1
0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 19:20:19 +01:00

Implement frame-work agnostic rollout and training workers (#137)

* Added checkpoint state file to coach checkpointing.

* Removed TF specific code from rollout_worker, training_worker, and s3_data_store
This commit is contained in:
Sina Afrooze
2018-11-23 18:05:44 -08:00
committed by Balaji Subramaniam
parent 4a6c404070
commit 5332013bd1
7 changed files with 350 additions and 117 deletions

View File

@@ -12,37 +12,26 @@ import os
import math
from rl_coach.base_parameters import TaskParameters, DistributedCoachSynchronizationType
from rl_coach.checkpoint import CheckpointStateFile, CheckpointStateReader
from rl_coach.core_types import EnvironmentSteps, RunPhase, EnvironmentEpisodes
from google.protobuf import text_format
from tensorflow.python.training.checkpoint_state_pb2 import CheckpointState
from rl_coach.data_stores.data_store import SyncFiles
def has_checkpoint(checkpoint_dir):
"""
True if a checkpoint is present in checkpoint_dir
"""
if os.path.isdir(checkpoint_dir):
if len(os.listdir(checkpoint_dir)) > 0:
return os.path.isfile(os.path.join(checkpoint_dir, "checkpoint"))
return False
def wait_for_checkpoint(checkpoint_dir, data_store=None, timeout=10):
"""
block until there is a checkpoint in checkpoint_dir
"""
chkpt_state_file = CheckpointStateFile(checkpoint_dir)
for i in range(timeout):
if data_store:
data_store.load_from_store()
if has_checkpoint(checkpoint_dir):
if chkpt_state_file.read() is not None:
return
time.sleep(10)
# one last time
if has_checkpoint(checkpoint_dir):
if chkpt_state_file.read() is not None:
return
raise ValueError((
@@ -54,21 +43,6 @@ def wait_for_checkpoint(checkpoint_dir, data_store=None, timeout=10):
))
def data_store_ckpt_load(data_store):
while True:
data_store.load_from_store()
time.sleep(10)
def get_latest_checkpoint(checkpoint_dir):
if os.path.exists(os.path.join(checkpoint_dir, 'checkpoint')):
ckpt = CheckpointState()
contents = open(os.path.join(checkpoint_dir, 'checkpoint'), 'r').read()
text_format.Merge(contents, ckpt)
rel_path = os.path.relpath(ckpt.model_checkpoint_path, checkpoint_dir)
return int(rel_path.split('_Step')[0])
def should_stop(checkpoint_dir):
return os.path.exists(os.path.join(checkpoint_dir, SyncFiles.FINISHED.value))
@@ -83,6 +57,7 @@ def rollout_worker(graph_manager, data_store, num_workers, task_parameters):
graph_manager.create_graph(task_parameters)
with graph_manager.phase_context(RunPhase.TRAIN):
chkpt_state_reader = CheckpointStateReader(checkpoint_dir, checkpoint_state_optional=False)
last_checkpoint = 0
act_steps = math.ceil((graph_manager.agent_params.algorithm.num_consecutive_playing_steps.num_steps)/num_workers)
@@ -97,20 +72,20 @@ def rollout_worker(graph_manager, data_store, num_workers, task_parameters):
elif type(graph_manager.agent_params.algorithm.num_consecutive_playing_steps) == EnvironmentEpisodes:
graph_manager.act(EnvironmentEpisodes(num_steps=act_steps))
new_checkpoint = get_latest_checkpoint(checkpoint_dir)
new_checkpoint = chkpt_state_reader.get_latest()
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type == DistributedCoachSynchronizationType.SYNC:
while new_checkpoint < last_checkpoint + 1:
while new_checkpoint is None or new_checkpoint.num < last_checkpoint + 1:
if should_stop(checkpoint_dir):
break
if data_store:
data_store.load_from_store()
new_checkpoint = get_latest_checkpoint(checkpoint_dir)
new_checkpoint = chkpt_state_reader.get_latest()
graph_manager.restore_checkpoint()
if graph_manager.agent_params.algorithm.distributed_coach_synchronization_type == DistributedCoachSynchronizationType.ASYNC:
if new_checkpoint > last_checkpoint:
if new_checkpoint is not None and new_checkpoint.num > last_checkpoint:
graph_manager.restore_checkpoint()
last_checkpoint = new_checkpoint
if new_checkpoint is not None:
last_checkpoint = new_checkpoint.num