peterpaniff

> Have you trained on COCO ? I followed the author's experiment settings, but AP50 on test-dev only have 31.6%, which is 40.4% in the paper, I am very confused...

> _No description provided._ do you have trained the model on COCO, let's chat

> I have the same problem. I trained on voc0712,tested in voc07, but mAP only have 55.1%. do you have trained the model on COCO? Let's chat

> > > Have you trained on COCO ? I followed the author's experiment settings, but AP50 on test-dev only have 31.6%, which is 40.4% in the paper, I am...

> > > > > Have you trained on COCO ? I followed the author's experiment settings, but AP50 on test-dev only have 31.6%, which is 40.4% in the paper,...

it contains DDB(depth-wise dense blocks) subnetworks, like denseNet, too much concatenation operations lead to the huge use of GPU memory, thus it slow down the training process.

> # step 2 > from openprompt.plms import load_plm plm, tokenizer, model_config, WrapperClass = load_plm("bert", "bert-base-cased") 如何修改代码，加载本地模型 在[huggingface](https://huggingface.co/)上搜索关键字bert-base-cased，进入https://huggingface.co/bert-base-cased/tree/main，将文件保存下来保存到某个目录，比如你保存到/home/123/bert_model/，这个路径根据你的需要进行修改，然后将这个路径替换掉"bert-base-cased"即可。