xaiunits/demo/examples/text.py at main · XAI-Units/xaiunits · GitHub

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
# %%
from xaiunits.datagenerator import TextTriggerDataset
from xaiunits.pipeline import Pipeline, Results
from xaiunits.methods import wrap_method_llm
from captum.attr import (
    KernelShap,
    FeatureAblation,
    ShapleyValueSampling,
    LayerIntegratedGradients,
    Lime,
)
import pandas as pd
from transformers.generation import GenerationConfig


def llm_attribute_input_generator(feature_inputs, y_labels, target, context, model):
    attribute_input = {}
    if target is not None:
        attribute_input["target"] = y_labels

    attribute_input["not_used"] = None
    attribute_input["gen_args"] = {
        "generation_config": GenerationConfig(pad_token_id=tokenizer.pad_token_id)
    }

    return attribute_input


suffix = ""
save = False
# Inference Only

if __name__ == "__main__":
    pass
    # %%
    dataset = TextTriggerDataset(
        (0, 1000),
        # model_name= "XAIUnits/TriggerLLM_Deterministic" # uncomment to get TriggerLLM_Deterministic
    )
    model, tokenizer = dataset.generate_model()

    methods = [
        wrap_method_llm(
            ShapleyValueSampling,
            tokenizer=tokenizer,
            input_generator_fns=llm_attribute_input_generator,
            other_inputs={"n_samples": 100},
        ),
        wrap_method_llm(
            FeatureAblation,
            tokenizer=tokenizer,
            input_generator_fns=llm_attribute_input_generator,
        ),
        wrap_method_llm(
            Lime,
            tokenizer=tokenizer,
            input_generator_fns=llm_attribute_input_generator,
            other_inputs={"n_samples": 100},
        ),
        wrap_method_llm(
            KernelShap,
            tokenizer=tokenizer,
            input_generator_fns=llm_attribute_input_generator,
            other_inputs={"n_samples": 100},
        ),
        wrap_method_llm(
            LayerIntegratedGradients,
            tokenizer=tokenizer,
            input_generator_fns=llm_attribute_input_generator,
            class_params={"layer": model.model.embed_tokens},
            other_inputs={"n_steps": 50, "internal_batch_size": 1},
        ),
    ]

    results = Results()

    Pipeline(
        models=model,
        datas=dataset,
        methods=methods,
        metrics=dataset.default_metric,
        batch_size=1,  # TextTokenInput (from captum) only support a single prompt, hence batch size 1
        method_seeds=[0],
        results=results,
        default_target="y_labels",
    ).run("cuda")

    # %%
    suffix = dataset.model_name if suffix == "" else suffix
    # print results
    try:
        results.print_stats()
    except:
        pass

    # Save results
    df = results.data
    if save:
        df.to_csv(
            f"results/results_text{suffix}.csv", sep=",", index=False, encoding="utf-8"
        )

    # %%
    latex = True
    if save:
        df = pd.read_csv(f"results/results_text{suffix}.csv", sep=",")
    # print(df)

    pd.options.display.float_format = "{:,.3f}".format
    pd.options.display.max_columns = 100

    metric = "attr_time"

    # print(df)
    mu = pd.pivot_table(
        df,
        index=["method"],
        values=[metric],
        aggfunc="mean",
    )

    std = pd.pivot_table(
        df,
        index=["method"],
        values=[metric],
        aggfunc="std",
    )

    combined = pd.pivot_table(
        pd.concat([mu, std], axis=0),
        index=["method"],
        values=[metric],
        aggfunc=[lambda col: " ± ".join([f"{r:.3f}" for r in col])],
    )

    if latex:
        combined = combined.to_latex()

    print(combined.replace("wrapper_", ""))

# %%