-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathopts.py
186 lines (159 loc) · 11.4 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
import argparse
parser = argparse.ArgumentParser(description="PyTorch implementation of Temporal Segment Networks")
parser.add_argument('dataset', type=str)
parser.add_argument('modality', type=str, choices=['RGB', 'Flow'])
parser.add_argument('--train_list', type=str, default="")
parser.add_argument('--val_list', type=str, default="")
parser.add_argument('--root_path', type=str, default="")
parser.add_argument('--store_name', type=str, default="")
# ========================= Model Configs ==========================
parser.add_argument('--arch', type=str, default="BNInception")
parser.add_argument('--num_segments', type=int, default=3)
# parser.add_argument('--consensus_type', type=str, default='avg')
parser.add_argument('--k', type=int, default=3)
parser.add_argument('--dropout', '--do', default=0.5, type=float,
metavar='DO', help='dropout ratio (default: 0.5)')
# parser.add_argument('--loss_type', type=str, default="nll",
# choices=['nll'])
parser.add_argument('--pretrain', type=str, default='imagenet')
parser.add_argument('--tune_from', type=str, default=None, help='fine-tune from checkpoint')
# ========================= Learning Configs ==========================
parser.add_argument('--epochs', default=50, type=int, metavar='N',
help='number of total epochs to run') # TODO(changed from 120 to 50)
parser.add_argument('-b', '--batch-size', default=128, type=int,
metavar='N', help='mini-batch size (default: 256)')
parser.add_argument('--lr', '--learning-rate', default=0.001, type=float,
metavar='LR', help='initial learning rate')
parser.add_argument('--lr_type', default='step', type=str,
metavar='LRtype', help='learning rate type')
parser.add_argument('--lr_steps', default=[20, 40], type=float, nargs="+", # TODO(changed from [50,100] to [20,40])
metavar='LRSteps', help='epochs to decay learning rate by 10')
parser.add_argument('--momentum', default=0.9, type=float, metavar='M',
help='momentum')
parser.add_argument('--weight-decay', '--wd', default=1e-4, type=float,
metavar='W', help='weight decay (default: 1e-4)') # TODO(changed from 5e-4 to 1e-4)
parser.add_argument('--clip-gradient', '--gd', default=20, type=float,
metavar='W', help='gradient norm clipping (default: disabled)') # TODO(changed from None to 20)
parser.add_argument('--no_partialbn', '--npb', default=False, action="store_true")
# ========================= Monitor Configs ==========================
parser.add_argument('--print-freq', '-p', default=20, type=int,
metavar='N', help='print frequency (default: 10)')
parser.add_argument('--eval-freq', '-ef', default=1, type=int,
metavar='N', help='evaluation frequency (default: 1)') # TODO(changed from 5 to 1)
# ========================= Runtime Configs ==========================
parser.add_argument('-j', '--workers', default=8, type=int, metavar='N',
help='number of data loading workers (default: 8)')
parser.add_argument('--resume', default='', type=str, metavar='PATH',
help='path to latest checkpoint (default: none)')
parser.add_argument('-e', '--evaluate', dest='evaluate', action='store_true',
help='evaluate model on validation set')
parser.add_argument('--snapshot_pref', type=str, default="")
parser.add_argument('--start-epoch', default=0, type=int, metavar='N',
help='manual epoch number (useful on restarts)')
parser.add_argument('--gpus', nargs='+', type=int, default=None)
parser.add_argument('--root_log', type=str, default='logs')
parser.add_argument('--root_model', type=str, default='checkpoint')
parser.add_argument('--dense_sample', default=False, action="store_true", help='use dense sample for video dataset')
# TODO(yue) ADAPTIVE RESEARCH HYPER-PARAMETERS
parser.add_argument('--exp_header', default="default", type=str, help='experiment header')
parser.add_argument('--rescale_to', default=224, type=int)
# TODO(yue) adaptive resolution and skipping (hardcoded version)
parser.add_argument('--ada_reso_skip', action='store_true', help='adaptively select scale and choose to skip')
parser.add_argument('--reso_list', default=[224], type=int, nargs='+', help="list of resolutions")
parser.add_argument('--skip_list', default=[], type=int, nargs='+', help="list of frames to skip")
parser.add_argument('--backbone_list', default=[], type=str, nargs='+', help="backbones for diff resos")
parser.add_argument('--shared_backbone', action='store_true', help="share same backbone weight")
parser.add_argument('--accuracy_weight', default=1., type=float)
parser.add_argument('--efficency_weight', default=0., type=float)
parser.add_argument('--show_pred', action='store_true')
# TODO(yue) multi-label cases (for activity-net-v1.3)
# Always provides (single + multi) mAPs. Difference is only in training
parser.add_argument('--loss_type', type=str, default="nll", choices=['nll', 'bce'])
# TODO(yue) for policy network:
parser.add_argument('--policy_backbone', default='mobilenet_v2', type=str, help="backbones for policy network")
parser.add_argument('--policy_input_offset', default=0, type=int, help="select which scale for policy network")
parser.add_argument('--hidden_dim', default=512, type=int, help="dimension for hidden state and cell state")
parser.add_argument('--offline_lstm_last', action='store_true', help="just using LSTM(last one), no policy")
parser.add_argument('--offline_lstm_all', action='store_true', help="just using LSTM(all average), no policy")
parser.add_argument('--random_policy', action='store_true', help="just using random policy there")
parser.add_argument('--all_policy', action='store_true', help="just using all feat there")
parser.add_argument('--eff_loss_after', default=-1, type=int, help="use eff loss after X epochs")
parser.add_argument('--save_freq', default=10, type=int, help="freq to save network model weight") # TODO(yue)
parser.add_argument('--model_paths', default=[], type=str, nargs="+", help='path to load models for backbones')
parser.add_argument('--policy_path', default="", type=str, help="path of the policy network")
# TODO(yue) maybe we want to use ImageNet pretrain or not, depending on the resolution
# TODO(yue) annealing
parser.add_argument('--exp_decay', action='store_true', help="type of annealing")
parser.add_argument('--init_tau', default=5.0, type=float, help="annealing init temperature")
parser.add_argument('--exp_decay_factor', default=-0.045, type=float, help="exp decay factor per epoch")
# TODO(yue) small tweak
parser.add_argument('--policy_from_scratch', action='store_true', help="policy network without pretraining")
parser.add_argument('--frozen_list', default=[], type=str, nargs="+", help='list of frozen part')
parser.add_argument('--policy_also_backbone', action='store_true', help="use policy as the last backbone")
parser.add_argument('--uniform_loss_weight', type=float, default=1e-6, help="loss to constraints all uses equal")
parser.add_argument('--lite_mode', action='store_true')
# TODO(yue) try different losses for efficiency terms
parser.add_argument('--use_gflops_loss', action='store_true') # TODO(yue) use flops as loss assignment
parser.add_argument('--head_loss_weight', type=float, default=1e-6) # TODO(yue) punish to the high resolution selection
parser.add_argument('--frames_loss_weight', type=float, default=1e-6) # TODO(yue) use num_frames as a loss assignment
# TODO(yue) finetuning and testing
parser.add_argument('--base_pretrained_from', type=str, default='', help='for base model pretrained path')
parser.add_argument('--skip_training', action='store_true') # TODO(yue) just doing eval
parser.add_argument('--freeze_policy', action='store_true') # TODO(yue) fix the policy
# TODO(yue) reproducibility
parser.add_argument('--random_seed', type=int, default=1007)
# TODO(yue) for FCVID or datasets where eval is too heavy
parser.add_argument('--partial_fcvid_eval', action='store_true')
parser.add_argument('--partial_ratio', type=float, default=0.2)
# TODO(yue) crop
parser.add_argument('--center_crop', action='store_true')
parser.add_argument('--random_crop', action='store_true')
# TODO(yue) oracle scsampler (from ListenToLook ideas)
parser.add_argument('--consensus_type', type=str, default='avg') # TODO can also use scsampler!
parser.add_argument('--top_k', type=int, default=10) # TODO can also use scsampler!
# TODO(yue) real SCSampler (we also use --top_k to select frames, and use consensus type=='scsampler')
parser.add_argument('--real_scsampler', action='store_true')
parser.add_argument('--sal_rank_loss', action='store_true')
parser.add_argument('--frame_independent', action='store_true') # TODO use tsn in models_ada
parser.add_argument('--freeze_backbone', action='store_true')
# TODO(yue) reproducibility
parser.add_argument('--test_from', type=str, default="")
# TODO(yue) adaptive-cropping (only 1, 5, 9)
parser.add_argument('--ada_crop_list', default=[], type=int, nargs="+", help='num of anchor points per scaling')
# TODO(yue) visualizations
parser.add_argument('--save_meta', action='store_true')
parser.add_argument('--ablation', action='store_true')
parser.add_argument('--remove_all_base_0', action='store_true')
parser.add_argument('--save_all_preds', action='store_true')
# TODO(yue) RL
parser.add_argument('--default_signal', default=0, type=int)
parser.add_argument('--uniform_cross_entropy', action='store_true')
parser.add_argument('--ignore_new_fc_weight', action='store_true')
parser.add_argument('--no_weights_from_linear', action='store_true')
parser.add_argument('--uno_reso', action='store_true')
parser.add_argument('--use_reinforce', action='store_true')
parser.add_argument('--separated', action='store_true')
parser.add_argument('--no_baseline', action='store_true')
parser.add_argument('--detach_reward', action='store_true')
parser.add_argument('--negative_loss', action='store_true')
# TODO(yue) Exps/data paths
parser.add_argument('--data_dir', type=str, default="../../datasets/activity-net-v1.3")
parser.add_argument('--log_dir', type=str, default="../../logs_tsm")
# TODO RSEE
parser.add_argument('--irte', action='store_true', help='use RSEE')
parser.add_argument('--irte_joint', action='store_true', help='train RSEE and policy')
parser.add_argument('--blstm', action='store_true', help="use BLSTM for temporal exiting")
parser.add_argument('--blstm_hidden_dim', type=int, default=512, help='hidden_dim for blstm')
parser.add_argument('--irte_final', action='store_true', help='final version of RSEE')
parser.add_argument('--exit_loss', action='store_true', help='use exit loss')
parser.add_argument('--exit_w', type=float, default=0.7, help='exit weight')
parser.add_argument('--frm', type=str, default='none', help='feature reuse module')
parser.add_argument('--online_train', action='store_true', help='whether online train')
parser.add_argument('--load_imagenet', action='store_true', help='load imagenet pretrained weights')
parser.add_argument('--multi_blstm', action='store_true', help='prepare a blstm for every reso')
parser.add_argument('--fake_r', action='store_true', help='use fake r_all to all choose 224')
parser.add_argument('--kd', action='store_true', help='use kd')
parser.add_argument('--stan_lstm', action='store_true', help='use stan lstm')
parser.add_argument('--baseline', action='store_true', help='train baseline')
parser.add_argument('--TE', type=float, default=0.9, help='temporal exit threshold')
parser.add_argument('--vanilla', action='store_true', help='use vanilla backbone')