5fa1a76
1
2
The text input is concatenated in the front of the visual embeddings in the embedding layer, and is expected to be bound by [CLS] and a [SEP] tokens, as in BERT.
