It is mentioned in paper that both Visformer-Ti and Visformer-S were trained with batch size 512. However, in this github repository, batch size for Visformer-Ti is setup to 1024 and Visformer-S to 512. Is there no final accuracy difference between training with 512 and 1024?
It is mentioned in paper that both Visformer-Ti and Visformer-S were trained with batch size 512. However, in this github repository, batch size for Visformer-Ti is setup to 1024 and Visformer-S to 512. Is there no final accuracy difference between training with 512 and 1024?