Spaces:
Sleeping
Sleeping
File size: 6,534 Bytes
f0c8e2c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 |
import torch
from transformers import AutoModel, AutoTokenizer
class QAEmbedder:
def __init__(self, model_name="paraphrase-MiniLM-L6-v2"):
"""
Defines a QA embedding model. This is, given a set of questions,
this class returns the corresponding embedding vectors.
Args:
model_name (`str`): Directory containing the necessary tokenizer
and model files.
"""
self.model = None
self.tokenizer = None
self.model_name = model_name
self.set_model(model_name)
def get_model(self, model_name):
"""
Loads a general tokenizer and model using pytorch
'AutoTokenizer' and 'AutoModel'
Args:
model_name (`str`): Directory containing the necessary tokenizer
and model files.
"""
model = AutoModel.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
return model, tokenizer
def set_model(self, model_name):
"""
Sets a general tokenizer and model using the 'self.get_model'
method.
Args:
model_name (`str`): Directory containing the necessary tokenizer
and model files.
"""
self.model, self.tokenizer = self.get_model(self.model_name)
def _mean_pooling(self, model_output, attention_mask):
"""
Internal method that takes a model output and an attention
mask and outputs a mean pooling layer.
Args:
model_output (`torch.Tensor`): output from the QA model
attention_mask (`torch.Tensor`): attention mask defined in the QA tokenizer
Returns:
The averaged tensor.
"""
token_embeddings = model_output[0]
input_mask_expanded = (
attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
)
pool_emb = torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(
input_mask_expanded.sum(1), min=1e-9
)
return pool_emb
def get_embeddings(self, questions, batch=32):
"""
Gets the corresponding embeddings for a set of input 'questions'.
Args:
questions (`list` of `str`): List of strings defining the questions to be embedded
batch (`int`): Performs the embedding job 'batch' questions at a time
Returns:
The embedding vectors.
"""
question_embeddings = []
for i in range(0, len(questions), batch):
# Tokenize sentences
encoded_input = self.tokenizer(
questions[i : i + batch],
padding=True,
truncation=True,
return_tensors="pt",
)
# Compute token embeddings
with torch.no_grad():
model_output = self.model(**encoded_input)
# Perform mean pooling
batch_embeddings = self._mean_pooling(
model_output, encoded_input["attention_mask"]
)
question_embeddings.append(batch_embeddings)
question_embeddings = torch.cat(question_embeddings, dim=0)
return question_embeddings
class QASearcher:
def __init__(self, model_name="paraphrase-MiniLM-L6-v2"):
"""
Defines a QA Search model. This is, given a new question it searches
the most similar questions in a set 'context' and returns both the best
question and associated answer.
Args:
model_name (`str`): Directory containing the necessary tokenizer
and model files.
"""
self.answers = None
self.questions = None
self.question_embeddings = None
self.embedder = QAEmbedder(model_name=model_name)
def set_context_qa(self, questions, answers):
"""
Sets the QA context to be used during search.
Args:
questions (`list` of `str`): List of strings defining the questions to be embedded
answers (`list` of `str`): Best answer for each question in 'questions'
"""
self.answers = answers
self.questions = questions
self.question_embeddings = self.get_q_embeddings(questions)
def get_q_embeddings(self, questions):
"""
Gets the embeddings for the questions in 'context'.
Args:
questions (`list` of `str`): List of strings defining the questions to be embedded
Returns:
The embedding vectors.
"""
question_embeddings = self.embedder.get_embeddings(questions)
question_embeddings = torch.nn.functional.normalize(
question_embeddings, p=2, dim=1
)
return question_embeddings.transpose(0, 1)
def cosine_similarity(self, questions, batch=32):
"""
Gets the cosine similarity between the new questions and the 'context' questions.
Args:
questions (`list` of `str`): List of strings defining the questions to be embedded
batch (`int`): Performs the embedding job 'batch' questions at a time
Returns:
The cosine similarity
"""
question_embeddings = self.embedder.get_embeddings(questions, batch=batch)
question_embeddings = torch.nn.functional.normalize(
question_embeddings, p=2, dim=1
)
cosine_sim = torch.mm(question_embeddings, self.question_embeddings)
return cosine_sim
def get_answers(self, questions, batch=32):
"""
Gets the best answers in the stored 'context' for the given new 'questions'.
Args:
questions (`list` of `str`): List of strings defining the questions to be embedded
batch (`int`): Performs the embedding job 'batch' questions at a time
Returns:
A `list` of `dict`'s containing the original question ('orig_q'), the most similar
question in the context ('best_q') and the associated answer ('best_a').
"""
similarity = self.cosine_similarity(questions, batch=batch)
response = []
for i in range(similarity.shape[0]):
best_ix = similarity[i].argmax()
best_q = self.questions[best_ix]
best_a = self.answers[best_ix]
response.append(
{
"orig_q": questions[i],
"best_q": best_q,
"best_a": best_a,
}
)
return response
|