-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathopts.py
256 lines (222 loc) · 14.5 KB
/
opts.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
import argparse
from distutils.util import strtobool
def parse_opt(s=None):
parser = argparse.ArgumentParser()
# me added
parser.add_argument('--reproducePaper', action='store_true', help='Only relevant for Infusion. Reproduces '
'models from the paper by using the provided '
'Infusion DB (contains original order of obj-ids '
'which impacts position of newly added objects)')
parser.add_argument('--filter_objects', type=str, default='original',
help='filter objects from img')
parser.add_argument('--filter_mode', type=str, default='none',
help='filter or select given list of objects per qid')
parser.add_argument('--use_bbox', action='store_true',
help='use bbox coordinates or not')
parser.add_argument('--gen_exp', type=str, default='none')
parser.add_argument('--nonsense_class', action='store_true')
parser.add_argument('--bottomup', action='store_true',
help='for reading bottomup features from original release')
parser.add_argument('--visfis', action='store_true',
help='run 101k OOD test from visfis paper')
parser.add_argument('--visfis_all', action='store_true',
help='run full OOD test from visfis paper')
parser.add_argument('--visfis_hatcp', action='store_true',
help='run OOD test for VQA-HAT from visfis paper')
parser.add_argument('--visfis_hatcp_removeNoAnswerQIDs', action='store_true')
parser.add_argument('--visfis_vlr', action='store_true')
parser.add_argument('--vlr_padding', type=int, default=0)
parser.add_argument('--oracle_aug', action='store_true')
parser.add_argument('--id_and_color', action='store_true')
parser.add_argument('--augmentation', action='store_true')
parser.add_argument('--augmentation_1inK', action='store_true')
parser.add_argument('--augmentation_type', type=str, default='none')
parser.add_argument('--feature_name', type=str, default='detectron',
help='which features to use')
parser.add_argument('--feature_vqahat_vf', action='store_true',
help='vf in vqahat features to use')
parser.add_argument('--feature_vqahat_vf_original', action='store_true',
help='original vf in vqahat features to use')
parser.add_argument('--feature_vqahat_vf_original_symbolic_features', action='store_true',
help='original vf in vqahat features to use')
parser.add_argument('--overwrite_hint', action='store_true')
parser.add_argument('--no_init_vision', action='store_true')
parser.add_argument('--filter_hint', action='store_true')
parser.add_argument('--infusion', action='store_true',
help='run with ground truth feature infusion for hint alignments')
parser.add_argument('--hint_path', type=str, default='.',
help='path containing hint files to be used')
parser.add_argument('--num_obj_max_manual_setting', type=int, default=100,
help='max number of allowed visual objects in the input')
parser.add_argument('--flooding', type=float, default=0.0,
help='add constant to loss')
parser.add_argument('--hints_reduction_to_detectron', action='store_true',
help='reduce used hints to hint qids available in detectron recognitions for better comparison')
parser.add_argument('--randomize_hints', type=str, default='none')
parser.add_argument('--individual_img', action='store_true')
# logging & saving
parser.add_argument('--saved_model_prefix', type=str, default='')
parser.add_argument('--print_every_batch', action='store_true')
parser.add_argument('--save_every_epoch', action='store_true')
## dataset
parser.add_argument('--discard_items_for_test', type=bool, default=False)
parser.add_argument('--portion_of_training', type=float, default=1.0)
parser.add_argument('--train_subset', type=str)
parser.add_argument('--val_subset', type=str)
parser.add_argument('--use_two_testsets', action='store_true')
parser.add_argument('--split_test_2', type=str, help='second test split')
# spatial feats
parser.add_argument('--spatial_type', type=str, default='none', choices=['none', 'simple', 'linear', 'mesh'])
parser.add_argument('--spatial_length', default=0, type=int)
# oracle exp
parser.add_argument('--oracle_type', type=str, default="none", choices=['none', 'simple', 'wordvec'])
parser.add_argument('--oracle_embed_size', type=int, default=512)
parser.add_argument('--oracle_threshold', type=float, default=0.85)
parser.add_argument('--use_input_mask', action='store_true')
parser.add_argument('--mask_type', type=str, default='impt', choices=['impt', 'non_impt'])
# model
parser.add_argument('--model_type', type=str, default='updn_ramen', choices=['updn', 'rn', 'ban', 'updn_ramen', 'lang_only', 'updn_ramen_new', 'lxmert', 'updn_ramen_transfer'])
parser.add_argument('--num_objects', type=int)
parser.add_argument('--lxmert_hid_dim', type=int, default=768)
parser.add_argument('--lxmert_small_model', action='store_true')
## Feature Importance
parser.add_argument('--model_importance', type=str, default='none', choices=['gradcam', 'expected_gradient', 'LOO', 'KOI', 'SHAP', 'avg_effect', 'updn_att', 'none'])
parser.add_argument('--gradcam_type', type=str, default='sum', choices=['sum', 'l1', 'l2']) # for expected gradient / gradcam
parser.add_argument('--num_sample_omission', type=int) # for all omission based methods
parser.add_argument('--num_sample_eg', type=int) # for expected gradient
parser.add_argument('--eval_FI', action='store_true') # model.eval for computing FI
parser.add_argument('--OBJ11', action='store_true') # ugly hack to make visfis work
parser.add_argument('--FI_predicted_class', dest='FI_predicted_class',
type=lambda x: bool(strtobool(x))) # FI on predicted class vs. gt class; required
parser.add_argument('--vqa_loss_type', type=str, default="softmax", choices=['softmax', 'sigmoid'])
parser.add_argument('--impt_threshold', type=float)
# control for random for VisFIS
parser.add_argument('--random_suff', action='store_true')
parser.add_argument('--random_unc', action='store_true')
parser.add_argument('--random_inv_FI', action='store_true')
parser.add_argument('--random_align', action='store_true')
## additional objectives
# hint
parser.add_argument('--use_relaxed_hint_loss', action='store_true')
parser.add_argument('--use_hint_loss', action='store_true')
parser.add_argument('--hint_loss_weight', type=float)
# invariance-FI=0
parser.add_argument('--use_zero_loss', action='store_true')
parser.add_argument('--zero_loss_weight', type=float)
parser.add_argument('--normalize_FI', type=bool, default=True) # only for zero_loss
# direct alignment
parser.add_argument('--use_direct_alignment', action='store_true') ## also use use align_loss_type
parser.add_argument('--alignment_loss_weight', type=float)
# data augment methods
parser.add_argument('--aug_type', type=str, default='none', choices=['none','suff-human','suff-random','invariance', 'uncertainty-NonImptOnly', 'suff-uncertainty', 'suff-uncertainty-Chang2021', 'uncertainty-uniform', "saliency-guided"])
parser.add_argument('--align_loss_type', type=str, choices=['l1', 'l2', 'kl', 'cossim'])
parser.add_argument('--aug_loss_weight', type=float) # for suff only
parser.add_argument('--replace_func', type=str, default='negative_ones', choices=['0s', 'negative_ones', 'random_sample', 'gaussian-singla', 'gaussian-ours', 'shuffle'])
# train / val
parser.add_argument('--lr_type', type=str, default='none', choices=['none','ramen','neg_analysis'],
help='decide what lr scheduler to use')
parser.add_argument('--pretrain_finetune', action='store_true')
parser.add_argument('--pretrain_epochs', type=int, default=50)
parser.add_argument('--finetune_lr', type=float, default=2e-5) # 2e-5 for HINT, 5e-5 for SCR
# calc metrics
parser.add_argument('--calc_dp_level_metrics', action='store_true')
parser.add_argument('--percentages', type=list, default=[0.1,0.25,0.5])
parser.add_argument('--RRR_only', action='store_true') # for metrics
parser.add_argument('--ACC_only', action='store_true') # for metrics
# Data input settings
parser.add_argument('--data_dir', type=str)
parser.add_argument('--dataset', type=str, default='vqacp2', choices=['xaicp', 'gqacp', 'hatcp', 'bottomup', 'vinvl'])
parser.add_argument('--do_not_discard_items_without_hints', action='store_true')
parser.add_argument('--split', type=str, help='training split')
parser.add_argument('--hint_type', type=str)
parser.add_argument('--split_test', type=str, help='test split')
# parser.add_argument('--rnn_size', type=int, default=1280,
# help='size of the rnn in number of hidden nodes in question gru')
parser.add_argument('--num_hid', type=int, default=1280,
help='size of the rnn in number of hidden nodes in question gru')
parser.add_argument('--num_layers', type=int, default=2,
help='number of GCN layers')
parser.add_argument('--rnn_type', type=str, default='gru',
help='rnn, gru, or lstm')
parser.add_argument('--v_dim', type=int, default=2048,
help='2048 for resnet, 4096 for vgg')
parser.add_argument('--logit_layers', type=int, default=1,
help='number of layers in the RNN')
parser.add_argument('--activation', type=str, default='ReLU',
help='number of layers in the RNN')
parser.add_argument('--norm', type=str, default='weight',
help='number of layers in the RNN')
parser.add_argument('--initializer', type=str, default='kaiming_normal',
help='number of layers in the RNN')
# Optimization: General
parser.add_argument('--max_epochs', type=int, default=40,
help='number of epochs')
parser.add_argument('--batch_size', type=int, default=384,
help='minibatch size')
parser.add_argument('--grad_clip', type=int, default=0.25,
help='clip gradients at this value') # neg analysis 0.25; ramen 50; lxmert 5
parser.add_argument('--dropC', type=float, default=0.5,
help='strength of dropout in the Language Model RNN')
parser.add_argument('--dropG', type=float, default=0.2,
help='strength of dropout in the Language Model RNN')
parser.add_argument('--dropL', type=float, default=0.1,
help='strength of dropout in the Language Model RNN')
parser.add_argument('--dropW', type=float, default=0.4,
help='strength of dropout in the Language Model RNN')
parser.add_argument('--dropout', type=float, default=0.2,
help='strength of dropout in the Language Model RNN')
# Optimization: for the Language Model
parser.add_argument('--optimizer', type=str, default='Adam',
help='what update to use? rmsprop|sgd|sgdmom|adagrad|Adam')
# parser.add_argument('--learning_rate', type=float, default=0.001,
# help='learning rate')
parser.add_argument('--learning_rate', type=float, help='learning rate')
parser.add_argument('--optim_alpha', type=float, default=0.9,
help='alpha for adam')
parser.add_argument('--optim_beta', type=float, default=0.999,
help='beta used for adam')
parser.add_argument('--optim_epsilon', type=float, default=1e-8,
help='epsilon that goes into denominator for smoothing')
parser.add_argument('--weight_decay', type=float, default=0,
help='weight_decay')
parser.add_argument('--seed', type=int, default=1000,
help='seed')
parser.add_argument('--ntokens', type=int, default=777,
help='ntokens')
parser.add_argument('--load_checkpoint_path')
parser.add_argument('--checkpoint_path', type=str, default='',
help='directory to store checkpointed models')
# Other params that probably don't need to be changed
parser.add_argument('--load_model_states', type=str, default=0,
help='which model to load')
parser.add_argument('--evaluate_every', type=int, default=300,
help='which model to load')
parser.add_argument('--test', action='store_true')
parser.add_argument('--predict', action='store_true')
parser.add_argument('--predict_checkpoint', type=str)
parser.add_argument('--change_scores_every_epoch', action='store_true')
parser.add_argument('--test_has_regularization_split', action='store_true')
parser.add_argument('--apply_answer_weight', action='store_true')
parser.add_argument('--ignore_counting_questions', action='store_true')
parser.add_argument('--num_workers', type=int, default=16)
parser.add_argument('--lr_milestones', type=int, nargs='+', default=[5, 10])
parser.add_argument('--lr_gamma', type=float, default=1)
# SCR loss parameters
parser.add_argument('--use_scr_loss', action='store_true')
parser.add_argument('--num_sub', type=int, default=5,
help='size of the proposal object set')
parser.add_argument('--bucket', type=int, default=4,
help='bucket of predicted answers')
parser.add_argument('--scr_hint_loss_weight', type=float, default=0,
help='Influence strength loss weights')
parser.add_argument('--scr_compare_loss_weight', type=float, default=0,
help='self-critical loss weights')
parser.add_argument('--reg_loss_weight', type=float, default=0.0,
help='regularization loss weights, set to zero in our paper ')
# Parameters for the main VQA loss
parser.add_argument('--vqa_loss_weight', type=float, default=1)
if s is not None:
args = parser.parse_args("")
else:
args = parser.parse_args()
return args