File size: 621 Bytes
794924b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
vqa_root: '/export/share/datasets/vision/VQA/Images/mscoco/' #followed by train2014/
vg_root: '/export/share/datasets/vision/visual-genome/'  #followed by image/
train_files: ['vqa_train','vqa_val','vg_qa']
ann_root: 'annotation'

# set pretrained as a file path or an url
pretrained: 'https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model*_vqa.pth'

# size of vit model; base or large
vit: 'base'
batch_size_train: 16 
batch_size_test: 32 
vit_grad_ckpt: False
vit_ckpt_layer: 0
init_lr: 2e-5

image_size: 480

k_test: 128
inference: 'rank'

# optimizer
weight_decay: 0.05
min_lr: 0
max_epoch: 10