抱抱脸变压器问题答案置信度得分

Question

如何从huggingface转换器问题答案的样例代码中获取答案置信度分数?我看到管道确实返回分数，但是下面的核心也可以返回置信度分数吗?

from transformers import AutoTokenizer, TFAutoModelForQuestionAnswering
import tensorflow as tf
tokenizer = AutoTokenizer.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
model = TFAutoModelForQuestionAnswering.from_pretrained("bert-large-uncased-whole-word-masking-finetuned-squad")
text = r"""
🤗 Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides general-purpose
architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet…) for Natural Language Understanding (NLU) and Natural
Language Generation (NLG) with over 32+ pretrained models in 100+ languages and deep interoperability between
TensorFlow 2.0 and PyTorch.
"""
questions = [
"How many pretrained models are available in Transformers?",
"What does Transformers provide?",
"Transformers provides interoperability between which frameworks?",
]
for question in questions:
inputs = tokenizer.encode_plus(question, text, add_special_tokens=True, return_tensors="tf")
input_ids = inputs["input_ids"].numpy()[0]
text_tokens = tokenizer.convert_ids_to_tokens(input_ids)
answer_start_scores, answer_end_scores = model(inputs)
answer_start = tf.argmax(
answer_start_scores, axis=1
).numpy()[0]  # Get the most likely beginning of answer with the argmax of the score
answer_end = (
tf.argmax(answer_end_scores, axis=1) + 1
).numpy()[0]  # Get the most likely end of answer with the argmax of the score
answer = tokenizer.convert_tokens_to_string(tokenizer.convert_ids_to_tokens(input_ids[answer_start:answer_end]))
print(f"Question: {question}")
print(f"Answer: {answer}n")

代码从https://huggingface.co/transformers/usage.html

Answer 1

分数只是应用softmax函数后，应答开始令牌应答结束令牌的对数的乘积。请看下面的例子:管道输出:

import torch
from transformers import AutoTokenizer, AutoModelForQuestionAnswering, pipeline
tokenizer = AutoTokenizer.from_pretrained("distilbert-base-cased-distilled-squad")
model = AutoModelForQuestionAnswering.from_pretrained("distilbert-base-cased-distilled-squad")
text = r"""
🤗 Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides general-purpose
architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet…) for Natural Language Understanding (NLU) and Natural
Language Generation (NLG) with over 32+ pretrained models in 100+ languages and deep interoperability between
TensorFlow 2.0 and PyTorch.
"""
question = "How many pretrained models are available in Transformers?"
question_answerer = pipeline("question-answering", model = model, tokenizer= tokenizer)
print(question_answerer(question=question, context = text))

输出:

{'score': 0.5254509449005127, 'start': 256, 'end': 264, 'answer': 'over 32+'}

没有管道:

inputs = tokenizer(question, text, add_special_tokens=True, return_tensors="pt")
outputs = model(**inputs)

首先，我们创建一个掩码，每个上下文令牌为1，否则为0(问题令牌和特殊令牌)。我们使用批编码。Sequence_ids方法:

non_answer_tokens = [x if x in [0,1] else 0 for x in inputs.sequence_ids()]
non_answer_tokens = torch.tensor(non_answer_tokens, dtype=torch.bool)
non_answer_tokens

输出:

tensor([False, False, False, False, False, False, False, False, False, False,
False, False, False,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
True,  True,  True, False])

我们使用这个掩码将特殊令牌和问题令牌的logits设置为负无穷，然后应用softmax(负无穷可以防止这些令牌影响softmax结果):

from torch.nn.functional import softmax
potential_start = torch.where(non_answer_tokens, outputs.start_logits, torch.tensor(float('-inf'),dtype=torch.float))
potential_end = torch.where(non_answer_tokens, outputs.end_logits, torch.tensor(float('-inf'),dtype=torch.float))
potential_start = softmax(potential_start, dim = 1)
potential_end = softmax(potential_end, dim = 1)
potential_start

输出:

tensor([[0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00,
0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00,
0.0000e+00, 1.0567e-04, 9.7031e-05, 1.9445e-06, 1.5849e-06, 1.2075e-07,
3.1704e-08, 4.7796e-06, 1.8712e-07, 6.2977e-08, 1.5481e-07, 8.0004e-08,
3.7896e-07, 1.6438e-07, 9.7762e-08, 1.0898e-05, 1.6518e-07, 5.6349e-08,
2.4848e-07, 2.1459e-07, 1.3785e-06, 1.0386e-07, 1.8803e-07, 8.1887e-08,
4.1088e-07, 1.5618e-07, 2.5624e-06, 1.8526e-06, 2.6710e-06, 6.8466e-08,
1.7953e-07, 3.6242e-07, 2.2788e-07, 2.3384e-06, 1.2147e-05, 1.6065e-07,
3.3257e-07, 2.6021e-07, 2.8140e-06, 1.3698e-07, 1.1066e-07, 2.8436e-06,
1.2171e-07, 9.9341e-07, 1.1684e-07, 6.8935e-08, 5.6335e-08, 1.3314e-07,
1.3038e-07, 7.9560e-07, 1.0671e-07, 9.1864e-08, 5.6394e-07, 3.0210e-08,
7.2176e-08, 5.4452e-08, 1.2873e-07, 9.2636e-08, 9.6012e-07, 7.8008e-08,
1.3124e-07, 1.3680e-06, 8.8716e-07, 8.6627e-07, 6.4750e-06, 2.5951e-07,
6.1648e-07, 8.7724e-07, 1.0796e-05, 2.6633e-07, 5.4644e-07, 1.7553e-07,
1.6015e-05, 5.0054e-07, 8.2263e-07, 2.6336e-06, 2.0743e-05, 4.0008e-07,
1.9330e-06, 2.0312e-04, 6.0256e-01, 3.9638e-01, 3.1568e-04, 2.2009e-05,
1.2485e-06, 2.4744e-06, 1.0092e-05, 3.1047e-06, 1.3597e-04, 1.5105e-06,
1.4960e-06, 8.1164e-08, 1.6534e-06, 4.6181e-07, 8.7354e-08, 2.2356e-07,
9.1145e-07, 8.8194e-06, 4.4202e-07, 1.9238e-07, 2.8077e-07, 1.4117e-05,
2.0613e-07, 1.2676e-06, 8.1317e-08, 2.2337e-06, 1.2399e-07, 6.1745e-08,
3.4725e-08, 2.7878e-07, 4.1457e-07, 0.0000e+00]],
grad_fn=<SoftmaxBackward>)

这些概率现在可以用来提取答案的开始和结束令牌，并计算答案分数:

answer_start = torch.argmax(potential_start)
answer_end = torch.argmax(potential_end)
answer = tokenizer.decode(inputs.input_ids.squeeze()[answer_start:answer_end+1])
print(potential_start.squeeze()[answer_start])
print(potential_end.squeeze()[answer_end])
print(potential_start.squeeze()[answer_start] *potential_end.squeeze()[answer_end])
print(answer)

输出:

tensor(0.6026, grad_fn=<SelectBackward>)
tensor(0.8720, grad_fn=<SelectBackward>)
tensor(0.5255, grad_fn=<MulBackward0>)
over 32 +

注::请记住，这个答案不包括任何特殊情况(结束令牌在开始令牌之前)。