|
train_file: ['vqa_train.json', |
|
'vqa_val.json', |
|
] |
|
|
|
test_file: ['vqa_test.json'] |
|
answer_list: 'answer_list.json' |
|
|
|
vqa_root: 'coco/' |
|
vg_root: 'visual_genome/' |
|
|
|
image_res: 224 |
|
batch_size_train: 64 |
|
batch_size_test: 4 |
|
k_test: 128 |
|
|
|
warm_up: True |
|
|
|
|
|
|
|
optimizer: {opt: adamW, lr: 2e-5, weight_decay: 0.02} |
|
schedular: {sched: cosine, lr: 2e-5, epochs: 8, min_lr: 1e-6, decay_rate: 1, warmup_lr: 1e-5, warmup_epochs: 4, cooldown_epochs: 0} |
|
|
|
use_vis_prefix: True |
|
start_layer_idx: 11 |
|
end_layer_idx: 23 |
|
|
|
injected_hidden_states: 1 |
|
|
|
lm_loss_weight: 0.1 |
|
|
|
unfreeze_text_layer_norm: False |
|
unfreeze_vision_layer_norm: False |
|
|
|
|
|
|
|
|
|
|
