Incremental learning method, system and device based on multi-model cooperation and capacity management, and storage medium

Abstract

The invention discloses an incremental learning method, system and device based on multi-model cooperation and capacity management and a storage medium. The method comprises the following steps: judging a current model pool state; sampling and mixing new and old data from the core set; finely adjusting the main model of the model pool; evaluating the performance on the old task verification set; judging performance attenuation; freezing the current model pool to train a new model, and adding the new model to the model pool; converting an unconverted model in the model pool, and inputting a to-be-detected image; executing multi-model parallel reasoning to obtain a final detection result; continuously judging a system state; reconstructing a system model, and collecting historical data and a core set; the lightweight student model is trained through distillation, and the model is used for replacing an original model in a model pool, so that the problems of disastrous forgetting, model performance bottleneck and calculation efficiency of incremental model training in the existing incremental learning technology are solved, and the calculation efficiency of the model is improved according to model capacity management adjusted by calculation resources. And the performance and efficiency of the target detection model in long-term incremental learning are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an incremental learning method, system, device, and storage medium based on multi-model collaboration and capacity management. Background Technology

[0002] With the development of deep learning technology, object detection models have been widely used in various fields. However, the data types and distributions in real-world application environments are often dynamically changing, requiring models to continuously learn new knowledge without forgetting what they have already learned. This places higher demands on the incremental learning capabilities of these models.

[0003] Existing incremental learning methods face two main challenges: catastrophic forgetting, where models overwrite or forget old knowledge when learning new knowledge; and model capacity limitations, where the representational power of a single model is limited and cannot meet the needs of long-term continuous learning. Traditional solutions include regularization-based methods, playback-based methods, and dynamic architecture-based methods, but these methods have limited effectiveness in object detection tasks, especially in addressing the performance degradation problem in long-term incremental learning. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide an incremental learning method that effectively solves the problems of catastrophic forgetting and model capacity limitation in object detection tasks by using dynamic management of model pools to achieve long-term sustainable learning; another purpose of this invention is to provide an incremental learning system, device and storage medium based on multi-model collaboration and capacity management.

[0005] Technical solution: The incremental learning method based on multi-model collaboration and capacity management described in this invention includes... (1) Determine the current model pool status. If the model pool is not empty, proceed to step (2); otherwise, proceed to step (5). (2) The preprocessed incremental sample data is mixed with the sampled data in the core set to obtain a new sample set. The core set is constructed by using a combination strategy based on feature space coverage and hard example mining on the old sample set. (3) Construct a weighted loss function based on the old sample set and the new sample set, and fine-tune the main model of the model pool; (4) Evaluate the performance of the fine-tuned master model on the old task validation set. If the performance degradation is not less than the set threshold, proceed to step (5); otherwise, proceed to step (6). (5) Freeze the current model pool, train a new model based on the core set and the new sample set, add the new model to the model pool, and proceed to step (6). (6) Convert the unconverted models in the model pool, input the image to be detected, perform multi-model parallel inference, and use a weighted fusion algorithm to obtain the final detection result; (7) Determine whether the system state has reached the reconstruction threshold. If it is not less than the reconstruction threshold, the system model is reconstructed. The model pool is used as the teacher model. The lightweight student model is trained by knowledge distillation technology. The lightweight student model replaces the model in the current model pool. Incremental learning ends. Otherwise, incremental learning ends.

[0006] Furthermore, in step (2), a core set is constructed from the old sample set using a combined strategy of feature space coverage and hard example mining, as follows: The K-Center Greedy algorithm is used to select the first representative sample from the old sample set; The strategy of identifying difficult examples is used to select diverse second samples from the old sample set. A core set is constructed based on the first and second samples using weighted averages.

[0007] Furthermore, in step (2), the preprocessing of incremental sample data includes size normalization, data augmentation and annotation format standardization. Data augmentation includes random cropping, color dithering and mosaic enhancement.

[0008] Furthermore, in step (4), the threshold is set to the mAP50 attenuation range of 3%-5%.

[0009] Furthermore, in step (6), converting the unconverted models in the model pool involves converting the unconverted models in the model pool into ONNX format.

[0010] Furthermore, in step (6), the weighted fusion algorithm is a confidence-based weighted box fusion algorithm, including... Clustering of multi-model prediction boxes; Confidence-weighted fusion of overlapping bounding boxes; Calculate a weighted average of the confidence levels.

[0011] Furthermore, in step (7), the reconstruction threshold is at least one of the upper limit of the number of models in the model pool or the system inference delay threshold.

[0012] The incremental learning system based on multi-model collaboration and capacity management described in this invention includes: The model pool status determination module is used to determine the current status of the model pool; The data processing module is used to mix the preprocessed incremental sample data with the sampled data in the core set to obtain a new sample set. The core set is constructed by using a combination strategy based on feature space coverage and hard example mining on the old sample set. The model pool master model fine-tuning module is used to construct a weighted loss function based on the old and new sample sets, and to fine-tune the model pool master model. The main model performance evaluation module is used to evaluate the performance of the fine-tuned main model on the old task validation set. The model pool update module is used to freeze the current model pool, train a new model based on the core set and the new sample set, and add the new model to the model pool. The multi-model parallel inference module is used to transform the untransformed models in the model pool, input the image to be detected, perform multi-model parallel inference, and use a weighted fusion algorithm to obtain the final detection result. The reconstruction judgment module determines whether the system state has reached the reconstruction threshold. If it is not less than the reconstruction threshold, the system model is reconstructed. The model pool is used as the teacher model, and a lightweight student model is trained using knowledge distillation technology. The lightweight student model replaces the model in the current model pool, and the incremental learning ends. Otherwise, the incremental learning ends.

[0013] The computer device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0014] The computer-readable storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0015] Beneficial Effects: Compared with existing technologies, the significant advantages of this invention are as follows: This invention preprocesses incremental sample data; determines the current model pool state; samples from the core set and mixes new and old data; fine-tunes the main model in the model pool; evaluates performance on the old task validation set; determines whether the performance decay is less than a threshold; freezes the current model pool and trains a new model; adds a new model to the model pool; transforms the untransformed models in the model pool and inputs the image to be detected; performs multi-model parallel inference and uses a weighted fusion algorithm to obtain the final detection result; outputs the image to be detected; continuously determines whether the system state has reached the reconstruction threshold; reconstructs the system model, collects historical data and the core set; distills and trains a lightweight student model, and uses this model to replace the original models in the model pool. This solves the problems of catastrophic forgetting, model performance bottlenecks, and computational efficiency issues in incremental model training that exist in existing incremental learning techniques. It achieves model capacity management adjusted according to computing resources, significantly improving the performance and efficiency of the object detection model in long-term incremental learning. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 A flowchart of the core set construction process; Figure 3 The flowchart for multi-model collaborative reasoning; Figure 4 A flowchart for collecting historical data and the core set during model pool reconstruction; Figure 5 Flowchart for training a lightweight student model for knowledge distillation; Figure 6 The figure shows the experimental results comparing the computational efficiency of this invention. Detailed Implementation

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] The incremental learning method based on multi-model collaboration and capacity management described in this invention includes the following steps: S1: On a server equipped with a domestically produced GPU, the incremental sample data is preprocessed by denoising and blurring image filtering, including normalizing the image size to 640×640 (adjusted according to the algorithm model requirements), applying random cropping, color dithering and mosaic data augmentation, specifically optimized for the target detection task.

[0019] S2: Determine the current model pool status. If the model pool is empty during the first run, proceed directly to the new model training process and go to S7. If it is not the first run, detect the category information and number of models contained in the model and go to S3.

[0020] S3: As Figure 2 As shown, sampling old data from the core set and mixing it with new data includes the following main sub-steps: The core set uses the K-Center Greedy algorithm based on feature space coverage to select diverse sample feature correspondences, and then selects the original image samples corresponding to the features; the analysis results are obtained using an inference model network, and samples with high confidence (greater than 0.9) and misidentification are selected; the training dataset is constructed by mixing the core set (weight 0.2), hard examples (weight 0.1), and the new dataset (weight 0.7) according to weights.

[0021] S4: Fine-tune the main model of the model pool, using the weighted loss function L. total =λ1*L new +λ2*L old The bounding box regression loss for older samples is assigned a lower weight. The weight coefficients λ1 and λ2 are dynamically adjusted according to the current training stage, with an initial loss weight of 0.5 for older knowledge and an additional 0.3 for older knowledge in the later stages of training.

[0022] S5: Evaluate performance on a validation set that includes all known categories, using mAP50 as the primary evaluation metric, and evaluate performance by calculating the average accuracy.

[0023] S6: Set the performance degradation threshold to 3%. When mAP50 drops below this threshold, a model update will be triggered.

[0024] S7: When performance degradation exceeds a threshold, freeze the current model pool and train a new model. The new model focuses on learning the new data distribution of the current incremental batch, adapting to the new data distribution by adjusting the weight of new and old samples (the weight of new samples is 0.95).

[0025] S8: Add the new model to the model pool as an independent parameter without affecting the parameters and performance of the original model.

[0026] S9: Convert the unconverted models in the model pool to ONNX format and input the image to be detected. Use the graphics optimization engine inside the ONNX Runtime and utilize the parallel communication mechanism gRPC to communicate with each inference model.

[0027] S10: As Figure 3 As shown, multi-model parallel inference is performed using a weighted box fusion algorithm, which includes the following main sub-steps: clustering multi-model predicted boxes; clustering predicted boxes with a confidence score greater than 0.9; if the IoU is greater than 0.55, they are added to the corresponding overlapping box set; otherwise, a new overlapping box set is created; calculating the coordinates of the fused predicted boxes using the confidence score as the weight; and taking the average confidence score of all boxes in the overlapping box set as the fused confidence score.

[0028] S11: The output detection results include the fused prediction boxes, class confidence, location information, uncertainty estimates for each prediction box, and model source information.

[0029] S12: Continuously monitor the system status and set the reconstruction threshold, including the following sub-steps: the number of models reaches 4; the inference latency exceeds 300ms.

[0030] S13: As Figure 4 As shown, when the reconstruction threshold is reached, a rolling window strategy is used to collect important samples from the most recent 7 training cycles. Inference is performed on the above dataset. The feature vector of the last layer of the backbone network is intercepted through a callback function. At the same time, the image size, ID and timestamp information are stored, as well as the inference results such as the predicted box and confidence score.

[0031] S14: Train a lightweight student model using a distillation method and replace the existing models in the model pool with this model. This includes the following steps: Freeze all model parameters in the model pool to form a teacher model, and use the data generated in S13 as the training dataset. A weighted bounding box fusion algorithm is used to generate the fused result from the training samples, and features are extracted at the same time. The traditional loss of the object detection algorithm, including bounding box regression loss and classification loss, is calculated using "real labels," which is the task loss. The predicted bounding box loss is calculated using Smooth L1, and the class loss is calculated using KL divergence. These two losses constitute the distillation loss. The feature loss is calculated using the features extracted by the student model and the feature vector in S13. Since the networks may be different, an adapter layer needs to be added to the feature map of the student model to obtain the feature loss. The total loss function of the student model is constructed using the task loss (weight 0.6), distillation loss (weight 0.2), and feature loss (weight 0.2), and training begins to obtain the new model.

[0032] The incremental learning system based on multi-model collaboration and capacity management described in this invention includes: The model pool status determination module is used to determine the current status of the model pool; The data processing module is used to mix the preprocessed incremental sample data with the sampled data in the core set to obtain a new sample set. The core set is constructed by using a combination strategy based on feature space coverage and hard example mining on the old sample set. The model pool master model fine-tuning module is used to construct a weighted loss function based on the old and new sample sets, and to fine-tune the model pool master model. The main model performance evaluation module is used to evaluate the performance of the fine-tuned main model on the old task validation set. The model pool update module is used to freeze the current model pool, train a new model based on the core set and the new sample set, and add the new model to the model pool. The multi-model parallel inference module is used to transform the untransformed models in the model pool, input the image to be detected, perform multi-model parallel inference, and use a weighted fusion algorithm to obtain the final detection result. The reconstruction judgment module determines whether the system state has reached the reconstruction threshold. If it is not less than the reconstruction threshold, the system model is reconstructed. The model pool is used as the teacher model, and a lightweight student model is trained using knowledge distillation technology. The lightweight student model replaces the model in the current model pool, and the incremental learning ends. Otherwise, the incremental learning ends.

[0033] The computer device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0034] The computer-readable storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0035] To demonstrate the computational efficiency advantage of this invention, during implementation, the traditional full-scale training method (using all new samples mixed with all old samples as the training set) was also employed to statistically analyze the computational resources used, such as... Figure 6 As shown in the figure, the computational cost of the traditional method gradually increases with the increase of the sample size, while the computational cost of the present invention gradually increases before the model pool is updated, but the growth rate is less than that of the traditional method. When the 8th sample is added, the model pool is reconstructed, and the computational cost increases sharply. However, after the reconstruction is completed, the computational cost decreases and the method cycle of the present invention is repeated, which has the advantages.

Claims

1. An incremental learning method based on multi-model collaboration and capacity management, characterized in that, include (1) Determine the current model pool status. If the model pool is not empty, proceed to step (2); otherwise, proceed to step (5). (2) The preprocessed incremental sample data is mixed with the sampled data in the core set to obtain a new sample set. The core set is constructed by using a combination strategy based on feature space coverage and hard example mining on the old sample set. (3) Construct a weighted loss function based on the old sample set and the new sample set, and fine-tune the main model of the model pool; (4) Evaluate the performance of the fine-tuned master model on the old task validation set. If the performance degradation is not less than the set threshold, proceed to step (5); otherwise, proceed to step (6). (5) Freeze the current model pool, train a new model based on the core set and the new sample set, add the new model to the model pool, and proceed to step (6). (6) Convert the unconverted models in the model pool, input the image to be detected, perform multi-model parallel inference, and use a weighted fusion algorithm to obtain the final detection result; (7) Determine whether the system state has reached the reconstruction threshold. If it is not less than the reconstruction threshold, the system model is reconstructed. The model pool is used as the teacher model. The lightweight student model is trained by knowledge distillation technology. The lightweight student model replaces the model in the current model pool. Incremental learning ends. Otherwise, incremental learning ends.

2. The incremental learning method based on multi-model collaboration and capacity management according to claim 1, characterized in that, In step (2), a core set is constructed from the old sample set using a combined strategy of feature space coverage and hard example mining, as follows: The K-Center Greedy algorithm is used to select the first representative sample from the old sample set; The strategy of identifying difficult examples is used to select diverse second samples from the old sample set. A core set is constructed based on the first and second samples using weighted averages.

3. The incremental learning method based on multi-model collaboration and capacity management according to claim 1, characterized in that, In step (2), the preprocessing of incremental sample data includes size normalization, data augmentation and annotation format standardization. Data augmentation includes random cropping, color dithering and mosaic enhancement.

4. The incremental learning method based on multi-model collaboration and capacity management according to claim 1, characterized in that, In step (4), the threshold is set to the mAP50 attenuation range of 3%-5%.

5. The incremental learning method based on multi-model collaboration and capacity management according to claim 1, characterized in that, In step (6), converting the unconverted models in the model pool involves converting the unconverted models in the model pool into ONNX format.

6. The incremental learning method based on multi-model collaboration and capacity management according to claim 1, characterized in that, In step (6), the weighted fusion algorithm is a confidence-based weighted box fusion algorithm, including... Clustering of multi-model prediction boxes; Confidence-weighted fusion of overlapping bounding boxes; Calculate a weighted average of the confidence levels.

7. The incremental learning method based on multi-model collaboration and capacity management according to claim 1, characterized in that, In step (7), the reconstruction threshold is at least one of the upper limit of the number of models in the model pool or the system inference delay threshold.

8. An incremental learning system based on multi-model collaboration and capacity management, characterized in that, include The model pool status determination module is used to determine the current status of the model pool; The data processing module is used to mix the preprocessed incremental sample data with the sampled data in the core set to obtain a new sample set. The core set is constructed by using a combination strategy based on feature space coverage and hard example mining on the old sample set. The model pool master model fine-tuning module is used to construct a weighted loss function based on the old and new sample sets, and to fine-tune the model pool master model. The main model performance evaluation module is used to evaluate the performance of the fine-tuned main model on the old task validation set. The model pool update module is used to freeze the current model pool, train a new model based on the core set and the new sample set, and add the new model to the model pool. The multi-model parallel inference module is used to transform the untransformed models in the model pool, input the image to be detected, perform multi-model parallel inference, and use a weighted fusion algorithm to obtain the final detection result. The reconstruction judgment module determines whether the system state has reached the reconstruction threshold. If it is not less than the reconstruction threshold, the system model is reconstructed. The model pool is used as the teacher model, and a lightweight student model is trained using knowledge distillation technology. The lightweight student model replaces the model in the current model pool, and the incremental learning ends. Otherwise, the incremental learning ends.

9. 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 method according to any one of claims 1 to 7.

10. 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 method according to any one of claims 1 to 7.

Images