DRV and congestion collaborative prediction method based on deep learning

Abstract

The invention provides a DRV and congestion collaborative prediction method based on deep learning. The method comprises the following steps: establishing a multi-task intensive prediction model CVNet; taking the congestion and the DRV as a specific task of the multi-task intensive prediction model, and obtaining a loss function of the specific task; and by minimizing the total loss in the loss function, enabling the multi-task intensive prediction model to perform collaborative prediction on congestion and DRV. According to the prediction method provided by the invention, global wiring congestion and detailed wiring DRV can be jointly predicted by using features collected before a layout stage as a center, and a multi-task learning (MLT) method is proposed, so that the prediction accuracy can be effectively improved.

Description

Technical Field

[0001] The present disclosure relates to the field of digital integrated circuit electronic design automation (EDA), and particularly to a method for collaborative prediction of DRV and congestion based on deep learning. Background Art

[0002] In modern integrated circuit (IC) design, routing congestion and design rule violations (DRV) play a crucial role in determining the routability of an IC layout. On an IC layout, routing congestion occurs when the demand for routing resources such as metal layers, vias, or routing tracks exceeds the available resources within a specific area. This imbalance restricts the available paths for interconnections, thus hindering the effective routing of connections between components, leading to challenges in meeting timing constraints, signal integrity issues, and overall design quality. On the other hand, DRV refers to specific locations where design rules are violated. These violations may cause manufacturing problems, such as short circuits in metal wires, pitch violations, or other electrical rule violations. Congestion and DRV need to be iterated multiple times during the highly time-consuming placement and routing process in order to correct them before the final design, as shown in Figure 1 Fig. a. An effective and efficient solution is to incorporate machine learning (ML) techniques into the design process. A supervised ML model can be trained based on historical data and patterns in the placement stage to predict the likelihood and severity of routing congestion or DRV problems in a new design, as shown in Figure 1 Fig. b. If there is too much congestion or DRV, placement will be guided by the prediction results. Without DRV and congestion, the subsequent routing results will follow the prediction results. By discovering potential problems through the ML model, designers or EDA tools can optimize the routing path and make more informed decisions. This method can ultimately improve the design quality and reduce the turnaround time cost.

[0003] Using a machine learning model to predict congestion or DRV based on the placement result can improve the design quality and reduce the turnaround time cost between the placement and routing stages. Most existing methods utilize supervised STL models and build and train single-task prediction models for congestion or DRV respectively. The advantage of using STL models is that they can adapt to the unique characteristics and requirements of congestion or DRV. This enables researchers to design specialized models and feature representations to improve the prediction accuracy of each model individually.

[0004] Most current methods rely on single-task learning (STL) models, which can predict congestion or DRV separately. These STL methods simplify the modeling and training by treating congestion and DRV as independent tasks and optimizing each task separately, rather than using a unified optimization strategy to consider the relationship between them, resulting in low prediction accuracy. Summary of the Invention

[0005] A DRV and congestion collaborative prediction method based on deep learning is proposed in this disclosure to solve at least one technical problem in the prior art. The specific steps are as follows:

[0006] A DRV and congestion collaborative prediction method based on deep learning includes:

[0007] Build a multi-task dense prediction model CVNet;

[0008] Regard congestion and DRV as specific tasks of the multi-task dense prediction model, and obtain the loss functions of the specific tasks;

[0009] Make the multi-task dense prediction model perform collaborative prediction on congestion and DRV by minimizing the total loss in the loss function.

[0010] The multi-task dense prediction model CVNet includes:

[0011] First, define the input and output of the multi-task dense prediction model:

[0012] The multi-task dense prediction model has dual inputs. For any task, the training dataset D i consists of n i samples, and the input is described as follows:

[0013]

[0014] Among them, is the j-th training instance; and are the true maps corresponding to the congestion and DRV binary labels respectively; indicates that each layout is divided into w i ×h i blocks; c i is the number of channels of the input feature map; and n1 represents the total number of training instances of the congestion task; n2 represents the total number of training instances of the DRV task; w1, h1 are the length and width of the true label of the congestion task respectively, and w2, h2 are the length and width of the true label of the DRV task respectively;

[0015] Assume X=(X 1 ,X 2 ),Y=(Y 1 ,Y 2 ), is an MTL model, and θ mtl represents its total trainable parameters. The learning problem of the multi-task dense prediction model is expressed as:

[0016]

[0017] Express the total loss function of the multi-task dense prediction model as The calculation formula is as follows:

[0018]

[0019] Where represents the losses of two specific tasks, represents their joint function.

[0020] The multi-task dense prediction model CVNet includes: the CVNet-P model;

[0021] The CVNet-P model contains a single input with 9 channels and generates two outputs, each with one channel, for predicting congestion and DRV respectively.

[0022] The CVNet-P model includes:

[0023] Two convolutional layers Conv1 for further downsampling;

[0024] Apply the third convolution Conv2, perform batch normalization and hyperbolic tangent activation to stabilize training and accelerate convergence;

[0025] Three cross-stitch units are integrated on the encoder backbone for predicting congestion and DRV, promoting information exchange and interaction between the congestion and DRV tasks.

[0026] The two convolutional layers Conv1 for further downsampling include:

[0027] Set two consecutive convolutional layers Conv1, where the convolutional layer Conv1 is a 2D convolution with a 3x3 kernel and a padding of 1;

[0028] Use the first convolutional layer Conv1 for batch normalization and Leaky ReLU activation, and then apply another 2D convolution with the same parameters to further refine the features;

[0029] Downsample the feature map through a max pooling layer with a kernel size of 2x2 and a stride of 2.

[0030] The three cross-stitch units integrated on the encoder backbone for predicting congestion and DRV include:

[0031] According to the two downsampled maps from the h th Max Pool layer (z for the congestion and DRV tasks 1 and z 2),The cross-stitch unit at position (i, j) is defined as follows:

[0032]

[0033] Where represent the outputs of the cross-stitch units for congestion and DRV respectively, and α 1,1 、α 1,2 、α 2,1 and α 2,2 are the weights for information sharing between the congestion and DRV tasks. Z 1 (i, j), Z 2 (i, j) represent the inputs of the cross-stitch units for congestion and DRV respectively;

[0034] By normalizing the original labels to [0, 1] and using a threshold for binary classification, a binary map output is obtained;

[0035] The threshold is 10% of the maximum DRV violation count.

[0036] The multi-task dense prediction model CVNet includes: the CVNet-C model;

[0037] The CVNet-C model has two outputs, corresponding to congestion and DRV predictions respectively, and adopts a cascaded network structure, depending on different inputs for congestion and DRV prediction tasks;

[0038] Among them, the input of the DRV task includes a 9+1 channel feature map, where the predicted congestion map is stacked with the initial 9-channel feature map to form a cascaded structure, and the congestion backbone and the DRV backbone are arranged and interconnected in sequence.

[0039] The multi-task dense prediction model includes:

[0040] The total loss of the multi-task dense prediction model is expressed as the geometric mean of the losses of each task, as follows:

[0041] are the loss values of the congestion prediction task and the DRV prediction task respectively;

[0042] Assume and are the predicted congestion value and the actual congestion value of the kth pixel respectively, and the loss function of the congestion task is defined as the mean square error between them:

[0043]

[0044] Where represents the number of pixels in the congestion map; n1 represents the total number of training instances of the congestion task;

[0045] Assume and are the predicted DRV value and the actual DRV value of the k-th pixel respectively, then the loss function of the DRV task is defined as the mean square error (MSE) between them:

[0046]

[0047] where represents the number of pixels in the DRV heatmap.

[0048] Beneficial effects:

[0049] The beneficial effects of the present disclosure at least include: The prediction method described in the present disclosure only uses a small amount of information to perform collaborative prediction of congestion and DRV. Through multi-task learning, CVNet can promote the performance of DRV prediction, and an encoder-decoder architecture is adopted to predict global routing congestion and detailed routing DRV. Both variants CVNet-P and CVNet-C promote the prediction performance of the DRV backbone through the congestion backbone; CVNet-P adopts a parallel network structure, and the two tasks share similar input features, while CVNet-C adopts a cascaded network structure, and both congestion and DRV predictions have separate inputs. The prediction method described in the present disclosure can center on jointly predicting global routing congestion and detailed routing DRV with the features collected before the layout stage, and a multi-task learning (MLT) method can effectively improve the prediction accuracy. Description of the drawings

[0050] Figure 1 is a comparison diagram based on the traditional layout and routing method and intelligent layout and routing;

[0051] Figure 2 is the true heatmap of congestion in the global layout stage and DRV in the detailed layout stage;

[0052] Figure 3 is the structure diagram of the CVNet-P model;

[0053] Figure 4 is the structure diagram of the CVNet-C model;

[0054] Figure 5 is the training and test loss curve diagram of the CVNet-P and CVNet-C models. Specific implementation manners

[0055] In the following, various embodiments of the present disclosure will be described more fully. The present disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of the present disclosure to the specific embodiments disclosed herein, but the present disclosure should be understood to cover all adjustments, equivalents, and / or alternative solutions that fall within the spirit and scope of the various embodiments of the present disclosure.

[0056] In the following, the term "comprising" or "may comprise" that may be used in various embodiments of the present disclosure indicates the presence of the disclosed functions, operations, or elements, and does not limit the addition of one or more functions, operations, or elements. In addition, as used in various embodiments of the present disclosure, the terms "comprising", "having", and their cognates are only intended to indicate a specific feature, number, step, operation, element, component, or combination of the foregoing items, and should not be construed as precluding the existence or addition of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing items first.

[0057] In various embodiments of the present disclosure, the expression "or" or "at least one of A or / and B" includes any combination or all combinations of the recited words. For example, the expression "A or B" or "at least one of A or / and B" may include A, may include B, or may include both A and B.

[0058] Expressions (such as "first", "second", etc.) used in various embodiments of the present disclosure may modify various constituent elements in the various embodiments, but do not limit the corresponding constituent elements. For example, the above expressions do not limit the order and / or importance of the elements. The above expressions are only for the purpose of distinguishing one element from other elements. For example, the first user device and the second user device indicate different user devices, although both are user devices. For example, without departing from the scope of the various embodiments of the present disclosure, the first element may be referred to as the second element, and similarly, the second element may also be referred to as the first element.

[0059] It should be noted that if it is described that one constituent element is "connected" to another constituent element, the first constituent element may be directly connected to the second constituent element, and a third constituent element may be "connected" between the first constituent element and the second constituent element. Conversely, when one constituent element is "directly connected" to another constituent element, it can be understood that there is no third constituent element between the first constituent element and the second constituent element.

[0060] The term "user" used in various embodiments of the present disclosure may indicate a person who uses an electronic device or a device that uses an electronic device (e.g., an artificial intelligence electronic device).

[0061] The terms used in the various embodiments of the present disclosure are only used for the purpose of describing specific embodiments and are not intended to limit the various embodiments of the present disclosure. As used herein, the singular form is intended to also include the plural form, unless the context clearly indicates otherwise. Unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as the meanings commonly understood by ordinary technicians in the field to which the various embodiments of the present disclosure belong. The terms (such as the terms defined in the generally used dictionary) will be interpreted as having the same meaning as the contextual meaning in the relevant technical field and will not be interpreted as having an idealized meaning or an overly formal meaning, unless clearly defined in the various embodiments of the present disclosure.

[0062] Congestion and DRV are interrelated because routing congestion can lead to design rule violations. When routing is severely congested, it can lead to violations of various design constraints, such as minimum spacing, width, or other rules specified for routing. In advanced processes, the relationship between congestion and DRV is more complex because there are more reasons, such as pin accessibility and new design rules. Figure 2 As shown in Figure 1, in some areas, congestion may occur simultaneously with DRV, and similarly, in some cases, DRV may occur simultaneously with congestion. However, no matter how complex the relationship is, the correlation between congestion and DRV does exist. Figure 1 In the design flow shown in (b), a machine learning model that can simultaneously predict congestion and DRV is crucial.

[0063] However, most current methods rely on single-task learning (STL) models that can predict congestion or DRV separately. These STL methods simplify modeling and training by treating congestion and DRV as independent tasks and optimizing each task separately, without utilizing a unified optimization strategy to consider the relationship between them. Considering this relationship and implementing a unified optimization strategy across various tasks simultaneously can improve the accuracy and efficiency of prediction. Therefore, machine learning models that can simultaneously predict congestion and DRV provide a possible alternative and improvement to existing STL models.

[0064] This disclosure focuses on jointly predicting global routing congestion and detailed routing DRV using features collected before the layout stage, and proposes a multi-task learning (MLT) approach that can effectively address the above challenges. Specific embodiment 1:

[0066] This embodiment provides a DRV and congestion collaborative prediction method CVNet based on deep learning. Through the CVNet model, the accuracy of congestion and DRVs prediction after layout can be improved and the time consumption can be reduced.

[0067] First, define the input and output of the CVNet model:

[0068] Both of the two structural variants in the CVNet model have dual inputs. For each task, the training dataset D i consists of n i samples, and the inputs are described as follows:

[0069]

[0070] where is the j-th training instance, whose structure is a multi-channel feature map of the layout (one feature corresponds to one channel), while and are the ground truth maps corresponding to the congestion and DRV binary labels respectively. Assume that represents that each layout is divided into w i ×h i blocks (each block corresponds to one pixel), and c i is the number of channels of the input feature map. Let and Then Equation (1) can be rewritten as:

[0071]

[0072] Assume X = (X 1 , X 2 ), Y = (Y 1 , Y 2 ), is an MTL model, and θ mtl represents its total trainable parameters. The multi-task learning problem can be formulated as:

[0073]

[0074] Assume that and are the specific task backbones for T1 and T2, and θ1 and θ2 represent their respective trainable parameters. Equation (3) can be elaborated as follows:

[0075]

[0076] Denote the total loss function of the multi-task model as The calculation formula is as follows:

[0077]

[0078] where represents the losses of two specific tasks, represents their joint function. By minimizing the total loss the multi-task model can make collaborative predictions for multiple tasks.

[0079] CVNet contains two architecture variants: CVNet-P and CVNet-C. Among them, CVNet-P adopts a parallel network design, while CVNet-C adopts a cascaded network architecture.

[0080] The model structure diagram of the CVNet-P variant is as Figure 3 shown. CVNet-P accepts a single input containing 9 channels and generates two outputs, each with one channel, respectively designated for predicting congestion and DRV. CVNet-P contains two parallel network structures, each of which is customized for congestion or DRV prediction. The backbones of the two specific tasks follow a similar pattern in their encoder and decoder structures.

[0081] Each encoder starts with Conv1, which is a convolutional layer that performs two consecutive operations: first, a 2D convolution with a 3x3 kernel and a padding of 1, then batch normalization and Leaky ReLU activation; then, another 2D convolution with the same parameters is applied.

[0082] Subsequently, a max pooling layer with a kernel size of 2x2 and a stride of 2 is used to downsample the feature map.

[0083] Next, another convolutional layer with operations similar to Conv1 refines the extracted features, and another max pooling layer then performs further downsampling.

[0084] Finally, a third convolution Conv2 with a 3×3 kernel, a stride of 1, and a padding of 1 is applied;

[0085] then batch normalization and hyperbolic tangent activation are performed to stabilize training and accelerate convergence.

[0086] Three cross-stitch units are integrated on the encoder backbones of specific tasks to facilitate information exchange and interaction between the congestion and DRV tasks. These cross-stitch units enable the two tasks to share information while retaining their individuality and autonomously learning specific task functions. According to the two downsampled maps (z th and z 1 for the congestion and DRV tasks respectively) from the h 2 -th max pooling layer, the cross-stitch unit at position (i, j) is defined as follows:

[0087]

[0088] where represent the outputs of the cross-stitch units for congestion and DRV respectively. In addition, α 1,1 , α 1,2 , α 2,1 and α2,2 is trainable and represents the weights for information sharing between two tasks.

[0089] Each decoder backbone starts with Conv1, which aims to refine the feature space for a specific task. Next, Upc is used, which is an upsampling layer that employs transposed 2D convolution with a kernel size of 4, a stride of 2, and a padding of 1. This layer, together with batch normalization and the Leaky ReLU activation function, enhances the spatial resolution. The process is repeated with another Conv1 layer to further refine the features, followed by another Upc layer for additional upsampling. Subsequently, Conv3 (a convolutional layer with a kernel size of 3 and a padding of 1) refines the features before generating the final output using the sigmoid activation function. This architecture allows for effective feature refinement and reconstruction while gradually increasing the spatial dimension of the desired output. Additionally, skip connections are included between the encoder and the decoder. These connections ensure smooth information transfer from early encoder layers to subsequent decoder layers, preserving fine-grained details and spatial information that might be lost during encoder downsampling. By bypassing certain layers, these skip connections provide access to high-level information and detailed spatial information, thus facilitating more accurate and detailed dense prediction.

[0090] Additionally, since the DRV prediction is a binary classification task, the original labels need to be normalized to [0, 1], and then a threshold is used for binary classification to obtain a binary map output, where the threshold is 10% of the maximum DRV violation count.

[0091] CVNet-C: As Figure 4 shown, CVNet-C is also a specialized deep neural network with two outputs corresponding to congestion and DRV prediction respectively. Compared with CVNet-P, it adopts a cascaded network structure and relies on different inputs for congestion and DRV prediction. In the backbones of the two specific tasks, the encoder is similar to that of CVNet-P, except without the cross-stitch unit, while the decoder has the same structure as the decoder in CVNet-P. The input for the congestion task consists of a 9-channel feature map, similar to the input of CVNet-P. However, the input for the DRV task includes a (9 + 1)-channel feature map, where the predicted congestion map (from the main line of the congestion task) is stacked with the initial 9-channel feature map. This creates a cascaded structure where the congestion backbone and the DRV backbone are arranged and interconnected in sequence. This sequential arrangement allows for progressive feature extraction and refinement, with each backbone specializing in different aspects of the input data or prediction tasks.

[0092] Since CVNet performs two tasks simultaneously, and the loss function plays a crucial role in the training process. To ensure robustness and scalability, this embodiment adopts a loss function that is almost unaffected by the scale of the data of a single task, and represents the total loss of the multi-task learning problem as the geometric mean of the losses of each task. Equation (5) can be elaborated as follows:

[0093]

[0094] Suppose and are the predicted congestion value and the actual congestion value of the k-th pixel respectively. The loss function of the congestion task is defined as the mean square error (MSE) between them:

[0095]

[0096] where represents the number of pixels in the congestion map.

[0097] Similarly, suppose and are the predicted DRV value and the actual DRV value of the k-th pixel respectively. Then the loss function of the DRV task is defined as the mean square error (MSE) between them.

[0098]

[0099] where represents the number of pixels in the DRV hotspot map.

[0100] Verification step:

[0101] This embodiment has been verified on the 28nm dataset of the open-source CircuitNet. This dataset uses commercial EDA tools, including Synopsys Design Compiler and Cadence Innovus, to obtain 10,242 samples from 6 designs and has different tool parameter settings.

[0102] The input features selected for the experiment are macro_region, cell_density, RUDY, RUDY_long, RUDY_short, RUDY_pin, RUDY_pin_long, congestion_eGR_vertical_util, congestion_eGR_horizontal, which are generated after global placement; the label used for congestion is congestion, which is generated after global routing, the label used for DRV is DRC violations, which is generated after detailed routing, and we have performed normalization and binary classification processing.

[0103] The experimental environment is an Intel Xeon Gold 6354 CPU, a 2TB hard disk capacity, and an Nvidia A100 graphics card. Among them, CVNet-P, CVNet-C, and the baseline STL (single-task) model are implemented in Python 3.9.16, PyTorch 2.0.1, and LibMTL (B. Lin and Y. Zhang, “LibMTL: A Python library for multi-task learning,” Journal of Machine Learning Research, vol. 24, no. 209, pp. 1–7, 2023.). CVNet-P and CVNet-C are trained with the same learning rate of 0.0001, while all STL models adopt a fixed learning rate of 0.0002. The training process uses a consistent batch size of 32 for all models and continues until convergence, adopting an early stopping strategy with a patience of 30. Throughout the training phase, the Adam optimizer is applied to all models.

[0104] Table 1. Dataset Design and Partition

[0105]

[0106] To evaluate the generalization ability of this patent, a cross-design scheme is adopted for data partitioning. As shown in Table 1, 7078 samples of 4 designs are selected for training, and the remaining 3164 samples of 2 designs are reserved for testing. This ensures that all designs in the test dataset are brand new and do not repeat the data in the training phase.

[0107] All feature maps are resized to a size of 256×256 pixels and then concatenated before being input into the model. The task-specific losses of CVNet are normalized, and then the total loss is calculated using Equation (8).

[0108] To evaluate the performance of CVNet, we use four evaluation metrics: PSNR (Peak Signal-to-Noise Ratio), NRMSE (Normalized Root Mean Square Error), F1 (F1 score), and AUC ROC (Area Under the ROC Curve). The first two are used to evaluate congestion prediction, while the last two are used to evaluate DRV prediction.

[0109] To evaluate the prediction performance and computational efficiency of CVNet-P and CVNet-C, the prediction results of six STL models were used as the baseline. For congestion prediction, two baselines were used: STLPROS and STL-C. Each model took a 9-channel feature map as input. For DRV prediction, there were four baseline models: STL-UNet, STL-V, STL-PROS|STL-V, and STL-C|STL-V. All models were trained using the training data and then their performance was evaluated only based on the test data, which included designs not present in the training set. The prediction performance and corresponding time costs of all models are shown in Table 2 and Table 3 respectively.

[0110] Table 2. Prediction performance of the models

[0111]

[0112] According to the data in Table 2, the congestion prediction results of CVNet-P and CVNet-C are almost the same as those of the STL models. However, it is obvious that their prediction performance for DRV is better than that of the STL models.

[0113] In addition, the data in Table 3 show that CVNet-P and CVNet-C are superior to all STL models in terms of training and test time costs. In summary, compared with the STL models, the two variants of CVNet show superior performance in terms of prediction accuracy and computational efficiency. In addition, CVNet-C has a slight advantage over CVNet-P in terms of prediction accuracy and computational efficiency.

[0114] Table 3. Running time of the models

[0115] Model Training Time(s)↓ Testing Time(s)↓ CVNet-P 7700 2310 CVNet-C 7620 2286 STL-C+STL-V 4804+5380＝10184 1441+1614＝3055 STL-PROS+STL-U-Net 10940+6250＝17190 3282+1875＝5157 STL-C|STL-V 4804+5620＝10424 1441+1686＝3127 STL-PROS|STL-U-Net 10940+9280＝20220 3282+2784＝6066

[0116] By comparing the total training and test losses of the two variants of CVNet, as Figure 5 shown, it is worth noting that their convergence speeds are similar. However, compared with the corresponding curve of CVNet-C, the loss curve of CVNet-P shows more obvious fluctuations. This can be attributed to the cascaded structure architecture of CVNet-C, where the output of the congestion-specific backbone contributes to the input of the DRV-specific backbone. This sequential information flow reduces the directional conflict between the gradients of the two different backbones, thus improving stability during training. Therefore, compared with CVNet-P with a parallel structure, CVNet-C shows a smoother loss curve with fewer fluctuations.

[0117] In summary, the method described in this embodiment only utilizes a small amount of information for collaborative prediction of congestion and DRV. Through multi-task learning, CVNet can promote the performance of DRV prediction, and adopts an encoder-decoder architecture to predict global routing congestion and detailed routing DRV. Both variants, CVNet-P and CVNet-C, promote the prediction performance of the DRV backbone through the congestion backbone; CVNet-P adopts a parallel network structure where the two tasks share similar input features, while CVNet-C adopts a cascaded network structure where congestion and DRV predictions have separate inputs. By using the GLS loss optimization strategy, both variants of CVNet accurately predict the combined impact of congestion and DRV on design quality. Cross-design experimental results show that both CVNet-P and CVNet-C outperform the STL model in terms of prediction accuracy and computational efficiency. CVNet-C has a slight advantage over CVNet-P in both prediction accuracy and computational efficiency. Specific Embodiment 2:

[0119] The present disclosure also provides an embodiment:

[0120] An electronic device includes: a storage medium and a processing unit; wherein, the storage medium is used to store a computer program; the processing unit exchanges data with the storage medium and is used to execute the computer program through the processing unit when performing congestion and DRV predictions, and perform the steps of the prediction method described in Specific Embodiment 1.

[0121] The above CPU can perform various appropriate actions and processes according to the program stored in the storage medium. The electronic device further includes the following peripherals, including an input part such as a keyboard and a mouse, and may also include an output part such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc. and a speaker.

[0122] The present disclosure also provides an embodiment:

[0123] A readable storage medium: a computer program is stored in the readable storage medium; when the computer program runs, it executes the steps of the prediction method described in Specific Embodiment 1.

[0124] In this embodiment, a computer-readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present disclosure, a computer-readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, which carries computer-readable program code. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination of the above.

[0125] The above-disclosed are only several specific implementation scenarios of the present disclosure. However, the present disclosure is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present disclosure. The above serial numbers of the present disclosure are only for description and do not represent the superiority or inferiority of the implementation scenarios.

Claims

1. A DRV and congestion collaborative prediction method based on deep learning, characterized in that, including: building a multi-task dense prediction model CVNet; taking congestion and DRV as specific tasks of the multi-task dense prediction model and obtaining the loss functions of the specific tasks; making the multi-task dense prediction model perform collaborative prediction on congestion and DRV by minimizing the total loss in the loss functions.

2. The DRV and congestion collaborative prediction method based on deep learning according to claim 1, wherein The multi-task dense prediction model CVNet includes: firstly defining the input and output of the multi-task dense prediction model: The multi-task intensive prediction model has dual inputs. For any task, the training dataset D i consists of n i samples, and the inputs are described as follows: Among them, is the j-th training instance; and are the true maps corresponding to the congestion and DRV binary labels respectively; indicates that each layout is divided into w i ×h i blocks; c i is the number of channels of the input feature map; and Suppose X=(X 1 ,X 2 ), Y=(Y 1 ,Y 2 ), is an MTL model, and θ mtl represents its total trainable parameters. The learning problem of the multi-task dense prediction model is formulated as: Express the total loss function of the multi-task intensive prediction model as The calculation formula is as follows: where represents the losses of two specific tasks, represents their joint function.

3. The method for collaborative prediction of DRV and congestion based on deep learning according to claim 1 or 2, characterized in that, The multi-task dense prediction model CVNet includes: the CVNet-P model; The CVNet-P model contains a single input of 9 channels and generates two outputs, each output having one channel, respectively for predicting congestion and DRV.

4. The method for collaborative prediction of DRV and congestion based on deep learning according to claim 3, wherein The CVNet-P model includes: two convolutional layers Conv1 for further downsampling; applying the third convolution Conv2, performing batch normalization and hyperbolic tangent activation to stabilize training and accelerate convergence; three cross-stitch units integrated on the encoder backbone for predicting congestion and DRV to promote information exchange and interaction between the congestion and DRV tasks.

5. The method for collaborative prediction of DRV and congestion based on deep learning according to claim 4, wherein The two convolutional layers Conv1 for further downsampling include: setting two consecutive convolutional layers Conv1, where the convolutional layer Conv1 is a 2D convolution with a 3x3 kernel and a padding of 1; using the first convolutional layer Conv1 for batch normalization and Leaky ReLU activation, and then applying another 2D convolution with the same parameters to further refine the features; using a max pooling layer with a kernel size of 2x2 and a stride of 2 for downsampling the feature map.

6. The DRV and congestion collaborative prediction method based on deep learning according to claim 4, characterized in that The three cross-stitch units integrated on the encoder backbone for predicting congestion and DRV include: According to two downsampled maps from the h th Max Pool layer (z for the congestion and DRV tasks 1 and z 2 ), the cross-stitch unit at position (i, j) is defined as follows: Among them respectively represent the outputs of the cross-stitch units for congestion and DRV, and α 1,1 , α 1,2 , α 2,1 and α 2,2 are the weights for information sharing between the congestion and DRV tasks.

7. The method for collaborative prediction of DRV and congestion based on deep learning according to claim 4, wherein: by normalizing the original label to [0,1] and using a threshold for binary classification to obtain a binary map output; the threshold is 10% of the maximum DRV violation count.

8. The method for collaborative prediction of DRV and congestion based on deep learning according to claim 1, characterized in that The multi-task dense prediction model CVNet includes: the CVNet-C model; The CVNet-C model has two outputs, corresponding to congestion and DRV predictions respectively, and adopts a cascaded network structure, relying on different inputs to perform congestion and DRV prediction tasks; wherein, the input of the DRV task includes a 9+1 channel feature map, where the predicted congestion map is stacked with the initial 9-channel feature map to form a cascaded structure, where the congestion backbone and the DRV backbone are arranged and interconnected in sequence.

9. The DRV and congestion collaborative prediction method based on deep learning according to claim 2, wherein The multi-task dense prediction model includes: representing the total loss of the multi-task dense prediction model as the geometric mean of the losses of each task, as follows: Hypothesis and are the predicted congestion value and the actual congestion value of the k-th pixel respectively. The loss function of the congestion task is defined as the mean square error between them: Among them represents the number of pixels in the congestion map; Hypothesis and are the predicted DRV value and the actual DRV value of the k-th pixel respectively. Then, the loss function of the DRV task is defined as the mean square error MSE between them: Among them, represents the number of pixels in the DRV hot spot map.

Images