The open source implementation of the cross attention mechanism from the paper: "JOINTLY TRAINING LARGE AUTOREGRESSIVE MULTIMODAL MODELS"
- Lucidrains
- Agorians
pip install cross-attn
import torch
from cross_attn.main import MultiModalCrossAttention
# Test the MultiModalCrossAttention module
dim = 512 # For example
num_heads = 8
cross_attn = MultiModalCrossAttention(dim, num_heads)
Hllm_sample = torch.randn(32, dim, dim) # Batch size = 32, Sequence length = 10
Himg_sample = torch.randn(32, dim, dim)
output = cross_attn(Hllm_sample, Himg_sample)
print(output)
print(output.shape) # Expected: [32, 10, 512]MIT
@misc{2309.15564,
Author = {Emanuele Aiello and Lili Yu and Yixin Nie and Armen Aghajanyan and Barlas Oguz},
Title = {Jointly Training Large Autoregressive Multimodal Models},
Year = {2023},
Eprint = {arXiv:2309.15564},
}