-
Notifications
You must be signed in to change notification settings - Fork 2
/
load_config.m
149 lines (118 loc) · 5.73 KB
/
load_config.m
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
function load_config(train_mode)
% Sets all configurations used in the ACTOR project
%
% options for train_mode: 'train', 'test', 'val'
% Init config
global CONFIG
CONFIG = {};
if ~exist('train_mode', 'var')
train_mode = 'train';
end
CONFIG.train_mode = train_mode;
CONFIG.always_reset_reconstruction = 0;
CONFIG.sequence_length_train = 10;
CONFIG.sequence_length_eval = 10;
fprintf('CONFIG MODE SET TO: (%s)\n', CONFIG.train_mode);
CONFIG.output_dir = strcat('../rl-results/', datestr(now, 'yyyy-mm-dd_HH.MM.SS'), '/');
mkdir(CONFIG.output_dir);
% Caffe / system parameters
helper = Helpers();
CONFIG.use_gpu = 1;
CONFIG.gpu_id = 0;
CONFIG.rng_seed_caffe = 0;
CONFIG.rng_seed = 0;
% ------------------ POSE PREDICTOR PARAMS - START ----------------------------------
CONFIG.predictor_collect_blobs_from = {'conv4_4_CPM'};
CONFIG.predictor_use_blob = 'conv4_4_CPM';
CONFIG.predictor_limited_caching = 0;
CONFIG.training_eps_per_scene = 1; % Number of training eps before changing scene.
CONFIG.triangulation_max_reproj_err = 100; % Limit for maximum reprojection error of triangulated points.
% ------------------ POSE PREDICTOR PARAMS - END ------------------------------------
% ------------------ TRAINING/EVALUATION PARAMS - START -------------------
CONFIG.training_agent_snapshot_ep = 5000;
CONFIG.training_agent_nbr_eps = 80000;
CONFIG.training_agent_eps_per_batch = 5;
CONFIG.training_agent_lr = 1e-6;
CONFIG.caffe_lr_policy = 'multistep';
CONFIG.training_agent_lr_update_factor = 0.5;
CONFIG.training_agent_lr_update_steps = [1500000];
CONFIG.training_agent_batch_size = 40;
CONFIG.training_agent_act_funcs = {'ReLU', 'ReLU', 'TanH', 'TanH', 'TanH', 'TanH'};
CONFIG.agent_ms = {{[1, 10], [25, 50], 1}};
helper.set_agent_m_values();
CONFIG.training_baselines = {'random'};
CONFIG.evaluation_mode = 'val';
CONFIG.evaluation_init_snapshot = 0; % eval at first iteration already? (set to 1 in hdock)
CONFIG.evaluation_nbr_actions = [10]; % Hdock winner decided by first entry
CONFIG.evaluation_compute_all_fixations = 0;
CONFIG.evaluation_baselines = {'random', 'max_azim', 'oracle'};
CONFIG.baselines_stop_triangulated = 0; % 0 --> follow agent, 1 --> tri-stop individually
% ------------------ TRAINING/EVALUATION PARAMS - END ---------------------
% ------------------ RL AGENT PARAMS - START ------------------------------
% Agent protos
CONFIG.agent_proto_folder = './models/agent/';
CONFIG.agent_proto_name = 'actor'; % leave blank '' if only "train.proto"
% Other agent settings
CONFIG.agent_solver_proto = helper.set_proto_name('solver');
CONFIG.agent_train_proto = helper.set_proto_name('train');
CONFIG.agent_deploy_proto = CONFIG.agent_train_proto;
CONFIG.agent_random_init = 1; % 1=random init, 0=load agent_weights
CONFIG.agent_weights = '';
CONFIG.agent_stop_triangulated = 1; % 1 --> skip other stopping rules and stop when all triangulated
CONFIG.agent_nbr_actions = 10; % <= 0: auto stop, >= 1: exactly that many cams
CONFIG.agent_max_traj_len = 10;
CONFIG.agent_step_penalty = -0.07;
CONFIG.agent_same_cam_penalty = -2.5;
CONFIG.agent_eval_same_cam_strategy = ''; % '', 'random' or 'continue'
% ------------------ RL AGENT PARAMS - END --------------------------------
% ------------------ DATASET/PANOPTIC PARAMS - START ----------------------
CONFIG.panoptic_joint_mapping = [9, 11, 1, 12, 13, 14, 5, 6, 7, 15, 16, 17, 2, 3, 4];
CONFIG.panoptic_scaling_factor = 0.1; % The ratio between cm and mm
CONFIG.panoptic_scene_filter_train = {}; % {'pose'} --> exclude all "pose" scenes
CONFIG.panoptic_scene_filter_eval = {};
CONFIG.panoptic_match_cost_thresh = 0.5;
CONFIG.panoptic_instance_cache_suffix = 'instance_tuned.mat';
CONFIG.panoptic_instance_cache_suffix_det = '_instance_40k_tuned_2k.mat';
CONFIG.panoptic_min_iou = 0.4;
CONFIG.panoptic_nbr_appearance_samples = 10;
CONFIG.appearance_gt_iou_occluded = 0.75; % if 1, never occluded
CONFIG.dataset_video_format = 'hd';
CONFIG.dataset_path = strcat('../panoptic-data-hd-final/', CONFIG.train_mode, '/');
CONFIG.dataset_cache = strcat('../panoptic-cache-hd-final-multi-id/', CONFIG.train_mode, '/');
CONFIG.panoptic_error_type = '3d';
% ------------------ DATASET/PANOPTIC PARAMS - END ------------------------
% ------------------ MISC PARAMS - START ----------------------------------
CONFIG.stats_ma_weight = 0.0002;
CONFIG.stats_defaults_show_plots = 0;
CONFIG.stats_save_plots = 1;
CONFIG.stats_print_iter = 10;
CONFIG.stats_disabled_stats = {};
CONFIG.use_recorder = 0;
CONFIG.recorder_save_video = 1;
CONFIG.recorder_reset_video = 0;
CONFIG.recorder_frame_rate = 1;
CONFIG.recorder_full_body = 0;
% ------------------ MISC PARAMS - END ------------------------------------
% ------------------ CONSTANTS - START ------------------------------------
CONFIG.dataset_img_dimensions = [1080, 1920]; % row, cols
CONFIG.panoptic_crop_margin = [50, 75]; % x, y
CONFIG.agent_elev_angle_max = 0.529450481839650; % can predict elev \in [-elev_max, elev_max]
% Angle canvas and camera rig canvas
CONFIG.agent_use_angle_canvas = 1;
CONFIG.agent_nbr_bins_azim = 9;
CONFIG.agent_nbr_bins_elev = 5;
CONFIG.agent_use_rig_canvas = 1;
CONFIG.agent_use_joints_triangulated = 1;
% ------------------ CONSTANTS - END --------------------------------------
% ------------------ DATASET PREPROCESSING - START ------------------------
% The flags below are constants that are used to balance and preprocess
% the datasets. Changing these is not necessary.
CONFIG.sequence_step = 5; % Divides dataset fps by 5
CONFIG.training_sequence_overlap_pose = 0.5;
CONFIG.training_sequence_overlap_mafia = 0.5;
CONFIG.training_sequence_overlap_ultimatum = 0.5;
CONFIG.evaluation_sequence_overlap = 0.50; % how much overlap in sequences for testing model?
% ------------------ DATASET PREPROCESSING - END --------------------------
helper.set_solver_proto();
helper.set_train_proto();
end