-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathdqn_image_representations_scale_range.py
173 lines (156 loc) · 5.72 KB
/
dqn_image_representations_scale_range.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
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
"""###IMP dummy_seed should always be last in the order in the OrderedDict below!!!
"""
import itertools
from ray import tune
from collections import OrderedDict
num_seeds = 100
var_env_configs = OrderedDict(
{
"state_space_size": [8], # , 10, 12, 14] # [2**i for i in range(1,6)]
"action_space_size": [8], # 2, 4, 8, 16] # [2**i for i in range(1,6)]
"delay": [0], # + [2**i for i in range(4)],
"sequence_length": [1], # , 2, 3, 4],#i for i in range(1,4)]
"reward_density": [0.25], # np.linspace(0.0, 1.0, num=5)
"make_denser": [False],
"terminal_state_density": [0.25], # np.linspace(0.1, 1.0, num=5)
"transition_noise": [0], # , 0.01, 0.02, 0.10, 0.25]
"reward_noise": [0], # , 1, 5, 10, 25] # Std dev. of normal dist.
"image_representations": [True],
"image_transforms": ["scale"], # , 'scale', 'flip', 'rotate'],
"image_scale_range": [
(0.8, 1.25),
(0.75, 1.3333333333333333),
(0.6666666666666666, 1.5),
(0.5, 2),
], # for 400x400: [(0.25, 4), (0.3333333333333333, 3), (0.5, 2), (0.75, 1.3333333333333333)], # 100x100, radius 20: [(0.8, 1.25), (0.75, 1.3333333333333333), (0.6666666666666666,1.5), (0.5, 2)], # 100x100, radius 20: [(0.8, 1.25), (0.6, 1.5), (0.4, 1.75), (0.2, 2)], # 100x100, radius 10: [(0.8,1.5), (0.6,2), (0.4,2.5), (0.2,3)]
"image_width": [100],
"image_height": [100],
"dummy_seed": [i for i in range(num_seeds)],
}
)
var_configs = OrderedDict({"env": var_env_configs})
env_config = {
"env": "RLToy-v0",
"horizon": 100,
"env_config": {
"seed": 0, # seed
"state_space_type": "discrete",
"action_space_type": "discrete",
"generate_random_mdp": True,
"repeats_in_sequences": False,
"reward_scale": 1.0,
"completely_connected": True,
},
}
algorithm = "DQN"
agent_config = {
"adam_epsilon": 1e-4,
"beta_annealing_fraction": 1.0,
"buffer_size": 1000000,
"double_q": False,
"dueling": False,
"exploration_final_eps": 0.01,
"exploration_fraction": 0.1,
"final_prioritized_replay_beta": 1.0,
"hiddens": None,
"learning_starts": 1000,
"lr": 1e-5, # "lr": grid_search([1e-2, 1e-4, 1e-6]),
"n_step": 1,
"noisy": False,
"num_atoms": 1,
"prioritized_replay": False,
"prioritized_replay_alpha": 0.5,
"sample_batch_size": 4, # Renamed from sample_batch_size in some Ray version
"schedule_max_timesteps": 20000,
"target_network_update_freq": 800,
"timesteps_per_iteration": 1000,
"min_iter_time_s": 0,
"train_batch_size": 32,
}
# formula [(W−K+2P)/S]+1; for padding=same: P = ((S-1)*W - S + K)/2
filters_84x84 = [
[
16,
[8, 8],
4,
], # changes from 84x84x1 with padding 4 to 22x22x16 (or 26x26x16 for 100x100x1)
[32, [4, 4], 2], # changes to 11x11x32 with padding 2 (or 13x13x32 for 100x100x1)
[
256,
[11, 11],
1,
], # changes to 1x1x256 with padding 0 (or 3x3x256 for 100x100x1); this is the only layer with valid padding in Ray!
]
filters_100x100 = [
[
16,
[8, 8],
4,
], # changes from 84x84x1 with padding 4 to 22x22x16 (or 26x26x16 for 100x100x1)
[32, [4, 4], 2], # changes to 11x11x32 with padding 2 (or 13x13x32 for 100x100x1)
[
64,
[13, 13],
1,
], # changes to 1x1x64 with padding 0 (or 3x3x64 for 100x100x1); this is the only layer with valid padding in Ray!
]
# [num_outputs(=8 in this case), [1, 1], 1] conv2d appended by Ray always followed by a Dense layer with 1 output
# filters_99x99 = [
# [16, [8, 8], 4], # 51x51x16
# [32, [4, 4], 2],
# [64, [13, 13], 1],
# ]
filters_50x50 = [
[16, [4, 4], 2],
[32, [4, 4], 2],
[64, [13, 13], 1],
]
filters_400x400 = [
[16, [32, 32], 16],
[32, [4, 4], 2],
[64, [13, 13], 1],
]
model_config = {
"model": {
"fcnet_hiddens": [256, 256],
# "custom_preprocessor": "ohe",
"custom_options": {}, # extra options to pass to your preprocessor
"conv_activation": "relu",
"conv_filters": filters_100x100,
# "no_final_linear": False,
# "vf_share_layers": True,
# "fcnet_activation": "tanh",
"use_lstm": False,
"max_seq_len": 20,
"lstm_cell_size": 256,
"lstm_use_prev_action_reward": False,
},
}
eval_config = {
"evaluation_interval": 1, # I think this means every x training_iterations
"evaluation_config": {
"explore": False,
"exploration_fraction": 0,
"exploration_final_eps": 0,
"evaluation_num_episodes": 10,
"horizon": 100,
"env_config": {
"dummy_eval": True, # hack Used to check if we are in evaluation mode or training mode inside Ray callback on_episode_end() to be able to write eval stats
"transition_noise": 0
if "state_space_type" in env_config["env_config"]
and env_config["env_config"]["state_space_type"] == "discrete"
else tune.function(lambda a: a.normal(0, 0)),
"reward_noise": tune.function(lambda a: a.normal(0, 0)),
"action_loss_weight": 0.0,
},
},
}
value_tuples = []
for config_type, config_dict in var_configs.items():
for key in config_dict:
assert isinstance(
var_configs[config_type][key], list
), "var_config should be a dict of dicts with lists as the leaf values to allow each configuration option to take multiple possible values"
value_tuples.append(var_configs[config_type][key])
cartesian_product_configs = list(itertools.product(*value_tuples))
print("Total number of configs. to run:", len(cartesian_product_configs))