-
Notifications
You must be signed in to change notification settings - Fork 298
/
ga.py
284 lines (243 loc) · 12.8 KB
/
ga.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
from .es import *
GATask = namedtuple('GATask', ['params', 'population', 'ob_mean', 'ob_std', 'timestep_limit'])
def setup(exp, single_threaded):
import gym
gym.undo_logger_setup()
from . import policies, tf_util
config = Config(**exp['config'])
env = gym.make(exp['env_id'])
if exp['env_id'].endswith('NoFrameskip-v4'):
from .atari_wrappers import wrap_deepmind
env = wrap_deepmind(env)
sess = make_session(single_threaded=single_threaded)
policy = getattr(policies, exp['policy']['type'])(env.observation_space, env.action_space, **exp['policy']['args'])
tf_util.initialize()
return config, env, sess, policy
def rollout_and_update_ob_stat(policy, env, timestep_limit, rs, task_ob_stat, calc_obstat_prob):
if policy.needs_ob_stat and calc_obstat_prob != 0 and rs.rand() < calc_obstat_prob:
rollout_rews, rollout_len, obs = policy.rollout(
env, timestep_limit=timestep_limit, save_obs=True, random_stream=rs)
task_ob_stat.increment(obs.sum(axis=0), np.square(obs).sum(axis=0), len(obs))
else:
rollout_rews, rollout_len, rollout_nov = policy.rollout(env, timestep_limit=timestep_limit, random_stream=rs)
return rollout_rews, rollout_len
def run_master(master_redis_cfg, log_dir, exp):
logger.info('run_master: {}'.format(locals()))
from . import tabular_logger as tlogger
logger.info('Tabular logging to {}'.format(log_dir))
tlogger.start(log_dir)
config, env, sess, policy = setup(exp, single_threaded=False)
master = MasterClient(master_redis_cfg)
noise = SharedNoiseTable()
rs = np.random.RandomState()
if isinstance(config.episode_cutoff_mode, int):
tslimit, incr_tslimit_threshold, tslimit_incr_ratio = config.episode_cutoff_mode, None, None
adaptive_tslimit = False
elif config.episode_cutoff_mode.startswith('adaptive:'):
_, args = config.episode_cutoff_mode.split(':')
arg0, arg1, arg2 = args.split(',')
tslimit, incr_tslimit_threshold, tslimit_incr_ratio = int(arg0), float(arg1), float(arg2)
adaptive_tslimit = True
logger.info(
'Starting timestep limit set to {}. When {}% of rollouts hit the limit, it will be increased by {}'.format(
tslimit, incr_tslimit_threshold * 100, tslimit_incr_ratio))
elif config.episode_cutoff_mode == 'env_default':
tslimit, incr_tslimit_threshold, tslimit_incr_ratio = None, None, None
adaptive_tslimit = False
else:
raise NotImplementedError(config.episode_cutoff_mode)
episodes_so_far = 0
timesteps_so_far = 0
tstart = time.time()
master.declare_experiment(exp)
best_score = float('-inf')
population = []
population_size = exp['population_size']
num_elites = exp['num_elites']
population_score = np.array([])
while True:
step_tstart = time.time()
theta = policy.get_trainable_flat()
assert theta.dtype == np.float32
if policy.needs_ob_stat:
ob_stat = RunningStat(
env.observation_space.shape,
eps=1e-2 # eps to prevent dividing by zero at the beginning when computing mean/stdev
)
curr_task_id = master.declare_task(GATask(
params=theta,
population=population,
ob_mean=ob_stat.mean if policy.needs_ob_stat else None,
ob_std=ob_stat.std if policy.needs_ob_stat else None,
timestep_limit=tslimit
))
tlogger.log('********** Iteration {} **********'.format(curr_task_id))
# Pop off results for the current task
curr_task_results, eval_rets, eval_lens, worker_ids = [], [], [], []
num_results_skipped, num_episodes_popped, num_timesteps_popped, ob_count_this_batch = 0, 0, 0, 0
while num_episodes_popped < config.episodes_per_batch or num_timesteps_popped < config.timesteps_per_batch:
# Wait for a result
task_id, result = master.pop_result()
assert isinstance(task_id, int) and isinstance(result, Result)
assert (result.eval_return is None) == (result.eval_length is None)
worker_ids.append(result.worker_id)
if result.eval_length is not None:
# This was an eval job
episodes_so_far += 1
timesteps_so_far += result.eval_length
# Store the result only for current tasks
if task_id == curr_task_id:
eval_rets.append(result.eval_return)
eval_lens.append(result.eval_length)
else:
assert result.returns_n2.dtype == np.float32
# Store results only for current tasks
if task_id == curr_task_id:
# Update counts
result_num_eps = result.lengths_n2.size
result_num_timesteps = result.lengths_n2.sum()
episodes_so_far += result_num_eps
timesteps_so_far += result_num_timesteps
curr_task_results.append(result)
num_episodes_popped += result_num_eps
num_timesteps_popped += result_num_timesteps
# Update ob stats
if policy.needs_ob_stat and result.ob_count > 0:
ob_stat.increment(result.ob_sum, result.ob_sumsq, result.ob_count)
ob_count_this_batch += result.ob_count
else:
num_results_skipped += 1
# Compute skip fraction
frac_results_skipped = num_results_skipped / (num_results_skipped + len(curr_task_results))
if num_results_skipped > 0:
logger.warning('Skipped {} out of date results ({:.2f}%)'.format(
num_results_skipped, 100. * frac_results_skipped))
# Assemble results + elite
noise_inds_n = list(population[:num_elites])
returns_n2 = list(population_score[:num_elites])
for r in curr_task_results:
noise_inds_n.extend(r.noise_inds_n)
returns_n2.extend(r.returns_n2)
noise_inds_n = np.array(noise_inds_n)
returns_n2 = np.array(returns_n2)
lengths_n2 = np.array([r.lengths_n2 for r in curr_task_results])
# Process returns
idx = np.argpartition(returns_n2, (-population_size, -1))[-1:-population_size-1:-1]
population = noise_inds_n[idx]
population_score = returns_n2[idx]
assert len(population) == population_size
assert np.max(returns_n2) == population_score[0]
print('Elite: {} score: {}'.format(population[0], population_score[0]))
policy.set_trainable_flat(noise.get(population[0][0], policy.num_params))
policy.reinitialize()
v = policy.get_trainable_flat()
for seed in population[0][1:]:
v += config.noise_stdev * noise.get(seed, policy.num_params)
policy.set_trainable_flat(v)
# Update number of steps to take
if adaptive_tslimit and (lengths_n2 == tslimit).mean() >= incr_tslimit_threshold:
old_tslimit = tslimit
tslimit = int(tslimit_incr_ratio * tslimit)
logger.info('Increased timestep limit from {} to {}'.format(old_tslimit, tslimit))
step_tend = time.time()
tlogger.record_tabular("EpRewMax", returns_n2.max())
tlogger.record_tabular("EpRewMean", returns_n2.mean())
tlogger.record_tabular("EpRewStd", returns_n2.std())
tlogger.record_tabular("EpLenMean", lengths_n2.mean())
tlogger.record_tabular("EvalEpRewMean", np.nan if not eval_rets else np.mean(eval_rets))
tlogger.record_tabular("EvalEpRewMedian", np.nan if not eval_rets else np.median(eval_rets))
tlogger.record_tabular("EvalEpRewStd", np.nan if not eval_rets else np.std(eval_rets))
tlogger.record_tabular("EvalEpLenMean", np.nan if not eval_rets else np.mean(eval_lens))
tlogger.record_tabular("EvalPopRank", np.nan if not eval_rets else (
np.searchsorted(np.sort(returns_n2.ravel()), eval_rets).mean() / returns_n2.size))
tlogger.record_tabular("EvalEpCount", len(eval_rets))
tlogger.record_tabular("Norm", float(np.square(policy.get_trainable_flat()).sum()))
tlogger.record_tabular("EpisodesThisIter", lengths_n2.size)
tlogger.record_tabular("EpisodesSoFar", episodes_so_far)
tlogger.record_tabular("TimestepsThisIter", lengths_n2.sum())
tlogger.record_tabular("TimestepsSoFar", timesteps_so_far)
num_unique_workers = len(set(worker_ids))
tlogger.record_tabular("UniqueWorkers", num_unique_workers)
tlogger.record_tabular("UniqueWorkersFrac", num_unique_workers / len(worker_ids))
tlogger.record_tabular("ResultsSkippedFrac", frac_results_skipped)
tlogger.record_tabular("ObCount", ob_count_this_batch)
tlogger.record_tabular("TimeElapsedThisIter", step_tend - step_tstart)
tlogger.record_tabular("TimeElapsed", step_tend - tstart)
tlogger.dump_tabular()
# if config.snapshot_freq != 0 and curr_task_id % config.snapshot_freq == 0:
if config.snapshot_freq != 0:
import os.path as osp
filename = 'snapshot_iter{:05d}_rew{}.h5'.format(
curr_task_id,
np.nan if not eval_rets else int(np.mean(eval_rets))
)
assert not osp.exists(filename)
policy.save(filename)
tlogger.log('Saved snapshot {}'.format(filename))
def run_worker(master_redis_cfg, relay_redis_cfg, noise, *, min_task_runtime=.2):
logger.info('run_worker: {}'.format(locals()))
assert isinstance(noise, SharedNoiseTable)
worker = WorkerClient(master_redis_cfg, relay_redis_cfg)
exp = worker.get_experiment()
config, env, sess, policy = setup(exp, single_threaded=True)
rs = np.random.RandomState()
worker_id = rs.randint(2 ** 31)
assert policy.needs_ob_stat == (config.calc_obstat_prob != 0)
while True:
task_id, task_data = worker.get_current_task()
task_tstart = time.time()
assert isinstance(task_id, int) and isinstance(task_data, GATask)
if policy.needs_ob_stat:
policy.set_ob_stat(task_data.ob_mean, task_data.ob_std)
if rs.rand() < config.eval_prob:
# Evaluation: noiseless weights and noiseless actions
policy.set_trainable_flat(task_data.params)
eval_rews, eval_length = policy.rollout(env) # eval rollouts don't obey task_data.timestep_limit
eval_return = eval_rews.sum()
logger.info('Eval result: task={} return={:.3f} length={}'.format(task_id, eval_return, eval_length))
worker.push_result(task_id, Result(
worker_id=worker_id,
noise_inds_n=None,
returns_n2=None,
signreturns_n2=None,
lengths_n2=None,
eval_return=eval_return,
eval_length=eval_length,
ob_sum=None,
ob_sumsq=None,
ob_count=None
))
else:
# Rollouts with noise
noise_inds, returns, signreturns, lengths = [], [], [], []
task_ob_stat = RunningStat(env.observation_space.shape, eps=0.) # eps=0 because we're incrementing only
while not noise_inds or time.time() - task_tstart < min_task_runtime:
if len(task_data.population) > 0:
seeds = list(task_data.population[rs.randint(len(task_data.population))]) + [noise.sample_index(rs, policy.num_params)]
else:
seeds = [noise.sample_index(rs, policy.num_params)]
v = noise.get(seeds[0], policy.num_params)
policy.set_trainable_flat(v)
policy.reinitialize()
v = policy.get_trainable_flat()
for seed in seeds[1:]:
v += config.noise_stdev * noise.get(seed, policy.num_params)
policy.set_trainable_flat(v)
rews_pos, len_pos = rollout_and_update_ob_stat(
policy, env, task_data.timestep_limit, rs, task_ob_stat, config.calc_obstat_prob)
noise_inds.append(seeds)
returns.append(rews_pos.sum())
signreturns.append(np.sign(rews_pos).sum())
lengths.append(len_pos)
worker.push_result(task_id, Result(
worker_id=worker_id,
noise_inds_n=noise_inds,
returns_n2=np.array(returns, dtype=np.float32),
signreturns_n2=np.array(signreturns, dtype=np.float32),
lengths_n2=np.array(lengths, dtype=np.int32),
eval_return=None,
eval_length=None,
ob_sum=None if task_ob_stat.count == 0 else task_ob_stat.sum,
ob_sumsq=None if task_ob_stat.count == 0 else task_ob_stat.sumsq,
ob_count=task_ob_stat.count
))