gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 19:20:19 +01:00
Code Issues Projects Releases Wiki Activity

Adding nfs pv, pvc, waiting for memory to be full

Browse Source
This commit is contained in:
Ajay Deshpande
2018-09-19 09:03:11 -07:00
committed by zach dwiel
parent 13d81f65b9
commit 98850464cc
4 changed files with 188 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
rl_coach/training_worker.py
View File

@@ -1,22 +1,19 @@
"""
"""
import argparse
import time
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
# 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):
num_steps = graph_manager.schedule_params.heatup_steps.num_steps
while len(graph_manager.agent_params.memory) < num_steps:
time.sleep(1)
def training_worker(graph_manager, checkpoint_dir):
"""
restore a checkpoint then perform rollouts using the restored model
@@ -29,10 +26,20 @@ def training_worker(graph_manager, checkpoint_dir):
# save randomly initialized graph
graph_manager.save_checkpoint()
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() < 100):
time.sleep(10)
# TODO: critical: wait for minimum number of rollouts in memory before training
# TODO: Q: training steps passed into graph_manager.train ignored?
# TODO: specify training steps between checkpoints (in preset?)
# TODO: replace while true with what? number of steps, convergence, time, ...
# TODO: low: move evaluate out of this process
# training loop
for _ in range(10):
for _ in range(40):
graph_manager.phase = core_types.RunPhase.TRAIN
graph_manager.train(core_types.TrainingSteps(1))
graph_manager.phase = core_types.RunPhase.UNDEFINED
@@ -72,5 +79,6 @@ def main():
checkpoint_dir=args.checkpoint_dir,
)
if __name__ == '__main__':
main()