Training config
ModelArguments:
base_model_name: Name of the model to be downloaded from HF
sign_input_dim (depricated, it is calculated automatically now): Dimmension of sign features on the model input (to be changed with multimodal input)
hidden_dropout_prob: Dropout probability of the input layer
max_length: Max sequence length to be generated by the model
# TrainingArguments are overwritten by not-None arguments
TrainingArguments:
project_name: Name of wandb project
model_name: Name of the run name (will affect wandb and checkpoint path)
output_dir: Output path for checkpoints
resume_from_checkpoint: Path to checkpoint to be loaded. The path needs to contain model.safetensors file
load_only_weights: True for loading only weights to use different scheduler or optimizer. False for loading the whole checkpoint
# Logging and saving
report_to: "wandb" for wandb report. Otherwise wandb not used
# Debugging
max_train_samples: Maximum size of training dataset for debugging. "none" for default size
max_val_samples: Maximum size of validation dataset for debugging. "none" for default size
# Data processing
max_sequence_length: Max input sequence length (decoder context window) - data sequence is cropped by this length
max_token_length: Dataloader tokenizer max_length
skip_frames: Use only each n-th frame. "True" for each 2nd frame
SignDataArguments:
data_dir: Prefix of the path to data including annotation file and metadatafile. This path is joined by paths bellow.
annotation_path: Path to train/dev annotation file
visual_features: Path to train/dev metadata file
pose:
enable_input: Whether to use this modality. Our code supports only "pose" so far
train: Path to train metafile
dev: Path to validation metafile
SignModelArguments: Dimmensions per each sign-features modality
Predict config
ModelArguments:
base_model_name: Name of the model to be downloaded from HF
sign_input_dim: Dimmension of sign features on the model input (to be changed with multimodal input)
max_length: Max sequence length to be generated by the model
EvaluationArguments:
output_dir: Output path for predictions
model_name: Only for logging purposes
skip_frames: Use only each n-th frame. "True" for each 2nd frame
# Data processing
split: test
max_sequence_length: Max input sequence length (decoder context window) - data sequence is cropped by this length
max_token_length: Dataloader tokenizer max_length
# Generation parameters
model_dir: Path to checkpoint to be evaluated. The path needs to contain model.safetensors file
# Debugging
max_val_samples: Maximum size of validation dataset for debugging. "none" for default size
SignDataArguments:
data_dir: Prefix of the path to data including annotation file and metadatafile. This path is joined by paths bellow.
annotation_path: Path to test annotation file
visual_features: Path to test metadata file
pose:
enable_input: Whether to use this modality. Our code supports only "pose" so far
test: Path to test metafile
SignModelArguments: Dimmensions per each sign-features modality
