gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-18 03:30:19 +01:00
Code Issues Projects Releases Wiki Activity

Adding initial interface for backend and redis pubsub (#19)

Browse Source 
* Adding initial interface for backend and redis pubsub

* Addressing comments, adding super in all memories

* Removing distributed experience replay
This commit is contained in:
Ajay Deshpande
2018-10-03 15:07:48 -07:00
committed by zach dwiel
parent a54ef2757f
commit 6b2de6ba6d
21 changed files with 459 additions and 444 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
rl_coach/training_worker.py
View File

@@ -2,42 +2,17 @@
"""
import argparse
import time
import json
from rl_coach.base_parameters import TaskParameters
from rl_coach.coach import expand_preset
from rl_coach import core_types
from rl_coach.utils import short_dynamic_import
from rl_coach.memories.non_episodic.distributed_experience_replay import DistributedExperienceReplay
from rl_coach.memories.memory import MemoryGranularity
from rl_coach.memories.backend.memory_impl import construct_memory_params
# Q: specify alternative distributed memory, or should this go in the preset?
# A: preset must define distributed memory to be used. we aren't going to take a non-distributed preset and automatically distribute it.
def heatup(graph_manager):
memory = DistributedExperienceReplay(max_size=(MemoryGranularity.Transitions, 1000000),
redis_ip=graph_manager.agent_params.memory.redis_ip,
redis_port=graph_manager.agent_params.memory.redis_port)
while(memory.num_transitions() < graph_manager.heatup_steps.num_steps):
time.sleep(1)
class StepsLoop(object):
"""StepsLoop facilitates a simple while loop"""
def __init__(self, steps_counters, phase, steps):
super(StepsLoop, self).__init__()
self.steps_counters = steps_counters
self.phase = phase
self.steps = steps
self.step_end = self._step_count() + steps.num_steps
def _step_count(self):
return self.steps_counters[self.phase][self.steps.__class__]
def continue(self):
return self._step_count() < count_end:
def training_worker(graph_manager, checkpoint_dir):
"""
@@ -51,12 +26,8 @@ def training_worker(graph_manager, checkpoint_dir):
# save randomly initialized graph
graph_manager.save_checkpoint()
# optionally wait for a specific number of transitions to be in memory before training
heatup(graph_manager)
# training loop
stepper = StepsLoop(graph_manager.total_steps_counters, RunPhase.TRAIN, graph_manager.improve_steps)
while stepper.continue():
while True:
graph_manager.phase = core_types.RunPhase.TRAIN
graph_manager.train(core_types.TrainingSteps(1))
graph_manager.phase = core_types.RunPhase.UNDEFINED
@@ -65,7 +36,6 @@ def training_worker(graph_manager, checkpoint_dir):
graph_manager.save_checkpoint()
# TODO: signal to workers that training is done
def main():
parser = argparse.ArgumentParser()
@@ -85,12 +55,18 @@ def main():
help="(int) Port of the redis server",
default=6379,
type=int)
parser.add_argument('--memory_backend_params',
help="(string) JSON string of the memory backend params",
type=str)
args = parser.parse_args()
graph_manager = short_dynamic_import(expand_preset(args.preset), ignore_module_case=True)
graph_manager.agent_params.memory.redis_ip = args.redis_ip
graph_manager.agent_params.memory.redis_port = args.redis_port
if args.memory_backend_params:
args.memory_backend_params = json.loads(args.memory_backend_params)
if 'run_type' not in args.memory_backend_params:
args.memory_backend_params['run_type'] = 'trainer'
graph_manager.agent_params.memory.register_var('memory_backend_params', construct_memory_params(args.memory_backend_params))
training_worker(
graph_manager=graph_manager,