features (N, 2048) = graph node features? & How to switch "mode='teacher_forcing" to mode='feedback' when testing about meshed-memory-transformer

[marcellacornia]

Hi @XuMengyaAmy,

1. Yes, we only used the image region features coming from Faster R-CNN trained on Visual Genome. We did not use any additional features to represent the global image.
2. The Mk and Mv are additional vectors learned during training that can represent the intrinsic relationships between image regions.
3. This is not possible. The generated captions change according to the images given as input to the model.
4. When you call the model.forward, the model uses the ground-truth words. When you call the model.beam_search instead, the model uses the previously generated words.
5. Since we are currently working on different projects, we do not plan to release the script at the moment.

Hope this answers all your questions.

Best,
Marcella