in dpr_scale/datamodule/dpr.py [0:0]
def __init__(
self,
transform,
# Dataset args
train_path: str,
val_path: str,
test_path: str,
batch_size: int = 2,
val_batch_size: int = 0, # defaults to batch_size
test_batch_size: int = 0, # defaults to val_batch_size
num_positive: int = 1, # currently, like the original paper only 1 is supported
num_negative: int = 7,
neg_ctx_sample: bool = True,
pos_ctx_sample: bool = False,
num_val_negative: int = 7, # num negatives to use in validation
num_test_negative: int = 0, # defaults to num_val_negative
drop_last: bool = False, # drop last batch if len(dataset) not multiple of bs
num_workers: int = 0, # increasing this bugs out right now
use_title: bool = False, # use the title for context passages
sep_token: str = " ", # sep token between title and passage
use_cross_attention: bool = False, # Use cross attention model
*args,
**kwargs,