sub-model.
In step 103, for each node information group in the node information group sequence, the training samples in the training sample set corresponding to the node information group are reconstructed based on the model structure information corresponding to the node information group. , generate a target training sample set.
In some embodiments, for each node information group in the node information group sequence, the execution entity, based on model structure information corresponding to the node information group,
The training samples in the training sample set corresponding to the node information group may be reconstructed to generate a target training sample set.
Here, the target training samples in the set of target training samples may be samples for training a corresponding sub-model.
As an example, the execution entity generates a target training sample set by reconstructing the training samples in the training sample set corresponding to the node information group based on the model input of the submodel corresponding to the model structure information. be able to. For example, the size of the model input for the submodel is 300*300. The execution entity may crop the image in the training sample so that the size of the image in the generated target training sample is 300*300. As another example, the size of the model input for the submodel is 20*20. The execution entity may crop the image in the training sample so that the size of the image in the generated target training sample is 20*20.
In step 104, for each submodel of each submodel, model training is performed on the submodel through a target training sample set corresponding to the submodel to generate a subtraining result.

In some embodiments, for each sub-model of each of the sub-models, the execution entity:
The sub-model may be model-trained to generate a sub-training result through a set of target training samples corresponding to the sub-model. Here, the sub-training result can characterize the loss value of the corresponding sub-model.
Alternatively, each of the above sub-models may include a face recognition model, a race recognition model, and a posture recognition model. Note that the face recognition model may be a model for recognizing a user's face in an image. The race identification model may be a model for identifying the race of the user corresponding to the image. The posture recognition model may be a model for recognizing a user posture in an image.
As an example, the above-mentioned face recognition model is a multi task convolution model (MTCN).
tional neural network) model. The posture recognition model may be an OpenPose model. The above racial identification model is
It can include two first racial identification networks, two second racial identification networks, and one racial classification network. The first racial identification network consists of two 3*3,
It includes a convolutional layer with a step size of 1 and one 2*2, max pooling layer with a step size of 1. The second race identification network includes three 3x3, step size 1 convolutional layers and one 2x2, step size 1 max pooling layer. The racial classification network may include one fully connected layer of 2048 neurons, one fully connected layer of 1024 neurons, and a racial classifier. The racial classifier may be a softmax classifier.

In one embodiment, a sub-training result is generated by an execution entity performing model training on a sub-model in each sub-model using a set of target training samples corresponding to the sub-model. It may include the following steps.
First step: In response to the fact that the determined sub-model is a face recognition model, model training is performed on the face recognition model using the target training sample set corresponding to the written face recognition model. Generate sub-training results corresponding to .
For example, first, the execution entity performs model training on the face recognition model by sequentially inputting target training samples in the target training sample set corresponding to the face recognition model to the face recognition model. Then, the execution entity determines a loss value corresponding to the face recognition model using a predetermined loss function. Here, the predetermined loss function may be a mean squared error loss function.
Second step: in response to the determined sub-model being a race-aware model, by performing model training on the race-aware model with a target training sample set corresponding to the race-aware model. , generate sub-training results corresponding to the race recognition model.
For example, first, the execution entity performs model training on the race recognition model by sequentially inputting target training samples in the target training sample set corresponding to the race recognition model to the race recognition model. Then, the execution entity determines a loss value corresponding to the race recognition model using a predetermined loss function. Here, the predetermined loss function may be a mean squared error loss function.
Third step: In response to the fact that the determined sub-model is a posture recognition model, the posture recognition model is created by performing model training on the posture recognition model using the target training sample set corresponding to the posture recognition model. Generate sub-training results corresponding to .
For example, first, the execution entity performs model training on the posture recognition model by sequentially inputting target training samples in the target training sample set corresponding to the posture recognition model to the posture recognition model. Then, the execution entity determines a loss value corresponding to the posture recognition model using a predetermined loss function. Here, the predetermined loss function may be a mean squared error loss function.
Step 105: In response to all the determined sub-training results converging, based on the current model parameter information corresponding to the sub-model in each sub-model,
Generate a trained initial model.
In some embodiments, in response to all of the determined sub-training results converging, the execution entity determines the trained initial model based on the current model parameter information corresponding to the sub-model in each sub-model. generate. Here, the current model parameter information may be information on the weight parameters of the corresponding sub-model when the sub-training results converge.
In some embodiments, in response to all of the determined sub-training results converging, the execution entity determines the trained initial model based on the current model parameter information corresponding to the sub-model in each sub-model. Generating may include the following steps.

First step: In response to the convergence of the sub-training results corresponding to the determined face recognition model, the current model parameter information corresponding to the face recognition model is determined by determining the current model parameters of the face recognition model. generate.
For example, the execution entity obtains the weight parameters of the face recognition model by back propagation in response to the convergence of the loss value corresponding to the determined face recognition model, and obtains the current model parameter information.
Second step: responding to the racial recognition model by determining the current model parameters of the racial face recognition model in response to the convergence of the sub-training results corresponding to the determined racial recognition model. Generate current model parameter information.
For example, in response to convergence of loss values corresponding to the determined race recognition model, the execution entity obtains weight parameters of the race recognition model through back propagation to obtain current model parameter information.
Third step: determining the current model parameters of the posture face recognition model in response to the convergence of the sub-training results corresponding to the determined posture recognition model; Generate information.
For example, in response to the convergence of the loss value corresponding to the determined posture recognition model, the execution entity obtains the weight parameters of the posture recognition model by back propagation, and obtains current model parameter information.
Fourth step: A candidate initial model is generated by updating the model parameters of the initial model based on the obtained plurality of current model parameter information.
Here, the candidate initial model may be a model in which parameters in the initial model are replaced. The execution entity generates a candidate initial model by replacing the parameters of the initial model based on the weight parameters included in the model parameter information in the obtained plurality of current model parameter information.
Fifth step: Generate target samples by extracting training samples from a plurality of training sample sets corresponding to each submodel, and obtain the target sample set.
Here, the target samples may be randomly extracted from a plurality of training sample sets and are used to train the candidate initial model.
Sixth step: Model training is performed on the candidate initial model using the target sample set.
For example, an execution entity can train a candidate initial model by sequentially inputting target samples in a set of target samples into the candidate initial model.
Seventh step: Generate a trained initial model in response to convergence of the target training results corresponding to the determined candidate initial model.
For example, in response to the convergence of the target training results corresponding to the determined candidate initial model, the execution entity determines the parameters of the current training as the parameters of the initial model, and generates the initial model for which training has been completed. Can be done.

Optionally, the execution entity may further perform the following processing steps.
First step: Obtain the target image.
Here, the target image includes the target object. The target object may be a full body image of the user.
For example, the target image may be an image captured by a bank's monitoring device. The monitoring device may be a camera.
Second step: Generate user identity information corresponding to the target image by inputting the target image into the trained initial model.
Here, the user identity information may be user identity information of a user corresponding to the target object included in the target image. User identity information includes face position information, race information, and user posture information. The face position information can express the position of the face in the target object included in the target image. The racial information may represent a user's racial category corresponding to the target included in the target image. The user posture information can express the positions of the limbs in the target object included in the target image.
Third step: A marked target image is generated by performing image marking on the target image based on face position information, race information, and user posture information included in the user identity information.
For example, the execution entity can mark the corresponding position in the target image based on the face position in the face position information, the racial category in the racial information, and the limb position in the user posture information.
Third step: Image storage is performed for the marked target image.
For example, an execution entity may store marked images on a magnetic disk.
The deep learning-based distributed heterogeneous data processing method of some embodiments of the present disclosure can improve node utilization efficiency of distributed nodes and improve model training efficiency. Specifically, the causes of low node usage efficiency and model training efficiency are, firstly, that the node configuration of each node included in the distributed formula is often different, and the model structure of the model that is trained multiple times. Because they are often different, training a model using a fixed distributed training scheme often results in less efficient node usage of the nodes in the distributed scheme. Second, if the model includes multiple sub-models and the model structure is complex, the whole model will be trained directly, the model training cycle will be long, and the model training efficiency will be reduced. Based on this, the deep learning-based distributed heterogeneous data processing method of some embodiments of the present disclosure firstly includes a node where one submodel of each submodel included in the initial model is configured. including determining an information group sequence. In reality, the node configurations of the nodes are often different, so the model structures of the submodels are often different. If you adopt a fixed distributed training method and configure submodels on preset fixed nodes, the model structure and node configuration may not match (for example, a model with a simple model structure may be configured on a highly configured node). ), the node utilization of the node may decrease. Therefore, by determining the sequence of node information groups, sub-models are configured into corresponding node groups (for example, a model with a simple model structure is configured into a node with a low configuration), and the node usage efficiency of the nodes is increased. Next, a set of training samples corresponding to the node information group is determined based on the node configuration information and model structure information corresponding to each node information group in the sequence of node information groups, and the model structure information is the node information Characterize the model structure of the submodels corresponding to the groups. In practice, the configurations of node groups corresponding to different node information groups are often different, and the structures of different sub-models are also different, so the model training speeds of different node groups are also different. Therefore, in order to approximate the training time of submodels in each node group and reduce the training time, the number of training samples required for a node information group is determined based on the node configuration information of the node information group and the structure information of the submodels. need to be determined. Then, for each node information group in the node information group sequence, the training samples in the training sample set corresponding to the node information group are sample-reconstructed based on the model structure information corresponding to the node information group. Then, a target training sample set is generated. In reality, model inputs may differ between submodels. Therefore, it is necessary to reconstruct the training samples based on the model input of the submodel corresponding to the model structure information to obtain a target training sample set. Next, for each submodel in each of the submodels, model training is performed on the submodel using a target training sample set corresponding to the submodel to generate a subtraining result. Thereby, by training each sub-model individually, the model training period is shortened and the model training efficiency is improved. Finally, in response to all of the determined sub-training results converging, an initial model for which training has been completed is generated based on the current model parameter information corresponding to the sub-model in each of the sub-models. As a result, a trained initial model is generated using the model parameter information obtained by completing the training of each sub-model in the distributed model, thereby improving the node usage efficiency of the nodes in the distributed model and improving the model training efficiency.

Further referring to FIG. 2, as an implementation of the method shown in each of the above-mentioned figures, the present disclosure corresponds to the embodiment of the method shown in FIG. 1 and is specifically applicable to various electronic devices. , provides several embodiments of a distributed heterogeneous data processing apparatus based on deep learning.
As shown in FIG. 2, a distributed heterogeneous data processing apparatus 200 based on deep learning according to some embodiments includes an information determination unit 201, a sample determination unit 202, a sample reconstruction unit 20
3. includes a model training section 204 and a model generation section 205. Here, the information determination unit 20
1 is configured to determine a sequence of node information groups, and the sample determination unit 20
2 corresponds to a model structure information group that characterizes the model structure of a submodel corresponding to the node information group based on the node configuration information and model structure information corresponding to each node information group in the sequence of node information groups. The sample reconstruction unit 203 is configured to determine a set of training samples, and for each node information group in the node information group sequence, the sample reconstruction unit 203 determines a set of training samples based on the model structure information corresponding to the node information group. The model training unit 204 is configured to sample-reconstruct the training samples in the training sample set corresponding to the information group to generate a target training sample set, and the model training unit 204 is configured to perform the above training sample set for each sub-model of the respective sub-models. The model generating unit 205 is configured to perform model training on the sub-model through a set of target training samples corresponding to the sub-model to generate sub-training results, and the model generating unit 205 generates all the training results of the determined plurality of sub-models. In response to convergence, the system is configured to generate a trained initial model based on current model parameter information corresponding to the submodel in each of the submodels.

It will be understood that the units described in device 200 correspond to the steps of the method described with reference to FIG. Therefore, the operations, features, and beneficial effects described for the method described above apply equally to the device 200 and the units included in the device 200, and will not be further described here.
Referring now to FIG. 3, electronic equipment (
For example, a structural schematic diagram of a computing device) 300 is shown. The electronic device shown in FIG. 3 is merely an example, and does not impose any limitations on the functions and scope of use of the embodiments of the present disclosure.
As shown in FIG. 3, electronic device 300 performs various appropriate operations and processes in accordance with programs stored in read-only memory (ROM) 302 or programs loaded into random access memory (RAM) 303 from storage device 308. A processing device (eg, a central processor, a graphics processor, etc.) 301 can be included.
The RAM 303 also stores various programs and data necessary for the operation of the electronic device 300. The processing device 301, ROM 302, and RAM 303 are connected to each other via a bus 304. An input/output (I/O) interface 305 is also connected to the bus 304 .
Generally, I/O interface 305 includes input devices 30 including touch screens, touch pads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.
6 is an output device 307 such as a liquid crystal display (LCD), a speaker, a vibrator, etc.
, for example, a storage device 308 including a magnetic tape, a hard disk, etc., and a communication device 309. Communication device 309 may enable electronic device 300 to communicate wirelessly or wired with other devices to exchange data. Although FIG. 3 depicts electronic device 300 with various devices, it should be understood that not all of the illustrated devices need be implemented or included. Alternatively, more or fewer devices may be implemented or included. Each block shown in FIG. 3 may represent one device, or may represent multiple devices as necessary.

Claims (2)

A distributed heterogeneous data processing method based on deep learning,
determining a node information group sequence, one submodel of each submodel included in the initial model is placed in a node group corresponding to each node information group in the node information group sequence;
determining a set of training samples corresponding to the node information group based on node configuration information and model structure information corresponding to each node information group in the node information group sequence; characterizing the model structure of submodels corresponding to information groups;
For each node information group in the node information group sequence, objective training is performed by reconstructing the training samples in the training sample set corresponding to the node information group based on the model structure information corresponding to the node information group. generating a sample set;
generating a sub-training result by performing model training on the sub-model in each of the sub-models using a target training sample set corresponding to the sub-model; in response to all of the sub-training results converging, generating an initial model for which training has been completed based on current model parameter information corresponding to the sub-model in each of the sub-models;
including,
The “determining the node information group sequence” includes:
obtaining an initial node information sequence;
The initial node information can include a node number and node configuration information, and the node configuration information
can characterize the configuration of a node, and the node configuration information is the hardware that corresponds to the node.
It can be characterized by the maximum amount of data processed per unit time depending on the software configuration,
determining computing power requirement information corresponding to the submodel in each of the submodels;
The computing power requirement information characterizes the computing power demand of the submodel, and the computing power demand is
can be obtained by quantizing the maximum data processing amount of each layer included in
and
For each submodel in each of the submodels, computing power requirement information corresponding to the submodel is provided.
an initial node corresponding to the sub-model from the initial node information sequence based on the initial node information sequence.
sorting information subsets as node information groups;
The amount of floating-point computation characterized by the computing power requirement information corresponding to the submodel is
Matches the maximum data processing amount of the node corresponding to the node information in the node information group,
including;
The node configuration information includes task scheduler configuration information, internal memory configuration information, and
data processor configuration information, and the internal memory configuration information includes memory size information and
and memory read/write speed information, and
The node corresponding to each node information group in the node information group sequence
Based on the node configuration information and model structure information, the training corresponding to the node information group is
``determining the set of samples'' means
Task scheduler configuration information included in the node configuration information corresponding to the node information group
Based on the information, schedule the task of the node group corresponding to the node information group.
determining capability information; and
Task scheduling capability information is provided by the central process of the nodes in the corresponding node group.
The size of the central processor's scheduling capacity can be characterized and
The magnitude of cursive power can be obtained by grade quantization, and task scheduling
The competency information may include a task scheduling competency grade;
The value of the performance grade is an integer between [1,100],
Included in the memory configuration information included in the node configuration information corresponding to the node information group
Based on the memory size information and memory read/write speed information, the node information group corresponding to the node information group is
determining memory caching capability information of the node group to be configured;
Memory caching capacity information indicates the memory cache of the nodes in the corresponding node group.
The size of the caching ability can be characterized, and the size of the memory caching ability can be characterized.
The size can be obtained by grade quantization, and the memory caching capacity information can be obtained by
may include a caching capability grade, and the value of the memory caching capability grade is
is an integer between [1,100],
Data processor configuration information included in the node configuration information corresponding to the node information group
data processing of a node group corresponding to said node information group based on
establishing competency information;
The data processing capacity information is the graphics of the node in the corresponding node group.
It can characterize the amount of data processing capability of the card and
The amount of data processing capability of the data can be obtained by grade quantization,
Processing capability information may include a data processing capability grade;
The value of the lossing ability grade is an integer between [1,100],
The task scheduling capability information, the memory caching capability information, and the data
The node information group is based on the data processing capacity information and the model structure information.
determining training sample request information for a node group corresponding to the node group;
Based on the training sample request information, the training corresponding to the node information group is
determining the sampling sample set; and
A processing method characterized by comprising: The processing method according to claim 1 ,
Each of the sub-models includes a face recognition model, a race recognition model, and a posture recognition model,
The above-mentioned "generating a sub-training result by performing model training on the sub-model in each of the sub-models using a target training sample set corresponding to the sub-model",
In response to the determined sub-model being the face recognition model, the face recognition model is trained by performing model training on the face recognition model using a target training sample set corresponding to the face recognition model. generating sub-training results corresponding to;
In response to the confirmed sub-model being the race recognition model, performing model training on the race recognition model using a target training sample set corresponding to the race recognition model; generating a sub-training result corresponding to a race recognition model; and in response to said determined sub-model being said pose recognition model, said training sample set corresponding to said pose recognition model; generating sub-training results corresponding to the posture recognition model by performing model training on the posture recognition model;
The above-mentioned "generating an initial model for which training has been completed based on the current model parameter information corresponding to the sub-model in each of the sub-models"
generating current model parameter information corresponding to the face recognition model by determining current model parameters of the face recognition model in response to convergence of sub-training results corresponding to the determined face recognition model;
In response to the convergence of sub-training results corresponding to the finalized racial recognition model,
Generating current model parameter information corresponding to the racial recognition model by determining current model parameters of the racial face recognition model;
In response to the convergence of the sub-training results corresponding to the determined pose recognition model,
determining the current model parameters of the posture face recognition model to generate current model parameter information corresponding to the posture recognition model; and based on the obtained plurality of current model parameter information, generating a candidate initial model by updating model parameters;
The above-mentioned "generating an initial model for which training has been completed based on current model parameter information corresponding to a sub-model in each of the sub-models" further includes:
generating target samples and obtaining a target sample set by extracting training samples from a plurality of training sample sets corresponding to each of the sub-models;
training the candidate initial model using the target sample set, and generating the trained initial model in response to convergence of the target training result corresponding to the determined candidate initial model; including,
The processing method further includes:
Obtaining a target image including the target object;
Generating user identity information including face position information, race information, and user posture information corresponding to the target image by inputting the target image to the trained initial model;
generating a marked target image by performing image marking on the target image based on face position information, race information, and user posture information included in the user identity information; performing image storage for the target image;
The said processing method characterized by comprising.