-
Notifications
You must be signed in to change notification settings - Fork 0
/
opts.py
136 lines (119 loc) · 9.97 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
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--model", default="futr", help='model type')
parser.add_argument("--mode", default="train_eval", help="select action: [\"train\", \
\"predict\", \"train_eval\"]")
parser.add_argument("--dataset", type=str, default='breakfast')
parser.add_argument('--predict', "-p", action='store_true', help="predict for whole videos mode")
parser.add_argument('--wandb', type=str, default='project name', help="wandb runs name")
parser.add_argument("--single_sample_eval", action='store_true', help='evaluating on single sample')
#Dataset
parser.add_argument("--mapping_file", default="./datasets/breakfast/mapping.txt")
parser.add_argument("--features_path", default="./datasets/breakfast/features/")
parser.add_argument("--gt_path", default="./datasets/breakfast/groundTruth/")
parser.add_argument("--split", default="1", help='split number')
parser.add_argument("--file_path", default="./datasets/breakfast/splits")
parser.add_argument("--model_save_path", default="./save_dir/models/transformer")
parser.add_argument("--results_save_path", default="./save_dir/results/transformer")
parser.add_argument("--task", type=str, help="Next Action Anticipation/long-term anticipation", default='long')
#Training options
parser.add_argument("--batch_size", type=int, default=32)
parser.add_argument("--test_batch_size", type=int, default=1)
parser.add_argument("--epochs", type=int, default=50)
parser.add_argument("--warmup_epochs", type=int, default=10)
parser.add_argument("--workers", type=int, default=1)
parser.add_argument("--lr", type=float, default=1e-3)
parser.add_argument("--weight_decay", type=float, default=5e-3)
parser.add_argument("-warmup", '--n_warmup_steps', type=int, default=500)
parser.add_argument("--cpu", action='store_true', help='run in cpu')
parser.add_argument("--sample_rate", type=int, default=6)
parser.add_argument("--obs_perc", default=30)
parser.add_argument("--n_query", type=int, default=8)
#FUTR specific parameters
parser.add_argument("--n_head", type=int, default=8)
parser.add_argument("--hidden_dim", type=int, default=512)
parser.add_argument("--n_encoder_layer", type=int, default=2)
parser.add_argument("--n_decoder_layer", type=int, default=1)
parser.add_argument("--dropout", type=float, default=0.5)
parser.add_argument("--input_dim", type=int, default=2048)
#Model parameters
parser.add_argument("--seg", action='store_true', help='action segmentation')
parser.add_argument("--anticipate", action='store_true', help='future anticipation')
parser.add_argument("--pos_emb", action='store_true', help='positional embedding')
parser.add_argument("--max_pos_len", type=int, default=2000, help='position embedding number for linear interpolation')
#Test on GT or decoded input
parser.add_argument("--input_type", default="i3d_transcript", help="select input type: [\"decoded\", \"gt\"]")
parser.add_argument("--runs", default=0, help="save runs")
################################
# Graph general
parser.add_argument('--device', type=str, default='cuda', help="Device used to train: gpu / cpu")
parser.add_argument('--first_time', action='store_true', help="Running to save data for the first time")
parser.add_argument('--affordance_only_propagation', action='store_true', help="Perform propagation of only affordances")
parser.add_argument('--kg_attn', default=True, help="Use knowledge guided attention")
parser.add_argument('--use_gsnn', default=True, help="Use GSNN based propagation")
# Graph specific
parser.add_argument('--importance_loss_weight', type=float, default=0.01, help="Weight of importance loss")
parser.add_argument('--frame_save_interval', type=int, default=50, help="Interval to save each frame of the video")
parser.add_argument('--vocab_size', type=int, default=75, help="Size of concept vocabulary: 74 for 50salads")
parser.add_argument('--init_conf', type=float, default=0.35, help="Initialization confidence")
parser.add_argument('--min_num_init', type=int, default=1, help="Minimum number of initialization nodes")
parser.add_argument('--gradient_clipping', type=float, default=50, help="Use gradient clipping, don't use if -1")
parser.add_argument('--num_inter_steps', type=int, default=1, help="Number of intermediate steps")
# Propagation
parser.add_argument('--encoder_dropout', type=float, default=0.2, help="Dropout for node encoder")
parser.add_argument('--encoder_share_weights', type=bool, default=False, help="Share weights in node encoder")
parser.add_argument('--state_dim', type=int, default=10, help="State dimension for propagation net")
parser.add_argument('--condition_propagation', action='store_true', help="Condition the propagation on observed video context")
parser.add_argument('--condition_propagation_dim', type=int, default=50, help="Dimensionality of the conditioning input for propagation")
parser.add_argument('--use_nodetypes', type=bool, default=True, help="Use nodetypes for propagation")
parser.add_argument('--use_modified_gat', type=bool, default=True, help="Use modified GATv2 for encoding")
parser.add_argument('--node_bias_size', type=int, default=2, help="Dimensionality for node bias")
parser.add_argument('--num_steps', type=int, default=1, help="Number of propagation steps")
# Context net
parser.add_argument('--context_architecture', type=str, default='gated', help="Architecture for content net (gated)")
parser.add_argument('--context_transfer_function', type=str, default='tanh', help="Activation for context net (tanh)")
parser.add_argument('--context_use_node_input', type=bool, default=True, help="Use node bias for context net")
parser.add_argument('--context_use_ann', type=bool, default=True, help="Use node annotation for context net")
parser.add_argument('--context_out_net_h_size', type=int, default=10, help="")
parser.add_argument('--context_out_net_num_layer', type=int, default=0, help="")
parser.add_argument('--context_dim', type=int, default=5, help="Dimensionality for context encoding")
# Importance net
parser.add_argument('--gamma', type=float, default=0.5, help="Discount rate for importance")
parser.add_argument('--importance_architecture', type=str, default='sigout', help="Architecture for importance net (sigout)")
parser.add_argument('--importance_transfer_function', type=str, default='tanh', help="Activation for importance net (tanh)")
parser.add_argument('--importance_use_node_input', type=bool, default=True, help="Use node bias for importance net")
parser.add_argument('--importance_use_ann', type=bool, default=False, help="Use node annotation for importance net")
parser.add_argument('--importance_out_net_h_size', type=int, default=-1, help="")
parser.add_argument('--importance_out_net_num_layer', type=int, default=0, help="")
parser.add_argument('--importance_out_net_use_mask', type=bool, default=False, help="")
parser.add_argument('--image_conditioned_importance', type=bool, default=False, help="")
parser.add_argument('--imp_threshold', type=float, default=0.5, help="Threshold for expansion based on importance")
parser.add_argument('--imp_threshold_learnable', action='store_true', help="Use learnable threshold for expansion based on importance")
parser.add_argument('--imp_masking', type=str, default='active_cond', help="Importance prediction masking approach: naive / active_cond")
parser.add_argument('--importance_over_predicted', action='store_true', help="Compute importance loss over only predicted values")
parser.add_argument('--append_glove_emb_imp', action='store_true', help="Append glove embeddings to the node bias of importance net")
parser.add_argument('--propagation_method', type=str, default='threshold_based', help="Method for selecting indices for propagation: top_k / threshold_based")
parser.add_argument('--top_k_prop', type=int, default=3, help="If top_k selected, then indices to propagate at every step")
parser.add_argument('--weighted_importance_loss', action='store_true', help="Use a sample based weighting for importance loss computation")
# TO REMOVE -- both
parser.add_argument('--importance_accumulation', action='store_true', help="Increase importance for objects previously identified as active")
parser.add_argument('--importance_accumulation_weight', type=float, default=0.1, help="Weight for the previously active node's importance increment")
# Rectification matrix
parser.add_argument('--rectification_method', type=str, default='diagonal', help="Method to compute the rectification matrix: diagonal / full / weighting")
parser.add_argument('--rectification_residual', action='store_true', help="Instead of learning the rectification matrix learn its residual")
# Feature extraction
parser.add_argument('--feature_extraction_mode', type=str, default='rgb', help="Type of feature extraction: RGB / flow / both")
parser.add_argument('--last_k_frames', type=int, default=21, help="Use last n frames for computing video features")
# Finetune
parser.add_argument('--finetune', action='store_true', help="Finetune action anticipation using demonstations")
parser.add_argument('--demo_predict', action='store_true', help="Evaluate action anticipation on collected demonstations")
parser.add_argument('--demo_data_path', type=str, default='./demo_trajectories/', help="Path to collected demonstration data")
parser.add_argument('--prob_demos', type=float, default=0.7, help="Probability of choosing a vector from collected demonstration")
parser.add_argument('--scene_objects', type=str, default=['bowl', 'plate', 'tomato', 'lettuce',
'cucumber', 'knife', 'salt', 'pepper', 'vinegar',
'oil', 'spoon', 'spatula', 'glass', 'cheese'],
help="Objects placed inside the scene")
# Real world HRC
parser.add_argument('--entropy_threshold', type=float, default=1., help="Threshold for entropy for certainty")
parser.add_argument('--min_pred_length', type=int, default=100, help="Min number frames, required for history collection")
parser.add_argument('--max_pred_length', type=int, default=1000, help="Max number frames, after this sliding window starts")