Warm start a stratey based on config provided seed conditions (facebo…

…okresearch#487)

Summary:
Pull Request resolved: facebookresearch#487

There is a desire to warm start a strategy with data filtered on master table values recorded from previous experiments.

Data can be filtered on any data other than master.unique_id. This means that the below are all valid filter criteria:
- experiment_name
- experiment_description
- experiment_id
- participant_id
- anything stored in extra_metadata

Criteria will follow AND logic between fields, but inclusive OR logic within the same field.

Data will be further filtered out if it does not meet these criteria:
- Each parameter matches on name and type
- Each outcome matches on name and type
- Stimuli per trial are the same

Differential Revision: D66731597

aepsych/config.py

@@ -149,16 +149,14 @@ def to_dict(self, deduplicate: bool = True) -> Dict[str, Any]:
                 _dict[section][setting] = self[section][setting]
         return _dict
 
-    # Turn the metadata section into JSON.
-    def jsonifyMetadata(self, only_extra: bool = False) -> str:
-        """Return a json string of the metadata section.
+    def get_metadata(self, only_extra: bool = False) -> Dict[Any, Any]:
+        """Return a dictionary of the metadata section.
 
         Args:
-            only_extra (bool): Only jsonify the extra meta data.
+            only_extra (bool, optional): Only gather the extra metadata. Defaults to False.
 
         Returns:
-            str: A json string representing the metadata dictionary or an empty string
-                if there is no metadata to return.
+            Dict[Any, Any]: a collection of the metadata stored in this conig.
         """
         configdict = self.to_dict()
         metadata = configdict["metadata"].copy()
@@ -172,7 +170,21 @@ def jsonifyMetadata(self, only_extra: bool = False) -> str:
             ]
             for name in default_metadata:
                 metadata.pop(name, None)
+
+        return metadata
+
+    # Turn the metadata section into JSON.
+    def jsonifyMetadata(self, only_extra: bool = False) -> str:
+        """Return a json string of the metadata section.
+
+        Args:
+            only_extra (bool): Only jsonify the extra meta data.
 
+        Returns:
+            str: A json string representing the metadata dictionary or an empty string
+                if there is no metadata to return.
+        """        
+        metadata = self.get_metadata(only_extra)
         if len(metadata.keys()) == 0:
             return ""
         else:

aepsych/database/db.py

@@ -128,7 +128,7 @@ def execute_sql_query(self, query: str, vals: Dict[str, str]) -> List[Any]:
             List[Any]: The results of the query.
         """
         with self.session_scope() as session:
-            return session.execute(query, vals).fetchall()
+            return session.execute(query, vals).all()
 
     def get_master_records(self) -> List[tables.DBMasterTable]:
         """Grab the list of master records.
@@ -239,14 +239,14 @@ def get_raw_for(self, master_id: int) -> Optional[List[tables.DbRawTable]]:
 
         return None
 
-    def get_all_params_for(self, master_id: int) -> Optional[List[tables.DbRawTable]]:
+    def get_all_params_for(self, master_id: int) -> Optional[List[tables.DbParamTable]]:
         """Get the parameters for all the iterations of a specific experiment.
 
         Args:
             master_id (int): The master id.
 
         Returns:
-            List[tables.DbRawTable] or None: The parameters or None if they don't exist.
+            List[tables.DbParamTable] or None: The parameters or None if they don't exist.
         """
         warnings.warn(
             "get_all_params_for is the same as get_param_for since there can only be one instance of any master_id",
@@ -266,14 +266,14 @@ def get_all_params_for(self, master_id: int) -> Optional[List[tables.DbRawTable]
 
         return None
 
-    def get_param_for(self, master_id: int) -> Optional[List[tables.DbRawTable]]:
+    def get_param_for(self, master_id: int) -> Optional[List[tables.DbParamTable]]:
         """Get the parameters for a specific iteration of a specific experiment.
 
         Args:
             master_id (int): The master id.
 
         Returns:
-            List[tables.DbRawTable] or None: The parameters or None if they don't exist.
+            List[tables.DbParamTable] or None: The parameters or None if they don't exist.
         """
         raw_record = self.get_raw_for(master_id)
 
@@ -284,14 +284,14 @@ def get_param_for(self, master_id: int) -> Optional[List[tables.DbRawTable]]:
 
         return None
 
-    def get_all_outcomes_for(self, master_id: int) -> Optional[List[tables.DbRawTable]]:
+    def get_all_outcomes_for(self, master_id: int) -> Optional[List[tables.DbOutcomeTable]]:
         """Get the outcomes for all the iterations of a specific experiment.
 
         Args:
             master_id (int): The master id.
 
         Returns:
-            List[tables.DbRawTable] or None: The outcomes or None if they don't exist.
+            List[tables.DbOutcomeTable] or None: The outcomes or None if they don't exist.
         """
         warnings.warn(
             "get_all_outcomes_for is the same as get_outcome_for since there can only be one instance of any master_id",
@@ -311,14 +311,14 @@ def get_all_outcomes_for(self, master_id: int) -> Optional[List[tables.DbRawTabl
 
         return None
 
-    def get_outcome_for(self, master_id: int) -> Optional[List[tables.DbRawTable]]:
+    def get_outcome_for(self, master_id: int) -> Optional[List[tables.DbOutcomeTable]]:
         """Get the outcomes for a specific iteration of a specific experiment.
 
         Args:
             master_id (int): The master id.
 
         Returns:
-            List[tables.DbRawTable] or None: The outcomes or None if they don't exist.
+            List[tables.DbOutcomeTable] or None: The outcomes or None if they don't exist.
         """
         raw_record = self.get_raw_for(master_id)
 

aepsych/server/message_handlers/handle_setup.py

@@ -11,6 +11,7 @@
 from aepsych.config import Config
 from aepsych.strategy import SequentialStrategy
 from aepsych.version import __version__
+from aepsych.database.data_fetcher import DataFetcher
 
 logger = utils_logging.getLogger(logging.INFO)
 
@@ -33,6 +34,10 @@ def _configure(server, config):
     server.strat = SequentialStrategy.from_config(config)
     server.strat_id = server.n_strats - 1  # 0-index strats
 
+    for strat in server.strat.strat_list: 
+        fetcher = DataFetcher.from_config(config, strat.name)
+        fetcher.warm_start_strat(server, strat)
+
     return server.strat_id
 
 
@@ -88,13 +93,11 @@ def handle_setup(server, request):
                 par_id = tempconfig["metadata"].get("participant_id", fallback=None)
 
                 # This may be populated when replaying
-                if server._db_master_record is not None:
+                exp_id = tempconfig["metadata"].get("experiment_id", fallback=None)
+                if exp_id is None and server._db_master_record is not None:
                     exp_id = server._db_master_record.experiment_id
-                else:
-                    exp_id = tempconfig["metadata"].get("experiment_id", fallback=None)
 
                 extra_metadata = tempconfig.jsonifyMetadata(only_extra=True)
-
                 server._db_master_record = server.db.record_setup(
                     description=exp_desc,
                     name=exp_name,

aepsych/server/server.py

@@ -61,7 +61,7 @@ def __init__(self, socket=None, database_path=None):
         self.exit_server_loop = False
         self._db_master_record = None
         self._db_raw_record = None
-        self.db = db.Database(database_path)
+        self.db: db.Database = db.Database(database_path)
         self.skip_computations = False
         self.strat_names = None
 

aepsych/strategy.py

@@ -542,6 +542,21 @@ def n_trials(self) -> int:
         )
         return self.min_asks
 
+    def pre_warm_model(self, x: torch.Tensor, y: torch.Tensor) -> None:
+        """
+        Adds new data points to the strategy, and normalizes the inputs.
+        We speceifically disregard the n return value of normalize_inputs here in order 
+        to stop warm start data from affecting the trials run length. 
+
+        Args:
+            x torch.Tensor: The input data points.
+            y torch.Tensor: The output data points.
+
+        """ 
+        # warming the model shouldn't affect strategy.n
+        self.x, self.y, n = self.normalize_inputs(x, y)
+        self._model_is_fresh = False
+
     def add_data(
         self, x: Union[np.ndarray, torch.Tensor], y: Union[np.ndarray, torch.Tensor]
     ) -> None:

aepsych/utils.py

@@ -431,3 +431,9 @@ def get_dims(config: Config) -> int:
     except KeyError:
         # Likely old style of parameter definition, fallback to looking at a bound
         return len(config.getlist("common", "lb", element_type=float))
+
+def generate_default_outcome_names(count: int) -> List[str]:
+    if count == 1:
+        return ["outcome"]
+
+    return ["outcome_" + i for i in range(count)]

tests/server/test_server.py

@@ -57,13 +57,18 @@
 
 
 class BaseServerTestCase(unittest.TestCase):
+    #so that this can be overridden for tests that require specific databases.
+    @property
+    def database_path(self):
+        return "./{}_test_server.db".format(str(uuid.uuid4().hex))
+
     def setUp(self):
         # setup logger
         server.logger = utils_logging.getLogger(logging.DEBUG, "logs")
         # random port
         socket = server.sockets.PySocket(port=0)
         # random datebase path name without dashes
-        database_path = "./{}_test_server.db".format(str(uuid.uuid4().hex))
+        database_path = self.database_path
         self.s = server.AEPsychServer(socket=socket, database_path=database_path)
         self.db_name = database_path.split("/")[1]
         self.db_path = database_path
@@ -88,7 +93,7 @@ def dummy_create_setup(self, server, request=None):
         )
 
 
-class ServerTestCase(BaseServerTestCase):
+class ServerTestCase(BaseServerTestCase):    
     def test_final_strat_serialization(self):
         setup_request = {
             "type": "setup",