GPU Out Of Memory - Tensorflow Object Detection API

I’m currently training an EfficientDetD7 to detect objects in images, using the ‘Tensorflow Object Detection API’ framework.
I’m using an NVIDIA H100 GPU with 80 Gb of VRAM. I can only train with a batch size equal to or less than 2. After some tests I realized that each image in the batch occupies 33 Gb. Therefore, a batch size equal to 1 consumes 33 Gb and a batch size equal to 2 consumes 66 Gb.
Is there any way to reduce memory consumption? I’m already using tensorflow’s Memory Growth.