Question about the number of image features(embeddings) projected into the input to text decoder.

[dlwogns0128]

Hi, I read paper and checked the code about the number of image features projected into the input to text decoder.
Is it right the only feature of [CLS] token from Vision Transformer used to be projected into the input to text decoder?
And then, Does text decoder decode with the only one token of projected image features and others tokens of text features?

GenerativeImage2Text/generativeimage2text/layers/CLIP/model.py

Line 270 in 777860e

x = self.ln_post(x[:, 0, :])

Thanks about your nice research!

[amsword]

no. output_grid will be true in this case

GenerativeImage2Text/generativeimage2text/layers/CLIP/model.py

Line 263 in 777860e

if self.output_grid: