The test evaluations is lower than the paper's

[Rayyoana]

I got {'BLEU': [0.8164844125702885, 0.6596750046143284, 0.5145357003585111, 0.3929742875865025], 'METEOR': 0.2901561761368199, 'ROUGE': 0.5838811960462188, 'CIDEr': 1.3468707161094793} with you provided "swintransformer_base_texthiproto2000-800_best.pth" file. Do I need to train the model on swintransformer_large to get the results in your paper?

[NovaMind-Z]

Hi,
thanks for following our work! The results you reproduced above is for the Swin-B configuration(Table 2 in our paper). If you want to get the result CIDEr 144.2, you should train the Swin-L version (swintransofmer_large).