Could you tell me where is the code implement of cross-attention in the paper VECO? about alicemind HOT 1 CLOSED

We only released the fine-tuning code of VECO. In the directory of NLG code which are implemented based on fairseq, we transferred the pre-trained model state_dict keys into the format of fairseq via a mapping when fine-tuning on NLG tasks.
The encoder_attn denotes the cross_attention. The detailed mapping is shown as below:

{'cross_attention.query.weight': 'encoder_attn.q_proj.weight',
'cross_attention.query.bias': 'encoder_attn.q_proj.bias',
'cross_attention.key.weight': 'encoder_attn.k_proj.weight',
'cross_attention.key.bias': 'encoder_attn.k_proj.bias',
'cross_attention.value.weight': 'encoder_attn.v_proj.weight',
'cross_attention.value.bias': 'encoder_attn.v_proj.bias',
'cross_attention.output.dense.weight': 'encoder_attn.out_proj.weight',
'cross_attention.output.dense.bias': 'encoder_attn.out_proj.bias',
'cross_attention.output.LayerNorm.gamma': 'encoder_attn_layer_norm.weight',
'cross_attention.output.LayerNorm.beta': 'encoder_attn_layer_norm.bias'}

from alicemind.