wassname/ray
1
0
Fork 0
mirror of https://github.com/wassname/ray.git synced 2026-07-28 11:25:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

[rllib] Importance Sampling and KL Loss for APPO (#5051)

Browse Source
This commit is contained in:
Michael Luo
2019-07-29 15:02:32 -07:00
committed by Richard Liaw
parent 3b00144e7d
commit 1337c98f02
10 changed files with 288 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/ray/rllib/optimizers/aso_learner.py
+4 -1
View File
@@ -49,7 +49,8 @@ class LearnerThread(threading.Thread):
inqueue=self.inqueue,
size=minibatch_buffer_size,
timeout=learner_queue_timeout,
num_passes=num_sgd_iter)
num_passes=num_sgd_iter,
init_num_passes=num_sgd_iter)
self.queue_timer = TimerStat()
self.grad_timer = TimerStat()
self.load_timer = TimerStat()
@@ -58,6 +59,7 @@ class LearnerThread(threading.Thread):
self.weights_updated = False
self.stats = {}
self.stopped = False
self.num_steps = 0
def run(self):
while not self.stopped:
@@ -72,5 +74,6 @@ class LearnerThread(threading.Thread):
self.weights_updated = True
self.stats = get_learner_stats(fetches)
self.num_steps += 1
self.outqueue.put(batch.count)
self.learner_queue_size.push(self.inqueue.qsize())