A color formula prediction method and device based on neural architecture search

Abstract

The application belongs to the technical field of deep learning, and discloses a color formula prediction method and device based on neural architecture search, which comprises the following steps: constructing a search space according to a plurality of neural network modules; obtaining a plurality of neural network architectures according to a diffusion generation model and the search space; processing each neural network architecture by using a preset intervention operation, evaluating the neural network architectures before and after processing according to a causal robustness evaluation algorithm, and obtaining evaluation results of each neural network architecture; screening each evaluation result based on a multi-objective optimization algorithm to obtain an optimal evaluation result; obtaining an optimal neural network architecture according to the optimal evaluation result; inputting color features into the optimal neural network architecture and processing the color features to obtain a color master formula proportion. The application can screen out an optimal neural network architecture, so that an accurate color master formula proportion can be obtained according to color features.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and particularly relates to a color formula prediction method and device based on neural architecture search. BACKGROUND

[0002] Color formula prediction is a key production link in multiple industries such as coatings. The specific steps of the prediction method are as follows: according to a given target color, a combination of proportions of several basic color materials is predicted, so as to generate a finished product color closest to the target color. In the traditional production process, formula design highly depends on the experience and repeated experiments of colorists. This method not only consumes time and effort, but also is difficult to guarantee the stability and consistency of the color formula. Moreover, in the current neural network training architecture, even if the neural network is trained for a certain time, the color formula cannot be completely and accurately obtained. SUMMARY

[0003] The present application provides a color formula prediction method and device based on neural architecture search. A diffusion generative model is introduced to search multiple neural network architectures in the problem of color formula prediction, so as to improve the exploration efficiency and quality of the neural network architecture. Moreover, a causal robustness evaluation condition is proposed, which is cooperated with a multi-objective optimization algorithm to filter out the optimal neural network architecture, so that the accurate color master formula proportion can be obtained according to the color characteristics.

[0004] In a first aspect, the embodiments of the present application provide a color formula prediction method based on neural architecture search, which comprises the following steps:

[0005] A search space is constructed according to multiple neural network modules;

[0006] Multiple neural network architectures are obtained according to a diffusion generative model and the search space;

[0007] Each neural network architecture is processed by using a preset intervention operation, and the neural network architectures before and after processing are evaluated according to a causal robustness evaluation algorithm to obtain evaluation results of each neural network architecture;

[0008] Each evaluation result is filtered out based on a multi-objective optimization algorithm to obtain an optimal evaluation result;

[0009] An optimal neural network architecture is obtained according to the optimal evaluation result;

[0010] Color characteristics are input into the optimal neural network architecture and processed to obtain a color master formula proportion.

[0011] Further, the multiple neural network modules comprise:

[0012] an input processing module, an intermediate modeling module and an output module;

[0013] The output module is the Softmax layer; the formula for the Softmax layer is:

[0014]

[0015] in, The N-dimensional vector output by the output module. Z represents the proportion of the i-th color masterbatch formula. i For the linear score of the i-th element in the output module, z j R is the linear score of the j-th element in the output module, where R is a real number and N is the number of alphanumeric characters.

[0016] Furthermore, based on the diffusion generation model and the search space, several neural network architectures are obtained, including:

[0017] The preset encoding in the search space is input into the diffusion generation model to obtain the first architecture;

[0018] Gaussian noise is input into the first architecture and processed to obtain the second architecture;

[0019] The denoising formula is input into the second architecture and processed to obtain multiple neural network architectures.

[0020] Furthermore, the pre-defined intervention procedures include:

[0021] Input intervention operations, network structure intervention operations, and input subgroup intervention operations;

[0022] Input intervention operations include Gaussian noise intervention, feature occlusion, principal component rotation, and distribution drift;

[0023] Network structure intervention operations include module removal, layer parameter freezing, and module type replacement;

[0024] Input subgroup interventions include color feature distribution interventions.

[0025] Furthermore, the evaluation results include validation error, model parameter count, and robustness penalty term; the calculation formula for the causal robustness evaluation algorithm is as follows:

[0026]

[0027] in, To verify the error, n is the number of training data, x(i) is the input of the i-th training data, and y(i) is the true label of the i-th training data;

[0028]

[0029] in, Where L is the number of model parameters, L is the number of layers in the neural network architecture, and ParamCount(l) is the number of parameters in the l-th layer.

[0030]

[0031] in, For robustness penalty items, The causal robustness score is given by K, where K is the total number of intervention trials. The error loss of the neural network architecture f on the original validation set. Let the validation loss be the value under the k-th intervention condition.

[0032] in, and The calculation formula is:

[0033]

[0034]

[0035] Among them, f k For the neural network architecture or the parameters of the neural network architecture modified under the k-th intervention condition, Let be the feature vector of the i-th sample after the k-th input perturbation, n be the number of training data, x(i) be the input of the i-th training data, and y(i) be the true label of the i-th training data.

[0036] Furthermore, based on a multi-objective optimization algorithm, the various evaluation results are filtered to obtain the optimal evaluation result, including:

[0037] Construct a three-objective search function based on the evaluation results;

[0038] A non-dominated sorting genetic algorithm is used to optimize the three-objective search function and determine the optimal evaluation result.

[0039] Furthermore, the three-objective search function includes:

[0040]

[0041] Where a is the candidate neural network architecture, f1(a) and The validation error for candidate neural network architectures;

[0042]

[0043] Where a is the candidate neural network architecture, f2(a) and The number of model parameters for the candidate neural network architecture;

[0044]

[0045] Where a represents the candidate neural network architecture, and f3(a) represents the non-robustness. The causal robustness score is given to the candidate neural network architecture.

[0046] Secondly, embodiments of this application provide a color recipe prediction device based on neural architecture search, the device comprising:

[0047] Modules are used to construct the search space based on multiple neural network modules;

[0048] The generation module is used to obtain multiple neural network architectures based on the diffusion generation model and the search space;

[0049] The evaluation module is used to process each neural network architecture using preset intervention operations, evaluate the neural network architecture before and after processing according to the causal robustness evaluation algorithm, and obtain the evaluation results of each neural network architecture.

[0050] The filtering module is used to filter the evaluation results based on a multi-objective optimization algorithm to obtain the optimal evaluation result.

[0051] The results module is used to obtain the optimal neural network architecture based on the best evaluation results;

[0052] The output module is used to input color features into the optimal neural network architecture and process them to obtain the color masterbatch formula ratio.

[0053] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the steps of a color recipe prediction method based on neural architecture search as described in any of the above embodiments.

[0054] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a color recipe prediction method based on neural architecture search as described in any of the above embodiments.

[0055] In summary, compared with the prior art, the beneficial effects of the technical solution provided in this application include at least the following:

[0056] This application provides a color recipe prediction method based on neural architecture search. The method introduces a diffusion generation model to search for multiple neural network architectures in the color recipe prediction problem, thereby improving the efficiency and quality of neural network architecture exploration. Furthermore, it proposes a causal robustness evaluation condition, which works in conjunction with a multi-objective optimization algorithm to screen out the optimal neural network architecture, thereby obtaining an accurate color masterbatch recipe ratio based on color features. Attached Figure Description

[0057] Figure 1 A flowchart of a color recipe prediction method based on neural architecture search is provided as an exemplary embodiment of this application.

[0058] Figure 2 A structural diagram of a color recipe prediction device based on neural architecture search is provided as an exemplary embodiment of this application. Detailed Implementation

[0059] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0060] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0061] Please see Figure 1 This application provides a color recipe prediction method based on neural architecture search, which specifically includes the following steps:

[0062] Step S1: Construct a search space based on multiple neural network modules.

[0063] Among them, the search space can effectively model high-dimensional input features; it can natively support output ratio constraints (non-negative + sum of 1); and it can express the structural dependencies between color numerators (such as complementarity and reinforcement).

[0064] The multiple neural network modules include: an input processing module, an intermediate modeling module, and an output module.

[0065] The input processing module includes: an MLP layer (Dense), which is a standard fully connected layer that processes dense color feature vectors; a 1D convolutional layer (1D-CNN), which is suitable for processing spectral sequence inputs; and a Transformer encoder, whose self-attention mechanism can capture the global relationships between different color channels.

[0066] In some embodiments, the hyperparameters of the input processing module can be: MLP: units = 128; 1D-CNN: kernel_size = 3, stride = 1; Transformer: num_heads = 4, hidden_dim = 64.

[0067] The intermediate modeling modules include: fully connected layers, used to expand or compress the feature space dimension; residual connections, used to maintain gradient flow and avoid degradation in deep networks; Dropout, used to suppress overfitting and enhance robustness; and BatchNorm / LayerNorm, used to improve convergence speed and training stability. These modules can be freely connected through a search process to form shallow / deep, serial / parallel structural topologies.

[0068] The output module is the Softmax layer; the formula for the Softmax layer is:

[0069]

[0070] in, The N-dimensional vector output by the output module. Z represents the proportion of the i-th color masterbatch formula. i For the linear score of the i-th element in the output module, z j R is the linear score of the j-th element in the output module, where R is a real number and N is the number of alphanumeric characters.

[0071] The search space is defined as a triple: in, The set of candidate module types (MLP, Conv1D, Attention, Norm, Softmax, etc.); It is a set of legal connection methods (forward join, residual, parallel branch); This is a set of hyperparameters (such as width, depth, dropout rate, etc.). Each neural network architecture... It is a specific instance of a triple.

[0072] Step S2: Based on the diffusion generation model and the search space, multiple neural network architectures are obtained.

[0073] Diffusion models have demonstrated superior performance in image generation, molecular generation, and structure generation in recent years. This application applies diffusion models to the generation of neural network architectures, combining the characteristics of color recipe tasks to construct a graph structure generator for generating high-quality neural network architectures in the search space.

[0074] Diffusion generative models are a class of generative models based on progressive noise addition and denoising. Their core consists of two processes: First, the forward diffusion process: gradually adding noise to the original data until approximate Gaussian noise is obtained; second, the reverse generation process: training a model to progressively "reconstruct" the original data (i.e., the generated neural network architecture) from the noise. During the generation of the neural network architecture, the neural network architecture diagram is represented as a "data form" that can be learned and reconstructed by the diffusion model.

[0075] In some embodiments, multiple neural network architectures are obtained based on the diffusion generation model and the search space, including:

[0076] Step S21: Input the preset encoding in the search space into the diffusion generation model to obtain the first architecture.

[0077] The first architecture is a directed acyclic graph.

[0078] Step S22: Input Gaussian noise into the first architecture and process it to obtain the second architecture.

[0079] The second architecture is X. t .

[0080] Step S23: Input the denoising formula into the second architecture and process it to obtain multiple neural network architectures.

[0081] Each candidate neural network architecture is represented as a directed acyclic graph (DAG):

[0082] G = (V, E)

[0083] Where G is a directed acyclic graph, and V = {v1, v2, ..., v} n Each node represents a module (such as a convolutional layer or an attention layer); Each edge represents a data flow connection between two modules. Each node v i The feature vector includes the following: module type (e.g., MLP, Conv1D, Transformer); hyperparameters (e.g., number of channels, kernel size); and hierarchical position (hierarchical order or branch index). This graph is encoded as an adjacency matrix A and an attribute tensor X, where A∈{0,1} n×n It represents a connection relationship; X∈R n×d , which represents node attributes (composed of one-hot encoding and numerical hyperparameters).

[0084] During the forward diffusion stage (noise addition stage), Gaussian noise is gradually added starting from the actual architecture diagram (A,X):

[0085]

[0086] Among them, X t The attribute representation of the node at step t (with noise added); α t For a predefined noise attenuation sequence; This is Gaussian noise. Similarly, a binary sampling strategy (such as gated Bernoulli noise addition) can be used for the adjacency matrix A of the structure graph.

[0087] In the reverse generation stage (denoising stage), a conditional generator network is trained to gradually reconstruct the structure graph from the noise:

[0088] Final output

[0089] Where f is the neural network architecture, whose parameters are learnable, and commonly used architectures include Conditional Variational Autoencoder (CVAE) and Diffusion Transformer; the final generated This constitutes a complete candidate architecture.

[0090] To generate neural network architectures that meet practical constraints and task requirements, "conditional control" signals are introduced during the diffusion process. Conditional generation is a method that guides the direction of generation in a diffusion-based generative model. For example: model complexity is limited to no more than 10M parameters; network width is limited to less than 256 for each layer; and the output structure type must end with Softmax.

[0091] These conditional embeddings are added to the input of the diffusion generative model, enabling it to learn to generate neural network architectures within a "constrained structural space." The mathematical expression for this is:

[0092]

[0093] Where c is a condition vector, such as model size target, structure category, etc.

[0094] After generation, the following validity checks are performed on the candidate neural network architectures:

[0095]

[0096]

[0097] Structures that do not meet the criteria will be discarded, and valid candidate neural network architectures will be retained for subsequent steps.

[0098] To ensure the comprehensiveness of the search, the following mechanisms are introduced in the generation of neural network architectures to improve structural diversity: multiple noise sampling initialization; different conditional vectors to control multiple structural styles (such as shallow vs. deep layers); use of distance metrics (such as structure graph Edit Distance) to avoid repeated sampling of neural network architectures; and retaining a certain proportion of "structural variation" samples for evolution.

[0099] Step S3: Process each neural network architecture using preset intervention operations, evaluate the neural network architecture before and after processing according to the causal robustness evaluation algorithm, and obtain the evaluation results of each neural network architecture.

[0100] The preset intervention operations include: input intervention operations, network structure intervention operations, and input subgroup intervention operations; input intervention operations include Gaussian noise intervention, feature occlusion, principal component rotation, and distribution drift; network structure intervention operations include module removal, layer parameter freezing, and module type replacement; input subgroup intervention operations include color feature distribution intervention.

[0101] The input intervention operations include:

[0102] For the input color vector x∈R M The artificial perturbations are shown in Table 1 below:

[0103] Table 1

[0104]

[0105]

[0106] The network structure intervention operation intervenes in some nodes of the directed acyclic graph G=(V,E) to test whether the modules inside the directed acyclic graph are redundant and whether they depend on a single path, as shown in Table 2 below:

[0107] Table 2

[0108]

[0109] The input subgroup intervention operation divides the validation set of the neural network architecture into multiple subgroups (with different color feature distributions), such as neutral colors and highly saturated colors, light colors and dark colors, continuous spectral smooth inputs and noisy samples. The consistency of the neural network architecture's prediction bias for different input groups is observed, yielding the following two results:

[0110] The first scenario is that the prediction errors are consistent (low bias) among different neural network architectures; this indicates that the neural network architecture has strong stability and generalization ability to changes in the distribution of input data; the performance of the neural network architecture does not depend on a specific color distribution (such as bright or dark colors), and has good generalization ability; the existing neural network architecture can continue to be used to enter the next stage of training and deployment; this structure will obtain a high CRS score in causal robustness evaluation. Generally, when the CRS score obtained by the neural network architecture is greater than 0.5, it indicates that the prediction errors among different subgroups are consistent.

[0111] The second scenario involves significantly different prediction errors (high bias) among different neural network architectures. This indicates that the neural network architecture has performance biases or structural sensitivities to certain input subgroups. It may exhibit overfitting to dark samples and failure to adapt to light samples, or unstable output for highly saturated samples. If the prediction error bias of a neural network architecture is high, such architectures are directly eliminated in the search to improve overall robustness.

[0112] The formula for calculating the causal robustness score is as follows:

[0113]

[0114] Wherein, CRS is the causal robustness score, f is the neural network architecture, and K is the total number of intervention experiments, usually 3 to 5. The error loss of the neural network architecture f on the original validation set. Let be the validation loss under the k-th intervention condition.

[0115] In some embodiments, it is also necessary to perform real training and intervention on some candidate neural network architectures to calculate the true causal robustness score. Specific training and intervention methods are shown in Table 3 below:

[0116] Table 3

[0117]

[0118] Among them, formulas (1) and (2) in Table 3 are respectively:

[0119]

[0120]

[0121] in, The loss is the original validation set loss. The loss under the k-th intervention condition; The original loss of the neural network architecture; Let be the loss under the k-th intervention in the neural network architecture.

[0122] The evaluation results include validation error, model parameter count, and robustness penalty term; the calculation formula for the causal robustness evaluation algorithm is as follows:

[0123]

[0124] in, To validate the error, n is the number of training data, x(i) is the input of the i-th training data, and y(i) is the true label of the i-th training data. The validation error can measure the prediction accuracy of the neural network architecture on the validation set, and the smaller the value, the better.

[0125]

[0126] in, The model parameter count is L, where L is the number of layers in the neural network architecture, and ParamCount(l) is the number of parameters in the l-th layer. The model parameter count can reflect the scale of the neural network architecture, with the goal of reducing computational overhead and deployment costs.

[0127]

[0128] in, For robustness penalty items, The causal robustness score is given by K, where K is the total number of intervention trials. The error loss of the neural network architecture f on the original validation set. is the validation loss under the k-th intervention condition; the robustness penalty term can represent the degree of fluctuation in the prediction performance of the neural network architecture under multiple interventions. The smaller this term is, the more robust and stable the model is.

[0129] in, and The calculation formula is:

[0130]

[0131]

[0132] Among them, f k For the neural network architecture or the parameters of the neural network architecture modified under the k-th intervention condition, Let be the feature vector of the i-th sample after the k-th input perturbation, n be the number of training data, x(i) be the input of the i-th training data, and y(i) be the true label of the i-th training data.

[0133] In some embodiments, the CRS metric can also be used as one of the multiple objectives in NAS, with the specific formula as follows:

[0134]

[0135] Step S4: Based on the multi-objective optimization algorithm, filter each evaluation result to obtain the optimal evaluation result.

[0136] In some embodiments, the optimal evaluation result is obtained by filtering the various evaluation results based on a multi-objective optimization algorithm, including:

[0137] A three-objective search function is constructed based on the evaluation results.

[0138] A non-dominated sorting genetic algorithm is used to optimize the three-objective search function and determine the optimal evaluation result.

[0139] The three-objective search functions include:

[0140]

[0141] Where a is the candidate neural network architecture, f1(a) and The prediction error for the candidate neural network architecture;

[0142]

[0143] Where a is the candidate neural network architecture, f2(a) and The total number of model parameters for candidate neural network architectures;

[0144]

[0145] Where a represents the candidate neural network architecture, and f3(a) represents the non-robustness. The causal robustness score is given to the candidate neural network architecture.

[0146] In multi-objective optimization algorithms, a structure *a* is called Pareto optimal if there is no other structure *b* such that *b* is better than *a* on all objectives. It is defined as follows: Structure *a* is Pareto optimal if and only if:

[0147]

[0148] The resulting set of structures is called the Pareto Front, which represents the optimal set of solutions that achieve different trade-offs among multiple objectives.

[0149] In some embodiments, mainstream multi-objective optimization algorithms can be used, such as NSGA-II (Non-Dominated Sorting Genetic Algorithm II), which is suitable for multi-objective discrete optimization problems. The main framework is as follows: Set up an initial population: a set of candidate architectures from the structure generator; Perform each iteration: (1) Crossover: Structure combination (module splicing, path swapping); (2) Mutation: Slight perturbation of the structure (replacing layer type, adding skip connections); (3) Non-dominated sorting: Pareto sorting based on a three-objective search function; (4) Crowding Distance: Maintaining structural diversity. The termination condition is reaching the maximum number of iterations or the structure set is stable.

[0150] The specific search process steps are as follows: First, initialize the population; use a diffusion model to sample N candidate structures, and calculate f1, f2, f3 for each structure; Second, multi-objective scoring; construct the objective vector F(a) = [f1, f2, f3], perform non-dominated sorting, and generate the current Pareto set; Third, structural evolution; select elite structures from the Pareto front, and perform structural crossover / mutation to generate the next generation of structural graphs; Fourth, repeated iterations; update the Pareto front structure set until convergence or round number limit is met.

[0151] Step S5: Obtain the optimal neural network architecture based on the optimal evaluation results.

[0152] After the search is complete, output a set of Pareto optimal architectures. This allows users' decision-makers to choose the appropriate structure based on their specific needs. For cloud deployment, the structure with the highest accuracy (minimum f1) is selected; for edge deployment, the structure with the lowest accuracy (minimum f2) is selected; and for robustness as a priority, the structure with the highest CRS (minimum f3) is selected. Alternatively, a weighted scoring function, Score(a), can be used, calculated as follows:

[0153] Score(a)=αf1(a)+βf2(a)+γf3(a)

[0154] The weights α, β, and γ can be dynamically set by the user or the search engine.

[0155] Step S6: Input the color features into the optimal neural network architecture and process them to obtain the color masterbatch formula ratio.

[0156] The color characteristics are derived from the colorimeter, and the total proportion of the color masterbatch formula is 1.

[0157] Specifically, the color features utilize multidimensional spectral reflectance data exported from a professional colorimeter. Complete spectral reflectance data is collected for each color sample using industrial colorimetric equipment (such as Datacolor, BYK, X-Rite, etc.). The preset wavelength range is typically 360nm to 780nm, with a sampling interval of 2nm or 5nm, resulting in an initial input vector of approximately 200–210 dimensions. Each dimension in the input vector represents the reflectance at a specific wavelength, with values ​​typically ranging from [0,1]. Zero-mean unit variance normalization is applied to all dimensions to accommodate the training requirements of the neural network architecture.

[0158] The output is a vector of colorant formulation proportions (the vector sums to 1, with over 100 dimensions). Each sample corresponds to a sparse vector of over 100 dimensions, with each dimension representing the percentage of a colorant in the color formulation; approximately 90% or more of the dimensions are 0, with only 3 to 8 non-zero components; all components are non-negative, and the sum is strictly equal to 1 (i.e., the output belongs to the probabilistic simplex).

[0159] In some embodiments, a multi-objective joint loss function is employed during the full training phase, including:

[0160] Mean square error (MSE) measures the numerical deviation between the predicted formulation and the actual formulation. The formula is:

[0161]

[0162] KL divergence The Kullback-Leibler Divergence measures the difference between the predicted probability distribution and the true probability distribution. The formula is:

[0163]

[0164] Here, ∈ is a smoothing constant that prevents the logarithm from being negative infinity; KL divergence encourages the alignment of distribution shapes, which can improve the consistency of perception.

[0165] Sparse constraint loss Encourage sparse output formulations, that is, use as few color masterbatches as possible, with the following formula:

[0166]

[0167] Controlling the number of non-zero terms in the predicted proportion using the L1 norm helps improve the cost-effectiveness and interpretability of the actual color grading process.

[0168] Total loss function The formula is:

[0169]

[0170] Where α, β, and γ are the weighting coefficients of the loss term, which are adjusted according to the requirements of the task.

[0171] After the multi-objective joint loss function is calculated, the optimal model is derived based on the training results on the dataset and then deployed online.

[0172] The above embodiments provide a color recipe prediction method based on neural architecture search. This method introduces a diffusion generative model to search for multiple neural network architectures in the color recipe prediction problem, thereby improving the efficiency and quality of neural network architecture exploration. Furthermore, it proposes a causal robustness evaluation condition, which works in conjunction with a multi-objective optimization algorithm to screen out the optimal neural network architecture, thereby obtaining an accurate color masterbatch recipe ratio based on color features.

[0173] In some embodiments, the inputs and outputs of the optimal neural network architecture are as follows:

[0174] The input is color features (trimmed from a 200+ dimensional colorimetric vector):

[0175]

[0176] The output is the colorant formula ratio (a sparse 100+ dimensional vector):

[0177]

[0178]

[0179] The above color masterbatch formula ratios meet the following conditions: all components are ≥0, and the total is strictly 1.0; it can reflect the sparsity of the actual color masterbatch used, which facilitates the implementation of the color matching process.

[0180] Please see Figure 2 Another embodiment of this application provides a color recipe prediction device based on neural architecture search, the device comprising:

[0181] Module 101 is used to construct the search space based on multiple neural network modules.

[0182] The generation module 102 is used to obtain multiple neural network architectures based on the diffusion generation model and the search space.

[0183] The evaluation module 103 is used to process each neural network architecture using preset intervention operations, evaluate the neural network architecture before and after processing according to the causal robustness evaluation algorithm, and obtain the evaluation results of each neural network architecture.

[0184] The filtering module 104 is used to filter the various evaluation results based on a multi-objective optimization algorithm to obtain the optimal evaluation result.

[0185] Result module 105 is used to obtain the optimal neural network architecture based on the optimal evaluation result.

[0186] The output module 106 is used to input color features into the optimal neural network architecture and process them to obtain the color masterbatch formula ratio.

[0187] The specific limitations of the color recipe prediction device based on neural architecture search provided in this embodiment can be found in the embodiment of the color recipe prediction method based on neural architecture search described above, and will not be repeated here. Each module in the aforementioned color recipe prediction device based on neural architecture search can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0188] This application provides a computer device that may include a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it causes the processor to perform the steps of a color recipe prediction method based on neural architecture search as described in any of the above embodiments.

[0189] The working process, working details, and technical effects of the computer device provided in this embodiment can be found in the embodiment of a color recipe prediction method based on neural architecture search described above, and will not be repeated here.

[0190] This application provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps of a color recipe prediction method based on neural architecture search as described in any of the above embodiments. The computer-readable storage medium refers to a data storage medium, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0191] The working process, working details, and technical effects of the computer-readable storage medium provided in this embodiment can be found in the embodiment of a color recipe prediction method based on neural architecture search described above, and will not be repeated here.

[0192] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0193] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0194] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A color recipe prediction method based on neural architecture search, characterized in that, The method includes: The search space is constructed based on multiple neural network modules; Multiple neural network architectures were obtained based on the diffusion generation model and the search space; Each neural network architecture is processed using a preset intervention operation, and the neural network architectures before and after processing are evaluated according to a causal robustness evaluation algorithm to obtain the evaluation results of each neural network architecture. The optimal evaluation result is obtained by filtering the evaluation results using a multi-objective optimization algorithm. This filtering process includes: constructing a three-objective search function based on the evaluation results; optimizing the three-objective search function using a non-dominated sorting genetic algorithm to determine the optimal evaluation result; the evaluation result includes validation error, model parameter count, and robustness penalty term; the calculation formula for the causal robustness evaluation algorithm is as follows: ;in, To verify the error, n is the number of training data, x(i) is the input of the i-th training data, and y(i) is the true label of the i-th training data; ;in, For the number of model parameters, The number of layers in the neural network architecture. For the first The number of parameters in the layer; ;in, For robustness penalty items, The causal robustness score is given by K, where K is the total number of intervention trials. For neural network architecture Error loss on the original validation set, For the first The validation loss under each intervention condition; among which, and The calculation formula is: ; ;in, In the first Modified neural network architectures or parameters of neural network architectures under various intervention conditions. In the first The first type of input perturbation The feature vector of each sample, n is the number of training data, x(i) is the input of the i-th training data, and y(i) is the true label of the i-th training data; The optimal neural network architecture is obtained based on the optimal evaluation results. The color features are input into the optimal neural network architecture and processed to obtain the color masterbatch formula ratio.

2. The color recipe prediction method based on neural architecture search according to claim 1, characterized in that, The plurality of neural network modules include: Input processing module, intermediate modeling module, and output module; The output module is a Softmax layer; the formula for the Softmax layer is: in, ; The N-dimensional vector output by the output module. For the first The proportions of the color masterbatch formula, This is the linear score of the i-th element in the output module. R is the linear score of the j-th element in the output module, where R is a real number and N is the number of alphanumeric characters.

3. The color recipe prediction method based on neural architecture search according to claim 1, characterized in that, The process of obtaining multiple neural network architectures based on the diffusion generation model and the search space includes: The preset encoding in the search space is input into the diffusion generation model to obtain the first architecture; Gaussian noise is input into the first architecture and processed to obtain the second architecture; The denoising formula is input into the second architecture and processed to obtain the multiple neural network architectures.

4. The color recipe prediction method based on neural architecture search according to claim 1, characterized in that, The preset intervention operations include: Input intervention operations, network structure intervention operations, and input subgroup intervention operations; The input intervention operations include Gaussian noise intervention, feature occlusion, principal component rotation, and distribution drift; The network structure intervention operations include module removal, layer parameter freezing, and module type replacement; The input subgroup intervention operation includes color feature distribution intervention.

5. The color recipe prediction method based on neural architecture search according to claim 1, characterized in that, The three-objective search function includes: Where a is the candidate neural network architecture, f1(a) and The validation error for candidate neural network architectures; Where a is the candidate neural network architecture, f2(a) and The number of model parameters for the candidate neural network architecture; Where 'a' represents the candidate neural network architecture. For non-robustness, The causal robustness score is given to the candidate neural network architecture.

6. A color recipe prediction device based on neural architecture search, characterized in that, The device includes: Modules are used to construct the search space based on multiple neural network modules; A generation module is used to obtain multiple neural network architectures based on the diffusion generation model and the search space; The evaluation module is used to process each of the neural network architectures using preset intervention operations, evaluate the neural network architectures before and after processing according to the causal robustness evaluation algorithm, and obtain the evaluation results of each neural network architecture. A screening module is used to screen each evaluation result based on a multi-objective optimization algorithm to obtain the optimal evaluation result. The screening of each evaluation result based on the multi-objective optimization algorithm to obtain the optimal evaluation result includes: constructing a three-objective search function based on each evaluation result; optimizing the three-objective search function using a non-dominated sorting genetic algorithm to determine the optimal evaluation result; the evaluation result includes validation error, model parameter count, and robustness penalty term; the calculation formula for the causal robustness evaluation algorithm is: ;in, To verify the error, n is the number of training data, x(i) is the input of the i-th training data, and y(i) is the true label of the i-th training data; ;in, For the number of model parameters, The number of layers in the neural network architecture. For the first The number of parameters in the layer; ;in, For robustness penalty items, The causal robustness score is given by K, where K is the total number of intervention trials. For neural network architecture Error loss on the original validation set, For the first The validation loss under each intervention condition; among which, and The calculation formula is: ; ;in, In the first Modified neural network architectures or parameters of neural network architectures under various intervention conditions. In the first The first type of input perturbation The feature vector of each sample, n is the number of training data, x(i) is the input of the i-th training data, and y(i) is the true label of the i-th training data; The results module is used to obtain the optimal neural network architecture based on the optimal evaluation results; The output module is used to input color features into the optimal neural network architecture and process them to obtain the color masterbatch formula ratio.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the color recipe prediction method based on neural architecture search as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the color recipe prediction method based on neural architecture search as described in any one of claims 1 to 5.

Images