-
Notifications
You must be signed in to change notification settings - Fork 4
/
feature_conversion_methods.py
267 lines (220 loc) · 8.25 KB
/
feature_conversion_methods.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
"""
Helper preprocessing functions.
Code formatted using https://github.com/psf/black
"""
import random
import numpy as np
random.seed(10)
def cose_explanation_to_label(
example,
index,
tokenizer,
pred_only=False,
predictions_file=None,
include_input=False,
):
# Format:
# if include_input:
# Input: "cos_e question: [question] choice: [choice_0] choice: [choice_1] choice: [choice_2] explanation: [abstractive_explanation]"
# if not include_input:
# Input: "cos_e choice: [choice_0] choice: [choice_1] choice: [choice_2] explanation: [abstractive_explanation]"
# Output: "[answer]"
if pred_only:
abstr_expl = predictions_file[index]
else:
abstr_expl = example["abstractive_explanation"]
if include_input:
question = example["question"]
input_string = (
f"cos_e question: {question} choice: "
+ " choice: ".join(example["choices"])
+ f" explanation: {abstr_expl}"
)
else:
input_string = (
f"cos_e choice: "
+ " choice: ".join(example["choices"])
+ f" explanation: {abstr_expl}"
)
answer_string = example["answer"]
# tokenizer takes care of model-specific special tokens
encodings = tokenizer.encode_plus(
input_string + tokenizer.eos_token,
return_attention_mask=True,
)
# note even with "labels.shift_right()", the decoder attention mask length is still correct since we remove the last token
dec = tokenizer.encode_plus(
answer_string + tokenizer.eos_token,
return_attention_mask=True,
)
encodings["labels"] = dec["input_ids"]
encodings["decoder_attention_mask"] = dec["attention_mask"]
encodings["question_encoding"] = encodings["input_ids"]
return encodings
def esnli_explanation_to_label(
example,
index,
tokenizer,
pred_only=False,
predictions_file=None,
include_input=False,
):
# Format:
# if include_input:
# Input: "nli hypothesis: [hypothesis] premise: [premise] explanation: [abstractive_explanation]"
# if not include_input:
# Input: "nli explanation: [abstractive_explanation]"
# Output: "[answer]"
hypothesis = example["hypothesis"]
premise = example["premise"]
if pred_only:
abstr_expl = predictions_file[index]
else:
abstr_expl = example["explanation_1"]
if include_input:
input_string = (
f"nli hypothesis: {hypothesis} premise: {premise} explanation: {abstr_expl}"
)
else:
input_string = f"nli explanation: {abstr_expl}"
if example["label"] == 0:
answer_string = "entailment"
elif example["label"] == 1:
answer_string = "neutral"
elif example["label"] == 2:
answer_string = "contradiction"
# tokenizer takes care of model-specific special tokens
encodings = tokenizer.encode_plus(
input_string + tokenizer.eos_token,
return_attention_mask=True,
)
# note even with "labels.shift_right()", the decoder attention mask length is still correct since we remove the last token
dec = tokenizer.encode_plus(
answer_string + tokenizer.eos_token,
return_attention_mask=True,
)
encodings["labels"] = dec["input_ids"]
encodings["decoder_attention_mask"] = dec["attention_mask"]
encodings["question_encoding"] = encodings["input_ids"]
return encodings
def input_to_explanation_plus_label(
example,
index,
tokenizer,
datasource=None,
expl_only=False,
label_only=False,
gradients=None,
threshold=None,
):
# CoS-E Format:
# Input: "explain cos_e question: [question] choice: [choice_0] choice: [choice_1] choice: [choice_2]"
# e-SNLI Format:
# Input: "explain nli hypothesis: [hypothesis] premise: [premise]"
# Output: "[answer] explanation: [abstractive_explanation]"
# Explanation-only output: "None explanation: [abstractive_explanation]"
# Label-only output: "[answer]"
assert datasource in {"cos_e", "esnli"}
if datasource == "cos_e":
input_string, answer_string = cose_wt5_format(
example, expl_only=expl_only, label_only=label_only
)
elif datasource == "esnli":
input_string, answer_string = esnli_wt5_format(
example, expl_only=expl_only, label_only=label_only
)
# tokenizer takes care of model-specific special tokens
encodings = tokenizer.encode_plus(
input_string + tokenizer.eos_token,
return_attention_mask=True,
)
if threshold is not None:
# compute top-k tokens
k_length = round(len(encodings["input_ids"]) * threshold)
if k_length > 0:
if gradients is not None:
# select tokens to drop from ranked gradient file
# if gradient object doesn't exist, throw out
try:
grads = gradients["label"]["instance_%d" % (index + 1)][
"grad_input_2"
]
except:
# index doesn't exist in gradients file because couldn't produce gradients due to bad decoding
return {}
assert len(encodings["input_ids"]) == len(grads)
# select tokens to drop based on absolute-value rank-importance
top_token_inxs = np.argsort(np.abs(grads))[-k_length:]
else:
# select random tokens to drop for baseline
top_token_inxs = random.sample(
[i for i in range(len(encodings["input_ids"]))], k_length
)
else:
top_token_inxs = []
assert len(top_token_inxs) == k_length
# first determine which top-tokens are spans
# replace these tokens with special "sentinel"/pad tokens
# there are 100 extra_ids from positions 32000-32099 in the tokenizer by default
tmp_lst = []
tmp_attns = []
curr = 1000
curr_extra_id = 32100
for i, (item, att) in enumerate(
zip(encodings["input_ids"], encodings["attention_mask"])
):
if i in top_token_inxs:
if i != curr + 1:
if curr_extra_id - 1 < 32000:
raise Exception("too small id value")
tmp_lst.append(curr_extra_id - 1)
tmp_attns.append(att)
curr_extra_id = curr_extra_id - 1
curr = i
else:
tmp_lst.append(item)
tmp_attns.append(att)
assert len(tmp_lst) == len(tmp_attns)
encodings["input_ids"] = tmp_lst
encodings["attention_mask"] = tmp_attns
# note even with "labels.shift_right()", the decoder attention mask length is still correct since we remove the last token
dec = tokenizer.encode_plus(
answer_string + tokenizer.eos_token,
return_attention_mask=True,
)
encodings["labels"] = dec["input_ids"]
encodings["decoder_attention_mask"] = dec["attention_mask"]
encodings["question_encoding"] = encodings["input_ids"]
return encodings
def cose_wt5_format(item, expl_only=False, label_only=False):
question = item["question"]
answer = item["answer"]
abstr_expl = item["abstractive_explanation"]
input_string = f"explain cos_e question: {question} choice: " + " choice: ".join(
item["choices"]
)
if expl_only:
answer_string = f"None explanation: {abstr_expl}"
elif label_only:
answer_string = f"{answer}"
else:
answer_string = f"{answer} explanation: {abstr_expl}"
return input_string, answer_string
def esnli_wt5_format(item, expl_only=False, label_only=False):
premise = item["premise"]
hypothesis = item["hypothesis"]
if item["label"] == 0:
answer = "entailment"
elif item["label"] == 1:
answer = "neutral"
elif item["label"] == 2:
answer = "contradiction"
abstr_expl = item["explanation_1"]
input_string = f"explain nli hypothesis: {hypothesis} premise: {premise}"
if expl_only:
answer_string = f"None explanation: {abstr_expl}"
elif label_only:
answer_string = f"{answer}"
else:
answer_string = f"{answer} explanation: {abstr_expl}"
return input_string, answer_string