open-mmlab / mmdetection

OpenMMLab Detection Toolbox and Benchmark
https://mmdetection.readthedocs.io
Apache License 2.0
29.67k stars 9.47k forks source link

Fine-tuning on MM Grounding DINO needs tree pre-trained models?微调mm-g-dino需要三个预训练模型? #11932

Open zxt6174 opened 2 months ago

zxt6174 commented 2 months ago

To train a mm_grodunding_dino ,we need to load both BERT and Swin two pre-trained models。 To fine-tune a mm_grounding_dino using my dataset, I need to load a pre-trained MM_Grounding_DINO and the config file ,which means I need to load BERT and Swin again.There are three models in total that I should load. Is this the right model fine-tuning process? Or maybe my method is wrong, how can I change it?

训练一个mm_grounding_dino需要调用BERT和Swin。 在自己的数据集上微调一个mm_grounding_dino,需要调用预训练的mm_g_dino权重和它的配置文件,但它的配置文件里需要调用BERT和Swin。一共需要调用三个预训练模型。 这是正确的微调流程吗?一定要调用三个模型吗?不能只调用一个mm_g_dino吗?还是我的方法出错了?

simranbajaj06 commented 2 months ago

what are the end to end steps you follow to finetune MM_Grounding_DINO ? i finetune the MM_Grounding_DINO on 12 epochs and on 100 epochs as well , but its not giving the results even when i visualize the predictions there is no single bbox predicted on any image @zxt6174 @Czm369

zxt6174 commented 1 month ago

I follwed the usage.md and the train.md to fine-tune on my datasets.

simranbajaj06 commented 1 month ago

what is the size of data you take to finetune , and for how many epochs you trained on , and what time it took?