Change from resnet152 to resnet50

I have perform an image duplication detection with pytorch pretrained model. I am using resnet152 for extracing embedding vector. I already embedding millions of record already. And resnet152 is too big which consume 16GB of GPU memory. I want to switch to resnet50 which use only 4GB of GPU memeory. The problem for me is they are not generate the same embedding output. Any suggestion that I can approach this better?