-
Notifications
You must be signed in to change notification settings - Fork 0
/
summarize_tokens.py
200 lines (175 loc) · 7.06 KB
/
summarize_tokens.py
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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
"""
Summarization of tokens based on attention values - qualitative!
"""
import json
import glob
import scipy.stats as stats
import numpy as np
import pandas as pd
from collections import defaultdict
import seaborn as sns
import matplotlib.pyplot as plt
import argparse
sns.set_palette("Set2")
from wordcloud import WordCloud
import tqdm
import matplotlib.pyplot as plt
from matplotlib import rc, font_manager
import glob
from scipy import integrate
from fuji_rankings import *
font_size = 11
font_properties = {'family': 'serif', 'serif': ['Computer Modern Roman'],
'weight': 'normal', 'size': font_size}
font_manager.FontProperties(family='Computer Modern Roman', style='normal',
size=font_size, weight='normal', stretch='normal')
rc('text', usetex=False)
rc('font', **font_properties)
sns.set_style("white")
# parser = argparse.ArgumentParser()
# parser.add_argument("--input_json",default="generated_json_inputs/example_bbc_explanations.json")
# args = parser.parse_args()
# input_json = args.input_json
top_k = 25
# all_inputs = ["generated_json_inputs/positive.json",
# "generated_json_inputs/negative.json",
# "generated_json_inputs/neutral.json"]
all_inputs = glob.glob("generated_json_inputs/*")
#names = ["Positive","Negative","Neutral"]
names = [x.split("/")[-1].replace(".json","") for x in all_inputs]
first_dataframe = []
for input_json, fname in zip(all_inputs, names):
head_instances = []
## construct
fx = open(input_json,"rt")
json_file = eval(fx.read())
fx.close()
dobj = defaultdict(list)
for head in json_file:
text = head['text']
sets = head['sets']
for el in sets:
attention_vectors = el['attention_vectors']
token_space = el['tokens']
vid = 0
for el2 in attention_vectors:
vid+=1
vecs = el2['vectors']
hname = el2['name']
vecs = [float(x) for x in vecs]
assert len(vecs) == len(token_space)
for a,b in zip(token_space, vecs):
if "[CLS]" in a:
continue
if "[SEP]" in a:
continue
dobj[a].append(b)
head_row = [a, b, hname]
head_instances.append(head_row)
for token, attention_values in dobj.items():
out_object = {}
out_object['token'] = token
out_object['mean-att'] = np.mean(attention_values)
out_object['max-att'] = np.max(attention_values)
out_object['min-att'] = np.min(attention_values)
out_object['std-att'] = np.std(attention_values)
out_object['dataset'] = fname
first_dataframe.append(out_object)
## analyse headspace
head_df = pd.DataFrame(head_instances)
head_df.columns = ['token','attention','head']
fig = plt.figure(figsize = (6,7))
fig.subplots_adjust(hspace=0.2, wspace=0.2)
attention_distributions = {}
for i in range(1, 11):
print(f"wc, head {i} {fname}")
ax = fig.add_subplot(4, 3, i)
shead_df = head_df[head_df['head'] == i]
tokenspace = shead_df['token'].values
attention_space = shead_df['attention'].values
attention_distributions[i] = np.abs(attention_space)
# vx = np.argsort(attention_space)[::-1][0:int(len(attention_space)/3)]
# tokenspace = tokenspace[vx]
# if len(tokenspace) > 0:
# text = " ".join(tokenspace)
# wordcloud = WordCloud(width = 3000, height = 2000, random_state=1, background_color='white', colormap='coolwarm_r', collocations=False).generate(text)
# plt.title("Head {}".format(i))
# plt.imshow(wordcloud)
# plt.axis("off")
# plt.tight_layout()
# plt.savefig("figures/multihead_{}.png".format(fname), dpi = 300)
plt.clf()
# heatmap_points = []
# for k0,v0 in tqdm.tqdm(attention_distributions.items()):
# for k1,v1 in attention_distributions.items():
# if k0 != k1:
# similarity = compute_similarity(v0, v1, "fuzzy_jaccard")
# heatmap_points.append([k0,k1,similarity[1]])
# heatmap_df = pd.DataFrame(heatmap_points)
# heatmap_df.columns = ['head x','head y','AUFuji']
# with sns.axes_style("white"):
# heatmap_df = heatmap_df.pivot("head x", "head y", "AUFuji")
# mask = np.zeros_like(heatmap_df.values, dtype=np.bool)
# mask[np.triu_indices_from(mask)] = True
# sns.heatmap(heatmap_df,linewidths=.4, cmap = "coolwarm", mask = mask, cbar_kws={'label': 'AUFuji'})
# plt.tight_layout()
# plt.savefig("figures/heatmap_{}.png".format(fname), dpi = 300)
## plot
first_df = pd.DataFrame(first_dataframe)
for fname in names:
first_subset = first_df[first_df['dataset'] == fname]
first_subset = first_subset.sort_values(by = ["mean-att"], ascending=False)
first_subset = first_subset.iloc[0:top_k]
sns.barplot(first_subset['mean-att'], first_subset.token, palette = "coolwarm_r", yerr = first_subset['std-att'])
sns.barplot(first_subset['max-att'], first_subset.token, palette = "coolwarm_r", alpha = 0.3)
plt.xlabel("Mean attention value (max in the background)")
plt.ylabel("Top tokens")
plt.xticks(rotation=90)
plt.tight_layout()
plt.savefig("figures/{}.png".format(fname), dpi = 300)
plt.clf()
first_df['max-att'] = first_df['max-att'].transform(lambda x: np.log(x))
first_df['mean-att'] = first_df['mean-att'].transform(lambda x: np.log(x))
first_df['min-att'] = first_df['min-att'].transform(lambda x: np.log(x))
## max dist
for fname in names:
subset = first_df[first_df['dataset'] == fname]
sns.distplot(subset['max-att'],label = fname)
plt.legend()
plt.xlabel("Attention value (log)")
plt.ylabel("Density")
plt.tight_layout()
plt.savefig("figures/max_dist{}.png".format(fname), dpi = 300)
plt.clf()
## mean dist
for fname in names:
subset = first_df[first_df['dataset'] == fname]
sns.distplot(subset['mean-att'],label = fname)
plt.legend()
plt.xlabel("Attention value (log)")
plt.ylabel("Density")
plt.tight_layout()
plt.savefig("figures/mean_dist{}.png".format(fname), dpi = 300)
plt.clf()
## statistical analysis
for fname in names:
subset = first_df[first_df['dataset'] == fname]
sns.distplot(subset['min-att'],label = fname)
plt.legend()
plt.xlabel("Attention value (log)")
plt.ylabel("Density")
plt.tight_layout()
plt.savefig("figures/min_dist{}.png".format(fname), dpi = 300)
plt.clf()
stats.probplot(first_df['max-att'], dist="norm", plot=sns.mpl.pyplot)
plt.tight_layout()
plt.savefig("figures/maxqq{}.png".format(fname), dpi = 300)
plt.clf()
stats.probplot(first_df['min-att'], dist="norm", plot=sns.mpl.pyplot)
plt.tight_layout()
plt.savefig("figures/minqq{}.png".format(fname), dpi = 300)
plt.clf()
stats.probplot(first_df['mean-att'], dist="norm", plot=sns.mpl.pyplot)
plt.tight_layout()
plt.savefig("figures/meanqq{}.png".format(fname), dpi = 300)
plt.clf()