forked from jurgisp/pydreamer
-
Notifications
You must be signed in to change notification settings - Fork 0
/
launch.py
312 lines (273 loc) · 10.7 KB
/
launch.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
import argparse
import json
import os
import time
from logging import info
from distutils.util import strtobool
from multiprocessing import Process, Queue, Manager, get_context
from typing import List
import generator
import inference
import train
from pydreamer.tools import (configure_logging, mlflow_log_params,
mlflow_init, print_once, read_yamls)
def launch():
configure_logging('[launcher]')
parser = argparse.ArgumentParser()
parser.add_argument('--configs', nargs='+', required=True)
args, remaining = parser.parse_known_args()
# Config from YAML
conf = {}
configs = read_yamls('./config')
for name in args.configs:
if ',' in name:
for n in name.split(','):
conf.update(configs[n])
else:
conf.update(configs[name])
# Override config from command-line
parser = argparse.ArgumentParser()
for key, value in conf.items():
type_ = type(value) if value is not None else str
if type_ == bool:
type_ = lambda x: bool(strtobool(x))
parser.add_argument(f'--{key}', type=type_, default=value)
conf = parser.parse_args(remaining)
# Mlflow
worker_type, worker_index = get_worker_info()
is_main_worker = worker_type is None or worker_type == 'learner'#No idea what this is for when implementing multi-node for DDP.
mlrun = None
#Can be modified somehow to use just single run in DDP...
nodeid = 0
nnodes = 1
generator_log_all = True
if(conf.learner_workers > 1):
generator_log_all = False
nodeid = int(os.environ["SLURM_NODEID"])
nnodes = int(os.environ["SLURM_JOB_NUM_NODES"])
mlrun = mlflow_init(wait_for_resume=not is_main_worker)
artifact_uri = mlrun.info.artifact_uri
mlflow_log_params(vars(conf))
subprocesses: List[Process] = []
# Launch policy inference thread
manager = Manager()
q_main = None
policy_main = "network"
if conf.policy_inference == "remote":
q_main = manager.Queue()
policy_main = "remote_network"
q_clients = []
# Launch train+eval generators
print(f"DEBUGG: {conf.generator_prefill_steps} {conf.generator_prefill_steps // (conf.generator_workers * nnodes)} {conf.generator_workers} {nnodes}")
for i in range(conf.generator_workers):
q_self = None
if conf.policy_inference == "remote":
q_self = manager.Queue()
q_clients.append(q_self)
info(f'generator {i} queue {q_self}')
if belongs_to_worker('generator', i):
index = i+nodeid*conf.generator_workers
info(f'Launching train+eval generator {index}')
p = launch_generator(
conf.env_id,
conf,
save_uri=f'{artifact_uri}/episodes/{index}',
save_uri2=f'{artifact_uri}/episodes_eval/{index}',
num_steps=conf.n_env_steps // conf.env_action_repeat // conf.generator_workers,
limit_step_ratio=conf.limit_step_ratio / conf.generator_workers,
worker_id=index,
local_rank=i,
policy_main=policy_main,
policy_prefill=conf.generator_prefill_policy,
num_steps_prefill=conf.generator_prefill_steps // (conf.generator_workers * nnodes),
split_fraction=0.05,
log_mlflow_metrics=(generator_log_all or index<conf.generator_workers),
q_main=q_main,
q_self=q_self,
)
subprocesses.append(p)
# Launch train generators
for i in range(conf.generator_workers_train):
q_self = None
if conf.policy_inference == "remote":
q_self = manager.Queue()
q_clients.append(q_self)
info(f'generator {i} queue {q_self}')
if belongs_to_worker('generator_train', i):
index = i+nodeid*conf.generator_workers_train
info(f'Launching train generator {index}')
p = launch_generator(
conf.env_id,
conf,
f'{artifact_uri}/episodes/{index}',
num_steps=conf.n_env_steps // conf.env_action_repeat // conf.generator_workers,
limit_step_ratio=conf.limit_step_ratio / conf.generator_workers,
worker_id=index,
local_rank=i,
policy_main=policy_main,
policy_prefill=conf.generator_prefill_policy,
num_steps_prefill=conf.generator_prefill_steps // (conf.generator_workers*nnodes),
log_mlflow_metrics=(generator_log_all or index<conf.generator_workers),
q_main=q_main,
q_self=q_self,
)
subprocesses.append(p)
# Launch eval generators
for i in range(conf.generator_workers_eval):
q_self = None
if conf.policy_inference == "remote":
q_self = manager.Queue()
q_clients.append(q_self)
info(f'generator {i} queue {q_self}')
if belongs_to_worker('generator_eval', i):
index = i+nodeid*conf.generator_workers_eval
info(f'Launching eval generator {index}')
p = launch_generator(
conf.env_id_eval or conf.env_id,
conf,
f'{artifact_uri}/episodes_eval/{index}',
worker_id=conf.generator_workers + index,
local_rank=conf.generator_workers + i,
policy_main=policy_main,
metrics_prefix='agent_eval',
log_mlflow_metrics=(generator_log_all or index<conf.generator_workers),
q_main=q_main,
q_self=q_self,
)
subprocesses.append(p)
# Launch inference process
for i in range(conf.inference_workers):
if conf.policy_inference == "remote":
info(f'Launching remote inference {i} {q_main}')
p = launch_inference(
conf.env_id,
conf,
worker_id=i,
policy_main=policy_main,
q_main=q_main,
q_clients=q_clients,
)
subprocesses.append(p)
# Launch learner
for i in range(conf.learner_workers):
if belongs_to_worker('learner', i):
info(f'Launching learner {i}')
p = launch_learner(
conf,
worker_id=i
)
subprocesses.append(p)
# Wait & watch
try:
while len(subprocesses) > 0:
check_subprocesses(subprocesses)
time.sleep(1)
finally:
for p in subprocesses:
p.kill() # Non-daemon processes (learner, inference) need to be killed
def launch_learner(conf, worker_id):
p = Process(target=train.run, daemon=False, args=[conf], kwargs=dict(worker_id=worker_id,))
p.start()
return p
def launch_inference(env_id,
conf,
policy_main='remote_network',
worker_id=0,
q_main=None,
q_clients=None,
):
context = get_context("spawn")
p = context.Process(target=inference.main,
daemon=False,
kwargs=dict(
env_id=env_id,
policy_main=policy_main,
worker_id=worker_id,
model_conf=conf,
q_main=q_main,
q_clients=q_clients,
))
p.start()
return p
def launch_generator(env_id,
conf,
save_uri,
save_uri2=None,
policy_main='network',
policy_prefill='random',
worker_id=0,
local_rank=0,
num_steps=int(1e9),
num_steps_prefill=0,
limit_step_ratio=0,
split_fraction=0.0,
metrics_prefix='agent',
log_mlflow_metrics=True,
q_main=None,
q_self=None,
):
p = Process(target=generator.main,
daemon=True,
kwargs=dict(
env_id=env_id,
save_uri=save_uri,
save_uri2=save_uri2,
env_time_limit=conf.env_time_limit,
env_action_repeat=conf.env_action_repeat,
env_no_terminal=conf.env_no_terminal,
limit_step_ratio=limit_step_ratio,
policy_main=policy_main,
policy_prefill=policy_prefill,
num_steps=num_steps,
num_steps_prefill=num_steps_prefill,
worker_id=worker_id,
local_rank=local_rank,
model_conf=conf,
log_mlflow_metrics=log_mlflow_metrics,
split_fraction=split_fraction,
metrics_prefix=metrics_prefix,
metrics_gamma=conf.gamma,
q_main=q_main,
q_self=q_self,
))
p.start()
return p
def check_subprocesses(subprocesses):
subp_finished = []
for p in subprocesses:
if not p.is_alive():
if p.exitcode == 0:
subp_finished.append(p)
info(f'Generator process {p.pid} finished')
else:
raise Exception(f'Generator process {p.pid} died with exitcode {p.exitcode}')
for p in subp_finished:
subprocesses.remove(p)
def belongs_to_worker(work_type, work_index):
"""
In case of distributed workers, checks if this work should execute on this worker.
If not distributed, return True.
"""
worker_type, worker_index = get_worker_info()
return (
(worker_type is None or worker_type == work_type) and
(worker_index is None or worker_index == work_index)
)
def get_worker_info():
worker_type = None
worker_index = None
if 'TF_CONFIG' in os.environ:
# TF_CONFIG indicates Google Vertex AI run
tf_config = json.loads(os.environ['TF_CONFIG'])
print_once('TF_CONFIG is set:', tf_config)
if tf_config['cluster'].get('worker'):
# If there are workers in the cluster, then it's a distributed run
worker_type = {
'chief': 'learner',
'worker': 'generator',
}[str(tf_config['task']['type'])]
worker_index = int(tf_config['task']['index'])
print_once('Distributed run detected, current worker is:', f'{worker_type} ({worker_index})')
return worker_type, worker_index
if __name__ == '__main__':
launch()