-
Notifications
You must be signed in to change notification settings - Fork 2
/
randomsearchMNLI.py
203 lines (156 loc) · 6.21 KB
/
randomsearchMNLI.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
# 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.
#
import torch
import pytorch_lightning as pl
from processors import MnliProcessor
from firebert_fct import FireBERT_FCT
from bert_base_model import LightningBertForSequenceClassification
num_gpus = -1 if torch.cuda.is_available() else None
import gc
import random
gc.enable()
import time
from torch.utils.data import TensorDataset
# prepare hyperparameters
max_steps = -1 # if -1 then calculate number of training steps based on the length of the train set
len_train_set = 392702
gradient_accumulation_steps = 1
learning_rate = 2e-5
weight_decay = 0.00
adam_epsilon = 1e-8
warmup_proportion = 0
num_train_epochs = 1
batch_size = 7
if max_steps > 0:
num_train_epochs = max_steps // (len_train_set // gradient_accumulation_steps) + 1
num_training_steps = max_steps
else:
num_training_steps = len_train_set // gradient_accumulation_steps * num_train_epochs
warmup_steps = num_training_steps // num_train_epochs * warmup_proportion
def load_examples(features_file):
features = torch.load(features_file)
# Convert to Tensors and build dataset
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
all_idxs = torch.tensor([i for i in range(len(all_input_ids))], dtype=torch.long)
dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels, all_idxs)
return dataset
def elapsed_time():
global t_start
t_now = time.time()
t = t_now-t_start
t_start = t_now
return t
use_full_example = True
use_USE = True # if random.randint(1, 2) == 1 else False
USE_method = 'rank' #'filter' if random.randint(1, 2) == 1 else 'rank'
USE_multiplier = 12 #random.randint(5, 15) # 3
stop_words = True
perturb_words = 9 #random.randint(4, 9) # 2
candidates_per_word = 10 #random.randint(5, 15) #10
total_alternatives = 4 #random.randint(4, 7) # 5
match_pos = True #if random.randint(1, 2) == 1 else False
leave_alone = 0
random_out_of = 0
judge_bert = False
hparams = { 'learning_rate': learning_rate,
'adam_epsilon': adam_epsilon,
'weight_decay': weight_decay,
'warmup_steps': warmup_steps,
'num_training_steps': num_training_steps,
'batch_size': batch_size,
'use_USE': use_USE,
'USE_method': USE_method,
'USE_multiplier': USE_multiplier,
'stop_words': stop_words,
'perturb_words': perturb_words,
'candidates_per_word': candidates_per_word,
'total_alternatives': total_alternatives,
'match_pos': match_pos,
'use_full_example': use_full_example,
'leave_alone': leave_alone,
'random_out_of': random_out_of,
'judge_bert': judge_bert
}
print(hparams)
proc_hparams = {}
# delete this next line to run full 100%
proc_hparams.update({'sample_percent': 3,
'randomize': True})
# instantiate the model used for SWITCH
switch_model = LightningBertForSequenceClassification(load_from = 'resources/models/MNLI/pytorch_model.bin',
processor = MnliProcessor(),
hparams = {'batch_size': 6 })
switch_model.cuda()
model = FireBERT_FCT(switch_model=switch_model, processor=MnliProcessor(hparams=proc_hparams), hparams=hparams)
processor = model.get_processor()
train_dataset, train_examples = processor.load_and_cache_examples("data/MNLI", example_set='train')
val_dataset, _ = processor.load_and_cache_examples("data/MNLI", example_set='dev')
test_dataset, _ = processor.load_and_cache_examples("data/MNLI", example_set='test')
model.set_train_dataset(train_dataset, train_examples)
model.set_val_dataset(val_dataset)
model.set_test_dataset(test_dataset)
amp_opt_level='O1'
max_grad_norm = 1.0
t_start = time.time()
trainer = pl.Trainer(gpus=num_gpus,
max_epochs = num_train_epochs, amp_level=amp_opt_level, gradient_clip_val=max_grad_norm,
max_steps = num_training_steps)
trainer.fit(model)
training_time = round(elapsed_time(),2)
trainer.test(model)
train_results = trainer.tqdm_metrics
del train_dataset
del train_examples
del val_dataset
del test_dataset
del trainer
gc.collect()
torch.cuda.empty_cache()
# compare how well the model does against adversarial samples
test_dataset = load_examples('data/MNLI/generated/mnli_adversarial_samples_for_dev')
model.set_test_dataset(test_dataset)
trainer = pl.Trainer(gpus=num_gpus)
trainer.test(model)
adv_results = trainer.tqdm_metrics
results = "training_time: " + str(training_time) + "s, " + \
"val_loss: " + str(train_results['val_loss']) + ", " + \
"val_acc: " + str(train_results['avg_val_acc']) + ", " + \
"adv_val_acc: " + str(adv_results['avg_test_acc']) + ", " + \
"use_USE: " + str(use_USE) + ", " + \
"USE_multiplier: " + str(USE_multiplier) + ", " + \
"USE_method: " + str(USE_method) + ", " + \
"perturb_words: " + str(perturb_words) + ", " + \
"candidates_per_word: " + str(candidates_per_word) + ", " + \
"total_alternatives: " + str(total_alternatives) + ", " + \
"match_pos: " + str(match_pos)
print(results)
fname = str("results/fct/mnli-hparams-results.txt")
f = open(fname, "a")
f.write(results)
f.write( "\n")
f.close()
del processor
del test_dataset
del model
del switch_model
del trainer
del train_results
del adv_results
del results
del fname
del f
gc.collect()
torch.cuda.empty_cache()