-
Notifications
You must be signed in to change notification settings - Fork 2
/
firebert_fve.py
367 lines (293 loc) · 13.8 KB
/
firebert_fve.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
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
# coding=utf-8
# Copyright 2020 FireBERT authors. All rights reserved.
#
# Licensed under the MIT license
# See https://github.com/FireBERT-author/FireBERT/blob/master/LICENSE for details
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
#
# FVE as a subclass
# overrides extend_batch_examples_test() used in forward()
#
import torch
from torch import nn
import numpy as np
import argparse
from switch import SWITCH
from firebert_base import FireBERT_base
class FireBERT_FVE(FireBERT_base):
def __init__(self, load_from=None, processor=None, hparams=None):
super(FireBERT_FVE, self).__init__(load_from=load_from, processor=processor, hparams=hparams)
# need SWITCH to tell us what the important words are
self.switch = SWITCH(hparams=hparams, model=self, tokenizer=self.tokenizer, device=self.device)
# merge passed hparams over default hparams
hdict = self.get_default_hparams()
hdict.update(hparams)
self.hparams = argparse.Namespace(**hdict)
self.actualize_hparams()
return
# methods to set hparams on the fly
def update_hparams(self, new_hparams):
hdict = vars(self.hparams)
hdict.update(new_hparams)
self.hparams = argparse.Namespace(**hdict)
self.switch.update_hparams(new_hparams)
super().update_hparams(new_hparams)
self.actualize_hparams()
def actualize_hparams(self):
# remember some hparams a little better
self.count = self.hparams.vector_count
self.perturb_words = self.hparams.perturb_words
self.std = self.hparams.std
return
#
# here are some useful defaults
#
def get_default_hparams(self):
d = FireBERT_base.get_default_hparams(self)
d.update({
# these are for SWITCH
'use_USE':False,
'stop_words':True,
'perturb_words':2,
# this is for base
'verbose':False,
'vote_avg_logits':True,
# this is for us
'std':0.05,
'vector_count':10
})
return d
# this fills in the hook prepared in the base class
def extend_batch_examples_eval(self, input_ids=None, attention_mask=None, token_type_ids=None,
position_ids=None, head_mask=None, inputs_embeds=None, example_idx=None):
group_ids = None
group_sizes = []
inputs_embeds_results = inputs_embeds
# we do need the examples for SWITCH
if example_idx is not None:
# let's get the embeddings for the original samples
inputs_embeds = self.bert.get_input_embeddings()(input_ids).detach()
inputs_embeds_results = inputs_embeds
# in groups, we keep track of samples tha belong together for voting
group_ids = list(range(0,len(example_idx)))
group_sizes = []
current_group = 0
# gotta go through one by one. Yes, batch logic is nicer, but SWICH is a one-by-one thing, anyway.
for i, idx in enumerate(example_idx):
example = self.processor.all_examples[idx]
# call the perturbation method individually for each example
perturbed_inputs_embeds, sample_attention_mask, sample_token_type_ids, _, _, _ = \
self.perturb_example(example,
sample_index = i,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids)
# Put tensors together
# print("perturbed examples: ", perturbed_inputs_embeds[0])
if perturbed_inputs_embeds is None:
# didn't get any important words to perturb, so no batch extensions
group_sizes.append(1) # pus one for the original
else:
inputs_embeds_results = torch.cat((inputs_embeds_results, perturbed_inputs_embeds), dim=0)
attention_mask = torch.cat((attention_mask, sample_attention_mask), dim=0)
token_type_ids = torch.cat((token_type_ids, sample_token_type_ids), dim=0)
group_ids += [current_group]*(self.count)
group_sizes.append(self.count+1) # plus one for the original
current_group += 1
# need to erase the tokens (input_ids) so that BERT will use the embeddings
input_ids = None
# we probably received these as None, but if we don't set them to that, we might have to
# adjust them for the new batch size and that would be tedious
head_mask = None
position_ids = None
# won't need these anymore
example_idx = None
return input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds_results, example_idx, group_ids, group_sizes
#
# this perturbs an example and returns it in a batch with others
#
def perturb_example(self,example, count=None, std=None, sample_index=0, input_ids=None, attention_mask=None, token_type_ids=None):
if count is None:
count = self.count
if std is None:
std = self.std
if input_ids is None:
# convert example into features
input_ids, attention_mask, token_type_ids, _ = self.processor.create_feature_tensors([example], device=self.device)
sample_index = 0
# use SWITCH to figure out word importance within the list
word_indices, token_indices, word_list = \
self.switch.get_important_indices_from_example(example,
input_ids[sample_index].unsqueeze(0),
token_type_ids[sample_index].unsqueeze(0),
attention_mask[sample_index].unsqueeze(0))
# filter out useless stuff
word_indices = list(filter(lambda x:x!=-1, word_indices))
token_indices = list(filter(lambda x:x!=-1, token_indices))
# identify most important words
important_words = [word_list[i] for i in word_indices[:self.perturb_words]]
token_indices = token_indices[:self.perturb_words]
# get embeddings from BERT for the whole sample (set of words)
embeddings = self.bert.get_input_embeddings()(input_ids[sample_index]).detach()
#print("embeddings:",embeddings.shape)
#print(embeddings[0,:20])
batch = None
points = None
for token_index in token_indices:
# get the embedding vector for the most important word
v = embeddings[token_index].clone().detach()
#print("vector:", v.shape)
#print(v)
# scale the single sample set of tokens/embeddings up to a whole batch
batch = embeddings.repeat(count,1,1)
# hopefully give some GPU memory back
#del embeddings
# scatter a region around this vector
points = self.region_around(v, std=std, count=count)
#print("region:", points.shape)
#print(points[0])
#print("batch embeddings before clobber:",batch.shape)
#print(batch[0,:20])
# clobber the tensor for the region of perturbed vectors in there
batch[:,token_index,:] = points
#print("batch embeddings after clobber:",batch.shape)
#print(batch[0,:20])
attention_mask = attention_mask[sample_index].repeat(count, 1)
token_type_ids = token_type_ids[sample_index].repeat(count, 1)
return batch, attention_mask, token_type_ids, points, important_words, None
#
# helper methods
#
#
# make a field of Gaussian-perturbed vectors around a given vector v
#
def region_around(self, vector, std, count, device=None):
vectors = vector.repeat(count, 1)
#print("vectors:",vectors.shape)
region = torch.normal(mean=vectors, std=std).cpu()
#print("region:",region.shape)
return region
def get_single_vector(self, word):
#tbd
return None
def get_hparam(self, name):
return self.hparams[name]
#
# Tests.
#
def test_FireBERT_FVE(task, set, reps=1, sample=1, hparams_default={}, hparams_lists=None, lightning=''):
# prepare hyperparameters
hparams = hparams_default
# load the right processor class
if task == "MNLI":
processor = MnliProcessor({'sample_percent':sample}) # negative number means abs number of samples, not percent
elif task == "IMDB":
processor = ImdbProcessor({'sample_percent':sample})
# now instantiate the models
model = FireBERT_FVE(load_from='resources/models/'+task+lightning+'/pytorch_model.bin',
processor=processor,
hparams=hparams_default)
processor.set_tokenizer(model.tokenizer)
dataset, examples = processor.load_and_cache_examples("data/"+task, example_set=set)
model.set_test_dataset(dataset, examples)
#adv set
# load the right processor class
if task == "MNLI":
adv_processor = MnliProcessor({'sample_percent':sample}) # negative number means abs number of samples, not percent
elif task == "IMDB":
adv_processor = ImdbProcessor({'sample_percent':sample})
model_adv = FireBERT_FVE(load_from='resources/models/'+task+lightning+'/pytorch_model.bin',
processor=processor,
hparams=hparams_default)
adv_processor.set_tokenizer(model.tokenizer)
dataset_adv, examples_adv = adv_processor.load_and_cache_examples("data/"+task, example_set="adv_"+set)
model_adv.set_test_dataset(dataset_adv, examples_adv)
for i in range(reps):
if hparams_lists is None:
print("FireBERT_FVE specific test", task, set)
else:
print("FireBERT_FVE hparam test", task, set)
print("{")
for item in hparams_lists.items():
key = item[0]
values = item[1]
hparams[key] = random.choice(values)
print(" '"+key+"':",str(hparams[key])+",")
print("}")
# set the new hparams
model.update_hparams(hparams)
model_adv.update_hparams(hparams)
trainer = pl.Trainer(gpus=(-1 if torch.cuda.is_available() else None))
trainer.test(model)
result1 = trainer.tqdm_metrics
trainer = pl.Trainer(gpus=(-1 if torch.cuda.is_available() else None))
trainer.test(model_adv)
result2 = trainer.tqdm_metrics
f = open("results/five/hparams-results.csv", "a+")
print(task, ",", "adv_"+set, ",", sample, ',"',hparams,'",',result1['avg_test_acc'],",",result2['avg_test_acc'], sep="", file=f)
f.close()
print("iteration",i,"logged.")
elapsed_time()
print()
if hparams_lists is None:
break
def elapsed_time():
global t_start
t_now = time.time()
t = t_now-t_start
print("elapsed time: ",round(t,2), "s")
t_start = t_now
return t
if __name__ == "__main__":
import random
import time
import pytorch_lightning as pl
from processors import MnliProcessor, ImdbProcessor
t_start = time.time()
# prepare hyperparameters
hparams_default = {
'batch_size':8,
# these are for SWITCH
'use_USE':False,
'stop_words':True,
'perturb_words':2,
# this is for base
'verbose':False,
'vote_avg_logits':True,
# this is for us
'std':0.05,
'vector_count':10
}
hparams_lists = {
# these are for SWITCH
'stop_words':[True, False],
'perturb_words':range(1,20),
# this is for base
'vote_avg_logits':[True, False],
# this is for us
'std':np.arange(0.1,10,0.01),
'vector_count':range(3,15)
}
# parameter search
sample = 15
#test_FireBERT_FVE("IMDB", "dev", reps=500, sample=sample, hparams_default=hparams_default, hparams_lists=hparams_lists)
#test_FireBERT_FVE("MNLI", "dev", reps=500, sample=sample, hparams_default=hparams_default, hparams_lists=hparams_lists)
# Monday, on lightning model
best_IMDB_lightning = {'batch_size': 8, 'use_USE': False, 'stop_words': True, 'perturb_words': 1,
'verbose': False, 'vote_avg_logits': False, 'std': 8.4, 'vector_count': 11}
best_mnli_lightning = {'batch_size': 8, 'use_USE': False, 'stop_words': False, 'perturb_words': 1,
'verbose': False, 'vote_avg_logits': False, 'std': 2.31, 'vector_count': 8}
# Tuesday, on paper model
best_mnli = {'batch_size': 8, 'use_USE': False, 'stop_words': True, 'perturb_words': 1,
'verbose': False, 'vote_avg_logits': True, 'std': 8.14, 'vector_count': 8}
best_IMDB={'batch_size': 8, 'use_USE': False, 'stop_words': True, 'perturb_words': 1,
'verbose': False, 'vote_avg_logits': True, 'std': 2.29, 'vector_count': 10}
# actual test runs
#test_FireBERT_FVE("IMDB", "test", reps=1, sample=100, hparams_default=best_IMDB)
test_FireBERT_FVE("MNLI", "test", reps=1, sample=100, hparams_default=best_mnli)