File size: 3,388 Bytes
f4a1b77
38c1d39
3b1a0aa
 
 
f4a1b77
 
3b1a0aa
f4a1b77
 
 
3b1a0aa
 
 
27717dd
2a965c2
f4a1b77
 
3b1a0aa
 
 
 
 
 
 
 
 
2a965c2
3b1a0aa
2a965c2
3b1a0aa
 
 
 
 
 
e9d81ac
3b1a0aa
 
 
 
 
 
e9d81ac
3b1a0aa
 
 
 
1b9dc66
e9d81ac
3b1a0aa
 
1b9dc66
e9d81ac
3b1a0aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f4a1b77
55d104b
 
 
 
 
8e10efe
 
 
55d104b
 
 
 
 
 
 
 
 
 
 
 
8e10efe
 
f4a1b77
3b1a0aa
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
import torch
import datasets
import gradio

from transformers import GPT2LMHeadModel, GPT2TokenizerFast


class CrowSPairsDataset(object):
    def __init__(self):
        super().__init__()

        self.df = (datasets
                .load_dataset("BigScienceBiasEval/crows_pairs_multilingual")["test"]
                .to_pandas()
                .query('stereo_antistereo == "stereo"')
                .drop(columns="stereo_antistereo")
            )

    def sample(self, bias_type, n=10):
        return self.df[self.df["bias_type"] == bias_type].sample(n=n)

    def bias_types(self):
        return self.df.bias_type.unique().tolist()


def run(bias_type):
    sample = dataset.sample(bias_type)
    result = "<table><tr style='color: white; background-color: #555'><th>index</th><th>more stereotypical</th><th>less stereotypical<th></tr>"
    for i, row in sample.iterrows():
        result += f"<tr><td>{i}</td>"
        more = row["sent_more"]

        more = tokenizer(more, return_tensors="pt")["input_ids"].to(device)
        with torch.no_grad():
            out_more = model(more, labels=more.clone())
            score_more = out_more["loss"]
            perplexity_more = torch.exp(score_more).item()

        less = row["sent_less"]
        less = tokenizer(less, return_tensors="pt")["input_ids"].to(device)
        with torch.no_grad():
            out_less = model(less, labels=less.clone())
            score_less = out_less["loss"]
            perplexity_less = torch.exp(score_less).item()
            if perplexity_more > perplexity_less:
                shade = round(
                    abs((perplexity_more - perplexity_less) / perplexity_more), 2
                )
                shade = (shade + 0.2) / 1.2
                result += f"<td style='padding: 0 1em;)'>{row['sent_more']}</td><td style='padding: 0 1em; background-color: rgba(255,0,255,{shade})'>{row['sent_less']}</td></tr>"
            else:
                shade = abs((perplexity_less - perplexity_more) / perplexity_less)
                shade = (shade + 0.2) / 1.2
                result += f"<td style='padding: 0 1em; background-color: rgba(0,255,255,{shade})'>{row['sent_more']}</td><td style='padding: 0 1em;'>{row['sent_less']}</td></tr>"
    result += "</table>"
    return result


if torch.cuda.is_available():
    device = torch.device("cuda")
else:
    device = torch.device("cpu")

model_id = "gpt2"
model = GPT2LMHeadModel.from_pretrained(model_id).to(device)
tokenizer = GPT2TokenizerFast.from_pretrained(model_id)
dataset = CrowSPairsDataset()

bias_type_sel = gradio.Dropdown(label="Bias Type", choices=dataset.bias_types())

with open("description.md") as fh:
    desc = fh.read()

with open("notice.md") as fh:
    notice = fh.read()
    
with open("results.md") as fh:
    results = fh.read()

with gradio.Blocks() as iface:
    gradio.Markdown(desc)
    with gradio.Row(equal_height=True):
        with gradio.Column(scale=4):
            inp = gradio.Dropdown(label="Bias Type", choices=dataset.bias_types())
        with gradio.Column(scale=1):
            but = gradio.Button("Sample")
    out = gradio.HTML()
    but.click(run, inp, out)
    with gradio.Accordion("A note about explainability models"):
        gradio.Markdown(notice)
    with gradio.Accordion("Results for English and French BERT language models"):
        gradio.Markdown(results)

iface.launch()