Fix memory usage (#751)

* clear cache of flow.tree

* categorical time period dtype

* add pydantic for tests

* use time_label_dtype only when available

* allow missing taz skim_dict

* recover tree when needed

* predigitized time periods

* pass sh_tree back again for tracing

* better error message

Author: jpn--

activitysim/abm/models/parking_location_choice.py

@@ -318,7 +318,7 @@ def parking_location(
         if "trip_period" not in trips_merged_df:
             # TODO: resolve this to the skim time period index not the label, it will be faster
             trips_merged_df["trip_period"] = network_los.skim_time_period_label(
-                trips_merged_df[proposed_trip_departure_period]
+                trips_merged_df[proposed_trip_departure_period], as_cat=True
             )
         model_settings["TRIP_DEPARTURE_PERIOD"] = "trip_period"
 

activitysim/abm/models/trip_mode_choice.py

@@ -73,7 +73,7 @@ def trip_mode_choice(
     # setup skim keys
     assert "trip_period" not in trips_merged
     trips_merged["trip_period"] = network_los.skim_time_period_label(
-        trips_merged.depart
+        trips_merged.depart, as_cat=True
     )
 
     orig_col = "origin"

activitysim/abm/models/util/logsums.py

@@ -75,10 +75,10 @@ def compute_logsums(
     # FIXME - are we ok with altering choosers (so caller doesn't have to set these)?
     if (in_period_col is not None) and (out_period_col is not None):
         choosers["in_period"] = network_los.skim_time_period_label(
-            choosers[in_period_col]
+            choosers[in_period_col], as_cat=True
         )
         choosers["out_period"] = network_los.skim_time_period_label(
-            choosers[out_period_col]
+            choosers[out_period_col], as_cat=True
         )
     elif ("in_period" not in choosers.columns) and (
         "out_period" not in choosers.columns
@@ -92,17 +92,21 @@ def compute_logsums(
                 and tour_purpose in model_settings["OUT_PERIOD"]
             ):
                 choosers["in_period"] = network_los.skim_time_period_label(
-                    model_settings["IN_PERIOD"][tour_purpose]
+                    model_settings["IN_PERIOD"][tour_purpose],
+                    as_cat=True,
+                    broadcast_to=choosers.index,
                 )
                 choosers["out_period"] = network_los.skim_time_period_label(
-                    model_settings["OUT_PERIOD"][tour_purpose]
+                    model_settings["OUT_PERIOD"][tour_purpose],
+                    as_cat=True,
+                    broadcast_to=choosers.index,
                 )
         else:
             choosers["in_period"] = network_los.skim_time_period_label(
-                model_settings["IN_PERIOD"]
+                model_settings["IN_PERIOD"], as_cat=True, broadcast_to=choosers.index
             )
             choosers["out_period"] = network_los.skim_time_period_label(
-                model_settings["OUT_PERIOD"]
+                model_settings["OUT_PERIOD"], as_cat=True, broadcast_to=choosers.index
             )
     else:
         logger.error("Choosers table already has columns 'in_period' and 'out_period'.")

activitysim/abm/models/util/mode.py

@@ -131,8 +131,12 @@ def run_tour_mode_choice_simulate(
     assert ("in_period" not in choosers) and ("out_period" not in choosers)
     in_time = skims["in_time_col_name"]
     out_time = skims["out_time_col_name"]
-    choosers["in_period"] = network_los.skim_time_period_label(choosers[in_time])
-    choosers["out_period"] = network_los.skim_time_period_label(choosers[out_time])
+    choosers["in_period"] = network_los.skim_time_period_label(
+        choosers[in_time], as_cat=True
+    )
+    choosers["out_period"] = network_los.skim_time_period_label(
+        choosers[out_time], as_cat=True
+    )
 
     expressions.annotate_preprocessors(
         state, choosers, locals_dict, skims, model_settings, trace_label

activitysim/abm/models/util/vectorize_tour_scheduling.py

@@ -185,6 +185,12 @@ def dedupe_alt_tdd(state: workflow.State, alt_tdd, tour_purpose, trace_label):
 
     logger.info("tdd_alt_segments specified for representative logsums")
 
+    if tdd_segments is not None:
+        # apply categorical dtypes
+        tdd_segments["time_period"] = tdd_segments["time_period"].astype(
+            alt_tdd["out_period"].dtype
+        )
+
     with chunk.chunk_log(
         state, tracing.extend_trace_label(trace_label, "dedupe_alt_tdd")
     ) as chunk_sizer:
@@ -328,11 +334,12 @@ def compute_tour_scheduling_logsums(
     assert "out_period" not in alt_tdd
     assert "in_period" not in alt_tdd
 
-    # FIXME:MEMORY
-    #  These two lines each generate a massive array of strings,
-    #  using a bunch of RAM and slowing things down.
-    alt_tdd["out_period"] = network_los.skim_time_period_label(alt_tdd["start"])
-    alt_tdd["in_period"] = network_los.skim_time_period_label(alt_tdd["end"])
+    alt_tdd["out_period"] = network_los.skim_time_period_label(
+        alt_tdd["start"], as_cat=True
+    )
+    alt_tdd["in_period"] = network_los.skim_time_period_label(
+        alt_tdd["end"], as_cat=True
+    )
 
     alt_tdd["duration"] = alt_tdd["end"] - alt_tdd["start"]
 
@@ -383,17 +390,28 @@ def compute_tour_scheduling_logsums(
 
         # tracing.log_runtime(model_name=trace_label, start_time=t0)
 
-        # redupe - join the alt_tdd_period logsums to alt_tdd to get logsums for alt_tdd
-        logsums = (
-            pd.merge(
-                alt_tdd.reset_index(),
-                deduped_alt_tdds.reset_index(),
-                on=[index_name] + redupe_columns,
-                how="left",
-            )
-            .set_index(index_name)
-            .logsums
-        )
+        logsums = pd.Series(data=0, index=alt_tdd.index, dtype=np.float64)
+        left_on = [alt_tdd.index]
+        right_on = [deduped_alt_tdds.index]
+        for i in redupe_columns:
+            if (
+                alt_tdd[i].dtype == "category"
+                and alt_tdd[i].dtype.ordered
+                and alt_tdd[i].dtype == deduped_alt_tdds[i].dtype
+            ):
+                left_on += [alt_tdd[i].cat.codes]
+                right_on += [deduped_alt_tdds[i].cat.codes]
+            else:
+                left_on += [alt_tdd[i].to_numpy()]
+                right_on += [deduped_alt_tdds[i].to_numpy()]
+
+        logsums.iloc[:] = pd.merge(
+            pd.DataFrame(index=alt_tdd.index),
+            deduped_alt_tdds.logsums,
+            left_on=left_on,
+            right_on=right_on,
+            how="left",
+        ).logsums.to_numpy()
         chunk_sizer.log_df(trace_label, "logsums", logsums)
 
         del deduped_alt_tdds

activitysim/abm/tables/landuse.py

@@ -23,12 +23,16 @@ def land_use(state: workflow.State):
 
     sharrow_enabled = state.settings.sharrow
     if sharrow_enabled:
+        err_msg = (
+            "a zero-based land_use index is required for sharrow,\n"
+            "try adding `recode_pipeline_columns: true` to your settings file."
+        )
         # when using sharrow, the land use file must be organized (either in raw
         # form or via recoding) so that the index is zero-based and contiguous
-        assert df.index.is_monotonic_increasing
-        assert df.index[0] == 0
-        assert df.index[-1] == len(df.index) - 1
-        assert df.index.dtype.kind == "i"
+        assert df.index.is_monotonic_increasing, err_msg
+        assert df.index[0] == 0, err_msg
+        assert df.index[-1] == len(df.index) - 1, err_msg
+        assert df.index.dtype.kind == "i", err_msg
 
     # try to make life easy for everybody by keeping everything in canonical order
     # but as long as coalesce_pipeline doesn't sort tables it coalesces, it might not stay in order

activitysim/core/flow.py

@@ -267,6 +267,7 @@ def skims_mapping(
     parking_col_name=None,
     zone_layer=None,
     primary_origin_col_name=None,
+    predigitized_time_periods=False,
 ):
     logger.info("loading skims_mapping")
     logger.info(f"- orig_col_name: {orig_col_name}")
@@ -337,6 +338,10 @@ def skims_mapping(
                 ),
             )
         else:
+            if predigitized_time_periods:
+                time_rel = "_code ->"
+            else:
+                time_rel = " @"
             return dict(
                 # TODO:SHARROW: organize dimensions.
                 odt_skims=skim_dataset,
@@ -347,16 +352,16 @@ def skims_mapping(
                 relationships=(
                     f"df._orig_col_name -> odt_skims.{odim}",
                     f"df._dest_col_name -> odt_skims.{ddim}",
-                    "df.out_period      @  odt_skims.time_period",
+                    f"df.out_period{time_rel} odt_skims.time_period",
                     f"df._dest_col_name -> dot_skims.{odim}",
                     f"df._orig_col_name -> dot_skims.{ddim}",
-                    "df.in_period       @  dot_skims.time_period",
+                    f"df.in_period{time_rel} dot_skims.time_period",
                     f"df._orig_col_name -> odr_skims.{odim}",
                     f"df._dest_col_name -> odr_skims.{ddim}",
-                    "df.in_period       @  odr_skims.time_period",
+                    f"df.in_period{time_rel} odr_skims.time_period",
                     f"df._dest_col_name -> dor_skims.{odim}",
                     f"df._orig_col_name -> dor_skims.{ddim}",
-                    "df.out_period      @  dor_skims.time_period",
+                    f"df.out_period{time_rel} dor_skims.time_period",
                     f"df._orig_col_name -> od_skims.{odim}",
                     f"df._dest_col_name -> od_skims.{ddim}",
                 ),
@@ -525,6 +530,15 @@ def new_flow(
 
         cache_dir = state.filesystem.get_sharrow_cache_dir()
         logger.debug(f"flow.cache_dir: {cache_dir}")
+        predigitized_time_periods = False
+        if "out_period" in choosers and "in_period" in choosers:
+            if (
+                choosers["out_period"].dtype == "category"
+                and choosers["in_period"].dtype == "category"
+            ):
+                choosers["out_period_code"] = choosers["out_period"].cat.codes
+                choosers["in_period_code"] = choosers["in_period"].cat.codes
+                predigitized_time_periods = True
         skims_mapping_ = skims_mapping(
             state,
             orig_col_name,
@@ -534,6 +548,7 @@ def new_flow(
             parking_col_name=parking_col_name,
             zone_layer=zone_layer,
             primary_origin_col_name=primary_origin_col_name,
+            predigitized_time_periods=predigitized_time_periods,
         )
         if size_term_mapping is None:
             size_term_mapping = {}
@@ -774,6 +789,9 @@ def apply_flow(
         it ever again, but having a reference to it available later can be useful
         in debugging and tracing.  Flows are cached and reused anyway, so it is
         generally not important to delete this at any point to free resources.
+    tree : sharrow.DataTree
+        The tree data used to compute the flow result.  It is seperate from the
+        flow to prevent it from being cached with the flow.
     """
     if sh is None:
         return None, None
@@ -800,7 +818,7 @@ def apply_flow(
                 logger.error(f"error in apply_flow: {err!s}")
                 if required:
                     raise
-                return None, None
+                return None, None, None
             else:
                 raise
         with logtime(f"{flow.name}.load", trace_label or ""):
@@ -822,7 +840,9 @@ def apply_flow(
                     logger.error(f"error in apply_flow: {err!s}")
                     if required:
                         raise
-                    return None, flow
+                    tree = flow.tree
+                    flow.tree = None
+                    return None, flow, tree
                 raise
             except Exception as err:
                 logger.error(f"error in apply_flow: {err!s}")
@@ -833,4 +853,6 @@ def apply_flow(
                 # Detecting compilation activity when in production mode is a bug
                 # that should be investigated.
                 tracing.timing_notes.add(f"compiled:{flow.name}")
-            return flow_result, flow
+            tree = flow.tree
+            flow.tree = None
+            return flow_result, flow, tree

activitysim/core/interaction_simulate.py

@@ -171,7 +171,7 @@ def replace_in_index_level(mi, level, *repls):
 
             timelogger.mark("sharrow preamble", True, logger, trace_label)
 
-            sh_util, sh_flow = apply_flow(
+            sh_util, sh_flow, sh_tree = apply_flow(
                 state,
                 spec_sh,
                 df,
@@ -187,10 +187,13 @@ def replace_in_index_level(mi, level, *repls):
                     index=df.index if extra_data is None else None,
                 )
                 chunk_sizer.log_df(trace_label, "sh_util", None)  # hand off to caller
+                if sharrow_enabled != "test":
+                    # if not testing sharrow, we are done with this object now.
+                    del sh_util
 
             timelogger.mark("sharrow flow", True, logger, trace_label)
         else:
-            sh_util, sh_flow = None, None
+            sh_util, sh_flow, sh_tree = None, None, None
             timelogger.mark("sharrow flow", False)
 
         if (
@@ -404,9 +407,9 @@ def to_series(x):
         if sh_flow is not None and trace_rows is not None and trace_rows.any():
             assert type(trace_rows) == np.ndarray
             sh_utility_fat = sh_flow.load_dataarray(
-                # sh_flow.tree.replace_datasets(
-                #     df=df.iloc[trace_rows],
-                # ),
+                sh_tree.replace_datasets(
+                    df=df.iloc[trace_rows],
+                ),
                 dtype=np.float32,
             )
             sh_utility_fat = sh_utility_fat[trace_rows, :]

activitysim/core/los.py

@@ -845,7 +845,9 @@ def get_tappairs3d(self, otap, dtap, dim3, key):
 
         return s.values
 
-    def skim_time_period_label(self, time_period, fillna=None):
+    def skim_time_period_label(
+        self, time_period, fillna=None, as_cat=False, broadcast_to=None
+    ):
         """
         convert time period times to skim time period labels (e.g. 9 -> 'AM')
 
@@ -873,6 +875,14 @@ def skim_time_period_label(self, time_period, fillna=None):
         assert 0 == model_time_window_min % period_minutes
         total_periods = model_time_window_min / period_minutes
 
+        try:
+            time_label_dtype = self.skim_dicts["taz"].time_label_dtype
+        except (KeyError, AttributeError):
+            # if the "taz" skim_dict is missing, or if using old SkimDict
+            # instead of SkimDataset, this labeling shortcut is unavailable.
+            time_label_dtype = str
+            as_cat = False
+
         # FIXME - eventually test and use np version always?
         if np.isscalar(time_period):
             bin = (
@@ -888,6 +898,12 @@ def skim_time_period_label(self, time_period, fillna=None):
                 result = self.skim_time_periods["labels"].get(bin, default=default)
             else:
                 result = self.skim_time_periods["labels"][bin]
+            if broadcast_to is not None:
+                result = pd.Series(
+                    data=result,
+                    index=broadcast_to,
+                    dtype=time_label_dtype if as_cat else str,
+                )
         else:
             result = pd.cut(
                 time_period,
@@ -898,8 +914,10 @@ def skim_time_period_label(self, time_period, fillna=None):
             if fillna is not None:
                 default = self.skim_time_periods["labels"][fillna]
                 result = result.fillna(default)
-            result = result.astype(str)
-
+            if as_cat:
+                result = result.astype(time_label_dtype)
+            else:
+                result = result.astype(str)
         return result
 
     def get_tazs(self, state):

activitysim/core/simulate.py

@@ -536,7 +536,7 @@ def eval_utilities(
         locals_dict.update(state.get_global_constants())
         if locals_d is not None:
             locals_dict.update(locals_d)
-        sh_util, sh_flow = apply_flow(
+        sh_util, sh_flow, sh_tree = apply_flow(
             state,
             spec_sh,
             choosers,
@@ -652,7 +652,7 @@ def eval_utilities(
         if sh_flow is not None:
             try:
                 data_sh = sh_flow.load(
-                    sh_flow.tree.replace_datasets(
+                    sh_tree.replace_datasets(
                         df=choosers.iloc[offsets],
                     ),
                     dtype=np.float32,
@@ -731,7 +731,7 @@ def eval_utilities(
                 )
                 print(f"{sh_util.shape=}")
                 print(misses)
-                _sh_flow_load = sh_flow.load()
+                _sh_flow_load = sh_flow.load(sh_tree)
                 print("possible problematic expressions:")
                 for expr_n, expr in enumerate(exprs):
                     closeness = np.isclose(