Implementation Details

[HYC01]

python run.py with data_root=./datasets/mmimdb
num_gpus=1
num_nodes=1
per_gpu_batchsize=4
task_finetune_mmimdb
load_path=./pretrained_weights/vilt_200k_mlm_itm.ckpt
exp_name=./output/mmimdb

Thanks for your impressive work,
Following the above training setting, I got the results "40.23" on mmimdb dataset. So l will appreciate if you can provide more details about training settings.
Thanks for your time!

[icedle]

Perhaps the issue with the evaluation script? I set test_ratio=0.7, test_type=both, and I got the result "42.05"

[AAASnow]

你们运行了food数据集吗？请问你们的mmimdb数据集运行了多久呢？我在使用food，2张4090大概跑了四天还没有结束，是不是出现了问题呢？这是我的运行指令
python run.py with \data_root=./datasets/Food101 \num_gpus=2 num_nodes=1 \per_gpu_batchsize=8 batch_size=8 \task_finetune_food101 \tokenizer=./bert-base-uncased \load_path=./models/vilt_irtr_coco.ckpt \exp_name=food101_cpu

[HYC01]

这是来自QQ邮箱的假期自动回复邮件。你好，这里是胡逸琛的邮箱，我已收到您的来信，我将尽快给您回复，如有急事，请电联我，谢谢！联系电话：13867538299