/
train.py
176 lines (137 loc) · 7.85 KB
/
train.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
import tensorflow as tf
import os
import logging
import click
import numpy as np
import time
import sys
from copy import deepcopy
from collections import deque
import signal
from config import configure_parallel
from interruptHandler import InterruptHandler
from policies.const import TRAIN, EVAL
class Train:
def __init__(self, scope='train', **kwargs):
self._scope = scope
self._ep = 0
self._env_steps = 0
self._parallel_rollout_manager, \
self._parallel_train_manager, \
self._basic_conf, self._parallel_conf, self._modules_fn, self._managed_memory, \
self._summary_writer = configure_parallel(**kwargs)
self._jobdir = self._basic_conf['jobdir']
self._max_env_steps = int(self._basic_conf['max_env_steps'])
self._restart_after = self._basic_conf['restart_after']
self._store_params_every = self._basic_conf['store_params_every']
self._params_cache_size = self._basic_conf['params_cache_size']
self._render = self._basic_conf['render']
self._useSurprise = self._basic_conf['use_surprise']
self._logger = logging.getLogger(self._scope)
if self._restart_after is not None and os.path.exists(os.path.join(self._jobdir, 'restart')):
self._load_train()
self._cache = []
self._logger.info(f'Job pid is {os.getpid()}')
def start_loop(self):
self._loop()
def _loop(self):
if os.path.exists(os.path.join(self._jobdir, 'restart')):
os.remove(os.path.join(self._jobdir, 'restart'))
start_time = time.time()
self._parallel_rollout_manager.set_ep(self._ep)
self._steps_since_last_write = 0
# TODO: Remove
self.debug_buffer = deque(maxlen=1000)
with InterruptHandler() as h, InterruptHandler(signal.SIGUSR2) as h2:
while True:
runtime = time.time() - start_time
self._logger.info(f'Ep {self._ep}, env steps: {self._env_steps}/{self._max_env_steps}, runtime: {runtime // 60} min'
f'{"" if self._restart_after is None else " (restart in {} min)".format(((int(self._restart_after) * 60) - runtime) // 60)}')
# f' " (restart in "{(runtime - (int(self._restart_after) * 60)) // 60} min)')
self._parallel_rollout_manager.set_global_params(self._parallel_train_manager.get_global_params())
self._parallel_rollout_manager.generate_rollouts(render=self._render, mode=TRAIN)
episode = self._managed_memory['episode']
self._env_steps += np.sum([ep.shape[0] for ep in episode['o']])
self._steps_since_last_write += np.sum([ep.shape[0] for ep in episode['o']])
if not self._useSurprise:
self._managed_memory['episode']['surprises'] = [np.zeros_like(surprises) for surprises in self._managed_memory['episode']['surprises']]
# TODO: Remove
for i in range(len(self._managed_memory['episode']['surprises'])):
if np.any(self._managed_memory['episode']['surprises'][i][:,1:] > 0):
self.debug_buffer.append([{key: deepcopy(value[i]) for key, value in self._managed_memory['episode'].items()},
{key: deepcopy(value[i]) for key, value in self._managed_memory['info'].items()}])
self._summary_writer.add_scalar('train/debug_buffer_size', len(self.debug_buffer), self._env_steps)
self._parallel_train_manager.store_transitions()
self._parallel_train_manager.train()
if self._steps_since_last_write > 100:
self._steps_since_last_write = 0
self._parallel_train_manager.write_tflogs(ep=self._ep, env_step=self._env_steps)
self._summary_writer.add_scalar('train/ep', self._ep, self._env_steps)
if self._basic_conf['eval_runs']:
for i in range(len(self._basic_conf['tasks_specs'])):
success = 0
for _ in range(1):
self._parallel_rollout_manager.generate_rollouts(render=self._render, mode=EVAL, final_tasks=i)
success += int(self._managed_memory['episode']['success'][0][-1])
self._summary_writer.add_scalar(f'Task{i}/eval/success_rate', success/10.0, self._env_steps)
self._save_params()
if self._restart_after is not None and runtime >= int(self._restart_after) * 60:
self._logger.info(f'Restarting training after {runtime // 60} min after ep {self._ep}')
self._save_train()
open(os.path.join(self._jobdir, 'restart'), 'w').close()
sys.exit(3)
if h.signal_received:
self._save_params(force=True)
h.reset()
if h2.signal_received:
np.save(os.path.join(self._jobdir, 'debug_buffer_latest'), self.debug_buffer)
h2.reset()
if self._env_steps >= self._max_env_steps-1:
open(os.path.join(self._jobdir, 'done'), 'w').close()
break
self._ep += 1
def _save_params(self, force=False):
if self._store_params_every is not None and self._ep % self._store_params_every == 0:
self._cache.append(self._parallel_train_manager.save_params(f'{self._ep:03}'))
if force:
self._cache.append(self._parallel_train_manager.save_global_params('latest'))
self._parallel_train_manager.save_buffer('latest')
if len(self._cache) >= self._params_cache_size or force:
for ep in self._cache:
for item in ep:
np.save(*item)
self._cache = []
def _save_train(self):
if self._ep > 0:
self._save_params(force=True)
np.save(os.path.join(self._jobdir, 'train_latest'), [self._ep+1, self._env_steps])
def _load_train(self):
if not os.path.exists(os.path.join(self._jobdir, 'restart')): return
params = np.load(os.path.join(self._jobdir, 'train_latest.npy'))
self._ep = int(params[0])
self._env_steps = int(params[1])
self._logger.info(f'Continue training from ep {self._ep}')
def __del__(self):
self._save_train()
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
tf.logging.set_verbosity(tf.logging.ERROR)
@click.command()
@click.option('--basic_params_path', default=None, help='Path to basic config params')
@click.option('--params-path', default=None, help='Path to stored params')
@click.option('--params-prefix', default=None, help='Params prefix')
@click.option('--clean/--no-clean', default=False, help='Clean job directory')
@click.option('--jobdir', default='/tmp/test', help='Output directory')
@click.option('--max-env-steps', default=None, help='Miximum number of environmental steps')
@click.option('--num-worker', default=None, help='Number of parallel rollout worker')
@click.option('--env', default=None, help='Env specification')
@click.option('--restart-after', default=None, help='Restart training after n minutes')
@click.option('--json-string', default=None, help='Provide json with additional parameter (highest prio)')
@click.option('--experiment', default=None, help='Provide json file with additional params (overrides command line argument)')
@click.option('--seed', default=None, help='Seed for numpy, tensorflow and evs')
@click.option('--store-params-every', default=50, help='Store module params ever n-th episode')
@click.option('--params-cache-size', default=10, help='Number of items stored in cache before writing to disc')
def main(**kwargs):
train = Train(**kwargs)
train.start_loop()
if __name__ == '__main__':
main()