def main():
parser = argparse.ArgumentParser()
add_encoder_params(parser)
add_training_params(parser)
add_tokenizer_params(parser)
add_reader_preprocessing_params(parser)
# reader specific params
parser.add_argument("--max_n_answers", default=10, type=int,
help="Max amount of answer spans to marginalize per singe passage")
parser.add_argument('--passages_per_question', type=int, default=2,
help="Total amount of positive and negative passages per question")
parser.add_argument('--passages_per_question_predict', type=int, default=50,
help="Total amount of positive and negative passages per question for evaluation")
parser.add_argument("--max_answer_length", default=10, type=int,
help="The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument('--eval_top_docs', nargs='+', type=int,
help="top retrival passages thresholds to analyze prediction results for")
parser.add_argument('--checkpoint_file_name', type=str, default='dpr_reader')
parser.add_argument('--prediction_results_file', type=str, help='path to a file to write prediction results to')
# training parameters
parser.add_argument("--eval_step", default=2000, type=int,
help="batch steps to run validation and save checkpoint")
parser.add_argument("--output_dir", default=None, type=str,
help="The output directory where the model checkpoints will be written to")
parser.add_argument('--fully_resumable', action='store_true',
help="Enables resumable mode by specifying global step dependent random seed before shuffling "
"in-batch data")
args = parser.parse_args()
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
setup_args_gpu(args)
set_seed(args)
print_args(args)
trainer = ReaderTrainer(args)
if args.train_file is not None:
trainer.run_train()
elif args.dev_file:
logger.info("No train files are specified. Run validation.")
trainer.validate()
else:
logger.warning("Neither train_file or (model_file & dev_file) parameters are specified. Nothing to do.")
def setup_reader(model_file):
global reader
parser = argparse.ArgumentParser()
add_encoder_params(parser)
add_training_params(parser)
add_tokenizer_params(parser)
add_reader_preprocessing_params(parser)
args = parser.parse_args()
setup_args_gpu(args)
set_seed(args)
print_args(args)
reader = Reader(args, model_file)
def main():
parser = argparse.ArgumentParser()
add_encoder_params(parser)
add_training_params(parser)
add_tokenizer_params(parser)
# biencoder specific training features
parser.add_argument(
"--eval_per_epoch",
default=1,
type=int,
help="How many times it evaluates on dev set per epoch and saves a checkpoint",
)
parser.add_argument(
"--global_loss_buf_sz",
type=int,
default=150000,
help='Buffer size for distributed mode representations al gather operation. \
Increase this if you see errors like "encoded data exceeds max_size ..."',
)
parser.add_argument("--fix_ctx_encoder", action="store_true")
parser.add_argument("--shuffle_positive_ctx", action="store_true")
# input/output src params
parser.add_argument(
"--output_dir",
default=None,
type=str,
help="The output directory where the model checkpoints will be written or resumed from",
)
# data handling parameters
parser.add_argument(
"--hard_negatives",
default=1,
type=int,
help="amount of hard negative ctx per question",
)
parser.add_argument(
"--other_negatives",
default=0,
type=int,
help="amount of 'other' negative ctx per question",
)
parser.add_argument(
"--train_files_upsample_rates",
type=str,
help="list of up-sample rates per each train file. Example: [1,2,1]",
)
# parameters for Av.rank validation method
parser.add_argument(
"--val_av_rank_start_epoch",
type=int,
default=10000,
help="Av.rank validation: the epoch from which to enable this validation",
)
parser.add_argument(
"--val_av_rank_hard_neg",
type=int,
default=30,
help="Av.rank validation: how many hard negatives to take from each question pool",
)
parser.add_argument(
"--val_av_rank_other_neg",
type=int,
default=30,
help="Av.rank validation: how many 'other' negatives to take from each question pool",
)
parser.add_argument(
"--val_av_rank_bsz",
type=int,
default=128,
help="Av.rank validation: batch size to process passages",
)
parser.add_argument(
"--val_av_rank_max_qs",
type=int,
default=10000,
help="Av.rank validation: max num of questions",
)
parser.add_argument(
"--checkpoint_file_name",
type=str,
default="dpr_biencoder",
help="Checkpoints file prefix",
)
# My Model specific params
parser.add_argument('--use_linear', default=False, action='store_true')
args = parser.parse_args()
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps
)
)
if args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
setup_args_gpu(args)
set_seed(args)
print_args(args)
trainer = BiEncoderTrainer(args)
if args.train_file is not None:
trainer.run_train()
elif args.model_file and args.dev_file:
logger.info(
"No train files are specified. Run 2 types of validation for specified model file"
)
trainer.validate_nll()
trainer.validate_average_rank()
else:
logger.warning(
"Neither train_file or (model_file & dev_file) parameters are specified. Nothing to do."
)
parser.add_argument('--db_path', type=str, default="/checkpoint/sewonmin/dpr/data/wikipedia_split/psgs_w100_seen_only.tsv")
# retrieval specific params
parser.add_argument('--dense_index_path', type=str, default="")
parser.add_argument('--tfidf_path', type=str, default="/checkpoint/sewonmin/dpr/drqa_retrieval_seen_only/db-tfidf-ngram=2-hash=16777216-tokenizer=simple.npz")
parser.add_argument('--match', type=str, default='string', choices=['regex', 'string'])
parser.add_argument('--n-docs', type=int, default=100)
#parser.add_argument('--batch_size', type=int, default=32, help="Batch size for question encoder forward pass")
parser.add_argument('--index_buffer', type=int, default=50000,
help="Temporal memory data buffer size (in samples) for indexer")
parser.add_argument("--hnsw_index", action='store_true', help='If enabled, use inference time efficient HNSW index')
parser.add_argument("--save_or_load_index", action='store_true', default=True, help='If enabled, save index')
# reader specific params
add_encoder_params(parser)
add_training_params(parser)
add_tokenizer_params(parser)
add_reader_preprocessing_params(parser)
parser.add_argument("--max_n_answers", default=10, type=int,
help="Max amount of answer spans to marginalize per singe passage")
parser.add_argument('--passages_per_question', type=int, default=2,
help="Total amount of positive and negative passages per question")
parser.add_argument('--passages_per_question_predict', type=int, default=50,
help="Total amount of positive and negative passages per question for evaluation")
parser.add_argument("--max_answer_length", default=10, type=int,
help="The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument('--eval_top_docs', nargs='+', type=int,
help="top retrival passages thresholds to analyze prediction results for")
