configs/vqa.yaml

hidden_size: &hidden_size 768
vocab_size: &vocab_size 30522
type_vocab_size: &type_vocab_size 2
max_position_embeddings: &max_position_embeddings 512
pad_token_id: &pad_token_id 0

seed: 42
world_size: 1
device: "cuda"
dist_url: "env://"
output_root: "./examples/albef/outputs"

datamodule_args:
  train_files: ["./examples/albef/data_files/vqa_train.json", "./examples/albef/data_files/vg_qa.json", "./examples/albef/data_files/vqa_val.json"]
  test_files: ["./examples/albef/data_files/vqa_test.json"]
  answer_list: "./examples/albef/data_files/answer_list.json"
  vqa_root: "./examples/albef/data_files/coco"
  vg_root: "./examples/albef/data_files/visual_genome"
  batch_size: 32
  num_workers: 8

vision_encoder_args:
  hidden_size: *hidden_size
  image_size: 384
  patch_size: 16
  num_hidden_layers: 12
  num_attention_heads: 12
  mlp_dim: 3072
  dropout: 0.0
  attention_dropout: 0.0
  layer_norm_eps: 1e-6

text_encoder_args:
  vocab_size: *vocab_size
  hidden_size: *hidden_size
  type_vocab_size: *type_vocab_size
  max_position_embeddings: *max_position_embeddings
  pad_token_id: *pad_token_id
  num_hidden_layers: 6
  num_attention_heads: 12
  intermediate_size: 3072
  layer_norm_eps: 1e-12
  dropout: 0.0

multimodal_encoder_args:
  hidden_size: *hidden_size
  num_hidden_layers: 6
  num_attention_heads: 12
  intermediate_size: 3072
  layer_norm_eps: 1e-12

text_embeddings_args:
  hidden_size: *hidden_size
  vocab_size: *vocab_size
  pad_token_id: *pad_token_id
  max_position_embeddings: *max_position_embeddings
  type_vocab_size: *type_vocab_size
  layer_norm_eps: 1e-12

prediction_head_args:
  hidden_size: *hidden_size
  vocab_size: *vocab_size
  layer_norm_eps: 1e-12

training_args:
  log_every_n_steps: 100
  alpha: 0.4
  weight_decay: 0.02
  lr: 2e-5
  min_lr: 1e-6
  max_epochs: 8
  step_size: 100
  warmup_steps: 4
  checkpoint_root: "./examples/albef/checkpoints"

eval_args:
  log_every_n_steps: 100
  k_test: 128