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('--qa_file', required=True, type=str, default=None,
help="Question and answers file of the format: question \\t ['answer1','answer2', ...]")
parser.add_argument('--ctx_file', required=True, type=str, default=None,
help="All passages file in the tsv format: id \\t passage_text \\t title")
parser.add_argument('--encoded_ctx_file', type=str, default=None,
help='Glob path to encoded passages (from generate_dense_embeddings tool)')
parser.add_argument('--out_file', type=str, default=None,
help='output .json file path to write results to ')
parser.add_argument('--match', type=str, default='string', choices=['regex', 'string'],
help="Answer matching logic type")
parser.add_argument('--n-docs', type=int, default=200,
help="Amount of top docs to return")
parser.add_argument('--validation_workers', type=int, default=16,
help="Number of parallel processes to validate results")
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', help='If enabled, save index')
args = parser.parse_args()
assert args.model_file, 'Please specify --model_file checkpoint to init model weights'
setup_args_gpu(args)
print_args(args)
main(args)
def setup_dpr(model_file,
ctx_file,
encoded_ctx_file,
hnsw_index=False,
save_or_load_index=False):
global retriever
global all_passages
global answer_cache
global answer_cache_path
parameter_setting = model_file + ctx_file + encoded_ctx_file
answer_cache_path = hashlib.sha1(
parameter_setting.encode("utf-8")).hexdigest()
if os.path.exists(answer_cache_path):
answer_cache = pickle.load(open(answer_cache_path, 'rb'))
else:
answer_cache = {}
parser = argparse.ArgumentParser()
add_encoder_params(parser)
add_tokenizer_params(parser)
add_cuda_params(parser)
args = parser.parse_args()
args.model_file = model_file
args.ctx_file = ctx_file
args.encoded_ctx_file = encoded_ctx_file
args.hnsw_index = hnsw_index
args.save_or_load_index = save_or_load_index
args.batch_size = 1 # TODO
setup_args_gpu(args)
print_args(args)
saved_state = load_states_from_checkpoint(args.model_file)
set_encoder_params_from_state(saved_state.encoder_params, args)
tensorizer, encoder, _ = init_biencoder_components(args.encoder_model_type,
args,
inference_only=True)
encoder = encoder.question_model
encoder, _ = setup_for_distributed_mode(encoder, None, args.device,
args.n_gpu, args.local_rank,
args.fp16)
encoder.eval()
# load weights from the model file
model_to_load = get_model_obj(encoder)
logger.info("Loading saved model state ...")
prefix_len = len("question_model.")
question_encoder_state = {
key[prefix_len:]: value
for (key, value) in saved_state.model_dict.items()
if key.startswith("question_model.")
}
model_to_load.load_state_dict(question_encoder_state)
vector_size = model_to_load.get_out_size()
logger.info("Encoder vector_size=%d", vector_size)
if args.hnsw_index:
index = DenseHNSWFlatIndexer(vector_size, 50000)
else:
index = DenseFlatIndexer(vector_size, 50000,
"IVF65536,PQ64") #IVF65536
retriever = DenseRetriever(encoder, args.batch_size, tensorizer, index)
# index all passages
ctx_files_pattern = args.encoded_ctx_file
input_paths = glob.glob(ctx_files_pattern)
index_path = "_".join(input_paths[0].split("_")[:-1])
if args.save_or_load_index and (os.path.exists(index_path) or
os.path.exists(index_path + ".index.dpr")):
retriever.index.deserialize_from(index_path)
else:
logger.info("Reading all passages data from files: %s", input_paths)
retriever.index.index_data(input_paths)
if args.save_or_load_index:
retriever.index.serialize(index_path)
# get questions & answers
all_passages = load_passages(args.ctx_file)
