Hi, I noticed the default per_device_train_batch_size in the training script is 1. When I increased it to 16 for better efficiency, the model performance (especially visual grounding accuracy) didn't improve and even slightly dropped. Any insights on why this happens? Thanks in advance for your help!
Hi, I noticed the default per_device_train_batch_size in the training script is 1. When I increased it to 16 for better efficiency, the model performance (especially visual grounding accuracy) didn't improve and even slightly dropped. Any insights on why this happens? Thanks in advance for your help!