pretrain-vcr-large-4gpu.json

{
    "model_config": "/src/config/uniter-large.json",
    "checkpoint": "/pretrain/uniter-large.pt",
    "output_dir": "/storage/pretrain/vcr_2nd_stage_large",
    "mrm_prob": 0.15,
    "max_txt_len": 220,
    "conf_th": 0.2,
    "max_bb": 100,
    "min_bb": 10,
    "num_bb": 36,
    "train_batch_size": 6144,
    "val_batch_size": 8000,
    "gradient_accumulation_steps": 4,
    "learning_rate": 3e-05,
    "valid_steps": 2000,
    "num_train_steps": 45000,
    "optim": "adamw",
    "betas": [
        0.9,
        0.98
    ],
    "dropout": 0.1,
    "weight_decay": 0.01,
    "grad_norm": -1,
    "warmup_steps": 4500,
    "seed": 42,
    "fp16": true,
    "n_workers": 4,
    "pin_mem": true,
    "train_datasets": [
        {
            "name": "vcr",
            "db": [
                "/txt/vcr_train.db/"
            ],
            "img": [
                "/img/vcr_gt_train/;/img/vcr_train/"
            ],
            "tasks": [
                "mlm",
                "mrfr",
                "mrckl"
            ],
            "mix_ratio": [
                10,
                1,
                1
            ]
        }
    ],
    "val_datasets": [
        {
            "name": "vcr",
            "db": [
                "/txt/vcr_val.db/"
            ],
            "img": [
                "/img/vcr_gt_val/;/img/vcr_val/"
            ],
            "tasks": [
                "mlm",
                "mrfr",
                "mrckl"
            ]
        }
    ],
    "rank": 0
}