stackrl/config.gin at master · menezesandre/stackrl · 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
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
# Set environment parameters

# envs.make.env='Stack-v0'
envs.make.n_parallel=2

# envs.make.episode_length=30
# envs.make.urdfs='[5-9]?'
# envs.make.object_max_dimension=0.125
# envs.make.use_gui=False
# envs.make.simulator=None
envs.make.sim_time_step=.0125
# envs.make.gravity=9.8
# envs.make.num_sim_steps=None
# envs.make.velocity_threshold=0.01
# envs.make.smooth_placing = True
# envs.make.observer=None
# envs.make.observable_size_ratio=4
# envs.make.resolution_factor=5
# envs.make.max_z=1
envs.make.rewarder='dor'
# envs.make.goal_size_ratio=.25
envs.make.reward_scale=None
# envs.make.reward_params=None
# envs.make.flat_action=True
# envs.make.dtype='float32'

# envs.make.curriculum = {'goals':[], 'urdfs':[]}

# eval/envs.make.episode_length = 40
# eval/envs.make.urdfs = 50
# eval/envs.make.curriculum = None
eval/envs.make.rewarder='iou'
eval/envs.make.reward_scale=1.
eval/envs.make.sim_time_step=0.01

# Set network parameters

# left/nets.unet.filters=32
# left/nets.unet.depth=3
# left/nets.unet.out_channels=None
# left/nets.unet.out_activation=None
# left/nets.unet.kernel_initializer=None

# right/nets.unet.filters=32
# right/nets.unet.depth=3
# right/nets.unet.out_channels=None
# right/nets.unet.out_activation=None
# right/nets.unet.kernel_initializer=None

# nets.correlation.parallel_iterations=None

# nets.pos_layers.filters=32
# nets.pos_layers.depth=2

nets.DeepQSiamFCN.left_filters = 16
nets.DeepQSiamFCN.dueling = True
nets.DeepQSiamFCN.dueling_units = 256
# nets.DeepQSiamFCN.corr_channels = 16
nets.DeepQSiamFCN.pos_filters = 16

# Schedules

# schedules.ExponentialDecay.initial_learning_rate = 0.001
# schedules.ExponentialDecay.decay_steps = 100000
# schedules.ExponentialDecay.decay_rate = 0.1
# schedules.ExponentialDecay.staircase = False

# schedules.PolynomialDecay.initial_learning_rate = 0.001
# schedules.PolynomialDecay.decay_steps = 100000
# schedules.PolynomialDecay.end_learning_rate = 0.0001
# schedules.PolynomialDecay.power = 1.0

exploration/schedules.PolynomialDecay.initial_learning_rate = 1.0
exploration/schedules.PolynomialDecay.decay_steps = 400000
exploration/schedules.PolynomialDecay.end_learning_rate = .1
exploration/schedules.PolynomialDecay.power = 1.0

beta/schedules.PolynomialDecay.initial_learning_rate = 0.4
beta/schedules.PolynomialDecay.decay_steps = 400000
beta/schedules.PolynomialDecay.end_learning_rate = 1.0
beta/schedules.PolynomialDecay.power = 1.0

# lr/schedules.PolynomialDecay.initial_learning_rate = 0.00025
# lr/schedules.PolynomialDecay.decay_steps = 400000
# lr/schedules.PolynomialDecay.end_learning_rate = 0.0000625
# lr/schedules.PolynomialDecay.power = 2.0

# Set optimizer parameters

optimizers.Adam.learning_rate = 0.0000625
# optimizers.Adam.learning_rate = @lr/tf.keras.optimizers.schedules.ExponentialDecay()
optimizers.Adam.beta_1 = 0.95
optimizers.Adam.beta_2 = 0.95
# optimizers.Adam.epsilon = 0.0000001
# optimizers.Adam.amsgrad = False

# Set agent parameters

agents.DQN.optimizer=@optimizers.Adam
# agents.DQN.learning_rate=None
# agents.DQN.huber_delta=1.
# agents.DQN.minibatch_size=32
agents.DQN.replay_memory_size=400000
agents.DQN.prefetch=3
# agents.DQN.target_update_period=10000
# agents.DQN.reward_scale=None
agents.DQN.discount_factor=.966667
# agents.DQN.collect_batch_size=None
agents.DQN.exploration=@exploration/schedules.PolynomialDecay()
agents.DQN.prioritization=0.6
agents.DQN.priority_bias_compensation=@beta/schedules.PolynomialDecay()
agents.DQN.double=True
# agents.DQN.n_step=None
# agents.DQN.graph=True

# Set initialize policy

# initialize/stackrl.Baseline.method='lowest'

# Set training parameters

stackrl.Training.env = @envs.make
stackrl.Training.eval_env = @eval/envs.make
stackrl.Training.net = @nets.DeepQSiamFCN
stackrl.Training.agent = @agents.DQN
stackrl.Training.train_reward_buffer_length = 200
stackrl.Training.eval_reward_buffer_length = 100
# stackrl.Training.net = @nets.PseudoSiamFCN
# stackrl.Training.agent = @agents.DQN
# stackrl.Training.train_reward_buffer_length = 10
# stackrl.Training.eval_reward_buffer_length = 10
# stackrl.Training.directory = '.'
stackrl.Training.save_evaluated_policies = True
# stackrl.Training.log_to_file = True
# stackrl.Training.log_interval = 100
# stackrl.Training.eval_interval = 10000
stackrl.Training.checkpoint_interval = 1000000
# stackrl.Training.goal_check_interval = 5
stackrl.Training.seed = 1

stackrl.Training.initialize.num_steps = 10000
# stackrl.Training.initialize.policy = @initialize/stackrl.Baseline()

# stackrl.Training.run.max_num_iters = 1000000
# stackrl.Training.run.stop_when_complete = False