Finetuning midddle layers (vision module layers = 22 to 26 +projector + language layers = 0 to 4)

HyperParameters of this model

lr: 2e-5

epochs: 10

batch size: 1

grad acumulation step: 4