-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
191 lines (164 loc) · 7.98 KB
/
main.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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
import gym
import json
import torch
from src.environment.env_utils import init_environment, init_skeleton_environment
from src.nn import create_nets_sac, create_optimizers_sac
from src.losses import QValueLoss, NStepQValueLoss, NStepQValueLossSeparateEntropy
from src.sac import SAC
from src.r2d2.segment_sampler import SegmentSampler
from src.r2d2.experience_replay import PrioritizedExperienceReplay, ExperienceReplay
from src.r2d2.writer import Writer
from src.r2d2.trainer import Trainer
def main(config_file):
# read all parameters
with open(config_file, 'r') as f:
parameters = json.load(f)
# ============================== init environment ==============================
env_parameters = parameters['environment']
env_num = env_parameters['env_num']
segment_len = env_parameters['segment_len']
difficulty = env_parameters['difficulty']
accuracy = env_parameters['accuracy']
frame_skip = env_parameters['frame_skip']
timestep_limit = env_parameters['timestep_limit']
weights = env_parameters['weights']
footstep_weight, effort_weight, v_tgt_weight = weights['reward_weights']
alive_bonus, death_penalty, task_bonus = weights['alive_death_task']
# # other gym environments
# train_env, test_env = init_environment(env_num, env_name, frame_skip, segment_len)
# observation_space = train_env.observation_space.shape[0]
# action_space = train_env.action_space.shape[0]
train_env, test_env = init_skeleton_environment(
env_num, segment_len, difficulty, accuracy,
frame_skip, timestep_limit,
footstep_weight, effort_weight, v_tgt_weight,
alive_bonus, death_penalty
)
observation_space = [2 * 11 * 11, 97]
action_space = 22
# ================================= init nets ==================================
# available model types: 'feed_forward', 'recurrent', 'attention', 'skeleton'
network_parameters = parameters['networks']
model_type = network_parameters['model_type']
device = torch.device(network_parameters['device_str'])
# actor parameters
actor_parameters = network_parameters['actor_parameters']
hidden_dims_actor = actor_parameters['hidden_dim']
noisy_actor = actor_parameters['noisy'] == 'True'
layer_norm_actor = actor_parameters['layer_norm'] == 'True'
afn_actor = actor_parameters['afn']
residual_actor = actor_parameters['residual'] == 'True'
drop_actor = actor_parameters['dropout']
actor_lr = actor_parameters['learning_rate']
normal = actor_parameters['normal'] == 'True'
# critic parameters
critic_parameters = network_parameters['critic_parameters']
hidden_dims_critic = critic_parameters['hidden_dim']
noisy_critic = critic_parameters['noisy'] == 'True'
layer_norm_critic = critic_parameters['layer_norm'] == 'True'
afn_critic = critic_parameters['afn']
residual_critic = critic_parameters['residual'] == 'True'
drop_critic = critic_parameters['dropout']
q_value_dim = critic_parameters['q_value_dim']
critic_lr = critic_parameters['learning_rate']
policy_net, q_net_1, q_net_2, target_q_net_1, target_q_net_2 = create_nets_sac(
model_type, observation_space, action_space,
hidden_dims_actor, noisy_actor, layer_norm_actor, afn_actor, residual_actor, drop_actor, normal,
hidden_dims_critic, noisy_critic, layer_norm_critic, afn_critic, residual_critic, drop_critic,
device, q_value_dim + 1 # WARNING: q_value_dim here is reward_dim + 1!
)
policy_optimizer, q_optim_1, q_optim_2 = create_optimizers_sac(
policy_net, q_net_1, q_net_2, actor_lr, critic_lr
)
# ================================= init agent =================================
agent_parameters = parameters['agent_parameters']
gamma = agent_parameters['gamma']
soft_tau = agent_parameters['soft_tau']
n_step_loss = agent_parameters['n_step_loss']
rescaling = agent_parameters['rescaling'] == 'True'
q_weights = agent_parameters['q_weights']
q_value_loss = NStepQValueLossSeparateEntropy(gamma, device, q_weights, n_steps=n_step_loss, rescaling=rescaling)
n_steps = agent_parameters['n_step_train'] # number of steps from tail of segment to learn from
# aka eta, priority = eta * max_t(delta) + (1 - eta) * mean_t(delta)
priority_weight = agent_parameters['priority_weight']
use_observation_normalization = agent_parameters['use_observation_normalization'] == 'True'
agent = SAC(
policy_net, q_net_1, q_net_2, target_q_net_1, target_q_net_2,
q_value_loss,
policy_optimizer, q_optim_1, q_optim_2,
soft_tau, device, action_space, n_steps,
priority_weight,
q_value_dim, q_weights,
use_observation_normalization
)
# ================== init segment sampler & experience replay ==================
replay_parameters = parameters['replay_parameters']
log_dir = replay_parameters['log_dir']
writer = Writer(log_dir, env_num)
segment_sampler = SegmentSampler(agent, train_env, segment_len, q_weights, writer)
replay_capacity = replay_parameters['replay_capacity'] # R2D2 -> 100k
actor_size = replay_parameters['actor_size']
critic_size = replay_parameters['critic_size']
prioritization = replay_parameters['prioritization'] == 'True'
exp_replay_init_fn = PrioritizedExperienceReplay if prioritization else ExperienceReplay
experience_replay = exp_replay_init_fn(
replay_capacity, segment_len,
observation_space, action_space, q_value_dim,
# (1, 2) for feed forward net, (hidden_size, hidden_size * 2) for lstm of mhsa
actor_size, critic_size
)
# ================================ init trainer ================================
trainer_parameters = parameters['trainer_parameters']
start_priority_exponent = trainer_parameters['start_priority_exponent']
end_priority_exponent = trainer_parameters['end_priority_exponent']
start_importance_exponent = trainer_parameters['start_importance_exponent']
end_importance_exponent = trainer_parameters['end_importance_exponent']
prioritization_steps = trainer_parameters['prioritization_steps']
exp_replay_checkpoint = trainer_parameters['exp_replay_checkpoint']
if exp_replay_checkpoint == 'None':
exp_replay_checkpoint = None
agent_checkpoint = trainer_parameters['agent_checkpoint']
if agent_checkpoint == 'None':
agent_checkpoint = None
load_full = trainer_parameters['load_full'] == 'True'
trainer = Trainer(
env_num, test_env,
segment_sampler, log_dir, writer,
agent, experience_replay,
start_priority_exponent, end_priority_exponent,
start_importance_exponent, end_importance_exponent,
q_value_dim
)
trainer.load_checkpoint(agent_checkpoint, load_full)
# ================================ train agent ================================
training_parameters = parameters['training_parameters']
min_experience_len = training_parameters['min_experience_len']
num_epochs = training_parameters['num_epochs']
epoch_size = training_parameters['epoch_size']
batch_size = training_parameters['batch_size']
train_steps = training_parameters['train_steps']
test_n = training_parameters['test_n']
render = training_parameters['render'] == 'True'
segment_file = training_parameters['segment_file']
pretrain_critic = training_parameters['pretrain_critic'] == 'True'
num_pretrain_epoch = training_parameters['num_pretrain_epoch']
if num_pretrain_epoch > 0:
trainer.pretrain_from_segments(
segment_file, num_pretrain_epoch, batch_size,
actor_size, critic_size
)
trainer.train(
min_experience_len,
num_epochs, epoch_size,
train_steps, batch_size,
test_n, render,
prioritization_steps,
pretrain_critic,
exp_replay_checkpoint
)
writer.close()
train_env.close()
test_env.close()
if __name__ == '__main__':
experiment_config = 'src/experiments/exp_8c_2.json'
main(experiment_config)