Faster-RCNN to extract the region feature

[sxf97]

I would like to know how to use Faster-RCNN to extract the region feature(only retain 36 regions with highest Confidence) as the input feature and the dimension of each region feature is 2048，Can you give me a small demo if possible?

[lyhuohuo]

You can follow the instructions in https://github.com/jiasenlu/vilbert_beta/tree/master/data. You just need to modify the code in file ./tools/generate_tsv.py. Specifically, change "image_ids" to the list of image paths you want to extract features.
If you only want to use the image features of MABSA downstream datasets, we provide the features in google drive https://drive.google.com/drive/folders/1rm0FtHOTMUfZfRjWIE9Ukn_1D5MDXQy3?usp=sharing.

[sxf97]

好的，谢谢你呀

[sxf97]

你好，我还想问一下这个代码的运行，您是在windows下还是在linux、ubuntu下运行的，caffa环境搭建好难

[lyhuohuo]

是在linux ubuntu环境下的，caffe配置是比较麻烦

[luckypeak-xyz]

You can follow the instructions in https://github.com/jiasenlu/vilbert_beta/tree/master/data. You just need to modify the code in file ./tools/generate_tsv.py. Specifically, change "image_ids" to the list of image paths you want to extract features.
If you only want to use the image features of MABSA downstream datasets, we provide the features in google drive https://drive.google.com/drive/folders/1rm0FtHOTMUfZfRjWIE9Ukn_1D5MDXQy3?usp=sharing.

请问后面预训练所需要的 _box _att 和 _cls_again 文件夹下的 .npy 和 .npz 文件都是通过这里得到的 .tsv 转换而来的吗？

[lyhuohuo]

是的