gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2026-02-23 10:35:46 +01:00
Code Issues Projects Releases Wiki Activity

Add documentation on distributed Coach. (#158)

Browse Source 
* Added documentation on distributed Coach.
This commit is contained in:
Balaji Subramaniam
2018-11-27 02:26:15 -08:00
committed by Gal Novik
parent e3ecf445e2
commit d06197f663
151 changed files with 5302 additions and 643 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
rl_coach/data_stores/nfs_data_store.py
View File

@@ -22,10 +22,21 @@ class NFSDataStoreParameters(DataStoreParameters):
class NFSDataStore(DataStore):
"""
An implementation of data store which uses NFS for storing policy checkpoints when using Coach in distributed mode.
The policy checkpoints are written by the trainer and read by the rollout worker.
"""
def __init__(self, params: NFSDataStoreParameters):
"""
:param params: The parameters required to use the NFS data store.
"""
self.params = params
def deploy(self) -> bool:
"""
Deploy the NFS server in an orchestrator if/when required.
"""
if self.params.orchestrator_type == "kubernetes":
if not self.params.deployed:
if not self.deploy_k8s_nfs():
@@ -43,6 +54,9 @@ class NFSDataStore(DataStore):
)
def undeploy(self) -> bool:
"""
Undeploy the NFS server and resources from an orchestrator.
"""
if self.params.orchestrator_type == "kubernetes":
if not self.params.deployed:
if not self.undeploy_k8s_nfs():
@@ -59,6 +73,9 @@ class NFSDataStore(DataStore):
pass
def deploy_k8s_nfs(self) -> bool:
"""
Deploy the NFS server in the Kubernetes orchestrator.
"""
from kubernetes import client as k8sclient
name = "nfs-server-{}".format(uuid.uuid4())
@@ -148,6 +165,9 @@ class NFSDataStore(DataStore):
return True
def create_k8s_nfs_resources(self) -> bool:
"""
Create NFS resources such as PV and PVC in Kubernetes.
"""
from kubernetes import client as k8sclient
pv_name = "nfs-ckpt-pv-{}".format(uuid.uuid4())
@@ -226,6 +246,9 @@ class NFSDataStore(DataStore):
return True
def delete_k8s_nfs_resources(self) -> bool:
"""
Delete NFS resources such as PV and PVC from the Kubernetes orchestrator.
"""
from kubernetes import client as k8sclient
del_options = k8sclient.V1DeleteOptions()

17
rl_coach/data_stores/s3_data_store.py
View File

@@ -23,7 +23,16 @@ class S3DataStoreParameters(DataStoreParameters):
class S3DataStore(DataStore):
"""
An implementation of the data store using S3 for storing policy checkpoints when using Coach in distributed mode.
The policy checkpoints are written by the trainer and read by the rollout worker.
"""
def __init__(self, params: S3DataStoreParameters):
"""
:param params: The parameters required to use the S3 data store.
"""
super(S3DataStore, self).__init__(params)
self.params = params
access_key = None
@@ -51,6 +60,10 @@ class S3DataStore(DataStore):
return True
def save_to_store(self):
"""
save_to_store() uploads the policy checkpoint, gifs and videos to the S3 data store. It reads the checkpoint state files and
uploads only the latest checkpoint files to S3. It is used by the trainer in Coach when used in the distributed mode.
"""
try:
# remove lock file if it exists
self.mc.remove_object(self.params.bucket_name, SyncFiles.LOCKFILE.value)
@@ -95,6 +108,10 @@ class S3DataStore(DataStore):
print("Got exception: %s\n while saving to S3", e)
def load_from_store(self):
"""
load_from_store() downloads a new checkpoint from the S3 data store when it is not available locally. It is used
by the rollout workers when using Coach in distributed mode.
"""
try:
state_file = CheckpointStateFile(os.path.abspath(self.params.checkpoint_dir))