def config_and_inputs(self):
question_encoder_tester = DPRModelTester(self)
dpr_config_and_inputs = question_encoder_tester.prepare_config_and_inputs(
)
generator_tester = T5ModelTester(self, vocab_size=1100, n_positions=30)
t5_config_and_inputs = generator_tester.prepare_config_and_inputs()
(question_encoder_config, input_ids, _, input_mask, _, _,
_) = dpr_config_and_inputs
(generator_config, _, decoder_input_ids, _, decoder_attention_mask,
_) = t5_config_and_inputs
config = RagConfig.from_question_encoder_generator_configs(
question_encoder_config,
generator_config,
n_docs=self.n_docs,
retrieval_vector_size=self.retrieval_vector_size,
max_combined_length=self.max_combined_length,
use_cache=False,
)
return {
"config": config,
"input_ids": input_ids,
"attention_mask": input_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
}
def get_rag_config(self):
question_encoder_config = AutoConfig.from_pretrained(
"facebook/dpr-question_encoder-single-nq-base")
generator_config = AutoConfig.from_pretrained(
"facebook/bart-large-cnn")
return RagConfig.from_question_encoder_generator_configs(
question_encoder_config,
generator_config,
bos_token_id=0,
decoder_start_token_id=2,
eos_token_id=2,
is_encoder_decoder=True,
pad_token_id=1,
vocab_size=50264,
title_sep=" / ",
doc_sep=" // ",
n_docs=5,
max_combined_length=300,
dataset="wiki_dpr",
dataset_split="train",
index_name="exact",
index_path=None,
use_dummy_dataset=True,
retrieval_vector_size=768,
retrieval_batch_size=8,
)
def config_and_inputs(self):
question_encoder_tester = DPRModelTester(self)
dpr_config_and_inputs = question_encoder_tester.prepare_config_and_inputs(
)
generator_tester = BartModelTester(self)
bart_config_and_inputs = generator_tester.prepare_config_and_inputs_for_common(
)
(question_encoder_config, input_ids, _, input_mask, _, _,
_) = dpr_config_and_inputs
(generator_config, bart_inputs_dict) = bart_config_and_inputs
decoder_input_ids, decoder_attention_mask = bart_inputs_dict[
"input_ids"], bart_inputs_dict["attention_mask"]
config = RagConfig.from_question_encoder_generator_configs(
question_encoder_config,
generator_config,
n_docs=self.n_docs,
retrieval_vector_size=self.retrieval_vector_size,
max_combined_length=self.max_combined_length,
use_cache=False,
)
return {
"config": config,
"input_ids": input_ids,
"attention_mask": input_mask,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": decoder_attention_mask,
}
