开放式 GPT2 文本生成解释
本笔记本演示了如何为用于开放式文本生成的 gpt2 输出获取解释。在这个演示中，我们使用 hugging face 提供的预训练 gpt2 模型（https://huggingface.co/gpt2）来解释 gpt2 生成的文本。我们进一步展示了如何为自定义生成的文本输出获取解释，并绘制任何输出生成令牌的全局输入令牌重要性。
[1]:
from transformers import AutoModelForCausalLM, AutoTokenizer
import shap
加载模型和分词器
[2]:
tokenizer = AutoTokenizer.from_pretrained("gpt2", use_fast=True)
model = AutoModelForCausalLM.from_pretrained("gpt2").cuda()
下面，我们设置某些模型配置。我们需要定义模型是解码器还是编码器-解码器。这可以通过模型配置文件中的 ‘is_decoder’ 或 ‘is_encoder_decoder’ 参数来设置。我们还可以设置自定义的模型生成参数，这些参数将在输出文本生成解码过程中使用。
[3]:
# set model decoder to true
model.config.is_decoder = True
# set text-generation params under task_specific_params
model.config.task_specific_params["text-generation"] = {
"do_sample": True,
"max_length": 50,
"temperature": 0.7,
"top_k": 50,
"no_repeat_ngram_size": 2,
}
定义初始文本
[4]:
s = ["I enjoy walking with my cute dog"]
创建一个解释器对象并计算SHAP值
[5]:
explainer = shap.Explainer(model, tokenizer)
shap_values = explainer(s)
Setting `pad_token_id` to `eos_token_id`:50256 for open-end generation.
可视化 shap 解释
[6]:
shap.plots.text(shap_values)
[0]
outputs
,
but
I
'm
not
sure
if
I
'll
ever
be
able
to
inputs
I
enjoy
walking
with
my
cute
dog
另一个例子…
[7]:
s = [
"Scientists confirmed the worst possible outcome: the massive asteroid will collide with Earth"
]
[8]:
explainer = shap.Explainer(model, tokenizer)
shap_values = explainer(s)
Setting `pad_token_id` to `eos_token_id`:50256 for open-end generation.
[9]:
shap.plots.text(shap_values)
[0]
outputs
in
the
coming
days
.
inputs
Scientists
confirmed
the
worst
possible
outcome
:
the
massive
asteroid
will
collide
with
Earth
自定义文本生成和调试偏差输出
下面我们演示如何使用模型解释在给定输入句子的情况下生成特定输出句子的可能性。例如，我们提出一个问题：在句子“I know many people who are [target].”中，哪个国家的居民（目标）在输出句子“They love their vodka!”中生成标记“vodka”的可能性较高？为此，我们首先定义输入-输出句子对
[10]:
# define input
x = [
"I know many people who are Russian.",
"I know many people who are Greek.",
"I know many people who are Australian.",
"I know many people who are American.",
"I know many people who are Italian.",
"I know many people who are Spanish.",
"I know many people who are German.",
"I know many people who are Indian.",
]
[11]:
# define output
y = [
"They love their vodka!",
"They love their vodka!",
"They love their vodka!",
"They love their vodka!",
"They love their vodka!",
"They love their vodka!",
"They love their vodka!",
"They love their vodka!",
]
我们将模型包装在一个 Teacher Forcing 评分类中，并创建一个文本掩码器
[12]:
teacher_forcing_model = shap.models.TeacherForcing(model, tokenizer)
masker = shap.maskers.Text(tokenizer, mask_token="...", collapse_mask_token=True)
创建一个解释器…
[13]:
explainer = shap.Explainer(teacher_forcing_model, masker)
生成 SHAP 解释值！
[14]:
shap_values = explainer(x, y)
既然我们已经生成了SHAP值，我们可以通过文本图查看输入中的标记对输出句子中标记“vodka”的贡献。注意：红色表示正贡献，蓝色表示负贡献，颜色的强度显示了其在相应方向上的强度。
[15]:
shap.plots.text(shap_values)
[0]
outputs
They
love
their
vodka
!
inputs
I
know
many
people
who
are
Russian
.
[1]
outputs
They
love
their
vodka
!
inputs
I
know
many
people
who
are
Greek
.
[2]
outputs
They
love
their
vodka
!
inputs
I
know
many
people
who
are
Australian
.
[3]
outputs
They
love
their
vodka
!
inputs
I
know
many
people
who
are
American
.
[4]
outputs
They
love
their
vodka
!
inputs
I
know
many
people
who
are
Italian
.
[5]
outputs
They
love
their
vodka
!
inputs
I
know
many
people
who
are
Spanish
.
[6]
outputs
They
love
their
vodka
!
inputs
I
know
many
people
who
are
German
.
[7]
outputs
They
love
their
vodka
!
inputs
I
know
many
people
who
are
Indian
.
为了查看哪些输入令牌影响（正面/负面）生成单词“vodka”的可能性，我们绘制了单词“vodka”的全局令牌重要性。
瞧！俄罗斯人喜欢他们的伏特加，不是吗？ :)
[16]:
shap.plots.bar(shap_values[0, :, "vodka"])
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！