-
Notifications
You must be signed in to change notification settings - Fork 10
/
boxing_ls_dqn.py
213 lines (195 loc) · 10.6 KB
/
boxing_ls_dqn.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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
import gym
# import argparse
import torch
import torch.optim as optim
import os
from tensorboardX import SummaryWriter
import utils.dqn_model as dqn_model
from utils.hyperparameters import HYPERPARAMS
from utils.agent import DQNAgent, TargetNet
from utils.actions import EpsilonGreedyActionSelector
import utils.experience as experience
import utils.utils as utils
from utils.srl_algorithms import ls_step, ls_step_dueling
import utils.wrappers as wrappers
import numpy as np
import random
import copy
if __name__ == "__main__":
params = HYPERPARAMS['boxing']
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# conditional_update:
# if true, test the updated weights before replacing the old ones,
# if the new weights perform better, then replace them (bool)
conditional_update = True
env = gym.make(params['env_name'])
env = wrappers.wrap_dqn(env)
if conditional_update:
test_env = gym.make(params['env_name'])
test_env = wrappers.wrap_dqn(test_env)
training_random_seed = 2019
save_freq = 50000
n_drl = 100000 # steps of DRL between SRL
n_srl = params['replay_size'] # size of batch in SRL step
num_srl_updates = 3 # number of to SRL updates to perform
use_double_dqn = False
use_dueling_dqn = True
use_boosting = True
use_ls_dqn = True
use_constant_seed = True # to compare performance independently of the randomness
save_for_analysis = False # save also the replay buffer for later analysis
change_threshold = 5 # if change in weights due to LS-UPDATE larger than this, do not update, too much deviation
num_test_rollouts = 20
debug_change = False
if debug_change:
w_adv_last_debug = None
lam = 1000 # regularization parameter
params['batch_size'] = 32
if use_ls_dqn:
print("using ls-dqn with lambda:", str(lam))
model_name = "-LSDQN-LAM-" + str(lam) + "-" + str(int(1.0 * n_drl / 1000)) + "K"
else:
model_name = "-DQN"
model_name += "-BATCH-" + str(params['batch_size'])
if use_double_dqn:
print("using double-dqn")
model_name += "-DOUBLE"
if use_dueling_dqn:
print("using dueling-dqn")
model_name += "-DUELING"
if use_boosting:
print("using boosting")
model_name += "-BOOSTING"
if use_ls_dqn and conditional_update:
print("using conditional update")
model_name += "-COND"
if use_constant_seed:
model_name += "-SEED-" + str(training_random_seed)
np.random.seed(training_random_seed)
random.seed(training_random_seed)
env.seed(training_random_seed)
if use_ls_dqn and conditional_update:
test_env.seed(training_random_seed)
torch.manual_seed(training_random_seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(training_random_seed)
print("training using constant seed of ", training_random_seed)
writer = SummaryWriter(comment="-" + params['run_name'] + model_name)
if use_dueling_dqn:
net = dqn_model.DuelingLSDQN(env.observation_space.shape, env.action_space.n).to(device)
else:
net = dqn_model.LSDQN(env.observation_space.shape, env.action_space.n).to(device)
tgt_net = TargetNet(net)
selector = EpsilonGreedyActionSelector(epsilon=params['epsilon_start'])
epsilon_tracker = utils.EpsilonTracker(selector, params)
agent = DQNAgent(net, selector, device=device)
exp_source = experience.ExperienceSourceFirstLast(env, agent, gamma=params['gamma'], steps_count=1)
buffer = experience.ExperienceReplayBuffer(exp_source, buffer_size=params['replay_size'])
optimizer = optim.Adam(net.parameters(), lr=params['learning_rate']) # TODO: change to RMSprop
utils.load_agent_state(net, optimizer, selector, load_optimizer=False, env_name='boxing',
path='./agent_ckpt/agent_ls_dqn_-boxing.pth')
frame_idx = 0
drl_updates = 0
with utils.RewardTracker(writer, params['stop_reward']) as reward_tracker:
while True:
frame_idx += 1
buffer.populate(1)
epsilon_tracker.frame(frame_idx)
new_rewards = exp_source.pop_total_rewards()
if new_rewards:
if reward_tracker.reward(new_rewards[0], frame_idx, selector.epsilon):
if save_for_analysis:
temp_model_name = model_name + "_" + str(frame_idx)
utils.save_agent_state(net, optimizer, frame_idx, len(reward_tracker.total_rewards),
selector.epsilon, save_replay=True, replay_buffer=buffer.buffer,
name=temp_model_name)
else:
utils.save_agent_state(net, optimizer, frame_idx, len(reward_tracker.total_rewards),
selector.epsilon, name='-boxing')
break
if len(buffer) < params['replay_initial']:
continue
optimizer.zero_grad()
batch = buffer.sample(params['batch_size'])
loss_v = utils.calc_loss_dqn(batch, net, tgt_net.target_model, gamma=params['gamma'],
device=device, double_dqn=use_double_dqn)
loss_v.backward()
optimizer.step()
drl_updates += 1
if debug_change and w_adv_last_debug is not None:
w_adv_last_debug_after = copy.deepcopy(net.fc2_adv.state_dict())
adv_weight_diff = torch.sum((w_adv_last_debug_after['weight'] - w_adv_last_debug['weight']) ** 2)
adv_bias_diff = torch.sum((w_adv_last_debug_after['bias'] - w_adv_last_debug['bias']) ** 2)
adv_total_weight_diff = torch.sqrt(adv_weight_diff + adv_bias_diff)
print("debug diff: ", adv_total_weight_diff)
# LS-UPDATE STEP
if use_ls_dqn and (drl_updates % n_drl == 0) and (len(buffer) >= n_srl):
# if use_ls_dqn and len(buffer) > 1:
print("performing ls step...")
batch = buffer.sample(n_srl)
if use_dueling_dqn:
if conditional_update:
w_adv_last_dict_before = copy.deepcopy(net.fc2_adv.state_dict())
w_val_last_dict_before = copy.deepcopy(net.fc2_val.state_dict())
ls_step_dueling(net, tgt_net.target_model, batch, params['gamma'], len(batch), lam=lam,
m_batch_size=256,
device=device,
use_boosting=use_boosting, use_double_dqn=use_double_dqn)
if debug_change:
w_adv_last_debug = copy.deepcopy(net.fc2_adv.state_dict())
if conditional_update:
print("comparing old and new weights...")
w_adv_last_dict_after = copy.deepcopy(net.fc2_adv.state_dict())
w_val_last_dict_after = copy.deepcopy(net.fc2_val.state_dict())
test_agent = copy.deepcopy(agent)
# test original
test_agent.dqn_model.fc2_adv.load_state_dict(w_adv_last_dict_before)
test_agent.dqn_model.fc2_val.load_state_dict(w_val_last_dict_before)
before_reward = utils.test_agent(test_env, test_agent)
# test new
test_agent.dqn_model.fc2_adv.load_state_dict(w_adv_last_dict_after)
test_agent.dqn_model.fc2_val.load_state_dict(w_val_last_dict_after)
after_reward = utils.test_agent(test_env, test_agent)
print("average reward:: original: %.3f" % before_reward, " least-squares: %.3f" % after_reward)
if (after_reward > before_reward) and (abs(after_reward - before_reward) < change_threshold):
net.fc2_adv.load_state_dict(w_adv_last_dict_after)
net.fc2_val.load_state_dict(w_val_last_dict_after)
print("using updated weights.")
else:
net.fc2_adv.load_state_dict(w_adv_last_dict_before)
net.fc2_val.load_state_dict(w_val_last_dict_before)
print("using original weights.")
else:
if conditional_update:
print("comparing old and new weights...")
w_last_before = copy.deepcopy(net.fc2.state_dict())
ls_step(net, tgt_net.target_model, batch, params['gamma'], len(batch), lam=lam,
m_batch_size=256, device=device, use_boosting=use_boosting,
use_double_dqn=use_double_dqn)
if conditional_update:
w_last_after = copy.deepcopy(net.fc2.state_dict())
test_agent = copy.deepcopy(agent)
# test original
test_agent.dqn_model.fc2.load_state_dict(w_last_before)
before_reward = utils.test_agent(test_env, test_agent, num_rollouts=num_test_rollouts)
# test new
test_agent.dqn_model.fc2.load_state_dict(w_last_after)
after_reward = utils.test_agent(test_env, test_agent, num_rollouts=num_test_rollouts)
print("average reward:: original: %.3f" % before_reward, " least-squares: %.3f" % after_reward)
if after_reward > before_reward:
net.fc2.load_state_dict(w_last_after)
print("using updated weights.")
else:
net.fc2.load_state_dict(w_last_before)
print("using original weights.")
if frame_idx % params['target_net_sync'] == 0:
tgt_net.sync()
if frame_idx % save_freq == 0:
if save_for_analysis and frame_idx % n_drl == 0:
temp_model_name = model_name + "_" + str(frame_idx)
utils.save_agent_state(net, optimizer, frame_idx, len(reward_tracker.total_rewards),
selector.epsilon, save_replay=True, replay_buffer=buffer.buffer,
name=temp_model_name)
else:
utils.save_agent_state(net, optimizer, frame_idx, len(reward_tracker.total_rewards),
selector.epsilon, name="-boxing")