File size: 7,601 Bytes
1bb8d13 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 |
import json
from random import uniform, seed
from statistics import mean
import pandas as pd
from datasets import load_dataset
from scipy.stats import spearmanr
with pd.option_context("max_colwidth", 1000):
# baselines
target = {
"flan-t5-xxl": "Flan-T5\textsubscript{XXL}",
"opt-13b": "OPT\textsubscript{13B}",
"davinci": "GPT-3\textsubscript{davinci}"
}
pretty_name = {
'is competitor/rival of': "Rival",
'is friend/ally of': "Ally",
'is influenced by': "Inf",
'is known for': "Know",
'is similar to': "Sim",
'average': "Avg",
}
# def get_iaa(scores_all):
# avg = [[mean(__s for _m, __s in enumerate(_s) if _m != _n) for _s in scores_all] for _n in range(7)]
# single = [[_s[_n] for _s in scores_all] for _n in range(7)]
# tmptmp = []
# ps = []
# for a, s in zip(avg, single):
# c = round(pd.DataFrame([a, s]).T.corr("spearman").values[0][1] * 100, 1)
# ps.append(spearmanr(a, s)[1] < 0.05)
# # if str(c) == "nan":
# # seed(0)
# # c_tmp = []
# # for _ in range(1000):
# # s_tmp = [_s + uniform(-0.5, 0.5) for _s in s]
# # c_tmp.append(round(pd.DataFrame([a, s_tmp]).T.corr("spearman").values[0][1] * 100, 1))
# # c = mean(c_tmp)
# tmptmp.append(c)
# list(zip(tmptmp, ps))
# return mean(tmptmp)
def format_text(_x, _y, _z):
bf = max(_x, _y, _z)
wf = str(min(_x, _y, _z))
_x = "\textcolor{blue}{" + str(_x) + "}" if _x == bf else str(_x)
_y = "\textcolor{blue}{" + str(_y) + "}" if _y == bf else str(_y)
_z = "\textcolor{blue}{" + str(_z) + "}" if _z == bf else str(_z)
_x = "\textcolor{red}{" + str(_x) + "}" if _x == wf else str(_x)
_y = "\textcolor{red}{" + str(_y) + "}" if _y == wf else str(_y)
_z = "\textcolor{red}{" + str(_z) + "}" if _z == wf else str(_z)
return f"{_x} / {_y} / {_z}"
data = load_dataset("cardiffnlp/relentless_full", split="test")
table_full = []
for prompt in ['qa', 'lc']:
output = []
for d in data:
for i in target.keys():
with open(f"experiments/results/lm_{prompt}/{i}/ppl.{d['relation_type'].replace(' ', '_').replace('/', '__')}.jsonl") as f:
ppl = [json.loads(x)['perplexity'] for x in f.read().split("\n") if len(x) > 0]
rank_map = {p: n for n, p in enumerate(sorted(ppl), 1)}
prediction = [rank_map[p] for p in ppl]
# get index
total_n = len(d['ranks'])
p = int(total_n/3)
top_n = [0, int(total_n * p / 100) + 1]
top_label = [x for x, _ in sorted(enumerate(d['ranks']), key=lambda x: x[1])[top_n[0]: top_n[1]]]
bottom_n = [total_n - int(total_n * p / 100), total_n]
bottom_label = [x for x, _ in sorted(enumerate(d['ranks']), key=lambda x: x[1])[bottom_n[0]: bottom_n[1]]]
mid_n = [top_n[1], bottom_n[0]]
mid_label = [x for x, _ in sorted(enumerate(d['ranks']), key=lambda x: x[1])[mid_n[0]: mid_n[1]]]
# top
top_pred = [x for x, _ in sorted(enumerate(prediction), key=lambda x: x[1])[top_n[0]: top_n[1]]]
top_acc = len(set(top_pred).intersection(set(top_label))) / len(top_label) * 100
# middle
mid_pred = [x for x, _ in sorted(enumerate(prediction), key=lambda x: x[1])[mid_n[0]: mid_n[1]]]
mid_acc = len(set(mid_pred).intersection(set(mid_label))) / len(mid_label) * 100
# top
bottom_pred = [x for x, _ in sorted(enumerate(prediction), key=lambda x: x[1])[bottom_n[0]: bottom_n[1]]]
bottom_acc = len(set(bottom_pred).intersection(set(bottom_label))) / len(bottom_label) * 100
output.append({"model": i, "relation_type": d['relation_type'], "top": round(top_acc, 1), "bottom": round(bottom_acc, 1), "middle": round(mid_acc, 1)})
for i in target.keys():
output.append({
"model": i, "relation_type": "average",
"top": round(mean([o['top'] for o in output if o['model'] == i]), 0),
"bottom": round(mean([o['bottom'] for o in output if o['model'] == i]), 0),
"middle": round(mean([o['middle'] for o in output if o['model'] == i]), 0)
})
df = pd.DataFrame(output)
df['accuracy'] = [format_text(x, y, z) for x, y, z in zip(df['top'], df['middle'], df['bottom'])]
table = df.pivot(index="relation_type", columns="model", values="accuracy")
table.columns.name = None
table.index.name = None
table = table[target.keys()]
table.columns = [target[i] for i in table.columns]
table.index = [pretty_name[i] for i in table.index]
table = table.T[list(pretty_name.values())]
table = table.T
table = table.to_latex(escape=False)
table = table.split(r"\midrule")[1].split(r"\bottomrule")[0]
table = r"\multicolumn{4}{l}{\emph{" + prompt.upper() + r" template}} \\ " + table
table_full.append(table)
table_full = "\midrule".join(table_full)
#
# output = []
# top_all = []
# mid_all = []
# bottom_all = []
#
# for d in data:
# if d['relation_type'] == "is influenced by":
# break
# total_n = len(d['ranks'])
# p = int(total_n / 3)
# top_n = [0, int(total_n * p / 100) + 1]
# top_label = [x for x, _ in sorted(enumerate(d['ranks']), key=lambda x: x[1])[top_n[0]: top_n[1]]]
# bottom_n = [total_n - int(total_n * p / 100), total_n]
# bottom_label = [x for x, _ in sorted(enumerate(d['ranks']), key=lambda x: x[1])[bottom_n[0]: bottom_n[1]]]
# mid_n = [top_n[1], bottom_n[0]]
# mid_label = [x for x, _ in sorted(enumerate(d['ranks']), key=lambda x: x[1])[mid_n[0]: mid_n[1]]]
#
# output.append({
# "model": "IAA", "relation_type": d['relation_type'],
# "top": round(get_iaa([d['scores_all'][_i] for _i in top_label]), 1),
# "bottom": round(get_iaa([d['scores_all'][_i] for _i in mid_label]), 1),
# "middle": round(get_iaa([d['scores_all'][_i] for _i in bottom_label]), 1)
# })
# top_all += [d['scores_all'][_i] for _i in top_label]
# mid_all += [d['scores_all'][_i] for _i in mid_label]
# bottom_all += [d['scores_all'][_i] for _i in bottom_label]
# output.append({
# "model": "IAA", "relation_type": "average",
# "top": round(get_iaa(top_all), 1),
# "bottom": round(get_iaa(mid_all), 1),
# "middle": round(get_iaa(bottom_all), 1)
# })
#
# df = pd.DataFrame(output)
# df['accuracy'] = [format_text(x, y, z) for x, y, z in zip(df['top'], df['middle'], df['bottom'])]
# table = df.pivot(index="relation_type", columns="model", values="accuracy")
# table.columns.name = None
# table.index.name = None
# table.index = [pretty_name[i] for i in table.index]
# table = table.T[list(pretty_name.values())]
# table = table.to_latex(escape=False)
# table = table.split(r"\midrule")[1].split(r"\bottomrule")[0]
# # table = r"\multicolumn{4}{l}{\emph{" + prompt.upper() + r" template}} \\ " + table
# table_full = table_full + table
print()
print()
print(table_full) |