Backbone Network Freezing for Consistent Image Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for constructing machine learning models in image processing fields, such as target detection, semantic segmentation, and key point detection, suffer from inconsistencies in training objects, image data types, and learning tasks between pre-training and fine-tuning, leading to suboptimal image processing performance.
Innovation Solution
A method involving pre-training a backbone network with single-object image data and then constructing a second model with multi-object data, keeping the backbone network parameters unchanged while updating other networks, ensuring consistency in training tasks and data types, thereby improving the overall image processing performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If pre-training and fine-tuning use different training objects, image data types, and learning tasks, then the model can be constructed with standard procedures, but the image processing performance becomes suboptimal due to inconsistencies
Solution Approach 1:
The patent enforces homogeneity by requiring that pre-training and fine-tuning stages use consistent training objects (both single-object and multi-object images), consistent image data types, and consistent learning tasks (object detection). This uniformity across training stages eliminates the performance degradation caused by inconsistencies while maintaining standardized model construction procedures.
2Reliability
If the backbone network is re-trained during fine-tuning, then the model adapts better to the specific task, but the training time and computational resources increase significantly
Solution Approach 1:
The patent applies preliminary action by performing comprehensive pre-training of the backbone network using both single-object and multi-object images before the fine-tuning stage. This preliminary training equips the backbone with robust feature extraction capabilities, reducing the need for extensive re-training during fine-tuning and thereby decreasing training time while maintaining task adaptation accuracy.
Solution Approach 2:
The patent implements partial action by freezing the backbone network parameters during the fine-tuning stage and only training the processing networks. This selective training approach achieves sufficient task adaptation without the computational overhead of re-training the entire model, effectively balancing reliability and training time.
3Productivity
If only the processing networks are trained during fine-tuning with frozen backbone, then the training efficiency improves, but the model may fail to adapt sufficiently to the specific image processing task
Solution Approach 1:
The patent utilizes parameter changes by carefully adjusting the learning rate and training iterations for the processing networks during fine-tuning. By optimizing these parameters, the model achieves sufficient task adaptation through processing network training alone, maintaining high reliability while preserving training efficiency gains from freezing the backbone.
Solution Approach 2:
The comprehensive pre-training of the backbone network with diverse single-object and multi-object images serves as a strong foundation, enabling the processing networks to adapt efficiently to specific tasks with minimal training. This preliminary preparation ensures that even with frozen backbone parameters, the model achieves high task adaptation accuracy.
4Adaptability or versatility
If inconsistent data types are used between pre-training and fine-tuning, then the model can be trained on diverse data, but the learning task consistency deteriorates leading to suboptimal performance
Solution Approach 1:
The patent enforces homogeneity by standardizing the learning task (object detection) and data types across both pre-training and fine-tuning stages. While the model is trained on diverse images (both single-object and multi-object), the consistent application of object detection tasks and uniform data processing procedures maintains learning task consistency, preventing performance degradation.
Data Source
AI summary
The present disclosure discloses a method, an apparatus, an electronic device and a computer-readable medium for constructing a model. The method includes: training a model to be processed using a first dataset to obtain a first model, constructing a second model according to the backbone network in the first model and training the second model using a second dataset, and constantly keeping network parameters of the backbone network in the second model unchanged during training of the second model so as to obtain a model to be used.