After trained model, the original classification accuracy decreases

[red-gezi]

你好，我想要在原模型可识别类别基础上新增我想要的类别 比如这个图，我想新增角落文字的识别，使用”OSD“这个标签 Hello, I want to add a new category to the original model based on the recognized categories. For example, in this image, I want to add the recognition of corner text using the label "OSD" 然后我在训练后发现，虽然能有效识别文字目标了，但是car无法识别了，person识别率也大幅度下降了 Then I found after training that although I could effectively recognize text targets, the car could not be recognized, and the person recognition rate also dropped significantly. 以下是我用自己的脚本按照项目指引生成的训练集和验证集 The following are the training and validation sets generated by my own script according to the project guidelines 训练集格式 Training set format 验证集格式 Validation set format 新特征列表 Validation set format 请问问题可能出在哪里 Please tell me where the problem may be

[BIGBALLON]

Hi, @red-gezi

Option 1: Fuse your data set with the open source data set and use the fused data set for fine-tuning Option 2: Use your data to train a model, and then perform weight ensemble with the original model (e.g. https://github.com/mlfoundations/wise-ft)

Although the above two solutions will slightly reduce the generalization of the model, it is obvious that just using your own data set for fine-tuning will seriously damage the generalization.

References:

• https://github.com/longzw1997/Open-GroundingDino/issues/2
• https://github.com/longzw1997/Open-GroundingDino/issues/24
• https://github.com/longzw1997/Open-GroundingDino/issues/33