primeqa.ir.dense.dpr_top.dpr.config.DPRTrainingArguments#

class primeqa.ir.dense.dpr_top.dpr.config.DPRTrainingArguments(output_dir: str, bsize: int = 8, collection: str = '', ctx_encoder_name_or_path: str = 'facebook/dpr-ctx_encoder-multiset-base', force_confict_free_batches: bool = False, encoder_gpu_train_limit: int = 8, learning_rate: float = 2e-05, max_grad_norm: float = 2.0, max_negatives: int = 0, max_hard_negatives: int = 1, epochs: int = 3, positive_pids: str = 'None', qry_encoder_name_or_path: str = 'facebook/dpr-question_encoder-multiset-base', queries: Optional[str] = None, resume_from_checkpoint: str = '', sample_negative_from_top_k: int = 1, train_dir: str = 'None', training_data_type: str = 'None', warmup_instances: int = 0, warmup_fraction: float = 0.0)#

Bases: object

Arguments used in training .. note:

here and in the other DPRXArguments classes:
(1) Some of the argument name (e.g. "bsize") are chosen to match between the training/indexing/search modalities, and also other IR engines in PrimeQA.
(2) The argument names are sorted alphabetically, with the mandatory arguments listed first.

Methods

Attributes

`bsize`
`collection`
`ctx_encoder_name_or_path`
`encoder_gpu_train_limit`
`epochs`
`force_confict_free_batches`
`learning_rate`
`max_grad_norm`
`max_hard_negatives`
`max_negatives`
`positive_pids`
`qry_encoder_name_or_path`
`queries`
`resume_from_checkpoint`
`sample_negative_from_top_k`
`train_dir`
`training_data_type`
`warmup_fraction`
`warmup_instances`
`output_dir`