-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathvqa-pretrain.sh
More file actions
28 lines (26 loc) · 930 Bytes
/
vqa-pretrain.sh
File metadata and controls
28 lines (26 loc) · 930 Bytes
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
export DATA_DIR=./X_COCO/
export LOG_DIR=./logs/vqa-pre
unset CUDA_VISIBLE_DEVICES
python -m paddle.distributed.launch --gpus "3" --log_dir $LOG_DIR run_pretrain.py \
--input_dir $DATA_DIR \
--output_dir $LOG_DIR \
--dataset vqa2 \
--model_type visualbert \
--model_name_or_path visualbert-vqa-pre \
--image_feature_type coco_detectron_fix_100 \
--train_batch_size 16 \
--learning_rate 1e-5 \
--num_train_epochs 1 \
--save_steps 5000
# python -m paddle.distributed.launch --gpus "3" --log_dir $LOG_DIR run_pretrain.py \
# --input_dir $DATA_DIR \
# --output_dir $LOG_DIR \
# --dataset vqa2 \
# --model_type visualbert \
# --model_name_or_path ./logs/vqa-pre/model_5000.pdparams \
# --image_feature_type coco_detectron_fix_100 \
# --train_batch_size 16 \
# --learning_rate 1e-5 \
# --num_train_epochs 1 \
# --init_from_ckpt \
# --save_steps 1000