Merge pull request #114 from googleinterns/warmup-model

Warmup model

app/visualize.py

@@ -26,7 +26,9 @@
 PIVOTS_COLORS = {"end_point":'red', "start_point":'green'} 
 ICON_TAGS = ['tourism','amenity', 'shop', 'leisure']
 
-icon_dir = os.path.abspath("./static/osm_icons")
+dirname = os.path.dirname(__file__)
+icon_dir = os.path.join(dirname, "./static/osm_icons")
+
 
 onlyfiles = [
   f for f in os.listdir(icon_dir) if os.path.isfile(os.path.join(icon_dir, f))]

cabby/evals/testdata/sample_test_evals_real.tsv

@@ -0,0 +1,2 @@
+0-0	40.73999276636397	-73.99324488535609	40.74374243749733	-73.98208530470956
+0-1	40.721682513136415	-74.00469558549776	40.717490300684155	-73.97684515379727

cabby/geo/geo_item.py

@@ -149,6 +149,7 @@ class RVSSample:
   geo_landmarks: Dict[str, gpd.GeoDataFrame] = attr.ib()
   geo_features: Dict[str, Any] = attr.ib()
   route_len: int = attr.ib()
+  route: list = attr.ib()
   instructions: str = attr.ib()
   id: int = attr.ib()
   version: float = attr.ib()
@@ -169,6 +170,7 @@ def to_rvs_sample(self,
               landmark_list,
               geo_entity.geo_features,
               route_length,
+              geo_entity.route.coords[:],
               instructions,
               id,
               VERSION,
@@ -205,6 +207,7 @@ def save(entities: Sequence[GeoEntity], path_to_save: str):
     mode = 'w'
 
   for geo_type, pivots_gdf in geo_types_all.items():
+    pivots_gdf = pivots_gdf.set_crs('epsg:4326')
     pivots_gdf.to_file(
         path_to_save, layer=geo_type, mode=mode, driver=_Geo_DataFrame_Driver)
 

cabby/geo/util.py

@@ -506,7 +506,7 @@ def get_line_length(line: LineString) -> float:
   return dist
 
 
-def point_from_list_coord(coord: Sequence) -> Point:
+def point_from_list_coord_yx(coord: Sequence) -> Point:
   '''Converts coordinates in list format (latitude and longtitude) to Point.
   E.g, of list [40.715865, -74.037258].
   Arguments:
@@ -519,16 +519,33 @@ def point_from_list_coord(coord: Sequence) -> Point:
 
   return Point(lon, lat)
 
+def point_from_list_coord_xy(coord: Sequence) -> Point:
+  '''Converts coordinates in list format (longtitude and latitude) to Point.
+  E.g, of list [-74.037258, 40.715865].
+  Arguments:
+    coord: A lat-lng coordinate to be converted to a point.
+  Returns:
+    A point.
+  '''
+  lat = coord[1]
+  lon = coord[0]
+
+  return Point(lon, lat)
+
 
 def point_from_str_coord_yx(coord_str: Text) -> Point:
   '''Converts coordinates in string format (latitude and longtitude) to Point.
-  E.g, of string '(40.715865, -74.037258)' or 'POINT(40.715865 -74.037258)'. 
+  E.g, of string '(40.715865, -74.037258)' or '[40.715865, -74.037258]' or 'POINT(40.715865 -74.037258)'. 
   Arguments:
     coord: A lat-lng coordinate to be converted to a point.
   Returns:
     A point.
   '''
-  list_coords_str = coord_str.replace("POINT", "").replace("(", "").replace(")", "").split(',')
+  coord_str = coord_str.replace("POINT", "").replace("(", "").replace(")", "")
+  coord_str = coord_str.replace("[", "").replace("]", "")
+
+  list_coords_str = coord_str.split(',')
+
   if len(list_coords_str)==1:
     list_coords_str = list_coords_str[0].split(' ')
 
@@ -540,9 +557,9 @@ def point_from_str_coord_yx(coord_str: Text) -> Point:
 
 def point_from_str_coord_xy(coord_str: Text) -> Point:
   '''Converts coordinates in string format (latitude and longtitude) to Point.
-  E.g, of string '(40.715865, -74.037258)' or 'POINT(40.715865 -74.037258)'. 
+  E.g, of string '(-74.037258, 40.715865)' or '[-74.037258, 40.715865]' or 'POINT(-74.037258 40.715865)'. 
   Arguments:
-    coord: A lat-lng coordinate to be converted to a point.
+    coord: A lng-lat coordinate to be converted to a point.
   Returns:
     A point.
   '''

cabby/geo/walk.py

@@ -1097,9 +1097,11 @@ def generate_and_save_rvs_routes(self,
 def load_entities(path: str) -> Sequence[geo_item.GeoEntity]:
   if not os.path.exists(path):
     return []
+
   geo_types_all = {}
   for landmark_type in LANDMARK_TYPES:
     geo_types_all[landmark_type] = gpd.read_file(path, layer=landmark_type)
+
   geo_types_all['route'] = gpd.read_file(path, layer='path_features')['geometry']
   geo_types_all['path_features'] = gpd.read_file(path, layer='path_features')
   geo_entities = []

cabby/model/dataset_item.py

@@ -116,18 +116,20 @@ class TextGeoSplit(torch.utils.data.Dataset):
   """
   def __init__(self, text_tokenizer, s2_tokenizer, data: pd.DataFrame, s2level: int, 
     unique_cells_df: pd.DataFrame, cellid_to_label: Dict[int, int], 
-    dprob: util.DistanceProbability, is_dist: Boolean = False):
+    model_type: str, dprob: util.DistanceProbability, is_dist: Boolean = False):
 
     self.text_tokenizer = text_tokenizer
     self.s2_tokenizer = s2_tokenizer
 
     self.is_dist = is_dist
 
-    data = data.assign(point=data.end_point)
+    data = data.assign(end_point=data.end_point)
 
-    data['cellid'] = data.point.apply(
+
+    data['cellid'] = data.end_point.apply(
       lambda x: gutil.cellid_from_point(x, s2level))
 
+
     data['neighbor_cells'] = data.cellid.apply(
       lambda x: gutil.neighbor_cellid(x))
 
@@ -149,11 +151,12 @@ def __init__(self, text_tokenizer, s2_tokenizer, data: pd.DataFrame, s2level: in
     data['far_cells'] = data.cellid.apply(
       lambda cellid: unique_cells_df[unique_cells_df['cellid']==cellid].far.iloc[0])
 
+
     cellids_array = np.array(data.cellid.tolist())
     neighbor_cells_array = np.array(data.neighbor_cells.tolist())
     far_cells_array = np.array(data.far_cells.tolist())
 
-    self.points = data.point.apply(
+    self.end_point = data.end_point.apply(
       lambda x: gutil.tuple_from_point(x)).tolist()
 
     self.labels = data.cellid.apply(lambda x: cellid_to_label[x]).tolist()
@@ -164,6 +167,48 @@ def __init__(self, text_tokenizer, s2_tokenizer, data: pd.DataFrame, s2level: in
     self.far_cells = self.s2_tokenizer(far_cells_array)
 
 
+    if 'T5' in model_type and 'landmarks' in data:
+      data['landmarks'] = data.landmarks.apply(
+        lambda l: [gutil.cellid_from_point(x, s2level) for x in l])
+
+      self.landmarks = self.s2_tokenizer(data.landmarks.tolist())
+
+
+    else:
+      self.landmarks = [0] * len(self.cellids)
+      logging.warning("Landmarks not processed")
+
+
+    if 'T5' in model_type and 'route' in data:
+      data['route'] = data.route.apply(
+        lambda l: [gutil.cellid_from_point(x, s2level) for x in l])
+      route_array = np.array(data.route.tolist())
+      self.route = self.s2_tokenizer(route_array)
+
+      data['route_fixed'] = data.route_fixed.apply(
+        lambda l: [gutil.cellid_from_point(x, s2level) for x in l])
+      self.route_fixed = self.s2_tokenizer(data.route_fixed.tolist())
+
+      start_point_cells = data.start_point.apply(
+        lambda x: gutil.cellid_from_point(x, s2level))
+
+
+      start_point_list = [
+        '; '.join(
+          [str(cellid_to_label[e]), str(cellid_to_label[s])]) for s, e in zip(
+            start_point_cells.tolist(), data.cellid.tolist())]
+      self.start_end = self.text_tokenizer(
+        start_point_list, truncation=True, padding=True, add_special_tokens=True)
+
+    else:
+      self.route = [0] * len(self.cellids)
+      self.route_fixed = [0] * len(self.cellids)
+      self.start_end = {
+        'attention_mask': [0] * len(self.cellids),
+        'input_ids': [0] * len(self.cellids)}
+      logging.warning("Route not processed")
+
+
   def __getitem__(self, idx: int):
     '''Supports indexing such that TextGeoDataset[i] can be used to get 
     i-th sample. 
@@ -175,18 +220,29 @@ def __getitem__(self, idx: int):
     '''
     text = {key: torch.tensor(val[idx])
         for key, val in self.encodings.items()}
-    cellid = self.cellids[idx]
+
+    cellid =  torch.tensor(self.cellids[idx])
+    landmarks =  torch.tensor(self.landmarks[idx])
+
+    route =  torch.tensor(self.route[idx])
+    route_fixed =  torch.tensor(self.route_fixed[idx])
+
+    start_end = {key: torch.tensor(val[idx])
+        for key, val in self.start_end.items()}
+
     neighbor_cells = torch.tensor(self.neighbor_cells[idx])
     far_cells = torch.tensor(self.far_cells[idx])
-    point = torch.tensor(self.points[idx])
+    end_point = torch.tensor(self.end_point[idx])
     label = torch.tensor(self.labels[idx])
     if self.is_dist:
       prob = torch.tensor(self.prob[idx])
     else:
       prob = torch.tensor([])
 
     sample = {'text': text, 'cellid': cellid, 'neighbor_cells': neighbor_cells, 
-      'far_cells': far_cells, 'point': point, 'label': label, 'prob': prob}
+      'far_cells': far_cells, 'end_point': end_point, 'label': label, 'prob': prob, 
+      'landmarks': landmarks, 'route': route, 'route_fixed': route_fixed, 
+      'start_end_input_ids': start_end['input_ids'], 'start_end_attention_mask': start_end['attention_mask']}
 
     return sample
 
@@ -195,6 +251,4 @@ def __len__(self):
 
 def calc_dist(start, unique_cells_df):
   return unique_cells_df.swifter.apply(
-    lambda end: gutil.get_distance_between_points(start, end.point), axis=1)
-
-
+    lambda end: gutil.get_distance_between_points(start, end.point), axis=1)