-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain.py
344 lines (274 loc) · 20.9 KB
/
train.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
import argparse
import importlib
from utils import *
def get_command_line_parser():
parser = argparse.ArgumentParser()
parser.add_argument('-project', type=str, default="orco")
parser.add_argument('-dataset', type=str, default='mini_imagenet',
choices=['mini_imagenet', 'cub200', 'cifar100'])
parser.add_argument('-dataroot', type=str, default="../datasets/")
parser.add_argument('-save_path_prefix', "-prefix", type=str, default="")
parser.add_argument('-gpu', default='0,1,2,3')
parser.add_argument('-num_workers', type=int, default=8)
parser.add_argument('-seed', type=int, default=1)
parser.add_argument('-debug', action='store_true')
parser.add_argument('-resume', action='store_true')
parser.add_argument('-momentum', type=float, default=0.9)
parser.add_argument('-gamma', type=float, default=0.1)
parser.add_argument('-temperature', type=int, default=16)
parser.add_argument('-start_session', type=int, default=1)
parser.add_argument('-model_dir', type=str, help='loading model parameter from a specific dir')
parser.add_argument('-epochs_base', type=int, default=100)
parser.add_argument('-epochs_joint', type=int, default=100)
parser.add_argument('-epochs_target_gen', type=int, default=1000)
parser.add_argument('-lr_base', type=float, default=0.1)
parser.add_argument('-lr_base_encoder', type=float, default=0.1)
parser.add_argument('-lr_new', type=float, default=0.1)
parser.add_argument('-optimizer', type=str, default='sgd',
choices=['sgd', 'adam', 'adamw', "mtadam"])
parser.add_argument('-optimizer_joint', type=str, default='sgd',
choices=['sgd', 'adam', 'adamw', "mtadam"])
parser.add_argument('-schedule', type=str, default='Step',
choices=['Step', 'Milestone', 'Cosine'])
parser.add_argument('-reserve_mode', type=str, default='all',
choices=["all", "full"])
parser.add_argument('-joint_schedule', type=str, default='Milestone',
choices=['Milestone', 'Cosine'])
parser.add_argument('-base_schedule', type=str, default='Milestone',
choices=['Milestone', 'Cosine'])
parser.add_argument('-decay', type=float, default=0.0005)
parser.add_argument('-decay_new', type=float, default=0)
parser.add_argument('-cos_n_lam', type=float, default=0.5)
parser.add_argument('-cos_b_lam', type=float, default=0.5)
parser.add_argument('-sup_lam', type=float, default=1)
parser.add_argument('-cos_lam', type=float, default=1)
parser.add_argument('-simplex_lam', type=float, default=1)
parser.add_argument('-perturb_epsilon_base', type=float, default=1e-2)
parser.add_argument('-perturb_epsilon_inc', type=float, default=1e-2)
parser.add_argument('-perturb_offset', type=float, default=0.5)
parser.add_argument('-base_mode', type=str, default='ft_dot',
choices=['ft_dot', 'ft_cos', 'ft_l2', "ft_dot_freeze"])
parser.add_argument('-fine_tune_backbone_base', action='store_true', help='')
parser.add_argument("-proj_type", type=str, default="proj",
choices=["proj", "proj_ncfscil"])
parser.add_argument('-batch_size_base', type=int, default=128)
parser.add_argument('-batch_size_sup_con', type=int, default=128)
parser.add_argument('-batch_size_new', type=int, default=0, help='set 0 will use all the availiable training image for new')
parser.add_argument('-batch_size_joint', type=int, default=0, help='set 0 will use all the availiable training image for new')
parser.add_argument('-batch_size_perturb', type=int, default=-1)
parser.add_argument('-test_batch_size', type=int, default=100)
parser.add_argument('-drop_last_batch', action="store_true", help="Drops the last batch if not equal to the assigned batch size")
parser.add_argument('-exemplars_count', type=int, default=-1)
parser.add_argument('-rand_aug_sup_con', action='store_true', help='')
parser.add_argument('-prob_color_jitter', type=float, default=0.8)
parser.add_argument('-min_crop_scale', type=float, default=0.2)
parser.add_argument('-warmup_epochs_base', type=int, default=3)
parser.add_argument('-warmup_epochs_inc', type=int, default=10)
# --- Other Params ---
# parser.add_argument('-test', action="store_true", help="Rerun the testing for the experiments")
# parser.add_argument('-experiment_path', type=str, default="")
# parser.add_argument('-max_inference', action="store_true", help="Run the new inference scheme and compute the accuracies")
# parser.add_argument('-t_scaling', action="store_true", help="run temperature scaling after the base session")
# parser.add_argument('-base_only', action="store_true", help="cross entropy with only base classes in the denominator")
# parser.add_argument('-balanced_testing', action="store_true", help="Do Balanced testing")
# # dino
# parser.add_argument('-all_crops', action="store_true", help="")
# parser.add_argument('-dino_outdim', type=int, default=60000)
# parser.add_argument('-dino_loss_weight', type=float, default=0.1)
# parser.add_argument('-dino_last_layer', type=str, default='nll',
# choices=['nll', 'unit', 'etf'])
# parser.add_argument('-use_multi_opt', action="store_true", help="")
# parser.add_argument('-dino_novel', action="store_true", help="")
# parser.add_argument('-dino_joint', action="store_true", help="")
# parser.add_argument('-class_level', action="store_true", help="Class level input for DINO vs instance level")
# parser.add_argument('-two_step', action="store_true", help="EASY, two step forward backward pass at each step")
# parser.add_argument('-cl2n', action="store_true", help="Feature transformation, centering l2 norm")
# parser.add_argument('-dino_over_fc', action="store_true", help="Feature transformation, centering l2 norm")
# parser.add_argument('-replace_mode', type=str, default='avg_cos', # CEC -> ft_cos
# choices=['ft_dot', 'ft_cos', 'avg_cos', 'ft_l2', 'dot']) # ft_dot means using linear classifier, ft_cos means using cosine classifier
# parser.add_argument('-mixup_start_epoch', type=int, default=100)
# parser.add_argument('-s2m2', action="store_true", help="applys2m2 training schedule for the base session")
# parser.add_argument('-classifier_last_layer', type=str, default='projection',
# choices=['linear', 'projection'])
# parser.add_argument('-classifier_hidden_dim', type=int, default=2048)
# parser.add_argument('-vis_encoder', action="store_true", help="Visualise encoder ")
# # Gaussian generator
# parser.add_argument('-use_vector_variance', action="store_true", help="Use vector variance for gaussian generation")
# parser.add_argument('-use_gaussian', action="store_true", help="Use the gaussian generated data for the joint session")
# parser.add_argument('-views_gaussian', type=int, default = 1)
# parser.add_argument('-aug_gaussian', action="store_true", help="augment the gaussian generated features by mixing them with other features from the same distribution")
# # ==== Scheduling parameters
# # teacher student temps
# parser.add_argument('-student_temp', type=float, default=0.1)
# parser.add_argument('-teacher_temp', type=float, default=0.07)
# # teacher temperature scheduling
# parser.add_argument('-teacher_scheduling', action="store_true", help="Scheudling the teacher temperature")
# parser.add_argument('-warmup_teacher_temp', type=float, default=0.04)
# parser.add_argument('-warmup_teacher_temp_epochs', type=int, default=25)
# # lr scheduling
# parser.add_argument('-schedule_lr', action="store_true", help="Cosine Scedhuling the lr ")
# parser.add_argument('-min_lr', type=float, default=0.0048)
# parser.add_argument('-warmup_epochs', type=int, default=10)
# # wd scheduling
# parser.add_argument('-schedule_wd', action="store_true", help="Cosine Scedhuling the wd")
# parser.add_argument('-weight_decay_end', type=float, default=0.0005)
# # teacher momentum scheduling
# parser.add_argument("-momentum_teacher", type=float, default=0.996)
# # contrast
# parser.add_argument('-ce_pretrain', action="store_true", help="Pretrain on supcon")
# # parser.add_argument('-sup_con_pretrain', action="store_true", help="Pretrain on supcon") # True
# parser.add_argument('-sup_con_base', action="store_true", help="Apply supcon in a multi task fashion during the base session") # False
# parser.add_argument('-use_sup_con_head', action="store_true", help="Have a projection mlp before the sup con loss for encoding projection") # True
# parser.add_argument('-sup_con_feat_dim', type=int, default=128) # 128
# parser.add_argument('-sup_con_freeze_backbone', action="store_true", help="freeze Backbone after pretraining on sup con") # True
# parser.add_argument('-sup_con_temp', type=float, default=0.07)
# parser.add_argument('-epochs_sup_con', type=int, default=50) # 128
# parser.add_argument('-lr_sup_con', type=float, default=0.05)
# parser.add_argument('-momentum_sup_con', type=float, default=0.9)
# parser.add_argument('-decay_sup_con', type=float, default=1e-4) # 0.0005 #1e-4
# parser.add_argument('-cosine_sup_con', action="store_true")
# parser.add_argument('-lr_decay_rate_sup_con', type=float, default=0.1) # 128
# parser.add_argument('-lr_decay_epochs_sup_con', type=str, default='100,200',
# help='where to decay lr, can be a list')# True
# parser.add_argument('-warm_sup_con', action="store_true")
# parser.add_argument('-margin_sup_con', type=float, default=0)
# parser.add_argument('-proto_init_sup_con', action="store_true")
# parser.add_argument('-mixup_sup_con', action="store_true")
# parser.add_argument('-train_base_in_novel', action="store_true")
# parser.add_argument('-train_novel_in_joint', action="store_true")
# parser.add_argument('-novel_high_layers', action="store_true")
# parser.add_argument('-reserve_method', type=str, default='dist',
# choices=['dist', 'angle', 'mixup', "hemisphere"])
# parser.add_argument('-skip_encode_norm', action='store_true', help='')
# parser.add_argument('-skip_sup_con_head', action='store_true', help='')
# parser.add_argument('-skip_nonlinearity', action='store_true', help='')
# parser.add_argument('-skip_novel', action='store_true', help='')
# parser.add_argument('-new_head_hidden_dim', type=int, default=-1)
# parser.add_argument('-incremental_on_supcon_head', action='store_true', help='')
# parser.add_argument('-fine_tune_backbone_joint', action='store_true', help='')
# parser.add_argument('-nesterov_new', action='store_true', help='')
# parser.add_argument('-joint_loss', type=str, default='ce_even',
# choices=['ce_even', 'ce_inter', "ce", "ce_weighted", "tsc"])
# parser.add_argument('-proto_method', type=str, help='Method used for prototype calculation', default="mean", choices=["mean", "gm"])
# parser.add_argument('-hm_patience', type=int, default=15)
# parser.add_argument('-compute_hardnegative', action="store_true")
# # about dataset and network
# parser.add_argument('-sweep_json', type=str, default="sweep_config/dino_sweep.json")
# parser.add_argument('-freeze_backbone', action="store_true", help="Backbone of the resnet model is frozen")
# parser.add_argument('-skip_wandb', action="store_true", help="Skip syncing with wandb")
# # about data augmentation
# parser.add_argument('-color_jitter', action="store_true", help="Apply Color Jitter")
# parser.add_argument('-data_aug', action="store_true", help="Perform 2 image data augmentation")
# parser.add_argument('-rand_aug', action="store_true", help="Perform random augmentation on CUB200")
# parser.add_argument('-milestones', nargs='+', type=int, default=[60, 70]) # Cub200 -> [60, 120, 200]
# parser.add_argument('-step', type=int, default=40)
# parser.add_argument('-T_0_den', type=int, default=2, help="number of iterations for the first restart")
# parser.add_argument('-eta_min', type=float, default=1e-2, help="Minimum learning rate when using cosine restart scheduler")
# parser.add_argument('-joint_loading_method', type=str, default='old',
# choices=['old', 'new', 'online'])
# parser.add_argument('-not_data_init', action='store_true', help='using average data embedding to init or not')
# parser.add_argument('-label_smoothing', type=float, default=0.0)
# parser.add_argument('-label_smoothing_novel', type=float, default=0.0)
# parser.add_argument('-label_smoothing_joint', type=float, default=0.0)
# parser.add_argument('-new_mode', type=str, default='ft_dot', # CEC -> avg_cos
# choices=['ft_dot', 'ft_cos', 'avg_cos', 'ft_l2', 'dot']) # ft_dot means using linear classifier, ft_cos means using cosine classifier, avg_cos means using average data embedding and cosine classifier
# # for episode learning
# parser.add_argument('-train_episode', type=int, default=50)
# parser.add_argument('-episode_shot', type=int, default=1)
# parser.add_argument('-episode_way', type=int, default=15)
# parser.add_argument('-episode_query', type=int, default=15)
# # parser.add_argument('-validation_metric', type=str, default='hm', choices=["hm", "loss", "acc", "none"])
# # parser.add_argument('-validation_metric_novel', type=str, default='none', choices=["hm", "loss", "acc", "none"])
# # parser.add_argument('-validation_metric_joint', type=str, default='none', choices=["hm", "loss", "acc", "none"])
# parser.add_argument('-schedule_joint', action='store_true', help='Schedule the learning rate using cosine scheduler')
# parser.add_argument('-rand_aug_joint', action='store_true', help='Schedule the learning rate using cosine scheduler')
# parser.add_argument('-rand_aug_novel', action='store_true', help='Schedule the learning rate using cosine scheduler')
# # for cec
# parser.add_argument('-lrg', type=float, default=0.1) #lr for graph attention network
# parser.add_argument('-low_shot', type=int, default=1)
# parser.add_argument('-low_way', type=int, default=15)
# parser.add_argument('-set_no_val', action='store_true', help='set validation using test set or no validation')
# # Reserve vector
# parser.add_argument('-apply_bnce', action='store_true')
# parser.add_argument('-use_ncfscil', action='store_true')
# parser.add_argument('-instance_mixup', action='store_true')
# parser.add_argument('-instance_mixup_joint', action='store_true')
# parser.add_argument('-strong_transform', action='store_true')
# parser.add_argument('-criterion', type=str, default='xent', # CEC -> ft_cos
# choices=["xent", "cosine", "xent+cosine", "cosine-squared", "none"]) # ft_dot means using linear classifier, ft_cos means using cosine classifier
# parser.add_argument('-radial_label_smoothing', type=float, default=0)
# parser.add_argument('-xent_weight', type=float, default=0.1)
# parser.add_argument('-kp_lam', type=float, default=0.0)
# parser.add_argument('-lam_at', type=float, default=0)
# # parser.add_argument('-joint_supcon', action='store_true')
# parser.add_argument('-testing_freq', type=float, default=5)
# # parser.add_argument('-online_assignment', action='store_true')
# # parser.add_argument("-assignment_mode_base", type=str, default="max",
# # choices=["min", "random", "max"])
# # parser.add_argument("-assignment_mode_novel", type=str, default="max",
# # choices=["min", "random", "max", "cosine_penalty"])
# # parser.add_argument('-assign_flip', action="store_true", help="flip assignment mode between min and max every session")
# # parser.add_argument('-reserve_init', type=str, default='randn', # CEC -> ft_cos
# # choices=["randn", "proto"]) # ft_dot means using linear classifier, ft_cos means using cosine classifier
# parser.add_argument('-novel_bias', action="store_true")
# parser.add_argument('-queue_pull', action="store_true")
# parser.add_argument('-warmup_epochs_base', type=int, default=3)
# parser.add_argument('-warmup_epochs_inc', type=int, default=10)
# parser.add_argument('-warmup_epochs_simplex', type=int, default=0)
# # parser.add_argument('-assign_similarity_metric', type=str, default='cos', # CEC -> ft_cos
# # choices=["cos", "euclidean", "mahalanobis", "cos_odd_inv"]) # ft_dot means using linear classifier, ft_cos means using cosine classifier
# parser.add_argument('-base_target_sampling', action="store_true")
# parser.add_argument('-target_sampling', action="store_true")
# parser.add_argument('-supcon_views', type=int, default=2)
# parser.add_argument('-supcon_even', action="store_true")
# parser.add_argument('-reserve_vector_count', type=int, default=-1)
# parser.add_argument('-proj_output_dim', type=int, default=-1)
# parser.add_argument('-proj_hidden_dim', type=int, default=2048)
# parser.add_argument('-init_proj_random', action="store_true")
# parser.add_argument('-skip_base_ft', action="store_true")
# parser.add_argument('-append_hard_positives', action="store_true")
# # parser.add_argument('-mix_lam', type=float, default=0)
# parser.add_argument("-pull_loss_mode", type=str, default="default",
# choices=["default"])
# parser.add_argument('-cos_b_margin', type=float, default=0)
# parser.add_argument('-cos_n_margin', type=float, default=0)
# parser.add_argument('-dist_lam', type=float, default=0.5)
# parser.add_argument('-proj_ema_update', action="store_true")
# parser.add_argument('-proj_ema_mode', type=str, default="prev",
# choices=["prev", "all", "base"])
# parser.add_argument('-proj_ema_beta', type=float, default=0.9999)
# parser.add_argument('-proj_ema_update_every', type=int, default=10)
# # Extra Experiments
# parser.add_argument('-heavy_inter_aug', action="store_true")
# parser.add_argument('-skip_perturbation', action="store_true")
# parser.add_argument('-skip_prep', action="store_true")
# parser.add_argument('-skip_orth', action="store_true")
# parser.add_argument('-spread_aware', action="store_true", help="Pull the class features in the direction of the target offset by the original offset to retain geometry")
# parser.add_argument('-class_aug_base', action="store_true", help="Class augmentation using mixup during base session")
# parser.add_argument('-class_aug_inc', action="store_true", help="Class augmentation using mixup during incremental session")
# parser.add_argument('-skip_perturbation_base', action="store_true", help="Class augmentation using mixup during incremental session")
# parser.add_argument('-skip_perturbation_inc', action="store_true", help="Class augmentation using mixup during incremental session")
# parser.add_argument('-remove_pert_numerator', action="store_true", help="Remove perturbations from numerator")
# parser.add_argument('-remove_curr_features', action="store_true", help="Removes current task features from sup con loss")
#
# # parser.add_argument('-apply_tbnce', action="store_true", help="Apply tbnce in all sessions")
# parser.add_argument('-pull_delay', type=float, default=0)
# parser.add_argument('-novel_bias_schedule', type=int, default=-1)
# # parser.add_argument("-perturb_mode", type=str, default="inc-curr-base",
# # choices=["inc-curr-base", "all", "inc+curr-base", "inc-curr+base", "all-ua", "inc+curr-base-ua"])
# # parser.add_argument("-perturb_dist", type=str, default="uniform",
# # choices=["uniform", "gaussian"])
# # parser.add_argument("-perturb_mode_base", type=str, default="inc",
# # choices=["all", "inc"])
# # parser.add_argument("-perturb_mode_inc", type=str, default="inc-curr-base",
# # choices=["inc-curr-base", "all", "inc+curr-base", "inc-curr+base"])
return parser
if __name__ == '__main__':
# Parse Arguments
parser = get_command_line_parser()
args = parser.parse_args()
set_seed(args.seed)
args.num_gpu = set_gpu(args)
# Trainer initialization
trainer = importlib.import_module('models.%s.fscil_trainer' % (args.project)).FSCILTrainer(args)
trainer.train()