config_base.yaml

data_dir: "" # directory where data is stored
track: True # "True" if you want to track the model using wandb
run_name: "" 
run_type: "joint" # one of ["joint", "fitb", "fc"]
input_json: # json file containing the caption data
save_path: # directory where the model is saved

seed: 0
learning_rate: 5e-5
batch_size: 16
pre_nepoch: 30 # number of epochs


# feature directories (all relative to the data_dir)
input_fc_dir: # path to action features
input_clip_dir: # path to clip features
input_arc_face_dir: # path to arc_face features
input_arc_face_clusters: # path to cluster indexes for face features
use_bert_embedding: True # whether to use bert embeddings (features are large, so set to false when debugging)
bert_embedding_dir: # path to bert embeddings
fc_feat_size: 1024

# additonal files
input_label_h5: # path to h5 file of labels

# Model settings
encoding_size: 512
tokenizer_path: # path to tokenizer (relative to data_dir)
bert_size: 1536
nvid: 5
nsegments: 5
max_sent_num: 5 


# optimizer settings
optim: "adam"
optim_alpha: 0.9
optim_beta: 0.999
optim_epsilon: 1e-8
weight_decay: 0
grad_clip: 0.1

# overfit run
overfit: False # whether to run the model on a small subset of data as sanity check

# checkpoint settings
use_checkpoint: False # whether to use a checkpoint
ckpt_path: # path to checkpoint (relative to data_dir)

# Fillin metric options (don't change)
split: "val"
val_id: ""
eval_accuracy: 1