A cross-task image classification method based on fast network architecture search
By alternately updating weights and structures on multi-task datasets, combining evolutionary algorithms and parallelization technology, the problem of poor generalization performance and low efficiency of neural network architecture search on multi-task and small-scale datasets is solved, and rapid Search and efficient adaptability.
Patent Information
- Application Number
- CN202011441402.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-08
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2040-12-08
AI Technical Summary
Existing neural network architecture search methods perform well on a single task, but poor generalization performance on multiple tasks or small-scale datasets and low search efficiency, especially in small-scale datasets, it is difficult to quickly find excellent networks on small-scale datasets. structure.
Using a fast architectural search method based on meta-learning and neural architecture search, the network structure is optimized to adapt to multitasks by alternately performing weight parameter updates and structure updates on multitask datasets, combining evolutionary algorithms and random forest models, and introducing training parallelism parallelizes the test to improve efficiency.
It realizes a network structure with excellent fast search performance on small-scale data sets, improves the network's adaptability to multi-tasks, reduces the computational volume and search time, can perform meta-architecture learning on large-resolution pictures, and in new tasks Quickly update the network structure on.
Smart Images

Figure CN112508104B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision and relates to a cross-task image classification method based on fast network architecture search Background Art
[0002] An excellent neural network architecture can significantly improve the efficiency and accuracy of computer vision tasks (image classification, object detection, image segmentation, etc.). Therefore, designing a high-performance neural network architecture is an important research topic in current deep learning. However, manually designed network architectures are severely limited by the prior knowledge of experts and the research energy of researchers, and can no longer meet the needs of current computer vision development. Based on this situation, Neural Architecture Search (NAS) was proposed, which refers to searching for an optimal neural network architecture in an automated manner based on a given search space while minimizing human intervention. At present, neural network architecture search has achieved results that are superior to manually designed network architectures in multiple fields (image classification, image segmentation, object detection, and natural language processing, etc.).
[0003] Neural network architecture search algorithms generally consist of three parts: search space, search strategy, and performance evaluation strategy. The search space defines the search range and determines the search speed. Combined with expert experience in network architecture, the search performance can be significantly improved by reducing the search space and simplifying the search process. The evaluation strategy generally trains and verifies an architecture in the usual training method to obtain the architecture, but the computational cost of the evaluation is very expensive and limits the search space. The search strategy defines how to search. Generally, there are search methods based on reinforcement learning, search methods based on evolutionary algorithms, and search methods based on gradients. In the paper Neural architecture search with reinforcement learning (Zoph B, Le Q V. Neural architecture search with reinforcement learning [J / OL]. arXiv preprint arXiv: 1611.01578, 2016.), Zoph B et al. converted the search space into candidate items for the reinforcement learning agent to choose, used RNN as a controller to control the process of combining and selecting to form the network architecture, and then used the policy gradient optimization method to maximize the expected reward of the controller sampling network, realizing the early neural network architecture search concept and proving the feasibility of NAS. However, this method takes 3 to 4 weeks to run on 800 GPUs, which is very inefficient. In the paper Darts: Differentiable architecture search (Liu H, Simonyan K, Yang Y. Darts: Differentiable architecture search [J / OL]. arXiv preprint arXiv: 1806.09055, 2018.), Liu et al. proposed a differentiable architecture search algorithm, which transforms the network architecture search into an optimization problem in continuous space. Therefore, the gradient descent method can be used to solve it, and the search network structure and the optimization network weight parameters are put in the same training process, which can efficiently complete the neural network architecture search task. In the patent (Zhang Xu, Gu Bo, Chen Junzhou, Lin Ziqi, Ding Beichen, Han Yu. A neural network architecture search method based on convolution kernel prediction [P]. Guangdong Province: CN111723914A, 2020-09-29), Zhang Xu et al. disclosed a neural network architecture search method based on convolution kernel prediction. Based on the concept of knowledge distillation, a teacher network is introduced as a guide, and the student network uses the convolution kernel prediction module to predict the optimal convolution kernel to achieve efficient architecture search.However, most current neural network architecture search methods usually perform well on a single task, but the network architecture obtained from a single task search has very poor generalization performance for multiple data sets or multiple learning tasks. The reason is that a single-task large data set can effectively help the algorithm distinguish the pros and cons of candidate networks, but the results of candidate networks in small data training sets of multiple tasks are similar and have a high probability of overfitting. Summary of the invention
[0004] The present invention overcomes the above-mentioned shortcomings of the prior art and proposes a cross-task image classification method based on fast network architecture search.
[0005] The present invention proposes a new fast architecture search method based on meta-learning and neural architecture search (NAS) to complete the cross-task neural network structure search task. The present invention solves the technical problem of how to quickly complete the neural architecture search on a small-scale data set and obtain a network structure with excellent performance.
[0006] A cross-task image classification method based on fast neural network architecture search, the specific steps are as follows:
[0007] Step 1: Obtain a multi-task dataset for neural network architecture search and divide the dataset into D containing training data mtrain and D containing test data mtest , whether D mtrain Or D mtest , randomly sample pictures from the data set to form a task set. Each task can be regarded as a task sample instance. The sample contains N categories, each category has K0 training samples and K1 test samples. For example, the first task extracted is {1, 3, 4, 6, 2}, and the next task extracted is {9, 5, 8, 10, 12}. The set of task training sets consisting of N×K0 samples is called the support set The set of task test sets consisting of N×K1 samples is called the query set. D mtrain A single support set in The training set D is used as the training set for individuals in a single task. mtrain A single support set in The test set is used as an individual test set in a single task. The specific data set is as follows Figure 2 As shown;
[0008] Step 2: Construct a super network containing the candidate network weights and a structure population encoded by the candidate network structure. The super network is stacked by basic unit blocks. The unit block structure is as follows: Figure 3As shown in the figure, the candidate operations of the unit block include 3×3 separable convolution, 5×5 separable convolution, 7×7 separable convolution, and Xception structure. The corresponding candidate operations are set inside the unit block according to the search space. Each unit block only selects one candidate operation, so that the entire convolutional neural network will randomly form a network structure. In each step of the super network optimization process, only one subnet weight is updated. The weight update formula of the super network is as follows:
[0009]
[0010] Γ(A) is the prior distribution, using uniform sampling, W represents the super network weight, W a Represents the subnet weight of structure a. The weight update formula of the subnet is as follows:
[0011]
[0012] The possible network structures are encoded to form individuals. At this time, the collection of all individuals is the initial structure population.
[0013] Step 3: Perform cross-task search on the multi-task dataset. The search results include meta-super networks and meta-structure populations. The search process consists of two stages: weight parameter update stage and structure update stage. These two stages are performed alternately. The overall search process is as follows: Figure 4 As shown;
[0014] Step 3-1: Weight parameter update process Figure 5 As shown, each unit in the super network randomly selects a candidate operation to form an optimized network individual Individuals composed of other candidate operations will not participate in the optimization, and the individual weight parameter W i , the super network weight parameter W, based on step 1, gets the support set Query Set In the individual During the training process, X represents the input data (image). At this time, the network's prediction is The loss on the jth task can be expressed as in is the evaluation metric (different tasks may be different), the individual network will learn m times for each task, where m is a small integer, and the individual Update the weights on the jth task according to formula (1.1):
[0015]
[0016] When all individuals have completed all tasks, calculate all sub-networks The difference vector obtained in the mth step, the weight parameters of the super network are updated according to formula (1.2), and a super network that adapts to multiple tasks is obtained:
[0017]
[0018] The parameter ∈ decays after each iteration, and the decay formula is as follows:
[0019]
[0020] Where n represents the current number of training times, and N represents the total number of training times;
[0021] Step 3-2: While updating the weight parameters, we use the evolutionary algorithm to update the structural parameters A. The structural update process is as follows: Figure 6 As shown in Figure 2, optimizing the network structure based on the evolutionary algorithm includes testing, screening, crossover and mutation, and building a new population. The optimization process is as follows: Figure 7 As shown in Figure 2, after several rounds of updating of the weight parameters, we perform a performance test on each individual in the population, thereby retaining some individuals with better performance for crossover and mutation, and mtest Randomly select a batch of tasks Then let each individual in the population update this batch of tasks M times. In order to screen out individuals that can converge quickly, M is a relatively small number (for example, 10 epochs), and then calculate the performance of each individual on this batch of tasks. In order to speed up the test, use the random forest model to predict the accuracy of each individual in the population;
[0022] Screening: After testing, each individual will have a performance score, and those with high scores will be retained, while those with low scores will be eliminated from the population;
[0023] Crossover and mutation: The retained individuals have a certain probability of undergoing crossover and mutation operations. The structural code of the individual is regarded as the gene of the individual, and the unit in the code is regarded as the chromosome. The crossover operation is to select two individuals from the retained population, cut off the chromosomes of the two individuals at a certain common position, and cross the two strings to form two new chromosomes. The mutation operation is that when the retained individual is replicated, there is a very small probability that a part of the chromosome may mutate into other chromosomes (that is, after replication, it may be exactly the same as the original individual, or some chromosomes may be different);
[0024] Construct a new population: After crossover and mutation, continue to randomly generate a batch of new network structures, add them to the population to keep the population size unchanged, and jump to step 3-1;
[0025] Step 3-2 and step 3-3 are executed repeatedly until the preset number of cycles is reached, and finally a super network and a network architecture population with good adaptability to multiple tasks are obtained;
[0026] Step 4: When facing a new task, directly use the meta-population obtained by searching on the metadata dataset as the initialization population for the new task, and then quickly update the meta-population based on the continuous evolution algorithm to obtain the optimal network structure for the new task. The search process is as follows: Figure 8 As shown in the figure, unlike the operation in step 3, the population during migration no longer randomly generates new individuals, but only the individuals in the structured population undergo crossover mutation to obtain new individuals, and the number of the population is reduced by half after each evolution, and the individuals remaining at the end of the evolution are the optimal network;
[0027] Furthermore, in step 3-1, training parallelization is introduced, such as Fig. 9 As shown, the training task pool is divided into K parts in advance, and the number of processes is also set to K. Then each process obtains the tasks belonging to the current process and the same network structure. After that, the network is updated slightly (the network parameters of each process will not be communicated across processes when updating). After the network on each process has learned its own task, the gradients of the last step of the network update on each process are collected, and then the gradients are averaged and distributed to each process for super network parameter update;
[0028] Furthermore, in step 3-2, test parallelization is introduced, such as Fig.10 As shown in the figure, similar to training parallelization, the test task pool is evenly divided into K processes, and the network parameter updates on each process do not affect each other. Finally, only the fitness values obtained on each process are averaged and used as the individual fitness mean of a certain structure.
[0029] Compared with the existing neural network structure search method, the cross-task image classification method based on fast network architecture search provided by the present invention has the following beneficial effects:
[0030] (1) By training on a multi-task dataset and implementing a parameter update strategy in a task-sample two-layer dataset structure, the neural network architecture search method disclosed in the present invention can learn information from multiple tasks to obtain good adaptability to multiple tasks.
[0031] (2) The super network uses single-path sampling (a unit block selects only one candidate operation at a time) and a continuous evolutionary algorithm to update the neural network structure parameters, which greatly reduces the amount of video memory and computation required for network forward reasoning. The neural network architecture search method of the present invention can directly perform meta-architecture learning on tasks that require large-resolution images as input.
[0032] (3) Compared with other architecture search methods, the present invention introduces a network performance predictor based on a random forest model, training parallelization and test parallelization methods, which improves the network search efficiency without affecting the final results. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is a schematic flow chart of the method of the present invention.
[0034] Figure 2 This is a diagram of a multi-task dataset for image classification.
[0035] Figure 3 It is a schematic diagram of the unit block of the super network and the internal candidate operations
[0036] Figure 4 This is a schematic diagram of the overall search process.
[0037] Figure 5 This is the neural network weight update flowchart
[0038] Figure 6 This is the neural network structure update flow chart
[0039] Figure 7 This is the flow chart of the evolutionary algorithm to optimize the neural network structure parameters
[0040] Figure 8 This is a flowchart for searching neural network architectures for new tasks.
[0041] Fig. 9 This is a parallel diagram of the distributed training process.
[0042] Fig.10 This is a parallel diagram of the distributed testing process.
[0043] Fig.11 It is the optimal structure searched on the Cifar-10 dataset. DETAILED DESCRIPTION
[0044] In order to more clearly understand the purpose, features and advantages of the present invention, the technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0045] The present invention is a cross-task image classification method based on fast network architecture search, and the specific process is as follows:
[0046] Step 1: Obtain a multi-task dataset for neural network architecture search. The specific steps are as follows: The basic dataset of the embodiment of the present invention is ImageNet-1k, which contains 1000 categories, a total of 1.2 million training images, 50,000 verification images, and 150,000 test images. The image size is 224×224. Generate a dataset D including training images mtrain and D including the verification image mtest , whether D mtrain Or D mtest , randomly sample pictures from 10 categories in the data set to form a task set. Each task can be regarded as a task sample instance. The sample contains 10 categories, each category has 200 training samples and 50 test samples. The entire training task pool and test task pool are composed of 500 tasks and 100 tasks respectively. The pictures in the task set use standard data enhancement methods, including random cropping (Random Crop), random flip (Horizontal Flip), and image enhancement;
[0047] Step 2: Build a super network containing the candidate network weights and a structure population encoded by the candidate network structure. The super network is composed of 20 basic unit blocks and other auxiliary layers stacked together. The unit block candidate operations are 3×3 separable convolution, 5×5 separable convolution, 7×7 separable convolution, and Xception structure. The specific structure of the candidate operation is as follows: Figure 3 As shown in Table 1, the super network architecture is shown in Table 1.
[0048] Table 1 Super network architecture
[0049] Input size Unit Block Number of channels Repetitions Step Length 224×224×3 3×3conv 16 1 2 112×112×16 CB 64 4 2 56×56×64 CB 160 4 2 28×28×160 CB 320 8 2 14×14×320 CB 640 4 2 7×7×640 1×1conv 1024 1 1 7×7×1024 GAP - 1 - 1024 fc 1000 1 -
[0050] Each unit block selects only one candidate operation, so that the entire convolutional neural network will randomly form a network structure. In each step of the super network optimization process, only one subnetwork weight is updated. The weight update formula of the super network is as follows:
[0051]
[0052] Γ(A) is the prior distribution, using uniform sampling, W represents the super network weight, W a Represents the subnet weight of structure a. The weight update formula of the subnet is as follows:
[0053]
[0054] The possible network structures are encoded to form individuals. At this time, the set of all individuals is the initial structure population. 2048 network structures are randomly sampled from the super network. The network structure is encoded into a 20-bit sequence according to the code number of the operation in each unit block. The encoding table of candidate operations is shown in Table 2. The 2048 individuals after the network structure encoding are used as the initial population of the evolutionary algorithm.
[0055] Table 2 Candidate operation number table
[0056]
[0057]
[0058] Step 3: Perform cross-task search on the multi-task dataset. The search results include meta-super networks and meta-structure populations. The search process includes two stages: weight parameter update stage and structure update stage. These two stages are performed alternately. The training, testing, and evolution three-step cycle in the entire search process is iterated 50 times:
[0059] Step 3-1: Weight parameter update process Figure 5 As shown, each unit in the super network randomly selects a candidate operation to form an optimized network individual Individuals composed of other candidate operations will not participate in the optimization, and the individual weight parameter W i , the super network weight parameter W, based on step 1, gets the support set Query Set In the individual During the training process, X represents the input data, and the network's prediction is The loss on the jth task can be expressed as in is the evaluation metric (different tasks may be different), the individual network will learn m times for each task, where m is a small integer, and the individual On the jth task, the weights are updated according to formula (1.1), the supernet learning rate α is fixed to 0.01, and each task is learned 10 times:
[0060]
[0061] When all individuals have completed all tasks, calculate all sub-networks The difference vector obtained in step 10, the weight parameters of the super network are as follows according to formula (1.2):
[0062]
[0063] The initial value of parameter ∈ is set to 0.1, and decay occurs after each iteration. The decay formula is as follows:
[0064]
[0065] Where n represents the current training number, and N represents the total training number, which is 50 in this example;
[0066] Step 3-2: While updating the weight parameters, we use the evolutionary algorithm to update the structural parameters A. The structural update process is as follows: Figure 6 As shown in Figure 1, the optimization of network structure based on evolutionary algorithm includes testing, screening, crossover and mutation, and building a new population. We perform performance tests on each individual in the population, thereby retaining some individuals with better performance for crossover and mutation. mtest Randomly select a batch of tasks Then let each individual in the population update this batch of tasks 10 times, and use the random forest model to predict the accuracy of each individual in the population;
[0067] Screening: After testing, each individual will have a performance score, i.e., accuracy. Individuals with high scores will be retained, while individuals with low scores will be removed from the population.
[0068] Crossover and mutation: The retained individuals have a certain probability of undergoing crossover and mutation operations. The structural code of the individual is regarded as the gene of the individual, and the unit in the code is regarded as the chromosome. The crossover operation is to select two individuals from the retained population, and the chromosomes of the two individuals are cut at a certain common position. The front and back strings are respectively crossed and combined to form two new chromosomes. The individuals that undergo crossover account for 50% of the number of individuals in the remaining population. The mutation operation is that when the retained individuals are replicated, there is a very small probability that a part of the chromosomes may mutate into other chromosomes (that is, after replication, it may be exactly the same as the original individual, or some chromosomes may be different). The number of individuals that undergo mutation accounts for 50% of the number of individuals in the remaining population, and the mutation probability is 0.1; Construct a new population: After crossover and mutation, continue to randomly generate a batch of new network structures, add them to the population to keep the population unchanged, and jump to step 3-1;
[0069] Step 4: The target dataset is Cifar-10, which contains 60,000 32×32 color images of 10 categories, with 6,000 images in each category. During the search, the size of each input image needs to be uniformly set to 224×224, and the data set is enhanced by standard data. The training process uses the traditional supervised training method, the loss function uses the cross entropy loss function, and the stochastic gradient descent method is used to update. The learning rate is fixed at 0.01, and then the best population searched on the multi-task data set is used as the initialization population for the new task according to step 2, with a total of 2048 network structures. The population is then quickly updated based on the continuous evolution algorithm. The super network learns 20 epochs. The training method follows the super network and subnetwork update method given in step 2. The population is updated every 2 epochs, and the population is evolved after each learning. The test phase of population evolution also uses the traditional supervised test method. Each evolution is halved according to the population reduction strategy. The population no longer randomly generates new individuals, but only the individuals in the structured population undergo crossover mutation to obtain new individuals, and the number of populations will be reduced by half after each evolution. The hyperparameter setting of the evolution process is the same as step 3-2. The individuals remaining at the end of evolution are the optimal network.
[0070] Furthermore, in steps 3-1 and 3-2, 8 processes are set for parallel search in training and testing parallelism, the batch size of each process is set to 4, and each task is learned 5 times.
[0071] The evaluation indicators of the neural network architecture include the test accuracy of the optimal architecture searched under the limitation of computing power, the time to search for the optimal architecture, and the existing capacity occupied during the search.
[0072] The comparison results of the search effects of the present invention and the existing methods on the Cifar-10 dataset are shown in Table 1. The optimal structure searched on Cifar-10 is as follows: Fig.11 shown.
[0073] Table 1. Search results of each method on the Cifar-10 dataset
[0074]
[0075] From the above experimental results, it can be seen that through the cross-task image classification method based on fast neural network architecture search provided by the embodiment of the present invention, the set of neural network architectures searched on the imagenet-1k dataset can be used to quickly perform neural network architecture search on cifar-10, and the search time is much shorter than other methods. The model of this embodiment of the present invention has a higher accuracy rate than the models of other methods.
[0076] The contents described in the embodiments of this specification are merely examples of implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A cross-task image classification method based on fast network architecture search, the specific steps are as follows: Step 1: Obtain a multi-task dataset for neural network architecture search and divide the dataset into D containing training data mtrain and D containing test data mtest , whether D mtrain Or D mtest , randomly sample pictures from the data set to form a task set. Each task is regarded as a task sample instance. The sample contains N categories, each category has K0 training samples and K1 test samples. The set of task training sets consisting of N×K0 samples is called the support set The set of task test sets consisting of N×K1 samples is called the query set. D mtrain A single support set in As the training set for individuals to be trained in a single task, D mtrain A single support set in Then it is a test set where individuals are tested in a single task; Step 2: Construct a super network containing candidate network weights and a structure population encoded by candidate network structures. The super network is composed of stacked unit blocks. The candidate operations of the unit blocks include 3×3 separable convolution, 5×5 separable convolution, 7×7 separable convolution, and Xception structure. The corresponding candidate operations are set inside the unit block according to the search space. Each unit block only selects one candidate operation, so that the super network randomly forms a network structure. In each step of the super network optimization process, only one subnetwork weight is updated. The weight update formula of the optimal super network is as follows: Γ(A) is the prior distribution, using uniform sampling, W represents the super network weight, W a Represents the subnet weight of structure a, and the structural update formula of the subnet is as follows: The possible network structures are encoded to form individuals. At this time, the collection of all individuals is the initial structure population. Step 3: Perform cross-task search on the multi-task dataset. The search results include meta-super networks and meta-structure populations. The search process includes two stages: the super network weight parameter update stage and the structure update stage. These two stages are performed alternately. Step 3-1: In the super network weight parameter update phase, each unit in the super network randomly selects a candidate operation to form an optimized network individual Individuals composed of other candidate operations will not participate in the optimization, and the individual weight parameter W i , the super network weight parameter W, based on step 1, gets the support set Query Set In the individual During the training process, X represents the input data, that is, the picture; at this time, the network's prediction is The loss on the jth task is expressed as in is the evaluation indicator, different tasks Different from that, the individual network learns each task m times, where m is a small integer. Update the weight on the jth task according to formula (3): When all individuals have completed all tasks, calculate all sub-networks The difference vector obtained in the mth step, the weight parameters of the super network are updated according to formula (4) to obtain a super network that adapts to multiple tasks; The parameter ∈ decays after each iteration, and the decay formula is as follows: Where n represents the current number of training times, and N represents the total number of training times; Step 3-2: While updating the weight parameters of the super network, use the evolutionary algorithm to update the structural parameter A. Optimize the network structure based on the evolutionary algorithm. The specific steps include testing, screening, crossover and mutation, and building a new population. After several rounds of updates of the weight parameters of the super network, we perform performance tests on each individual in the population, thereby retaining some individuals with better performance for crossover and mutation, and then select the best individuals from D. mtest Randomly select a batch of test sets from As a task, each individual in the population is asked to update this batch of tasks M times respectively; then the performance of each individual on this batch of tasks is calculated; in order to speed up the test, the random forest model is used to predict the accuracy of each individual in the population; Screening: After testing, each individual is given a performance score, and those with high scores are retained, while those with low scores are removed from the population; Crossover and mutation: The retained individuals have a certain probability of undergoing crossover and mutation operations. The structural code of the individual is regarded as the gene of the individual, and the unit in the code is regarded as the chromosome. The crossover operation is to select two individuals from the retained population, cut off the chromosomes of the two individuals at a certain common position, and cross the two strings to form two new chromosomes. The mutation operation is that when the retained individuals are replicated, there is a very small probability that a part of the chromosomes may mutate into other chromosomes, that is, after replication, it may be exactly the same as the original individual, or some chromosomes may be different; Construct a new population: After crossover and mutation, continue to randomly generate a batch of new candidate network structures, add them to the population to keep the population size unchanged, and jump to step 3-1; Step 3-1 and step 3-2 are executed repeatedly until the preset number of cycles is reached, and finally a super network and a network architecture population with good adaptability to multiple tasks are obtained; Step 4: When facing a new task, directly use the meta-population obtained by searching on the metadata set as the initialization population for the new task, and then quickly update the meta-population based on the continuous evolution algorithm to obtain the optimal network structure for the new task. The difference from the operation in step 3 is that the population during migration no longer randomly generates new individuals, but only the individuals in the structured population undergo crossover mutation to obtain new individuals, and after each evolution, the number of populations will be reduced by half, and the individuals remaining at the end of evolution are the optimal network.
2. The method according to claim 1, characterized in that In step 3-1, training parallelization is introduced. The training task pool is divided into K parts in advance, and the number of processes is also set to K. Then each process obtains the tasks belonging to the current process and the same network structure. After that, the network is updated slightly. The network parameters of each process will not be communicated across processes when updating. After the network on each process has learned its own task, the gradients of the last step of the network update on each process are collected, and then the gradients are averaged and distributed to each process for super network parameter update.
3. The method according to claim 1, characterized in that In step 3-2, the method introduces test parallelism, which is similar to training parallelism. The test task pool is evenly distributed to K processes. The network parameter updates on each process do not affect each other. Finally, only the fitness values obtained on each process are averaged and used as the individual fitness mean of the currently searched network structure.
