135 lines
5.4 KiB
Python
135 lines
5.4 KiB
Python
import os
|
|
from typing import List, Union, Optional
|
|
|
|
from FlagEmbedding.inference.embedder.model_mapping import (
|
|
EmbedderModelClass,
|
|
AUTO_EMBEDDER_MAPPING, EMBEDDER_CLASS_MAPPING
|
|
)
|
|
|
|
from FlagEmbedding import FlagAutoModel
|
|
|
|
from agent import GPTAgent, LLMAgent, LLMInstructAgent
|
|
|
|
prompt_template = """\
|
|
Given a retrieval task and a query, your mission is to generate a brief {answer_type} for the query in the context of the retrieval task.
|
|
Please generate without any explanation.
|
|
|
|
Task: {task}
|
|
|
|
Query: {query}
|
|
|
|
Your output:"""
|
|
|
|
class Reinforced_IR_Model():
|
|
def __init__(
|
|
self,
|
|
model_name_or_path: str,
|
|
model_class: Optional[Union[str, EmbedderModelClass]] = None,
|
|
normalize_embeddings: bool = True,
|
|
use_fp16: bool = True,
|
|
query_instruction_for_retrieval: Optional[str] = None,
|
|
devices: Optional[Union[str, List[str]]] = None,
|
|
pooling_method: Optional[str] = None,
|
|
trust_remote_code: Optional[bool] = None,
|
|
query_instruction_format: Optional[str] = None,
|
|
generator_model_name_or_path: Optional[str] = None,
|
|
temperature: float = 1.0,
|
|
gpu_memory_utilization: float = 0.5,
|
|
tensor_parallel_size: int = None,
|
|
top_p: float = 1.0,
|
|
max_tokens: int = 300,
|
|
api_key: Optional[str] = None,
|
|
base_url: Optional[str] = None,
|
|
model_type: str = "llm_instruct",
|
|
**kwargs,
|
|
):
|
|
self.model_name_or_path = model_name_or_path
|
|
self.model_class = model_class
|
|
self.normalize_embeddings = normalize_embeddings
|
|
self.use_fp16 = use_fp16
|
|
self.query_instruction_for_retrieval = query_instruction_for_retrieval
|
|
self.devices = devices
|
|
self.pooling_method = pooling_method
|
|
self.trust_remote_code = trust_remote_code
|
|
self.query_instruction_format = query_instruction_format
|
|
self.generator_model_name_or_path = generator_model_name_or_path
|
|
self.temperature = temperature
|
|
self.gpu_memory_utilization = gpu_memory_utilization
|
|
self.tensor_parallel_size = tensor_parallel_size
|
|
self.top_p = top_p
|
|
self.max_tokens = max_tokens
|
|
self.model_type = model_type
|
|
self.api_key = api_key
|
|
self.base_url = base_url
|
|
self.kwargs = kwargs
|
|
|
|
self.generator = None
|
|
self.retriever = None
|
|
|
|
def load_retriever(self):
|
|
if self.retriever is None:
|
|
self.retriever = FlagAutoModel.from_finetuned(
|
|
model_name_or_path=self.model_name_or_path,
|
|
model_class=self.model_class,
|
|
normalize_embeddings=self.normalize_embeddings,
|
|
use_fp16=self.use_fp16,
|
|
query_instruction_for_retrieval=self.query_instruction_for_retrieval,
|
|
devices=self.devices,
|
|
pooling_method=self.pooling_method,
|
|
trust_remote_code=self.trust_remote_code,
|
|
query_instruction_format=self.query_instruction_format,
|
|
**self.kwargs,
|
|
)
|
|
self.offload_generator()
|
|
|
|
def load_generator(self):
|
|
if self.generator_model_name_or_path is not None:
|
|
self.offload_retriever()
|
|
if self.generator is None and self.generator_model_name_or_path is not None:
|
|
if self.model_type == 'llm':
|
|
self.generator = LLMAgent(model_name=self.generator_model_name_or_path,
|
|
gpu_memory_utilization=self.gpu_memory_utilization,
|
|
tensor_parallel_size=self.tensor_parallel_size)
|
|
elif self.model_type == 'llm_instruct':
|
|
self.generator = LLMInstructAgent(generate_model_path=self.generator_model_name_or_path,
|
|
gpu_memory_utilization=self.gpu_memory_utilization,
|
|
tensor_parallel_size=self.tensor_parallel_size)
|
|
else:
|
|
self.generator = GPTAgent(model_name=self.generator_model_name_or_path,
|
|
api_key=self.api_key,
|
|
base_url=self.base_url)
|
|
|
|
def offload_retriever(self):
|
|
if self.retriever is not None:
|
|
del self.retriever
|
|
self.retriever = None
|
|
|
|
def offload_generator(self):
|
|
if self.generator is not None:
|
|
del self.generator
|
|
self.generator = None
|
|
|
|
def encode_queries(self, task_instruction, answer_type, queries, **kwargs):
|
|
prompts = [prompt_template.format(
|
|
answer_type=answer_type,
|
|
task=task_instruction,
|
|
query=query
|
|
) for query in queries]
|
|
self.load_generator()
|
|
if self.generator is not None:
|
|
augmented_queries = self.generator.generate(prompts, **kwargs)
|
|
print(augmented_queries)
|
|
augmented_queries = ['Generate the topic about this passage: ' + e for e in augmented_queries]
|
|
self.load_retriever()
|
|
if self.generator is not None:
|
|
return self.retriever.encode_corpus(augmented_queries, **kwargs) * 0.2 + \
|
|
self.retriever.encode_queries(queries, **kwargs) * 0.8
|
|
return self.retriever.encode_queries(queries, **kwargs)
|
|
|
|
def encode_corpus(self, corpus, **kwargs):
|
|
self.load_retriever()
|
|
return self.retriever.encode_corpus(corpus, **kwargs)
|
|
|
|
def encode(self, corpus, **kwargs):
|
|
self.load_retriever()
|
|
return self.retriever.encode(corpus, **kwargs) |