Spaces:
Sleeping
Sleeping
File size: 5,151 Bytes
eaf2e33 |
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 |
import time
import torch
from math import ceil
from src.drl.rep_mem import ReplayMem
from src.utils.filesys import getpath
from src.env.environments import AsyncOlGenEnv, SingleProcessOLGenEnv
class AsyncOffpolicyTrainer:
def __init__(self, rep_mem:ReplayMem=None, update_per=1, batch=256):
self.rep_mem = ReplayMem() if rep_mem is None else rep_mem
self.update_per = update_per
self.batch = batch
self.loggers = []
self.start_time = 0.
self.steps = 0
self.num_trans = 0
self.num_updates = 0
pass
def train(self, env:AsyncOlGenEnv, agent, budget, path, check_points=None):
if check_points is None: check_points = []
check_points.sort(reverse=True)
self._reset()
o = env.reset()
for logger in self.loggers:
if logger.__class__.__name__ == 'GenResLogger':
logger.on_episode(env, agent, 0)
while self.steps < budget:
agent.actor.eval()
a = agent.make_decision(o)
o, done = env.step(a)
self.steps += 1
if done:
model_credits = ceil(1.25 * env.eplen / self.update_per)
# agent.actor.train()
self._update(model_credits, env, agent)
if len(check_points) and self.steps >= check_points[-1]:
torch.save(agent.actor.net, getpath(f'{path}/policy{self.steps}.pth'))
check_points.pop()
self._update(0, env, agent, close=True)
torch.save(agent.actor.net, getpath(f'{path}/policy.pth'))
def _update(self, model_credits, env, agent, close=False):
transitions, rewss = env.close() if close else env.rollout()
self.rep_mem.add_transitions(transitions)
self.num_trans += len(transitions)
if len(self.rep_mem) > self.batch:
t = 1
while self.num_trans >= (self.num_updates + 1) * self.update_per:
agent.update(*self.rep_mem.sample(self.batch))
self.num_updates += 1
if not close and t == model_credits:
break
t += 1
for logger in self.loggers:
loginfo = self._pack_loginfo(rewss)
if logger.__class__.__name__ == 'GenResLogger':
logger.on_episode(env, agent, self.steps)
else:
logger.on_episode(**loginfo, close=close)
def _pack_loginfo(self, rewss):
return {
'steps': self.steps, 'time': time.time() - self.start_time, 'rewss': rewss,
'trans': self.num_trans, 'updates': self.num_updates
}
def _reset(self):
self.start_time = time.time()
self.steps = 0
self.num_trans = 0
self.num_updates = 0
def set_loggers(self, *loggers):
self.loggers = loggers
class SinProcOffpolicyTrainer:
def __init__(self, rep_mem:ReplayMem=None, update_per=2, batch=256):
self.rep_mem = ReplayMem() if rep_mem is None else rep_mem
self.update_per = update_per
self.batch = batch
self.loggers = []
self.start_time = 0.
self.steps = 0
self.num_trans = 0
self.num_updates = 0
def train(self, env:SingleProcessOLGenEnv, agent, budget, path):
self.__reset()
o = env.reset()
for logger in self.loggers:
if logger.__class__.__name__ == 'GenResLogger':
logger.on_episode(env, agent, 0)
while self.steps < budget:
agent.actor.eval()
a = agent.make_decision(o)
o, _, done, info = env.step(a)
self.steps += 1
if done:
self.__update(env, agent, info)
self.__update(env, agent, {'transitions': [], 'rewss': []}, True)
torch.save(agent.actor.net, getpath(f'{path}/policy.pth'))
def __update(self, env, agent, info, close=False):
transitions, rewss = info['transitions'], info['rewss']
# print(rewss)
self.rep_mem.add_transitions(transitions)
self.num_trans += len(transitions)
if len(self.rep_mem) > self.batch:
t = 1
while self.num_trans >= (self.num_updates + 1) * self.update_per:
agent.update(*self.rep_mem.sample(self.batch))
self.num_updates += 1
t += 1
for logger in self.loggers:
loginfo = self.__pack_loginfo(rewss)
if logger.__class__.__name__ == 'GenResLogger':
logger.on_episode(env, agent, self.steps)
else:
logger.on_episode(**loginfo, close=close)
def __pack_loginfo(self, rewss):
if len(rewss):
rewss = [rewss]
return {
'steps': self.steps, 'time': time.time() - self.start_time, 'rewss': rewss,
'trans': self.num_trans, 'updates': self.num_updates
}
def __reset(self):
self.start_time = time.time()
self.steps = 0
self.num_trans = 0
self.num_updates = 0
def set_loggers(self, *loggers):
self.loggers = loggers
|