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, }