C3D
C3D copied to clipboard
low accuracy for higher batch size
Hi, I have a problem regarding the batch size. I am finetuning resnet_UCF101.caffemodel with my own data set. I tried different batch sizes(4,8,16,24) but the network is not learning. But when I changed batchsize=1 the network is learning so fast.
What is the reason for this behaviour? Is there anything wrong in my implementation ? If I use batch size=1 is the results okay or does it affect normalisation? Can somebody help me please?