-
Notifications
You must be signed in to change notification settings - Fork 17
/
attention_figures1.py
301 lines (281 loc) · 13.9 KB
/
attention_figures1.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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
"""
Creates figures for attention intervention analysis from JSON files:
- Stacked bar chart with direct/indirect/total effects
- Heatmap for head-level effects
- Barplot for layer-level effects
- Combined heatmap/barplot for head- and layer-level effects
"""
import json
import os
import numpy as np
import pandas as pd
import seaborn as sns
from matplotlib import pyplot as plt
from matplotlib.colors import LinearSegmentedColormap
from mpl_toolkits.axes_grid1.axes_divider import make_axes_locatable
from mpl_toolkits.axes_grid1.colorbar import colorbar
from attention_utils import topk_indices
def save_figures(data, source, model_version, filter, suffix, k=10):
# Load data from json obj
results = data['results']
df = pd.DataFrame(results)
# Aggregate by head
# Convert column to 3d ndarray (num_examples x num_layers x num_heads)
indirect_by_head = np.stack(df['indirect_effect_head'].to_numpy())
direct_by_head = np.stack(df['direct_effect_head'].to_numpy())
# Average by head
mean_indirect_by_head = indirect_by_head.mean(axis=0)
mean_direct_by_head = direct_by_head.mean(axis=0)
# Select top k heads by indirect effect
topk_inds = topk_indices(mean_indirect_by_head, k)
# Aggregate by layer
# Convert column to 2d ndarray (num_examples x num_layers)
indirect_by_layer = np.stack(df['indirect_effect_layer'].to_numpy())
direct_by_layer = np.stack(df['direct_effect_layer'].to_numpy())
mean_indirect_by_layer = indirect_by_layer.mean(axis=0)
mean_direct_by_layer = direct_by_layer.mean(axis=0)
n_layers = indirect_by_layer.shape[1]
plt.rc('figure', titlesize=20)
# Plot stacked bar chart
palette = sns.color_palette()#('muted')
plt.figure(num=1, figsize=(5, 2))
topk_direct = []
topk_indirect = []
labels = []
for ind in topk_inds:
layer, head = np.unravel_index(ind, mean_indirect_by_head.shape)
topk_indirect.append(mean_indirect_by_head[layer, head])
topk_direct.append(mean_direct_by_head[layer, head])
labels.append(f'{layer}-{head}')
width = 0.6
inds = range(k)
p1 = plt.bar(inds, topk_indirect, width, linewidth=0, color=palette[1])
p2 = plt.bar(inds, topk_direct, width, bottom=topk_indirect, linewidth=0, color=palette[0])
plt.ylabel('Effect', size=11)
plt.title('Effects of top heads', fontsize=11)
plt.xticks(inds, labels, size=10)
plt.yticks(size=10)
p3 = plt.axhline(data['mean_total_effect'], linestyle='--')
plt.legend((p3, p2[0], p1[0]), ('Total', 'Direct', 'Indirect'), loc='upper right', fontsize=11,
bbox_to_anchor=(.99, 0.90))
sns.despine()
path = 'results/attention_intervention/stacked_bar_charts'
if not os.path.exists(path):
os.makedirs(path)
plt.savefig(f'{path}/{source}_{model_version}_{filter}_'
f'{suffix}.pdf', format='pdf')
plt.close()
annot = False
# Plot heatmap for direct and indirect effect
for effect_type in ('indirect', 'direct'):
if effect_type == 'indirect':
mean_effect = mean_indirect_by_head
else:
mean_effect = mean_direct_by_head
ax = sns.heatmap(mean_effect, rasterized=True, annot=annot, annot_kws={"size": 9}, fmt=".2f", square=True)
ax.set(xlabel='Head', ylabel='Layer', title=f'Mean {effect_type.capitalize()} Effect')
plt.figure(num=1, figsize=(7, 5))
path = f'results/attention_intervention/heat_maps_{effect_type}'
if not os.path.exists(path):
os.makedirs(path)
plt.savefig(f'{path}/{source}_{model_version}_{filter}_'
f'{suffix}.pdf', format='pdf')
plt.close()
# Plot layer-level bar chart for indirect and direct effects
for effect_type in ('indirect', 'direct'):
if effect_type == 'indirect':
mean_effect = mean_indirect_by_layer
else:
mean_effect = mean_direct_by_layer
plt.figure(num=1, figsize=(5, 5))
ax = sns.barplot(x=mean_effect, y=list(range(n_layers)), orient="h", color="#4472C4")
ax.set(ylabel='Layer', title=f'Mean {effect_type.capitalize()} Effect')
path = f'results/attention_intervention/layer_{effect_type}'
if not os.path.exists(path):
os.makedirs(path)
plt.savefig(f'{path}/{source}_{model_version}_{filter}_'
f'{suffix}.pdf', format='pdf')
plt.close()
# Plot combined heatmap and barchart for direct and indirect effects
for do_sort in False, True:
for effect_type in ('indirect', 'direct'):
if effect_type == 'indirect':
effect_head = mean_indirect_by_head
effect_layer = mean_indirect_by_layer
if do_sort:
effect_head = -np.sort(-effect_head) # Sort indirect effects within each layer in descending order
else:
if do_sort:
continue
effect_head = mean_direct_by_head
effect_layer = mean_direct_by_layer
fig = plt.figure(figsize=(3, 2.2))
if model_version == 'distilgpt2':
ax1 = plt.subplot2grid((100, 85), (0, 0), colspan=62, rowspan=99)
ax2 = plt.subplot2grid((100, 85), (32, 69), colspan=17, rowspan=35)
elif model_version in ('gpt2', 'gpt2_random'):
ax1 = plt.subplot2grid((100, 85), (0, 0), colspan=65, rowspan=99)
ax2 = plt.subplot2grid((100, 85), (12, 70), colspan=15, rowspan=75)
elif model_version == 'gpt2-medium':
ax1 = plt.subplot2grid((100, 85), (0, 5), colspan=55, rowspan=99)
ax2 = plt.subplot2grid((100, 85), (2, 64), colspan=17, rowspan=95)
elif model_version == 'gpt2-large':
ax1 = plt.subplot2grid((100, 85), (0, 5), colspan=55, rowspan=96)
ax2 = plt.subplot2grid((100, 85), (0, 62), colspan=17, rowspan=97)
elif model_version == 'gpt2-xl':
ax1 = plt.subplot2grid((100, 85), (0, 5), colspan=55, rowspan=96)
ax2 = plt.subplot2grid((100, 85), (0, 62), colspan=17, rowspan=97)
elif model_version == 'transfo-xl-wt103':
ax1 = plt.subplot2grid((100, 85), (0, 5), colspan=55, rowspan=96)
ax2 = plt.subplot2grid((100, 85), (12, 64), colspan=17, rowspan=72)
elif model_version == 'xlnet-base-cased':
ax1 = plt.subplot2grid((100, 85), (0, 0), colspan=65, rowspan=99)
ax2 = plt.subplot2grid((100, 85), (12, 70), colspan=15, rowspan=75)
elif model_version == 'xlnet-large-cased':
ax1 = plt.subplot2grid((100, 85), (0, 5), colspan=55, rowspan=99)
ax2 = plt.subplot2grid((100, 85), (2, 64), colspan=17, rowspan=95)
elif model_version == 'bert-base-uncased':
ax1 = plt.subplot2grid((100, 85), (0, 0), colspan=65, rowspan=99)
ax2 = plt.subplot2grid((100, 85), (12, 70), colspan=15, rowspan=75)
elif model_version == 'bert-large-uncased':
ax1 = plt.subplot2grid((100, 85), (0, 5), colspan=55, rowspan=99)
ax2 = plt.subplot2grid((100, 85), (2, 64), colspan=17, rowspan=95)
elif model_version == 'distilbert-base-uncased':
ax1 = plt.subplot2grid((100, 85), (0, 0), colspan=62, rowspan=99)
ax2 = plt.subplot2grid((100, 85), (32, 69), colspan=17, rowspan=35)
elif model_version == 'roberta-base':
ax1 = plt.subplot2grid((100, 85), (0, 0), colspan=65, rowspan=99)
ax2 = plt.subplot2grid((100, 85), (12, 70), colspan=15, rowspan=75)
elif model_version == 'roberta-large':
ax1 = plt.subplot2grid((100, 85), (0, 5), colspan=55, rowspan=99)
ax2 = plt.subplot2grid((100, 85), (2, 64), colspan=17, rowspan=95)
heatmap = sns.heatmap(effect_head, center=0.0, ax=ax1, annot=annot, annot_kws={"size": 9}, fmt=".2f", square=True, cbar=False, linewidth=0.1, linecolor='#D0D0D0',
cmap = LinearSegmentedColormap.from_list('rg', ["#F14100", "white", "#3D4FC4"], N=256))
plt.setp(heatmap.get_yticklabels(), fontsize=7)
plt.setp(heatmap.get_xticklabels(), fontsize=7)
heatmap.tick_params(axis='x', pad=1, length=2)
heatmap.tick_params(axis='y', pad=1, length=2)
heatmap.yaxis.labelpad = 2
heatmap.invert_yaxis()
if model_version != 'gpt2-xl':
for i, label in enumerate(heatmap.xaxis.get_ticklabels()):
if i%2 == 1:
label.set_visible(False)
for i, label in enumerate(heatmap.yaxis.get_ticklabels()):
if i%2 == 1:
label.set_visible(False)
if do_sort:
heatmap.axes.get_xaxis().set_ticks([])
else:
if model_version == 'gpt2-xl':
every_nth = 2
for n, label in enumerate(ax1.xaxis.get_ticklabels()):
if n % every_nth != 0:
label.set_visible(False)
for n, label in enumerate(ax1.yaxis.get_ticklabels()):
if n % every_nth != 0:
label.set_visible(False)
# split axes of heatmap to put colorbar
ax_divider = make_axes_locatable(ax1)
if model_version in ('gpt2-large', 'gpt2-xl'):
cax = ax_divider.append_axes('left', size='7%', pad='45%')
else:
cax = ax_divider.append_axes('left', size='7%', pad='33%')
# # make colorbar for heatmap.
# # Heatmap returns an axes obj but you need to get a mappable obj (get_children)
cbar = colorbar(ax1.get_children()[0], cax=cax, orientation='vertical')
cax.yaxis.set_ticks_position('left')
cbar.solids.set_edgecolor("face")
cbar.ax.tick_params(labelsize=7, length=4, pad=2)
ax1.set_title('Head Effect', size=9)
ax1.set_xlabel('Head', size=8)
ax1.set_ylabel('Layer', size=8)
for _, spine in ax1.spines.items():
spine.set_visible(True)
ax2.set_title(' Layer Effect', size=9)
bp = sns.barplot(x=effect_layer, ax=ax2, y=list(range(n_layers)), color="#3D4FC4", orient="h")
plt.setp(bp.get_xticklabels(), fontsize=7)
bp.tick_params(axis='x', pad=1, length=3)
ax2.invert_yaxis()
ax2.set_yticklabels([])
ax2.spines['top'].set_visible(False)
ax2.spines['right'].set_visible(False)
ax2.spines['left'].set_visible(False)
ax2.xaxis.set_ticks_position('bottom')
ax2.axvline(0, linewidth=.85, color='black')
path = f'results/heat_maps_with_bar_{effect_type}{"_sorted" if do_sort else ""}'
if not os.path.exists(path):
os.makedirs(path)
fname = f'{path}/{source}_{model_version}_{filter}_{suffix}.pdf'
plt.savefig(fname, format='pdf')
plt.close()
def main():
sns.set_context("paper")
sns.set_style("white")
model_versions = ['distilgpt2', 'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl']
filters = ['filtered', 'unfiltered']
# For testing:
#
# model_version = 'gpt2'
# split = 'dev'
# filter = 'filtered'
# fname = f"winobias_data/attention_intervention_{model_version}_{filter}_{split}.json"
# with open(fname) as f:
# data = json.load(f)
# save_figures(data, 'winobias', model_version, filter, split)
# return
#
# Other models:
# model_versions = [
# 'transfo-xl-wt103',
# 'xlnet-base-cased',
# 'xlnet-large-cased',
# 'bert-base-uncased',
# 'bert-large-uncased',
# 'distilbert-base-uncased',
# 'roberta-base',
# 'roberta-large',
# ]
# for mv in model_versions:
# filter = 'filtered'
# split = 'dev'
# stat = 'bergsma'
# autoregressive = mv.startswith('transfo-xl') or mv.startswith('xlnet')
# for ma in [-1] if autoregressive else [1, 2, 3, 4, 5, 6]:
# # Process winobias
# fname = (f'winobias_data/attention_intervention_{mv}_{filter}_' +
# (f'{split}_{ma}.json' if ma > 0 else f'{split}.json'))
# with open(fname) as f:
# suffix = split + (f'_{ma}' if ma > 0 else '')
# save_figures(json.load(f), 'winobias', mv, filter, suffix)
# # Process winogender
# fname = (f'winogender_data/attention_intervention_bergsma_{mv}_' +
# (f'{filter}_{ma}.json' if ma > 0 else f'{filter}.json'))
# with open(fname) as f:
# suffix = stat + (f'_{ma}' if ma > 0 else '')
# save_figures(json.load(f), 'winogender', mv, filter, suffix)
# return
# Process winobias
for model_version in model_versions:
for filter in filters:
for split in ['dev', 'test']:
fname = f"winobias_data/attention_intervention_{model_version}_{filter}_{split}.json"
if not os.path.exists(fname):
print('File does not exist:', fname)
continue
with open(fname) as f:
data = json.load(f)
save_figures(data, 'winobias', model_version, filter, split)
# Process winogender
for model_version in model_versions:
for filter in filters:
for stat in ['bergsma', 'bls']:
fname = f"winogender_data/attention_intervention_{stat}_{model_version}_{filter}.json"
if not os.path.exists(fname):
print('File does not exist:', fname)
continue
with open(fname) as f:
data = json.load(f)
save_figures(data, 'winogender', model_version, filter, stat)
if __name__ == '__main__':
main()