Hannibal046 / SDDS

请问代码中low_encoder和high_encoder中的low和high代表什么意思呢？

low: low-level feature(token)
high: high-level feature(utterance)

明白了，感谢回复。另外想问一下，Static-Dynamic Fusion Module是在GraphTransformerMultiHeadAttentionLayer中体现吗？

Hi, 具体的代码实现可以参考:

Lines 194 to 197 in 8ba2a10

    
           energy_ls = [] 
        
           for idx in range(self.n_heads): 
        
               energy_ls.append(self.combine_conv(torch.stack([energy[:,idx,:,:],feature_conv_output[:,idx,:,:]],dim=1)).squeeze(1)) 
        
           energy = self.conv_activation_fn(torch.stack(energy_ls,dim=1)) # bs,n_head,q_len,k_len

明白了，感谢

BART的源代码我也是用的Huggingface的版本: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bart/modeling_bart.py
Baseline代码可以参考:
https://github.com/huggingface/transformers/tree/main/examples/pytorch/summarization

好的感谢

	energy_ls = []
	for idx in range(self.n_heads):
	energy_ls.append(self.combine_conv(torch.stack([energy[:,idx,:,:],feature_conv_output[:,idx,:,:]],dim=1)).squeeze(1))
	energy = self.conv_activation_fn(torch.stack(energy_ls,dim=1)) # bs,n_head,q_len,k_len

求解答