run.sh

# Add the following line to enable test mode; otherwise, it defaults to validation mode
export TEST_MODE=True

# Baseline: greedy decoding, dataset could be "Controlled_Images_A", "Controlled_Images_B", "COCO_QA_one_obj", "COCO_QA_two_obj", "VG_QA_one_obj", "VG_QA_two_obj", "VSR"
# For "Controlled_Images_A", "Controlled_Images_B", "COCO_QA_one_obj", "COCO_QA_two_obj", use "four" option.
# For "VG_QA_one_obj" and "VG_QA_two_obj", use "six" option. 
python3 main_aro.py --dataset=Controlled_Images_A --model-name='llava1.5' --download --method=base  --option=four

# For Scaling_Vis on Controlled_A, a weight of 0.8 is used. 
# For Scaling_Vis on Controlled_B, a weight of 0.8 is used.
# For Scaling_Vis on COCO_QA_one_obj, a weight of 1.2 is used.
# For Scaling_Vis on COCO_QA_two_obj, a weight of 1.2 is used.
# For Scaling_Vis on VG_QA_one_obj, a weight of 2.0 is used.
# For Scaling_Vis on VG_QA_two_obj, a weight of 2.0 is used.
# For Scaling_Vis on VSR, a weight of 0.5 is used.
python3 main_aro.py --dataset=Controlled_Images_A --model-name='llava1.5' --download --method=scaling_vis  --weight=0.8  --option=four

# For Adapt_Vis on Controlled_A, weight1 is set to 0.5, weight2 to 1.5, and threshold to 0.4
# For Adapt_Vis on Controlled_B, weight1 is set to 0.5, weight2 to 1.5, and threshold to 0.35
# For Adapt_Vis on COCO_QA_one_obj, weight1 is set to 0.5, weight2 to 1.2, and threshold to 0.3
# For Adapt_Vis on COCO_QA_two_obj, weight1 is set to 0.5, weight2 to 1.2, and threshold to 0.3
# For Adapt_Vis on VG_QA_one_obj, weight1 is set to 0.5, weight2 to 2.0, and threshold to 0.2
# For Adapt_Vis on VG_QA_two_obj, weight1 is set to 0.5, weight2 to 2.0, and threshold to 0.2
# For Adapt_Vis on Controlled_B, weight1 is set to 0.5, weight2 to 1.2, and threshold to 0.64
python3 main_aro.py --dataset=Controlled_Images_A --model-name='llava1.5' --download --method adapt_vis --weight1 0.5  --weight2 1.5 --threshold 0.4 --option=four