-
Notifications
You must be signed in to change notification settings - Fork 0
/
config.py
136 lines (98 loc) · 3.27 KB
/
config.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
# configurations
from easydict import EasyDict as edict
__C = edict()
cfg = __C
__C.training_epochs = 30 #300
# DataIO
__C.use_xyz = True
__C.use_phi_theta = False
__C.use_yaw_pitch_roll = False
__C.use_cos_sin = False
__C.process_in_seconds = True
__C.batch_size = 32
__C.fps = 30
__C.include_own_history = True
__C.own_history_only = True
if __C.process_in_seconds:
__C.predict_len = 2 #predict the future len
__C.running_length = 2 #history in seconds
__C.predict_step = 2 #predict n seconds during testing
else:
__C.predict_len = 10 #predict the future frames
__C.running_length = 10
__C.predict_step = 10
__C.skip_frame_gap = 10 #not consecutive frames, with a gap
__C.OUTPUT_DIR = './model/'
__C.fix_target_user = False
__C.target_uer_ID = None
__C.use_more_video = True #train on multiple videos except test_video_ind
__C.test_video_ind = None#3#8
__C.add_residual_link = False #output resdiual link
__C.add_layer_residual_link = False #residual within the RNN
__C.has_reconstruct_loss = False
__C.stuff_zero = False
__C.stuff_last = False
__C.change_xyz2xxxyyyzzz = False #the same for fc-lstm
__C.concat_state = False
# __C.LEARNING_RATE = 0.001
__C.LEARNING_RATE = 1e-5
__C.lr_epoch_step = 10
__C.clip_gradient = True
__C.pop_alpha = 0.5
__C.add_xyz_sum1 = False
__C.use_CNN_data_format = False
__C.use_convLSTM_data_format = True
__C.use_decoder = False
__C.shuffle_data = False
__C.stateful_across_batch = False
__C.dropout_rate = 0.3
__C.conv_kernel_size = 5
__C.recurrent_dropout_rate = 0.3 #for convLSTM
__C.predict_mean_var = False
__C.sample_and_refeed = True
__C.use_GMM = True #using mixture density output layer
__C.input_mean_var = False
__C.teacher_forcing = False
__C.use_one_hot = False
__C.predict_eos = False
__C.use_residual_input = False #step difference
__C.normalize_residual = False
__C.subsample_datadb = False
__C.berrnoulli_loss_weight = 10
__C.use_mixed_dataset = False #Tsinghua+icme_saliency
__C.use_overlapping_chunks = True
if __C.use_overlapping_chunks:
if __C.process_in_seconds:
__C.data_chunk_stride=10 #second
else:
__C.data_chunk_stride=10 #frame
else:
if __C.process_in_seconds:
__C.data_chunk_stride=__C.running_length*__C.fps
else:
__C.data_chunk_stride=__C.running_length
__C.linear_mode = None # 'presistence'
if __C.linear_mode!=None:
__C.linear_mode_residual = True
__C.need_split = False
__C.dilation_rate = 1 #for convlstm
__C.use_saliency = False
if __C.use_saliency:
__C.only_use_sal_local = True # only use local motion, since global motion is not very accurate
__C.use_cropped_sal = False
__C.cut_data_head = False
__C.purelly_testing = False #used for testing on PC dataset, no gt future for the last few seconds
if __C.process_in_seconds:
__C.use_heat_sum = True
else:
__C.use_heat_sum = False
__C.time_shift = False # 9 seconds in encoder, the 10th second is used as the first step of decoder
__C.enc_last_out_as_dec_in = False
__C.use_embedding_AME = False
#use an embedding first, instead of directly copying hidden states from the encoder to decoder
__C.embed_frame_state_enc2dec = False
__C.rescale_input = False #from [-1,1] to [0,1]
__C.include_time_ind = False
__C.target_user_only = False
# __C.thu_tag=''
__C.thu_tag='_thu_'