Dimension mismatch error: While using it on my custom dataset for zeroshot multilabel case

Hi, thanks for your excellent work. I am trying to apply it on my custom dataset with 4 classes with 2 seen and 2 unseen classes . I m tryng to us tresnet_l pretrained weights but i am constantly getting error RuntimeError: mat1 and mat2 shapes cannot be multiplied (3x4 and 300x768). Can you guide what am i missing here.

Regards