I am currently working on image classification using transfer learning. I found underfitting in the accuracy curve. My supervisor suggested me to increase the node number of the preexisting networks of Keras application (vgg 16 , resnet50 e.t.c). Is there any way to do this? Please suggest.