how to produce embedding by bert

[mayuelala]

您好，我在复现您的工作的时候，有一个问题，关于使用bert 如何生成embedding

1. 您使用bert产生的embedding 是1536维的，官方默认给定的bert是768维的，您是用什么方式扩展维度的呢？是使用fc层？

[mayuelala]

2. 输入到 verb对应的object时，您是直接使用的object？还是用something代替的呢？

[mayuelala]

3.您在代码中使用的语义特征维度是1536，但是您PaStaNet论文中使用的维度是2304，请问应该是哪个呢

[hwfan]

The relation between dim 768, 1536 and 2304

1. For the original implementation of PaStaNet, we jointly process a <human-object-interaction> triplet, for each element of this triplet we extract its word embedding from BERT. That's to say, for an HOI triplet the dimension number is 768 x 3 = 2304.
2. In the implementation of A2V we simplify the model to <human-interaction> for adaptation to more general action understanding scenarios. Therefore the dimension number of word embedding is 768 x 2 = 1536.
3. For the word embedding of interacted object, please refer to the implementation of HAKE-Action for more details.