khanrc / honeybee

Official implementation of project Honeybee (CVPR 2024)
Other
428 stars 19 forks source link

Question abut think token #17

Closed Yi-Qi638 closed 7 months ago

Yi-Qi638 commented 7 months ago

Hi~ There is eos token when build C-Abstractor

def build_eos_tokens(config: HoneybeeVisualProjectorConfig, output_hidden_size: int):

think tokens

num_eos_tokens = config.num_eos_tokens if num_eos_tokens: eos_tokens = torch.nn.Parameter(torch.randn(1, num_eos_tokens, output_hidden_size)) nn.init.truncnormal(eos_tokens, mean=0.0, std=config.initializer_range) else: eos_tokens = None return eos_tokens

This is interesting and Is it necessary or important?

khanrc commented 7 months ago

We initially thought this eos tokens might be beneficial when conducting the experiments, but it turned out they weren't necessary. In fact, we found that removing this think or image indicator tokens actually improved performance. Table 10 in Appendix shows the results.