Make distributed coach work end-to-end.

- With data store, memory backend and orchestrator interfaces.

rl_coach/training_worker.py

@@ -4,15 +4,19 @@ import argparse
 import time
 import json
 
+from threading import Thread
+
 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.backend.memory_impl import construct_memory_params
+from rl_coach.data_stores.data_store_impl import get_data_store, construct_data_store_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 data_store_ckpt_save(data_store):
+    while True:
+        data_store.save_to_store()
+        time.sleep(10)
 
 def training_worker(graph_manager, checkpoint_dir):
     """
@@ -58,16 +62,26 @@ def main():
     parser.add_argument('--memory_backend_params',
                         help="(string) JSON string of the memory backend params",
                         type=str)
+    parser.add_argument('--data_store_params',
+                        help="(string) JSON string of the data store params",
+                        type=str)
     args = parser.parse_args()
 
     graph_manager = short_dynamic_import(expand_preset(args.preset), ignore_module_case=True)
 
     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'
+        args.memory_backend_params['run_type'] = 'trainer'
         graph_manager.agent_params.memory.register_var('memory_backend_params', construct_memory_params(args.memory_backend_params))
 
+    if args.data_store_params:
+        data_store_params = construct_data_store_params(json.loads(args.data_store_params))
+        data_store_params.checkpoint_dir = args.checkpoint_dir
+        graph_manager.data_store_params = data_store_params
+        # data_store = get_data_store(data_store_params)
+        # thread = Thread(target = data_store_ckpt_save, args = [data_store])
+        # thread.start()
+
     training_worker(
         graph_manager=graph_manager,
         checkpoint_dir=args.checkpoint_dir,