pretrain model载入后 verb分类输出全是0.5左右
whqwill opened this issue · 5 comments
activity2vec模型中，pretrain model载入后 verb分类输出全是0.5左右，看起来就像是随机输出没有训练过。
The pretrained model is used for finetuning, not inference. You need to finetune the pretrained model on HAKE or other custom dataset to obtain the best performance.
So what kind of dataset is used for pretrain? For my understanding, using pretrained model can at least infer some actions (in 156 class) more or less.
所以预训练模型是怎么一个过程？是训的什么呢
Our pretrained model is used as a basic backbone to the extract the activity embeddings and is trained on our HAKE-Large dataset.
It should be noted that now we only provide the pretrained weights of backbone ResNet50, which means the other redundant weights are randomly generated. These weights have been removed from pretrained checkpoint in the newest version of this project.
OK. Thanks.