prismer/configs/pretrain.yaml

datasets: ['cc12m', 'cc3m_sgu', 'coco', 'vg']

cc12m_data_path: '/workspace_dataset/cc12m'
cc3m_data_path: '/home/datasets/cc3m'
coco_data_path: '/workspace_dataset/dataset_vqa'
vg_data_path: '/home/datasets/vqa'
label_path: '/workspace_dataset/dataset_experts'

experts: ['depth', 'normal', 'seg_coco', 'edge', 'obj_detection', 'ocr_detection']    # 'none' for PrismerZ

image_resolution: 224
prismer_model: 'prismer_base'  # 'prismer-large' for Prismer(Z)-Large
freeze: 'freeze_lang_vision'
batch_size_train: 32  #   for 4 * 8 nodes [effective batch-size: 1024]

max_epoch: 20
weight_decay: 0.05
init_lr: 3e-4  # 1e-4 for prismer_large
min_lr: 1e-6
warmup_lr: 1e-6
warmup_steps: 2000