Pedestrian attribute recognition training acceleration method based on node merging and path prediction

By establishing a multi-dimensional performance cost model and graph search space, combining node merger and path prediction of computing power heterogeneous perception, the problem of the pedestrian attribute recognition model training time and high computing resources in the monitoring scenario is solved, and the pedestrian attribute recognition model is realized Quick training and efficient execution.

CN113822173BActive Publication Date: 2024-11-26HANGZHOU DIANZI UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202111019973.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-01
Publication Date
2024-11-26
Estimated Expiration
2041-09-01

AI Technical Summary

Technical Problem

In monitoring scenarios, pedestrian attribute recognition model training takes a long time and has high computing resource cost. It is difficult for the existing technology to effectively solve the problems of memory limitation and low search efficiency in parallel training of large-scale complex neural networks.

Method used

Using a method based on node merger and path prediction, by establishing a multi-dimensional performance cost model and graph search space, introducing node merger with heterogeneous computing power perception, shortening the scheduling length of key paths, and improving the performance of distributed parallel strategies.

Benefits of technology

It realizes rapid training of pedestrian attribute recognition model, accelerates the execution time of deep neural networks, reduces the cost of strategy search time, and expands the limitations of training scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113822173B_ABST
    Figure CN113822173B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for accelerating pedestrian attribute recognition training based on node merging and path prediction. The present invention uses a heuristic-based graph search algorithm to calculate the optimal parallelization strategy that consumes less computing resources in a short time, thereby realizing automatic search and tuning of parallel strategies for large-scale complex neural network models. The present invention first constructs a multidimensional performance cost model to guide the automatic search and tuning of parallel strategies, solving the problem of single evaluation of parallel efficiency in existing methods. Secondly, a graph search space is constructed to transform the operator placement and scheduling problems of the DNN model into a DAG scheduling problem, and then heterogeneous computing power-aware node merging is used to reduce the waiting dependency time between operators. Finally, when selecting devices, the critical path is optimized based on path prediction, thereby further improving parallel performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a pedestrian attribute recognition training acceleration method based on node merging and path prediction, which solves the problem of accelerating the training of pedestrian attribute recognition models in monitoring scenarios and provides an optimal model parallel training solution for large-scale complex neural networks. Background Art

[0002] In recent years, thanks to the rapid development of AI algorithms, hardware computing power and data sets, deep neural networks have made significant progress in natural language processing, computer vision and search recommendation. At the same time, the application of deep learning in pedestrian attribute recognition in surveillance scenarios has also put forward higher requirements for the training accuracy and speed of neural network algorithm models in mobile applications. The common characteristics of these applications are the use of deep, complex, hierarchical models and the need to rely on large amounts of data.

[0003] For dense models, the largest DNN models may involve hundreds of millions to billions of parameters, and for sparse models, they may even involve trillions of parameters. Therefore, as models become more complex and datasets become larger, the scale of machine learning models and training datasets is limited by GPU device memory. For example, a common implementation of a 1,000-layer residual network requires 48GB of memory, and even after further optimization to reduce memory costs, the DNN model still requires at least 7GB of memory. This makes it extremely difficult to run the entire model on a single device with limited memory. Therefore, in order to solve the high computing cost problem and single-device memory limitation caused by deep neural network model training, it has become an inevitable trend to implement model parallelism across multiple GPU devices in distributed scenarios.

[0004] The industry usually divides neural network computational graphs horizontally by layers, vertically across layers, or randomly and schedules them to different devices for execution. However, these methods rely heavily on expert experience, have unreasonable segmentation methods, low cluster utilization, and usually have large communication overhead.

[0005] The automatic search strategy based on machine learning is one of the methods to guide model parallel training. It uses learning methods to capture the characteristics of the computational graph and the cluster device topology graph, proposes a reasonable parallel strategy search space and establishes a strategy evaluation feedback model. It continuously iterates and searches for the optimal strategy in the search space through a specific sampling method. The optimization goal is to place the model to minimize the execution time of each round of iterative training.

[0006] At first, Hao et al. proposed the AutoSync adaptive framework based on traditional machine learning methods. Based on the data parallel mode, they used machine learning methods to predict the execution time of training strategies to guide the search of synchronous parallel strategies, but this method was limited to synchronous data parallelism. Later, Google Brain proposed a reinforcement learning-based automatic search parallel framework REINFORCE. This framework uses the real distributed execution time of the scheduling strategy as a penalty factor and feeds it back to the LSTM model to calculate the policy gradient to learn a better parallel strategy. However, the computing resources and search time required for the method proposed by Google are very expensive, and the practicality is not great. Placeto was first proposed by Ravichandran at NIPS. Unlike REINFORCE, it supports a more advanced Graph Embedding method to capture the computational graph structure and the dependencies between nodes, and can be generalized to other similar unknown computational graphs without retraining. At the same time, Placeto adopts an incremental iterative reinforcement learning method, that is, it returns to the real runtime immediately after completing the single operator scheduling strategy, thereby improving the algorithm execution efficiency. The above method optimizes the search efficiency, but due to the iterative search process based on learning and feedback, it still takes a long time to generate parallel strategies. However, in actual production scenarios, developers often need to deploy and train quickly, and the computing resource cost is also very high.

[0007] In order to make up for the shortcomings of the time-consuming and high computing resource cost of the search method based on machine learning, the automatic search strategy method based on heuristic methods such as dynamic programming and graph search has gradually become another research direction. Its biggest feature is that it can search for the optimal or suboptimal distributed parallel strategy in a constant time. At first, Wang Li et al. proposed a model splitting strategy automatic generation method for model parallel training, which automatically generated model splitting strategies based on static network models to achieve task allocation of network layers on different AI accelerators; Pellegrini et al. proposed the Scotch static mapping method, which used the method of finding the shortest path in graph theory to guide the segmentation of static graphs. Although the above methods can greatly shorten the strategy search time, the parallel strategy performance obtained is far lower than that of the expert design method, and the practicality is not high. Later, Jia et al. proposed OptCNN, a deep learning automatic parallel framework based on graph search, constructed a cost model to quantitatively evaluate the runtime performance of the training network, and searched for parallel strategies based on dynamic programming. Subsequently, Tofu used recursive segmentation and graph coarsening based on OptCNN dynamic programming to reduce the strategy search time and expand the limitations of the network model. The above methods are mainly based on coarse-grained hierarchical division and the execution time is estimated, so the performance improvement is limited. To solve this problem, Yi et al. proposed the FastT algorithm based on DAG scheduling, which uses fine-grained operator priority and critical path to place and schedule operators, but it does not consider the dynamic memory usage during device training and is not suitable for dynamic RNN. However, the above heuristic methods still have the following problems: 1) The cost model is constructed based on expert experience, and the dimension is single, which has limitations in characterizing parallel performance. 2) The DNN model task graph search problem is an NP-hard problem, and its search space increases with the increasing depth of the network. The search for large-scale complex deep neural networks has problems such as complex search schemes and long search time. Summary of the invention

[0008] In view of the deficiencies in the prior art, the present invention provides a pedestrian attribute recognition training acceleration method based on node merging and path prediction, which is used to accelerate the training of pedestrian attribute recognition models in monitoring scenarios, thereby solving the technical problem of inputting monitoring pedestrian pictures and taking a long time to train the model.

[0009] The present invention first establishes a multi-dimensional performance cost model to guide the automatic search and tuning of deep learning distributed parallel strategies; secondly, it extracts the features of the original calculation graph and the device topology graph and constructs a graph search space; then, based on heterogeneous computing power perception, it merges the nodes that meet the merging conditions to reduce the graph search space. Finally, the device is scheduled through the graph search algorithm, the priority is set for the operation to determine the critical path and placement order, and the optimal placement of the key node is predicted through the path from the parent node of the key node to the exit node, thereby shortening the scheduling length of the critical path, realizing the adaptive segmentation of the deep neural network, improving the performance of the distributed parallel strategy, and enhancing its scalability.

[0010] The method of the present invention specifically comprises:

[0011] Step 1: Analyze the impact of the pedestrian attribute recognition neural network model structure, operator attributes, etc. on performance, extract the key factors that affect the model parallel execution efficiency, and establish cost sub-models for computational cost, communication cost, and memory occupancy based on iteration and linear regression methods. On this basis, a multidimensional performance cost model is established to guide the automatic search and tuning of deep learning distributed parallel strategies. cost , measure the comprehensive performance of distributed parallel strategies. The steps to build a multi-dimensional performance cost model that can fully reflect the performance of distributed training are as follows:

[0012] 1) Analyze the key factors that affect the parallel execution efficiency of the model, and calculate the cost model based on iteration and linear regression methods. Communication cost model and memory usage model Three cost sub-models.

[0013] 2) In the memory sub-model, M d Under the condition of <C, the performance of the model parallel strategy is characterized from the two dimensions of computational cost and communication cost, and a multi-dimensional performance cost model is established. in, The value represents the device memory usage threshold set based on manual experience in the actual training environment. i,d , C i,j Respectively represent the computational cost and communication cost sub-models, in the memory sub-model M d satisfy The communication and computation costs are minimized under the constraints of .

[0014] Step 2: Construct a DAG graph search space based on the structural characteristics of the pedestrian attribute recognition neural network model and the cluster topology. The graph search space displays in detail the DNN model operator structure, cluster topology, communication of each operator on different devices, computing overhead, etc., and establishes a high-dimensional search space.

[0015] First, based on the DAG computation graph and the device topology graph, a graph search space G = (O, E) is constructed. O represents the set of all nodes, O = {o1, o2, ...o n}, each node represents an operation (such as convolution, dot product, etc.). pred(o i ) means that in the graph search space, node o i The set of direct predecessors, succ(o i ) represents node o i The set of direct successors of . If Then the node is called the entry node, denoted by o entry , Then the node is called the exit node, denoted by o exit If there are multiple entry nodes or exit nodes in the DAG computation graph, add a virtual entry and exit node with zero communication time and computing time. E is a set of directed edges between nodes, indicating the dependency relationship between nodes. o i After the execution is completed j To start execution.

[0016] Secondly, based on the multi-dimensional cost sub-model, the weights of nodes and edges are constructed. The weight of each node represents the i At the execution cost time of device d, use o i,d Indicates that the weight of each edge represents the node o i and node o j The communication cost time is c i,j express.

[0017] Step 3: Introduce node merging based on the awareness of computing power heterogeneity, merge nodes based on the operator's sensitivity to computing power heterogeneity, and reduce the search space.

[0018] First, a node merging method based on breadth search is introduced to reduce the DAG graph search space: for the computation node o of the DAG computation graph i , for its successor node o j Perform a breadth-traversal search. If operator o i The output tensor of is only used by the next operator o j Use, or operator o i The core idea is that if the communication cost between two computing nodes is too high, they can be merged and ablated into the same node in the graph search algorithm, which can effectively reduce the number of operator nodes and thus speed up the search for the best scheduling strategy.

[0019] Secondly, a computing power heterogeneity-aware node merging method is proposed to characterize the differences in the execution time of operators on different devices through computing power heterogeneity sensitivity. i and j They are the only parent and child nodes to each other. When the communication time between them is greater than o j The difference between the maximum execution time and the minimum execution time on any GPU is merged. The core idea is that when the communication time is greater than the maximum time that can be saved by placing the device, there is no need to separately j Node scheduling can be merged into o i It is scheduled as a node to speed up the search for the best scheduling strategy.

[0020] Step 4: To address the scheduling length problem of the computational graph, a graph search algorithm based on path prediction is proposed to search for strategies in the simplified search space. The optimal placement of key nodes is predicted through the path from the parent node of the key node to the exit node, thereby shortening the scheduling length of the critical path and reducing the execution time of the DAG computational graph.

[0021] First, set the operation priority for each operator, and obtain the upward weight priority rank by recursively calculating the DAG calculation graph from bottom to top. u , which represents node o i The longest path to the exit node is obtained by recursively calculating the downward weight priority rank by traversing the DAG calculation graph from top to bottom d , which represents node o i The longest path to the entry node.

[0022] Secondly, the critical path is determined based on the operator priority, and an optimistic table matrix is ​​introduced to represent the maximum value of the shortest path from the child node to the exit node.

[0023] Finally, a graph search algorithm based on path prediction is used for equipment scheduling. The optimal placement of key nodes is predicted through the path from the parent node of the key node to the exit node, thereby shortening the scheduling length of the key path and accelerating the training of the pedestrian attribute recognition model. The core idea is to minimize the execution time of the operators on the key path, so that the execution time of the entire calculation graph is also the shortest.

[0024] The beneficial effects of the present invention are: the establishment of a multi-dimensional performance cost model accurately depicts the computing power and communication load of the equipment in the actual training environment, broadens the limitations of the training scenario, and enhances its scalability; the introduction of heterogeneous computing power-aware node merging can reduce the graph search space and reduce the time cost of strategy search for large-scale complex deep networks; the use of a graph search algorithm based on path prediction predicts the optimal placement of key nodes through the path from the parent node of the key node to the exit node, effectively shortening the scheduling length of the key path, thereby reducing the execution time of the DAG calculation graph. At the same time, the multi-dimensional performance cost model and heterogeneous computing power-aware node merging in the present invention can be extended to other adaptive search scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the search method based on node merging and path prediction;

[0026] Figure 2 It is a schematic diagram of constructing a graph search space;

[0027] Figure 3 It is a schematic diagram of node merging;

[0028] Figure 4 This is a schematic diagram of graph search based on path prediction. DETAILED DESCRIPTION

[0029] The present invention will be further described below with reference to the accompanying drawings and specific implementation steps:

[0030] The present invention comprises the following steps:

[0031] Step 1: Construct a multi-dimensional performance cost model E cost , measure the comprehensive performance of distributed parallel strategies to guide the automatic search and tuning of distributed parallel strategies for deep learning models.

[0032] First, we analyze the factors that affect the performance of the pedestrian attribute recognition neural network. The performance indicators of distributed parallel training mainly consider the following three aspects: 1) Computational cost. Due to the differences in operator attributes (such as convolution, dot multiplication) and the heterogeneity of device computing power, the execution time of operators on each device will be different. Therefore, the actual execution time of operators on each device is an important reference for device selection; 2) Communication cost. When the communication time consumed by parameter synchronization between operators is too large, the communication overhead brought by parallelism will reduce the parallel performance. Therefore, the communication overhead should be reduced as much as possible; 3) Memory occupancy rate. Since large-scale memory read and write access will affect the device response time, the model parameters of each device are balanced to reduce the device memory usage, thereby speeding up the device response time.

[0033] Calculation cost: The end time of the actual execution process of the operator on the device minus the start time of execution is used to record and Respectively represent the operator o i The start and end times of execution on device d, For iterative mean calculation, the cost model is calculated as follows:

[0034]

[0035] Communication cost: Taking the ideal communication time (tensor size / communication bandwidth) as the independent variable and the communication cost as the dependent variable, the communication cost model is obtained by linear regression fitting. Indicates o i and j The size of the space tensor, b i,j represents the communication bandwidth between the operators placed on the devices, θ B represents the linear regression regularization term, and the communication cost is expressed as follows:

[0036]

[0037] Memory usage: The physical storage size of the model parameters in the memory of the current device d and the total memory of the device m d The ratio of is expressed as follows:

[0038]

[0039] Then, in the memory usage M d Meet the threshold Considering the minimum communication and computational costs under the constraint of , a multi-dimensional performance cost model is established to guide the automatic search and tuning of deep learning distributed parallel strategies, which is expressed as follows:

[0040]

[0041] Among them, E cost The smaller the value, the better the strategy performance. The memory usage is used as a node attribute. The memory peak is set to control the GPU memory usage. The above computational cost and communication cost are the node weight (execution time) and edge weight (communication time) of the graph search space construction.

[0042] Step 2: Based on the original computation graph and the device topology graph, construct a graph search space G = (O, E).

[0043] Abstract computational graph based on pedestrian attribute recognition neural network model structure characteristics and cluster topology and cluster device resource groups Construct a DAG graph search space, as shown in the attached Figure 2 As shown. Where O represents the set of all nodes, O = {o1, o2, ...on}, each node represents an operation (such as convolution, dot product, etc.). pred(o i ) means that in the graph search space, node o i The set of direct predecessors. If Then the node is called the exit node, denoted by o exit ; If there are multiple entry nodes or exit nodes in the DAG calculation graph, then add a virtual entry and exit node with zero communication time and calculation time. E is a set of directed edges between nodes, indicating the dependency relationship between nodes, that is, o i After the execution is completed j To start execution.

[0044] Secondly, based on the multi-dimensional cost sub-model, the weights of nodes and edges are constructed. The weight of each node represents the i At the execution cost time of device d, use o i,d Indicates that the weight of each edge represents the node o i and node o j The communication cost time is c i,j express.

[0045] Finally, define the earliest completion time EFT(o i , d j ), which represents node o i In device d j The earliest completion time on the node is determined by the earliest start time EST(o i , d j ) plus the node execution time ET(o i , d j )get.

[0046]

[0047] Among them, D Available (d j ) indicates device d j The earliest available time at which execution of a new operator can begin. Represents node o i The time when all predecessor nodes have completed execution. ), EST(o entry , d j )=0.

[0048] Step 3: Introduce node merging based on the awareness of computing power heterogeneity, merge nodes based on the operator's sensitivity to computing power heterogeneity, and reduce the search space.

[0049] First, a node merging method based on breadth search is introduced to reduce the DAG graph search space. The core idea is: if the communication cost between two computing nodes is too high, they can be merged and ablated into the same node scheduling in the graph search algorithm, which can effectively reduce the number of operator nodes and thus speed up the search for the best scheduling strategy. i , for its successor node o j Perform a breadth traversal search and merge nodes if the following rules are met:

[0050] (1)o j The only predecessor node of the node is o i , and no node merging operation is performed;

[0051] (2) i The width is 1(o j is its only successor node), and no node merging operation has been performed.

[0052] (3) i and j The communication cost between nodes is greater than o j The average computation cost on the device.

[0053] If the operator o i The output tensor of is only used by the next operator o j Use, or operator o i If the communication cost is not enough to offset the benefits of parallel computing of operators, then the two operations should be scheduled to execute on the same device. Since the execution time of operators that are not sensitive to computing power heterogeneity is small between different devices, the communication cost is small and the average computing cost is large, so the merger is ignored. For example, assuming that the communication cost is 45ms, the computing costs of the child nodes are 75, 85, and 95 respectively, the node communication cost c i,j =45m, average node computing cost ET avg (o j )=85, then if the original condition c i,j >ET avg (o j ), they cannot be merged.

[0054] Therefore, the present invention proposes a computing power heterogeneity perception node merging method, which uses computing power heterogeneity sensitivity to characterize the differences in operator execution time on different devices, as shown in the attached Figure 3 As shown: When o i and j They are the only parent and child nodes to each other. When the communication time between them is greater than o jThe difference between the maximum execution time and the minimum execution time on any GPU is merged. The core idea is that when the communication time is greater than the maximum time that can be saved by placing the device, there is no need to separately j Node scheduling can be merged into o i As a node scheduling. The above node merging method based on computing power heterogeneity perception is converted into the formula as follows:

[0055]

[0056] Among them, succ(o i ) means o i The set of successor nodes; pred(o j ) means o j The predecessor node set of isMerg(o j )=false means o j Has not been merged with other nodes in the DAG computation graph; c i,j Represents node o i and j The communication cost between d (o j ) is node o j The computation cost on device d, ET avg (o j ) is represented as node o j Average computation cost across all devices.

[0057] Step 4: To address the scheduling length problem of the computational graph, a graph search algorithm based on path prediction is proposed to search for strategies in the simplified search space. The optimal placement of key nodes is predicted through the path from the parent node of the key node to the exit node, thereby shortening the scheduling length of the critical path and reducing the execution time of the DAG computational graph.

[0058] First, set the operation priority for each operator, and obtain the upward weight priority rank by recursively calculating the DAG calculation graph from bottom to top. u , which represents node o i The longest path to the exit node is obtained by recursively calculating the downward weight priority rank by traversing the DAG calculation graph from top to bottom d , which represents node o i The longest path to the entry node, and DAG task scheduling based on operation priority. The specific calculation formula is as follows:

[0059]

[0060] Among them, ET avg (o i ) is node oi Average computation time on all devices; pre(o i ) represents node o i The set of all direct predecessor nodes, succ(o i ) represents node o i The set of all direct successor nodes of i,j Represents the slave node o i To node o j The average communication load cost; for the exit node o exit , rank u (o exit )=ET avg (o exit ); For the entry node o entry , rank d (o entry )=0.

[0061] Secondly, since operator placement and scheduling based on conventional priority order DAG scheduling cannot effectively shorten the execution time of the computational graph, a critical path is introduced and path prediction is used to optimize scheduling. The optimal placement of key nodes is predicted through the path from the parent node of the key node to the exit node, thereby shortening the scheduling length of the critical path and achieving the purpose of accelerating the execution of the computational graph.

[0062] Critical Node (CN) refers to the node that determines the length of DAG scheduling. u and rank d The sum is equal to the entry node, and the key node is determined by the condition:

[0063] rank u (o i )+rank d (o i )=rank u (o entry )+rank d (o entry ).

[0064] CNP(o i ) represents node o i Whether it is the parent node of the key node, the specific definition formula is as follows:

[0065]

[0066] At the same time, in order to calculate the path length from the child node to the exit, the concept of optimistic table matrix is ​​introduced, whose rows represent the number of nodes m and columns represent the number of devices n. Each element value OCT(o i , dj ) means that node o i Dispatch to device d j When i The maximum value of the shortest path from the child node to the exit node is calculated as follows:

[0067]

[0068] Therefore, the equipment placement diagram is as shown in the attached Figure 4 As shown in the figure, the process is to first calculate the earliest completion time (EFT) of the current node on each device. Then, calculate the EFT of the current node * Value, where EFT * Represents the earliest predicted completion time. If the node is the father node of a key node, EFT * The value of is EFT; otherwise, its value is the sum of EFT and OCT, as shown in formula (10). The purpose of this is to schedule to d when the node is the father node of the key node. j The device is not optimally placed for the current node, but can make its child node (i.e., key node) finish in a shorter time. Even if the current node is not the parent node of the key node, it can be placed in the EFT * The value can predict the scheduling length from its child node to the exit, which can advance the completion time of the child node to a certain extent. In summary, the core idea is to complete the operators on the critical path as early as possible and shorten the scheduling length of the calculation graph, so that the execution time of the entire calculation graph is also the shortest.

[0069]

[0070]

[0071]

[0072] The flowchart of the pedestrian attribute recognition training acceleration method based on node merging and path prediction is as follows: Figure 1 ,The specific implementation process is shown in Algorithm 1: First, lines 1-6 will traverse the entire data flow graph and search for each operator to see if it meets the conditions for merging heterogeneous computing power-aware nodes. If so, it will be merged (it should be noted here that the merged data flow graph is only used for subsequent graph search and scheduling. In the actual execution environment, the merged operators are still executed as independent operators). Then, lines 7-8 of this algorithm will calculate the CNP,rank of all nodes in the merged DAG graph according to the multi-dimensional performance cost model. u ,rank dand OCT value, wherein the present invention maintains a ready-list and places the entry node at the top of the ready-list. In the While loop 9-24, each iteration of the method schedules the node with the maximum rank u When scheduling the current node, this method needs to calculate the EFT of the node on all devices. * Value. When the device d k If the memory usage of the memory is greater than the set memory peak value, EFT(o i , d j )=+∞, thus ignoring the scheduling selection of the device. In line 19, the minimum EFT is obtained * (o i , d k ), then place the node on device d k Finally, the algorithm returns the scheduling strategy S, scheduling order queue Q and corresponding pedestrian attribute recognition effect diagram of all nodes in the DAG graph.

Claims

1. A pedestrian attribute recognition method based on node merging and path prediction. This method uses a pedestrian attribute recognition neural network model for recognition. Its input is a monitored pedestrian image. It is characterized in that The pedestrian attribute recognition neural network model is trained using the following steps: (1) Analyze the structure of the pedestrian attribute recognition neural network model and the impact of operator attributes on performance, and extract the key factors that affect the parallel execution efficiency of the model, including computational cost, communication cost, and memory usage; Based on iteration and linear regression methods, cost sub-models are established for computational cost, communication cost, and memory occupancy respectively. On this basis, a multi-dimensional performance cost model is established to guide the automatic search and tuning of deep learning distributed parallel strategies. The computational cost refers to the actual execution time of the operator extracted when executing the computational graph, the communication cost refers to the operator communication time calculated by iterative calculation of the communication load model with the ratio of the communication volume to the communication bandwidth as the independent variable, and the memory usage rate refers to the ratio of the physical storage size of the current device model parameters to the total memory of the device; (2) Constructing a DAG graph search space based on the structural characteristics of the pedestrian attribute recognition neural network model and the cluster topology. The DAG graph search space is used to display the DNN model operator structure, cluster topology, and the communication and computing overhead of each operator on different devices; (3) Introducing node merging based on the awareness of computing power heterogeneity, merging nodes based on the operator’s sensitivity to computing power heterogeneity to reduce the search space; (4) The graph search algorithm based on path prediction searches for the optimal parallel strategy in the simplified search space and predicts the optimal placement of key nodes through the path from the parent node of the key node to the exit node, thereby shortening the scheduling length of the critical path and reducing the execution time of the DAG calculation graph.

2. The pedestrian attribute recognition method based on node merging and path prediction according to claim 1, Features: The multi-dimensional performance cost model in step (1) is established through the following two steps: (1) Analyze the key factors that affect the parallel execution efficiency of the model, and establish the computational cost model E based on iteration and linear regression methods. i,d , Communication cost model C i,j and memory usage model M d Three cost sub-models; (2) In the memory sub-model M d Meet the threshold Under the constraints of , the computational cost and communication cost are used to characterize the performance of the model parallel strategy, and a multi-dimensional performance cost model is established. in, The value represents the device memory usage threshold set based on manual experience in the actual training environment. d Meet the threshold The communication and computation costs are minimized under the constraints of .

3. The pedestrian attribute recognition method based on node merging and path prediction according to claim 1, Features: The DAG graph search space described in step (2) refers to: a graph search space established by extracting features of operator attributes based on a computational graph in the form of a data flow graph, a device topology graph, and a multi-dimensional cost model.

4. The pedestrian attribute recognition method based on node merging and path prediction according to claim 1, Features: The node merging based on computing power heterogeneity awareness in step (3) refers to: merging nodes based on the operator's sensitivity to computing power heterogeneity; wherein the computing power heterogeneity sensitivity represents the difference in the execution time of the operator on different devices.

5. The pedestrian attribute recognition method based on node merging and path prediction according to claim 1, Features: Step (4) is specifically: First, set the operation priority for each operator, and recursively calculate the upward weight priority by traversing the DAG calculation graph from bottom to top, and recursively calculate the downward weight priority by traversing the DAG calculation graph from top to bottom; Secondly, the critical path is determined based on the operator priority, and an optimistic table matrix is ​​introduced to represent the maximum value of the shortest path from the child node to the exit node; Finally, a graph search algorithm based on path prediction is used for equipment scheduling. The optimal placement of key nodes is predicted through the path from the parent node of the key node to the exit node, thereby shortening the scheduling length of the critical path and accelerating the training of the pedestrian attribute recognition model.

Citation Information

Patent Citations

  • A pedestrian appearance attribute identification method based on Inception V3 multi-data-set joint training

    CN109886154A

  • Pedestrian attribute recognition method based on generative adversarial learning

    CN112016490A

  • Adaptive distributed parallel training method for neural network based on reinforcement learning

    CN113128702A

  • Pedestrian attribute identification and positioning method and convolutional neural network system

    US20200272902A1