-
Notifications
You must be signed in to change notification settings - Fork 0
/
ddpg_alg_spinup.py
340 lines (271 loc) · 13.8 KB
/
ddpg_alg_spinup.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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
import numpy as np
import tensorflow as tf
import gym
import time
import ddpg_core_spinup as core
from ddpg_core_spinup import get_vars
from spinup.utils.logx import EpochLogger
import matplotlib.pyplot as matplt
import pickle
from ddpg_noise import OrnsteinUhlenbeckActionNoise
class ReplayBuffer:
"""
A simple FIFO experience replay buffer for DDPG agents.
"""
def __init__(self, obs_dim, act_dim, size):
self.obs1_buf = np.zeros([size, obs_dim], dtype=np.float32)
self.obs2_buf = np.zeros([size, obs_dim], dtype=np.float32)
self.acts_buf = np.zeros([size, act_dim], dtype=np.float32)
self.rews_buf = np.zeros(size, dtype=np.float32)
self.done_buf = np.zeros(size, dtype=np.float32)
self.ptr, self.size, self.max_size = 0, 0, size
def store(self, obs, act, rew, next_obs, done):
self.obs1_buf[self.ptr] = obs
self.obs2_buf[self.ptr] = next_obs
self.acts_buf[self.ptr] = act
self.rews_buf[self.ptr] = rew
self.done_buf[self.ptr] = done
self.ptr = (self.ptr+1) % self.max_size
self.size = min(self.size+1, self.max_size)
def sample_batch(self, batch_size=32):
idxs = np.random.randint(0, self.size, size=batch_size)
return dict(obs1=self.obs1_buf[idxs],
obs2=self.obs2_buf[idxs],
acts=self.acts_buf[idxs],
rews=self.rews_buf[idxs],
done=self.done_buf[idxs])
def restore_buffer_from_file(self):
fileop = open("saved_buffer.pickle", "rb")
[self.obs1_buf, self.obs2_buf, self.acts_buf, self.rews_buf, self.done_buf, self.ptr, self.size, self.max_size] = pickle.load(fileop)
def store_buffer_into_file(self):
fileop = open("saved_buffer.pickle", "wb")
dump_buffer = [self.obs1_buf, self.obs2_buf, self.acts_buf, self.rews_buf, self.done_buf, self.ptr, self.size, self.max_size]
pickle.dump(dump_buffer, fileop)
"""
Deep Deterministic Policy Gradient (DDPG)
"""
def ddpg(env, actor_critic=core.mlp_actor_critic, ac_kwargs=dict(), seed=0,
steps_per_epoch=5000, epochs=100, replay_size=int(1e6), gamma=0.99,
polyak=0.995, pi_lr=1e-3, q_lr=1e-3, batch_size=1000, start_steps=10000,
act_noise=1.0, max_ep_len=100, logger_kwargs=dict(), save_freq=1, fresh_learn_idx=True,):
"""
Args:
env_fn : A function which creates a copy of the environment.
The environment must satisfy the OpenAI Gym API.
actor_critic: A function which takes in placeholder symbols
for state, ``x_ph``, and action, ``a_ph``, and returns the main
outputs from the agent's Tensorflow computation graph:
=========== ================ ======================================
Symbol Shape Description
=========== ================ ======================================
``pi`` (batch, act_dim) | Deterministically computes actions
| from policy given states.
``q`` (batch,) | Gives the current estimate of Q* for
| states in ``x_ph`` and actions in
| ``a_ph``.
``q_pi`` (batch,) | Gives the composition of ``q`` and
| ``pi`` for states in ``x_ph``:
| q(x, pi(x)).
=========== ================ ======================================
ac_kwargs (dict): Any kwargs appropriate for the actor_critic
function you provided to DDPG.
seed (int): Seed for random number generators.
steps_per_epoch (int): Number of steps of interaction (state-action pairs)
for the agent and the environment in each epoch.
epochs (int): Number of epochs to run and train agent.
replay_size (int): Maximum length of replay buffer.
gamma (float): Discount factor. (Always between 0 and 1.)
polyak (float): Interpolation factor in polyak averaging for target
networks. Target networks are updated towards main networks
according to:
.. math:: \\theta_{\\text{targ}} \\leftarrow
\\rho \\theta_{\\text{targ}} + (1-\\rho) \\theta
where :math:`\\rho` is polyak. (Always between 0 and 1, usually
close to 1.)
pi_lr (float): Learning rate for policy.
q_lr (float): Learning rate for Q-networks.
batch_size (int): Minibatch size for SGD.
start_steps (int): Number of steps for uniform-random action selection,
before running real policy. Helps exploration.
act_noise (float): Stddev for Gaussian exploration noise added to
policy at training time. (At test time, no noise is added.)
max_ep_len (int): Maximum length of trajectory / episode / rollout.
logger_kwargs (dict): Keyword args for EpochLogger.
save_freq (int): How often (in terms of gap between epochs) to save
the current policy and value function.
fresh_learn_idx (boolean): Whether this time is fresh/first learn
if True, we collect the data from true env and save the results,
if False, we try to load the results(prev saved) and do no-interact learning
"""
tf.reset_default_graph()
logger = EpochLogger(**logger_kwargs)
logger.save_config(locals())
tf.set_random_seed(seed)
np.random.seed(seed)
#env, test_env = env_fn(), env_fn()
#env = env_fn()
test_env = env
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.shape[0]
# Action limit for clamping: critically, assumes all dimensions share the same bound!
act_high_limit = env.action_space.high[0]
act_low_limit = env.action_space.low[0]
# Share information about action space with policy architecture
ac_kwargs['action_space'] = env.action_space
# Inputs to computation graph
x_ph, a_ph, x2_ph, r_ph, d_ph = core.placeholders(obs_dim, act_dim, obs_dim, None, None)
# Main outputs from computation graph
with tf.variable_scope('main'):
pi, q, q_pi = actor_critic(x_ph, a_ph, **ac_kwargs)
# Target networks
with tf.variable_scope('target'):
# Note that the action placeholder going to actor_critic here is
# irrelevant, because we only need q_targ(s, pi_targ(s)).
pi_targ, _, q_pi_targ = actor_critic(x2_ph, a_ph, **ac_kwargs)
# Experience buffer
replay_buffer = ReplayBuffer(obs_dim=obs_dim, act_dim=act_dim, size=replay_size)
if not fresh_learn_idx:
replay_buffer.restore_buffer_from_file() ################################################################################################
# Count variables
var_counts = tuple(core.count_vars(scope) for scope in ['main/pi', 'main/q', 'main'])
print('\nNumber of parameters: \t pi: %d, \t q: %d, \t total: %d\n'%var_counts)
# Bellman backup for Q function
backup = tf.stop_gradient(r_ph + gamma*(1-d_ph)*q_pi_targ)
# DDPG losses
pi_loss = -tf.reduce_mean(q_pi)
q_loss = tf.reduce_mean((q-backup)**2)
# Separate train ops for pi, q
pi_optimizer = tf.train.AdamOptimizer(learning_rate=pi_lr)
q_optimizer = tf.train.AdamOptimizer(learning_rate=q_lr)
train_pi_op = pi_optimizer.minimize(pi_loss, var_list=get_vars('main/pi'))
train_q_op = q_optimizer.minimize(q_loss, var_list=get_vars('main/q'))
# Polyak averaging for target variables
target_update = tf.group([tf.assign(v_targ, polyak*v_targ + (1-polyak)*v_main)
for v_main, v_targ in zip(get_vars('main'), get_vars('target'))])
# Initializing targets to match main variables
target_init = tf.group([tf.assign(v_targ, v_main)
for v_main, v_targ in zip(get_vars('main'), get_vars('target'))])
sess = tf.Session()
sess.run(tf.global_variables_initializer())
sess.run(target_init)
# Setup model saving
logger.setup_tf_saver(sess, inputs={'x': x_ph, 'a': a_ph}, outputs={'pi': pi, 'q': q})
def get_action(o, noise_scale):
a = sess.run(pi, feed_dict={x_ph: o.reshape(1,-1)})[0]
a += noise_scale * np.random.randn(act_dim)
#action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(act_dim), sigma=float(noise_scale) * np.ones(act_dim))
#a = np.add(a, action_noise())
return np.clip(a, act_low_limit, act_high_limit)
# action_noise = NormalActionNoise(mu=np.zeros(act_dim), sigma=float(noise_scale) * np.ones(act_dim))
# action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(act_dim), sigma=float(noise_scale) * np.ones(act_dim))
def test_agent(n=1):
o, r, d, ep_ret, ep_len, ep_act = test_env.reset(), 0, False, 0, 0, []
while not(d or (ep_len == max_ep_len)):
# Take deterministic actions at test time (noise_scale=0)
a = get_action(o, 0)
o, r, d, _ = test_env.step(a)
ep_ret += r
ep_len += 1
ep_act.append(a)
logger.store(TestEpRet=ep_ret, TestEpLen=ep_len)
return ep_ret, ep_act
start_time = time.time()
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
total_steps = steps_per_epoch * epochs
total_reward, total_action = [], []
decay_ratio = np.exp(np.log(1e-3)/total_steps)
# Main loop: collect experience in env and update/log each epoch
for t in range(total_steps):
"""
Until start_steps have elapsed, randomly sample actions
from a uniform distribution for better exploration. Afterwards,
use the learned policy (with some noise, via act_noise).
"""
if t > start_steps:
a = get_action(o, act_noise)
else:
a = env.action_space.sample()
act_noise *= decay_ratio
# Step the env
o2, r, d, info = env.step(a)
ep_ret += r
ep_len += 1
# Ignore the "done" signal if it comes from hitting the time
# horizon (that is, when it's an artificial terminal signal
# that isn't based on the agent's state)
d = False if ep_len==max_ep_len else d
# Store experience to replay buffer
if fresh_learn_idx:
replay_buffer.store(o, a, r, o2, d)
# Super critical, easy to overlook step: make sure to update
# most recent observation!
o = o2
if d or (ep_len == max_ep_len):
"""
Perform all DDPG updates at the end of the trajectory,
in accordance with tuning done by TD3 paper authors.
"""
for _ in range(ep_len):
batch = replay_buffer.sample_batch(batch_size)
feed_dict = {x_ph: batch['obs1'],
x2_ph: batch['obs2'],
a_ph: batch['acts'],
r_ph: batch['rews'],
d_ph: batch['done']
}
# Q-learning update
outs = sess.run([q_loss, q, train_q_op], feed_dict)
logger.store(LossQ=outs[0], QVals=outs[1])
# Policy update
outs = sess.run([pi_loss, train_pi_op, target_update], feed_dict)
logger.store(LossPi=outs[0])
logger.store(EpRet=ep_ret, EpLen=ep_len)
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
# End of epoch wrap-up
if t > 0 and t % steps_per_epoch == 0:
epoch = t // steps_per_epoch
# Save model
if (epoch % save_freq == 0) or (epoch == epochs-1):
logger.save_state({'env': env}, None)
# Test the performance of the deterministic version of the agent.
test_ret, test_act = test_agent()
total_reward.append(test_ret)
total_action.append(test_act)
# Log info about epoch
logger.log_tabular('Epoch', epoch)
logger.log_tabular('EpRet', with_min_and_max=True)
logger.log_tabular('TestEpRet', with_min_and_max=True)
logger.log_tabular('EpLen', average_only=True)
logger.log_tabular('TestEpLen', average_only=True)
logger.log_tabular('TotalEnvInteracts', t)
logger.log_tabular('QVals', with_min_and_max=True)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('LossQ', average_only=True)
logger.log_tabular('Time', time.time()-start_time)
logger.dump_tabular()
# post process #################################################
if fresh_learn_idx:
replay_buffer.store_buffer_into_file() # save the buffer for further training
matplt.figure()
matplt.plot(total_reward)
matplt.savefig('slice_training'+str(int(time.time()))+'.png')
#matplt.show()
sess.close()
return total_reward[-1], np.transpose(total_action[-1]) # for reshape the results
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--env', type=str, default='HalfCheetah-v2')
parser.add_argument('--hid', type=int, default=300)
parser.add_argument('--l', type=int, default=1)
parser.add_argument('--gamma', type=float, default=0.99)
parser.add_argument('--seed', '-s', type=int, default=0)
parser.add_argument('--epochs', type=int, default=50)
parser.add_argument('--exp_name', type=str, default='ddpg')
args = parser.parse_args()
from spinup.utils.run_utils import setup_logger_kwargs
logger_kwargs = setup_logger_kwargs(args.exp_name, args.seed)
ddpg(lambda : args.env, actor_critic=core.mlp_actor_critic,
ac_kwargs=dict(hidden_sizes=[args.hid]*args.l),
gamma=args.gamma, seed=args.seed, epochs=args.epochs,
logger_kwargs=logger_kwargs)