-
Notifications
You must be signed in to change notification settings - Fork 6
/
config_our.py
187 lines (146 loc) · 6.61 KB
/
config_our.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
import os
import sys
import json
import datetime
import argparse
import math
import time
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
import pandas as pd
from transformers import AdamW, get_linear_schedule_with_warmup
#from fairseq.models.bart import BARTModel
#from transformers import BartTokenizer
from config_m2m import *
from utils_en import *
from model_utils import FocalLoss, LDAMLoss
from model import BERTClassifier
#def parse_args():
# return parser.parse_args()
#ARGS = parse_args()
print('== Arguments ==')
print(ARGS)
# legacy code - Needs to be refactored
set_global_variables(ARGS.gpu, ARGS.device, ARGS.MAX_LEN, ARGS.batch_size)
# Real = [ATIS, TREC], Imbalance manipulation = [SNIPS]
if ARGS.dataset == 'ATIS':
N_SAMPLES = -1
elif ARGS.dataset == 'SNIPS':
N_SAMPLES = 1000
elif ARGS.dataset == 'TREC':
N_SAMPLES = 1000
else:
print('Dataset:', ARGS.dataset)
raise NotImplementedError()
random.seed(ARGS.random_seed)
np.random.seed(ARGS.random_seed)
torch.manual_seed(ARGS.random_seed)
torch.cuda.manual_seed_all(ARGS.random_seed)
# To be refactored
model_name = '%s_%s_%s_classifier_%s_%s_%s_%s.ckpt' % (ARGS.dataset, ARGS.data_setting, str(ARGS.imbalanced_ratio), ARGS.loss_type, ARGS.learning_rate, ARGS.data_augment, ARGS.gmodel)
print('\n== Load data ==')
bert, tokenizer = get_bert()
labeled_train_data, labeled_valid_data, labeled_test_data, HEAD_labels, TAIL_labels = load_data(ARGS.dataset, data_setting=ARGS.data_setting, imbalanced_ratio=ARGS.imbalanced_ratio, head_sample_num=N_SAMPLES, min_valid_data_num=ARGS.min_valid_data_num)
if ARGS.data_augment:
if ARGS.gmodel == 'bart':
assert ARGS.cmodel == None
augment_data_filename = './data/%s/aug_%s_%s_%s.csv' % (ARGS.dataset, ARGS.data_setting, ARGS.cmodel, ARGS.gmodel)
augment_data_filename = './data/%s/aug_%s_%s_%s.csv' % (ARGS.dataset, ARGS.data_setting, ARGS.cmodel, ARGS.gmodel)
print('data_augment:', ARGS.data_augment)
print('cmodel:', ARGS.cmodel)
print('gmodel:', ARGS.gmodel)
labeled_aug_data = augment_data(augment_data_filename, tokenizer)
labeled_train_data = pd.concat([labeled_aug_data, labeled_train_data])
##
#N_SAMPLES_PER_CLASS_BASE = [int(N_SAMPLES)] * N_CLASSES
#if ARGS.imb_type == 'longtail':
# N_SAMPLES_PER_CLASS_BASE = make_longtailed_imb(N_SAMPLES, N_CLASSES, ARGS.ratio)
#elif ARGS.imb_type == 'step':
# for i in range(ARGS.imb_start, N_CLASSES):
# N_SAMPLES_PER_CLASS_BASE[i] = int(N_SAMPLES * (1 / ARGS.ratio))
#elif ARGS.imb_type == 'all':
# for i in range(ARGS.imb_start, N_CLASSES):
# N_SAMPLES_PER_CLASS_BASE[i] = -1
#N_SAMPLES_PER_CLASS_BASE = tuple(N_SAMPLES_PER_CLASS_BASE)
#print(N_SAMPLES_PER_CLASS_BASE)
###
df_train_data_filtered = labeled_train_data[['sentence', 'label']]
df_train_sent_group_by_label = df_train_data_filtered.groupby(['label']).size().reset_index(name='counts')
classes, class_counts = np.unique(df_train_sent_group_by_label['label'].tolist(), return_counts=True)
##
N_SAMPLES_PER_CLASS = tuple(df_train_sent_group_by_label.counts.tolist())
N_SAMPLES_PER_CLASS_T = torch.Tensor(N_SAMPLES_PER_CLASS)
print(N_SAMPLES_PER_CLASS)
#print(N_SAMPLES_PER_CLASS_T)
##
def get_oversampled_data(dataset, num_sample_per_class, random_seed=0):
"""
Return a list of imbalanced indices from a dataset.
Input: A dataset (e.g., CIFAR-10), num_sample_per_class: list of integers
Output: oversampled_list ( weights are increased )
"""
length = dataset.__len__()
num_sample_per_class = list(num_sample_per_class)
num_samples = list(num_sample_per_class)
selected_list = []
indices = list(range(0,length))
# tmp_l = []
# for i in range(0, length):
# _, _, _, label = dataset.__getitem__(i)
# label = label[0]
# tmp_l.append(label)
# print(set(tmp_l))
# assert 1 == 2
for i in range(0, length):
index = indices[i]
_, _, _, label = dataset.__getitem__(index)
label = label[0]
if num_sample_per_class[label] > 0:
selected_list.append(1 / num_samples[label])
num_sample_per_class[label] -= 1
return selected_list
def get_data_dict(x_dict, y_dict, train=False, shuffle=False, skip=False):
for t in ['train', 'valid', 'test']:
x_dict['HEAD_'+t], x_dict['TAIL_'+t], y_dict['HEAD_'+t], y_dict['TAIL_'+t] = split_HEAD_TAIL(x_dict[t], y_dict[t])
data_dict = {}
loader_dict = {}
# JW (temp)
iter_list = ['', 'HEAD', 'TAIL']
if skip: iter_list = ['']
for t1 in iter_list:
for t2 in ['train', 'valid', 'test']:
key = t2
if t1 != '':
key = t1+'_'+key
data_dict[key] = Dataset(x_dict[key], y_dict[key])
if train or t2 != 'train':
if t2 == 'train':
from torch.utils.data.sampler import SubsetRandomSampler, WeightedRandomSampler
train_in_idx = get_oversampled_data(data_dict[key], num_of_class_samples, ARGS.random_seed)
loader_dict[key] = DataLoader(dataset=data_dict[key],
sampler=WeightedRandomSampler(train_in_idx, len(train_in_idx)),
batch_size=ARGS.batch_size, collate_fn=collate_fn, shuffle=False)
else:
loader_dict[key] = DataLoader(dataset=data_dict[key], batch_size=ARGS.batch_size, collate_fn=collate_fn, shuffle=False)
return x_dict, y_dict, data_dict, loader_dict
x_dict, y_dict, y_class, class_dict = convert_data(labeled_train_data, labeled_valid_data, labeled_test_data, tokenizer)
num_of_classes = len(y_class)
num_of_class_samples = [0] * (max(y_class) + 1)
max_samples = df_train_sent_group_by_label['counts']
max_count = max(df_train_sent_group_by_label['counts'].tolist())
for index, row in df_train_sent_group_by_label.iterrows():
num_of_class_samples[row.label] = row.counts
num_of_class_samples_T = torch.tensor(num_of_class_samples, dtype=torch.float)
x_dict, y_dict, data_dict, loader_dict = get_data_dict(x_dict, y_dict, train=True, shuffle=True)
train_loader, val_loader, test_loader = loader_dict['train'], loader_dict['valid'], loader_dict['test']
if 'translation.py' in sys.argv[0] or 'data_generation_for_gmodel.py' in sys.argv[0]:
from fairseq.models.bart import BARTModel
from transformers import BartTokenizer
bart_tokenizer = BartTokenizer.from_pretrained('facebook/bart-large')
elif 'token_importance.py' in sys.argv[0] == False:
print('Tokenizer is deleted!')
del tokenizer
torch.cuda.empty_cache()