解释一个问答变压器模型

在这里，我们演示如何解释一个问答模型的输出，该模型预测上下文文本的哪个范围包含给定问题的答案。

[1]:

import numpy as np
import torch
import transformers

import shap

# load the model
pmodel = transformers.pipeline("question-answering")
tokenized_qs = None  # variable to store the tokenized data


# define two predictions, one that outputs the logits for the range start,
# and the other for the range end
def f(questions, tokenized_qs, start):
    outs = []
    for q in questions:
        idx = np.argwhere(
            np.array(tokenized_qs["input_ids"]) == pmodel.tokenizer.sep_token_id
        )[0, 0]  # this code assumes that there is only one sentence in data
        d = tokenized_qs.copy()
        d["input_ids"][:idx] = q[:idx]
        d["input_ids"][idx + 1 :] = q[idx + 1 :]
        out = pmodel.model.forward(**{k: torch.tensor(d[k]).reshape(1, -1) for k in d})
        logits = out.start_logits if start else out.end_logits
        outs.append(logits.reshape(-1).detach().numpy())
    return outs


def tokenize_data(data):
    for q in data:
        question, context = q.split("[SEP]")
        tokenized_data = pmodel.tokenizer(question, context)
    return tokenized_data  # this code assumes that there is only one sentence in data


def f_start(questions):
    return f(questions, tokenized_qs, True)


def f_end(questions):
    return f(questions, tokenized_qs, False)


# attach a dynamic output_names property to the models so we can plot the tokens at each output position
def out_names(inputs):
    question, context = inputs.split("[SEP]")
    d = pmodel.tokenizer(question, context)
    return [pmodel.tokenizer.decode([id]) for id in d["input_ids"]]


f_start.output_names = out_names
f_end.output_names = out_names

解释起始位置

这里我们解释模型的起始范围预测。注意，由于模型输出依赖于模型输入的长度，因此我们传递模型的原生分词器进行掩码处理非常重要，这样当我们隐藏文本的部分内容时，我们可以保留相同数量的标记，从而为每个输出位置保留相同的含义。

[2]:

data = [
    "What is on the table?[SEP]When I got home today I saw my cat on the table, and my frog on the floor.",
]  # this code assumes that there is only one sentence in data
tokenized_qs = tokenize_data(data)

explainer_start = shap.Explainer(
    f_start, shap.maskers.Text(tokenizer=pmodel.tokenizer, output_type="ids")
)
shap_values_start = explainer_start(data)

shap.plots.text(shap_values_start)

Partition explainer: 2it [00:32, 32.86s/it]

[0]

outputs

[CLS]

What

is

on

the

table

?

[SEP]

When

I

got

home

today

I

saw

my

cat

on

the

table

,

and

my

frog

on

the

floor

.

[SEP]

inputs

What

is

on

the

table

?

[SEP]

When

I

got

home

today

I

saw

my

cat

on

the

table

,

and

my

frog

on the floor

.

解释结束位置

这个过程与上面相同，但现在我们解释结束标记。

[3]:

explainer_end = shap.Explainer(f_end, pmodel.tokenizer)
shap_values_end = explainer_end(data)

shap.plots.text(shap_values_end)

[0]

outputs

[CLS]

What

is

on

the

table

?

[SEP]

When

I

got

home

today

I

saw

my

cat

on

the

table

,

and

my

frog

on

the

floor

.

[SEP]

inputs

What

is

on

the

table

?

[SEP]

When I

got home

today I saw my

cat on

the table

,

and

my frog

on the floor

.

解释一个匹配函数

在上面的例子中，我们直接解释了来自模型的输出logits。这要求我们确保只以保持长度的方法扰动输入，以免改变输出logits的含义。一个不那么详细但更灵活的方法是，只需评分模型是否产生了特定的答案。

[4]:

def make_answer_scorer(answers):
    def f(questions):
        out = []
        for q in questions:
            question, context = q.split("[SEP]")
            results = pmodel(question, context, topk=20)
            values = []
            for answer in answers:
                value = 0
                for result in results:
                    if result["answer"] == answer:
                        value = result["score"]
                        break
                values.append(value)
            out.append(values)
        return out

    f.output_names = answers
    return f


f_answers = make_answer_scorer(["my cat", "cat", "my frog"])
explainer_answers = shap.Explainer(f_answers, pmodel.tokenizer)
shap_values_answers = explainer_answers(data)

shap.plots.text(shap_values_answers)

[0]

outputs

my cat

cat

my frog

inputs

What

is

on

the

table

?

[SEP]

When I got home

today I

saw

my

cat

on

the

table

,

and

my frog

on the floor

.

有更多有用示例的想法吗？我们鼓励提交增加此文档笔记本的拉取请求！