wenkeyu
wenkeyu
You should go to the [official site](http://shannon.cs.illinois.edu/DenotationGraph) or somewhere else(google it) to download the raw images and put them into the 'images' folder.
Sorry about an error in previous sh code when evaluating on Flickr30K. I have modified them in the current version by removing "--fold" . You can try the code now:...
感谢您的关注! Two-models ensemble是把同一个网络训练两次,分别保存为两个模型(训练时由于随机种子训练结果也不同),分别推理得到image与text的相似度矩阵(在evaluation.py文件中会保存为.npy文件),将两个矩阵求平均得到最终测试的相似度矩阵。 简单地说,基于bert的模型训练两次进行ensemble;基于gru的模型训练两次进行ensemble。这两个是分开的。 Thank you for your attention! For two-models ensemble, we train the same network twice and save them as two models (due to different random seed training...
因为发现这个对结果影响不大
和scan一样的提取方法,每张图片100个框,详情可见https://github.com/LuoweiZhou/VLP
regional,应该是前三个
抱歉,这些feature不是我提的,我没有权利在我的百度网盘里共享
详情可以参考 https://github.com/peteanderson80/bottom-up-attention
是在evaluate的时候产生了oom,你可以尝试调低train或者eval的batch size
coco12h左右,f30k 4h左右