Transformer Image captioning model produces just padding rather than a caption

Question

I am trying to produce a model that will produce a caption for an image using resnet as the encoder, transformer as the decoder and COCO as the database.

After traini