radip/parameters.example.py at master · azyner/radip · GitHub

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
# Example parameters function
# To be renamed as parameters.py in local only, git set to ignore. This is such that I do not have to push param
# changes to git, they should all exist in a log file anyway
import numpy as np
import random
import os

parameters = {}
##### GLOBAL
parameters['AAA'] = "Example parameters file."

##### DATASETS
parameters['data_list'] = [
    'queen-hanks',
    'leith-croydon',
    'roslyn-crieff',
    'oliver-wyndora',
    'orchard-mitchell'
]
parameters['test_csv'] = 'oliver-wyndora'
parameters['data_filename'] = 'intersections-dataset' # or 'short-debug'  # comment out if reading directly from csvs
parameters['short_wrangle'] = True

# Preprocessing
parameters['keep_large_vehicles'] = True
#parameters["input_columns"] = ['easting', 'northing', 'heading', 'speed']
parameters['ibeo_data_columns'] = ["relative_x", "relative_y", "relative_angle", "AbsVelocity"] #_X","AbsVelocity_Y","ObjectPredAge"]
parameters['reject_stopped_vehicles_before_intersection_enable'] = False
parameters['reject_stopped_vehicles_before_intersection_speed'] = 1  # meters per second, 1 = 3.6kph, there is sensor noise meaning no true zero
parameters['reject_stopped_vehicles_before_intersection_duration'] = 1.0  # seconds

# Training Parameters
parameters["batch_size"] = 100
parameters["learning_rate"] = 0.0005
parameters["learning_rate_min"] = 0.00001
parameters["learning_rate_decay_factor"] = 0.9999
parameters['steps_per_checkpoint'] = 200
parameters['decrement_steps'] = 1000
parameters['long_training_time'] = 48*60  # Final training is for this long (minutes)
parameters['long_training_steps'] = 100000 #or steps, limit is whichever comes first
parameters["model_type"] = "MDN"
parameters['train_dir'] = 'train'

#NETWORK Hyperparams
parameters["observation_steps"] = 7
parameters["prediction_steps"] = 60
parameters["subsample"] = 2
parameters['embedding_size'] =256 # 512  # 64 for each input
parameters["rnn_size"] = 256
parameters["num_layers"] = 3
parameters["dropout_prob"] = 0.0
parameters["embedding_dropout"] = 0.0
parameters['RNN_cell'] = "sketch_hyper"

parameters['peephole_connections'] = True
parameters['l2_recurrent_decay'] = False
parameters['l2_lstm_input_decay'] = False
parameters["use_scaling"] = True

parameters['reg_embedding_beta'] = 0.0
parameters['l2_reg_beta'] = 0.0
parameters['input_mask'] = [1,1,1,1]
parameters["feed_future_data"] = False
parameters["first_loss_only"] = False
parameters['no_feedforward'] = False
parameters["max_gradient_norm"] = 10.0


parameters["num_mixtures"] = 6


parameters["velocity_threshold"] = 2.0
parameters["track_padding"] = True
parameters['sample_temperature'] = 1.0
parameters['padding_loss_logit_weight'] = 0.0000001  # 0.001
parameters['padding_loss_mixture_weight'] = 0.001 # 0.1
parameters['padding_loss_weight'] = 0.001 # 0.1 #deprecated

#Data augmentation (deprecated)
parameters['augmentation_chance'] = 0.0
parameters["random_bias"] = 0
parameters["random_rotate"] = False

##### HYPER SEARCH
parameters['early_stop_cf'] = 1*60  # Time in minutes for training one crossfold
parameters['hyper_search_folds'] = 0 # Number of hyper searching attempts.
parameters['hyper_search_step_cutoff'] = 80000
parameters['loss_decay_cutoff'] = 1e-20
parameters['hyper_rnn_size_fn'] = random.uniform
parameters['hyper_rnn_size_args'] = (128, 1025)
parameters['hyper_learning_rate_fn'] = random.uniform
parameters['hyper_learning_rate_args'] = (-10, -3)
parameters['aug_function'] = random.uniform
parameters['aug_range'] = (-3, 3)
parameters['evaluation_metric_type'] = 'euclidean_err_sum'  # "perfect_distance" / validation_accuracy
parameters['hyper_reg_embedding_beta_fn'] = random.uniform
parameters['hyper_reg_embedding_beta_args'] = (-10, -1)  # 10^X # OR None
parameters['hyper_reg_l2_beta_fn'] = random.uniform
parameters['hyper_reg_l2_beta_args'] = (-30, -8)  # 10^X # OR None
parameters['hyper_learning_rate_decay_args'] = (0.9995, 1.0)
parameters['hyper_learning_rate_min_args'] = (-10, -2)
parameters['hyper_padding_loss_logit_weight_args'] = (10, 0.0001)
parameters['hyper_padding_loss_mixture_weight_args'] = (10, 0.0001)
parameters['hyper_padding_loss_logit_weight_fn'] = random.uniform
parameters['hyper_padding_loss_mixture_weight_fn'] = random.uniform

#Clustering Parameters
parameters['cluster_mix_weight_threshold'] = 0.5
parameters['cluster_eps'] = 2
parameters['cluster_min_samples'] = 1

##### deprecated
#parameters['device'] = 'gpu:0'
parameters["n_folds"] = 5 #deprecated
parameters['debug'] = False  # Skip the metric computation to hasten looptime
parameters['d_thresh_top_n'] = 1   # How many samples to take that exist immediately before d_thresh

# IBEO
#parameters['ibeo_data_columns'] = ["Object_X","Object_Y","ObjBoxOrientation","AbsVelocity"]#_X","AbsVelocity_Y","ObjectPredAge"]
# Used to investigate the usefullness of an input parameter
parameters["data_format"] = "ibeo" # OR 'legacy'


#C hange this to 1 or zero to set the GPU to use
#os.environ["CUDA_VISIBLE_DEVICES"]="1"