visdial-bert
visdial-bert copied to clipboard
no running on batch_size=80
python train.py -batch_size 80 -batch_multiply 1 -lr 2e-5 -image_lr 2e-5 -mask_prob 0.1 -sequences_per_image 2 -start_path checkpoints-release/vqa_pretrained_weights
hi, I cannot use batch_size=80 for training. What configuration are you running?
Are you using 8 gpus?
2gpus
I train on 8 gpus based on pytorch=1.4, but the result is very poor, must pytorch=1.0?