some questions on evaluating

[ALR-alr]

How can the model evaluate on GLEU tasks？The tasks are text-pure， but in the paper it said “Similar to PLM, when prefix image is none, this task will degenerate into “text-to-image generation” task, forcing the model to generate an image with the input caption”， so how can the model complete text-pure tasks？

[ALR-alr]

Maybe only use the transformer encoder in DAVINCI？

[shizhediao]

Yes, you can just use the text encoder separately.

[ALR-alr]

Yes, you can just use the text encoder separately.

It is said in the paper that "We follow the practice of BART (Lewis et al., 2020) and feed the
same input to the encoder and decoder, and the hidden state of the final decoder token is fed into a
new multi-class linear classifier or regression head." In my understanding, isn't the decoder input here similar to the one in the transformer, where already generated tokens are used as decoder input to generate the next token through autoregression? Why is it said that the decoder input is manually passed in and is the same as the encoder input?

[ALR-alr]

How can the model evaluate on GLEU tasks？The tasks are text-pure， but in the paper it said “Similar to PLM, when prefix image is none, this task will degenerate into “text-to-image generation” task, forcing the model to generate an image with the input caption”， so how can the model complete text-pure tasks？

As in the code

DaVinci/models/model_glue.py

Line 36 in 283ea6f

gen_text=text,

,
same text is passed in DAVINCI model, but we need to predict the relationship between two sentences. Why is the hidden state from just one sentence can be passed in the classifier and predict two sentences' relationship?
I'm sorry that I have so many beginner level question...