STEM / code /3dSAGER /data_partition.py
eduzrh
Complete STEM benchmark results: all experiments + comprehensive README
693f42e
Raw
History Blame Contribute Delete
34.5 kB
import json
import os
import config
from utils import *
import pickle as pkl
import argparse
from time import time
from multiprocessing import Pool, cpu_count
import numpy as np
from scipy.spatial import KDTree
class DataPartitionGenerator:
def __init__(self, args, min_surfaces_num=10):
self.dataset_name = args.dataset_name
self.min_surfaces_num = min_surfaces_num
self.train_neg_samples_list = args.train_neg_samples_list
self.train_size_ratio_list = args.train_size_ratio_list
self.test_size_ratio_list = args.test_size_ratio_list
self.test_negative_samples_list = args.test_negative_samples_list
self.cands_ids, self.index_ids = self._get_cands_and_index_ids()
# self.create_dataset_partition_dict()
def _get_cands_and_index_ids(self):
dataset_config = json.load(open('dataset_configs.json'))[self.dataset_name]
main_object_dict = getattr(self, f'_read_objects_{self.dataset_name}')(dataset_config)
cands_ids = set(main_object_dict['cands'].keys())
index_ids = set(main_object_dict['index'].keys())
# Store centroids for grid split and spatial negative sampling
self.cands_centroids = {k: np.asarray(main_object_dict['cands'][k]['centroid'], dtype=np.float64)
for k in main_object_dict['cands']}
self.index_centroids = {k: np.asarray(main_object_dict['index'][k]['centroid'], dtype=np.float64)
for k in main_object_dict['index']}
# Store pre-computed grid cells if available (from preprocess_hague.py cache)
self.cands_grid_cells = {k: main_object_dict['cands'][k]['grid_cell']
for k in main_object_dict['cands']
if 'grid_cell' in main_object_dict['cands'][k]}
self.index_grid_cells = {k: main_object_dict['index'][k]['grid_cell']
for k in main_object_dict['index']
if 'grid_cell' in main_object_dict['index'][k]}
self.grid_meta = main_object_dict.get('grid_meta', {})
del main_object_dict
return cands_ids, index_ids
# ------------------------------------------------------------------
# Spatial grid helpers
# ------------------------------------------------------------------
@staticmethod
def _assign_grid_cells(centroids_dict: dict, cell_size: float) -> dict:
"""
Assign each building to a grid cell based on its 2D centroid.
Parameters
----------
centroids_dict : {id: np.ndarray([x, y, z])}
cell_size : float — cell side length in the same units as the CRS (meters)
Returns
-------
{building_id: (cell_x, cell_y)}
"""
if not centroids_dict:
return {}
coords = np.array([c[:2] for c in centroids_dict.values()])
x_min, y_min = coords[:, 0].min(), coords[:, 1].min()
cell_assignments = {}
for bid, centroid in centroids_dict.items():
cx = int(np.floor((centroid[0] - x_min) / cell_size))
cy = int(np.floor((centroid[1] - y_min) / cell_size))
cell_assignments[bid] = (cx, cy)
return cell_assignments
@staticmethod
def _grid_split(cell_assignments: dict, train_ratio: float, seed: int,
contiguous_test: bool = True):
"""
Assign grid cells to train or test and return per-building split.
Each building belongs to exactly one cell (assigned by centroid).
No building can appear in both splits.
If contiguous_test=True (default): test cells are selected via BFS from the
bottom-left corner of the grid, producing a single contiguous geographic block.
This keeps the demo app's visible area coherent (train region) and ensures the
test zone is a spatially distinct held-out area.
If contiguous_test=False: cells are assigned randomly (original behaviour).
Returns
-------
train_ids : set
test_ids : set
"""
from collections import deque
unique_cells = list(set(cell_assignments.values()))
n_train_cells = max(1, int(round(train_ratio * len(unique_cells))))
n_test_cells = len(unique_cells) - n_train_cells
if contiguous_test and n_test_cells > 0:
cell_set = set(map(tuple, unique_cells))
# BFS starting from the bottom-left corner (min cx+cy) to grow a
# contiguous test region
start = min(cell_set, key=lambda c: (c[0] + c[1]))
visited = {start}
queue = deque([start])
test_cells = []
while queue and len(test_cells) < n_test_cells:
cx, cy = queue.popleft()
test_cells.append((cx, cy))
for nc in [(cx + 1, cy), (cx - 1, cy), (cx, cy + 1), (cx, cy - 1)]:
if nc in cell_set and nc not in visited:
visited.add(nc)
queue.append(nc)
test_cells = set(test_cells)
train_cells = cell_set - test_cells
strategy = "contiguous BFS"
else:
rng = np.random.default_rng(seed)
rng.shuffle(unique_cells)
train_cells = set(map(tuple, unique_cells[:n_train_cells]))
test_cells = set(map(tuple, unique_cells[n_train_cells:]))
strategy = "random"
train_ids = {bid for bid, cell in cell_assignments.items() if cell in train_cells}
test_ids = {bid for bid, cell in cell_assignments.items() if cell in test_cells}
print(f"[GridSplit] {len(train_cells)} train cells / {len(test_cells)} test cells | "
f"{len(train_ids)} train buildings / {len(test_ids)} test buildings "
f"({strategy})")
return train_ids, test_ids
def _build_spatial_index(self, train_index_ids: set) -> None:
"""
Build a KDTree on train-split index building centroids for
neighborhood-aware negative sampling.
"""
ids_list = [bid for bid in train_index_ids if bid in self.index_centroids]
if not ids_list:
self._index_kdtree = None
self._index_ids_list = []
return
xy = np.array([self.index_centroids[bid][:2] for bid in ids_list], dtype=np.float64)
self._index_kdtree = KDTree(xy)
self._index_ids_list = ids_list
# ------------------------------------------------------------------
def create_dataset_partition_dict(self, seed):
self.seed = seed
cands_ids = self.cands_ids
index_ids = self.index_ids
# --- Spatial grid split ---
cell_size = config.DataPartition.grid_cell_size
train_ratio = config.DataPartition.train_ratio
intersection_ids = cands_ids.intersection(index_ids)
# Use pre-computed grid cells from the preprocessed cache when available
# (real-world EPSG:7415 coordinates → correct spatial splits)
if self.cands_grid_cells:
intersection_cell_assignments = {bid: self.cands_grid_cells[bid]
for bid in intersection_ids
if bid in self.cands_grid_cells}
print(f"[GridSplit] Using pre-computed grid cells "
f"(cell_size={self.grid_meta.get('cell_size', cell_size):.0f} m, "
f"{len(intersection_cell_assignments)} buildings)")
else:
intersection_centroids = {bid: self.index_centroids[bid]
for bid in intersection_ids if bid in self.index_centroids}
intersection_cell_assignments = self._assign_grid_cells(intersection_centroids, cell_size)
train_intersection_ids, test_intersection_ids = self._grid_split(
intersection_cell_assignments, train_ratio, seed,
contiguous_test=config.DataPartition.contiguous_test
)
# Build KDTree on train-split index buildings for neighborhood negative sampling
self._build_spatial_index(train_intersection_ids)
train_negative_sampling_dict = self._get_train_negative_sampling_dict(
train_intersection_ids, index_ids
)
test_dict = self._get_test_ids_dict(
test_intersection_ids, index_ids, train_negative_sampling_dict
)
dataset_partition_dict = {'train': {'negative_sampling': train_negative_sampling_dict}, 'test': test_dict}
self._save_dataset_partition_dict(dataset_partition_dict)
def _get_train_negative_sampling_dict(self, cands_ids, index_ids):
np.random.seed(self.seed)
train_ids_dict = {}
intersection_set = cands_ids.intersection(index_ids)
for train_size, ratio_val in self.train_size_ratio_list.items():
print(f"Creating training data ({train_size})")
train_ids_dict[train_size] = {}
train_ids_size_num = int(ratio_val * len(intersection_set))
train_ids_for_curr_size = set(np.random.choice(list(intersection_set), train_ids_size_num, replace=False))
for neg_samples_num in self.train_neg_samples_list:
train_ids_dict[train_size][neg_samples_num] = self._get_pairs_per_neg_samples(train_ids_for_curr_size,
index_ids,
neg_samples_num)
return train_ids_dict
def _generate_pairs_for_id(self, args):
"""
Generate one positive pair + neg_samples_num negative pairs for cand_id.
Negative sampling strategy (config.Blocking.neighborhood_neg_ratio):
- Up to `neighborhood_neg_ratio` fraction drawn from buildings within
`neighborhood_radius` meters (same neighborhood, hard negatives).
- Remainder filled randomly from the full index pool.
Falls back to fully random sampling if the KDTree is unavailable.
"""
cand_id, index_ids_list, neg_samples_num, seed = args
rng = np.random.default_rng(seed + hash(cand_id) % 1_000_000)
neighborhood_radius = config.Blocking.neighborhood_radius
neighborhood_neg_ratio = config.Blocking.neighborhood_neg_ratio
n_neighborhood = int(round(neg_samples_num * neighborhood_neg_ratio))
n_random = neg_samples_num - n_neighborhood
spatial_neg_ids = []
# Spatial candidates from KDTree
if (n_neighborhood > 0
and self._index_kdtree is not None
and cand_id in self.cands_centroids):
cand_xy = self.cands_centroids[cand_id][:2]
neighbor_rows = self._index_kdtree.query_ball_point(cand_xy, r=neighborhood_radius)
neighbor_ids = [self._index_ids_list[i] for i in neighbor_rows
if self._index_ids_list[i] != cand_id]
if len(neighbor_ids) >= n_neighborhood:
chosen = rng.choice(neighbor_ids, n_neighborhood, replace=False)
spatial_neg_ids = list(chosen)
else:
spatial_neg_ids = neighbor_ids # take all available
n_random = neg_samples_num - len(spatial_neg_ids)
# Random candidates to fill the remainder
exclude = set(spatial_neg_ids) | {cand_id}
remaining = [bid for bid in index_ids_list if bid not in exclude]
if n_random > 0 and remaining:
n_draw = min(n_random, len(remaining))
random_neg_ids = list(rng.choice(remaining, n_draw, replace=False))
else:
random_neg_ids = []
all_neg_ids = spatial_neg_ids + random_neg_ids
neg_pairs = [(cand_id, neg_id) for neg_id in all_neg_ids]
return [(cand_id, cand_id)] + neg_pairs
# def _get_pairs_per_neg_samples(self, ids_for_curr_size, index_ids, neg_samples_num):
# np.random.seed(self.seed)
# pos_pairs = [(cand_id, cand_id) for cand_id in ids_for_curr_size]
# neg_pairs = []
# for cand_id in ids_for_curr_size:
# neg_samples = set(np.random.choice(list(index_ids), neg_samples_num, replace=False))
# neg_pairs.extend([(cand_id, neg_sample) for neg_sample in neg_samples if neg_sample != cand_id])
# all_pairs = pos_pairs + neg_pairs
# np.random.shuffle(all_pairs)
# return all_pairs
def _init_seed(self):
np.random.seed(self.seed)
def _get_pairs_per_neg_samples(self, ids_for_curr_size, index_ids, neg_samples_num):
ids_for_curr_size_list = list(ids_for_curr_size)
index_ids_list = list(index_ids)
args = [
(cand_id, index_ids_list, neg_samples_num, self.seed)
for cand_id in ids_for_curr_size_list
]
with Pool(cpu_count(), initializer=self._init_seed, initargs=()) as pool:
results = pool.map(self._generate_pairs_for_id, args)
all_pairs = [pair for sublist in results for pair in sublist]
np.random.seed(self.seed)
np.random.shuffle(all_pairs)
return all_pairs
def _get_test_ids_dict(self, cands_ids, index_ids, train_ids_dict):
test_ids_dict = {}
intersection_set = cands_ids.intersection(index_ids)
test_ids_dict['matching'] = self._get_test_pairs_for_matching(index_ids, intersection_set, train_ids_dict)
test_ids_dict['blocking'] = self._get_test_data_for_blocking(cands_ids, index_ids,
intersection_set,
train_ids_dict)
return test_ids_dict
def _get_test_pairs_for_matching(self, index_ids, intersection_set, train_ids_dict):
print("Creating test data for matching")
test_matching_dict = {}
test_matching_dict['negative_sampling'] = self._get_negative_sampling_test_ids_dict(index_ids,
intersection_set,
train_ids_dict)
return test_matching_dict
def _get_negative_sampling_test_ids_dict(self, index_ids, intersection_set, train_ids_dict):
np.random.seed(self.seed)
local_test_ids_dict = {}
for test_size, ratio_val in self.test_size_ratio_list.items():
print(f"Creating test data for matching ({test_size})")
local_test_ids_dict[test_size] = {}
corresponding_train_cands_ids = set \
([pair[0] for pair in train_ids_dict[test_size][self.train_neg_samples_list[0]]])
potential_test_ids = intersection_set - corresponding_train_cands_ids
test_ids_size_num = int(ratio_val * len(potential_test_ids))
test_ids_for_curr_size = set(np.random.choice(list(potential_test_ids), test_ids_size_num, replace=False))
for test_neg_samples in self.test_negative_samples_list:
local_test_ids_dict[test_size][test_neg_samples] = self._get_pairs_per_neg_samples \
(test_ids_for_curr_size, index_ids, test_neg_samples)
return local_test_ids_dict
def _get_test_data_for_blocking(self, cands_ids, index_ids, intersection_set, train_ids_dict, non_matched_rat=0.2):
test_blocking_dict = defaultdict(dict)
for test_size, ratio_val in self.test_size_ratio_list.items():
print(f"Creating test data for blocking ({test_size})")
corresponding_train_cands_ids = set([pair[0] for pair in
train_ids_dict[test_size][self.train_neg_samples_list[0]]])
potential_cands_test_ids = intersection_set - corresponding_train_cands_ids
# non_matched_cands_ids = cands_ids - intersection_set
cands_test_ids = set(np.random.choice(list(potential_cands_test_ids),
int(ratio_val * len(potential_cands_test_ids)), replace=False))
# select non-matched_ratio of the intersection_set and remove them from index_ids
index_ids_to_remove = set(np.random.choice(list(cands_test_ids),
int(non_matched_rat * len(cands_test_ids)), replace=False))
# non_matched_cands_ids = set(np.random.choice(list(non_matched_cands_ids),
# int(non_matched_rat * len(non_matched_cands_ids)), replace=False))
# index_test_ids = cands_test_ids.copy()
# cands_test_ids.update(non_matched_cands_ids)
# remove from index_ids the ids that are in index_ids_to_remove
index_test_ids = index_ids - index_ids_to_remove
index_test_ids = set(np.random.choice(list(index_test_ids),
int(ratio_val * len(index_test_ids)), replace=False))
# index_test_ids.update(set(np.random.choice(list(index_ids),
# int(ratio_val * len(index_ids)),
# replace=False)))
test_blocking_dict[test_size] = {'cands': cands_test_ids, 'index': index_test_ids}
return test_blocking_dict
def _read_objects_bo_em(self, dataset_config):
objects_path_dict = read_object_path_dict(dataset_config)
object_dict = defaultdict(dict)
for objects_type, objects_path in objects_path_dict.items():
for filename in os.listdir(objects_path):
file_ind = int(filename.split('.')[0])
json_data = read_json(objects_path, file_ind)
object_dict = self._insert_polygon_mesh(object_dict, objects_type, json_data, file_ind)
object_dict[objects_type] = dict(sorted(object_dict[objects_type].items()))
return object_dict
@staticmethod
def _remove_train_objects_from_object_dict(object_dict, train_ids):
for objects_type in object_dict.keys():
object_dict[objects_type] = {
object_id: object_data
for object_id, object_data in object_dict[objects_type].items()
if object_id not in train_ids
}
return object_dict
def _read_objects_gpkg(self, dataset_config):
objects_path_dict = read_object_path_dict(dataset_config)
object_dict = defaultdict(dict)
for objects_type, objects_path in objects_path_dict.items():
for filename in os.listdir(objects_path):
file_ind = int(filename.split('.')[0])
json_data = read_json(objects_path, file_ind)
json_data = json.loads(json_data)
object_dict = self._insert_polygon_mesh(object_dict, objects_type, json_data, file_ind)
object_dict[objects_type] = dict(sorted(object_dict[objects_type].items()))
return object_dict
def _read_objects_delivery3(self, dataset_config):
objects_path_dict = read_object_path_dict(dataset_config)
object_dict = defaultdict(dict)
mapping_dict = defaultdict(dict)
inv_mapping_dict = defaultdict(dict)
for objects_type, objects_path in objects_path_dict.items():
for file_ind, file_name in enumerate(os.listdir(objects_path)):
file_name = file_name.split('.')[0]
json_data = read_json(objects_path, file_name)
object_dict = self._insert_polygon_mesh(object_dict, objects_type, json_data, file_ind)
mapping_dict[objects_type][file_ind] = file_name
inv_mapping_dict[objects_type][file_name] = file_ind
object_dict[objects_type] = dict(sorted(object_dict[objects_type].items()))
object_dict['mapping_dict'] = mapping_dict
object_dict['inv_mapping_dict'] = inv_mapping_dict
return object_dict
@staticmethod
def _process_cityjson_file(args):
file_path, objects_type, standardize_obj_key_fn, get_polygon_mesh_fn = args
local_object_dict = {}
try:
with open(file_path, 'r') as f:
data = json.load(f)
vertices = data['vertices']
for obj_key in data['CityObjects']:
try:
new_obj_key = standardize_obj_key_fn(obj_key, objects_type)
polygon_mesh_data = get_polygon_mesh_fn(data, obj_key, vertices)
if polygon_mesh_data is not None:
local_object_dict[new_obj_key] = polygon_mesh_data
except:
continue
except:
pass
return objects_type, local_object_dict
def _read_objects_Hague(self, dataset_config):
"""Load from preprocessed cache (produced by preprocess_hague.py).
The cache contains real-world EPSG:7415 coordinates and pre-computed
grid_cell assignments, so grid splits are spatially correct.
Run `python preprocess_hague.py` once before calling this.
"""
raw_cache_path = (f"{config.FilePaths.object_dict_path}"
f"{self.dataset_name}_raw.joblib")
if not os.path.exists(raw_cache_path):
raise FileNotFoundError(
f"Preprocessed cache not found: {raw_cache_path}\n"
f"Run: python preprocess_hague.py"
)
print(f"Loading preprocessed cache: {raw_cache_path}")
return joblib.load(raw_cache_path)
def _read_objects_Lyon_CT(self, dataset_config):
"""Load Lyon cross-time dataset from preprocessed cache."""
raw_cache_path = (f"{config.FilePaths.object_dict_path}"
f"{self.dataset_name}_raw.joblib")
if not os.path.exists(raw_cache_path):
raise FileNotFoundError(
f"Lyon cross-time cache not found: {raw_cache_path}")
print(f"Loading Lyon_CT preprocessed cache: {raw_cache_path}")
return joblib.load(raw_cache_path)
def _read_objects_Lyon_CT_09_15(self, dataset_config):
"""Load Lyon 2009→2015 cross-time dataset (6-year gap)."""
raw_cache_path = (f"{config.FilePaths.object_dict_path}"
f"{self.dataset_name}_raw.joblib")
if not os.path.exists(raw_cache_path):
raise FileNotFoundError(
f"Lyon CT 09→15 cache not found: {raw_cache_path}")
print(f"Loading Lyon_CT_09_15 cache: {raw_cache_path}")
return joblib.load(raw_cache_path)
def _read_objects_Lyon_CT_Full(self, dataset_config):
"""Load Lyon 2009→2012 cross-time dataset (all confidence tiers)."""
raw_cache_path = (f"{config.FilePaths.object_dict_path}"
f"{self.dataset_name}_raw.joblib")
if not os.path.exists(raw_cache_path):
raise FileNotFoundError(f"Lyon CT Full cache not found: {raw_cache_path}")
print(f"Loading Lyon_CT_Full cache: {raw_cache_path}")
return joblib.load(raw_cache_path)
def _read_objects_Lyon_CT_12_15(self, dataset_config):
"""Load Lyon 2012→2015 cross-time dataset (3-year gap, CityGML 1.0→2.0)."""
raw_cache_path = (f"{config.FilePaths.object_dict_path}"
f"{self.dataset_name}_raw.joblib")
if not os.path.exists(raw_cache_path):
raise FileNotFoundError(
f"Lyon CT 12→15 cache not found: {raw_cache_path}")
print(f"Loading Lyon_CT_12_15 cache: {raw_cache_path}")
return joblib.load(raw_cache_path)
@staticmethod
def _generate_object_dict(results):
object_dict = defaultdict(dict)
for objects_type, obj_dict in results:
object_dict[objects_type].update(obj_dict)
intersection_keys = set(object_dict['cands'].keys()).intersection(object_dict['index'].keys())
object_dict['cands'] = {k: object_dict['cands'][k] for k in intersection_keys}
return object_dict, intersection_keys
@staticmethod
def _print_object_dict_info(object_dict, intersection_keys):
print(f"Number of overlapping objects: {len(intersection_keys)}")
print(f"Number of cands: {len(object_dict['cands'])}")
print(f"Number of index: {len(object_dict['index'])}")
return
@staticmethod
def _update_object_dict_mapping(object_dict, objects_path_dict):
for objects_type in objects_path_dict:
keys = list(object_dict[objects_type].keys())
object_dict['mapping_dict'][objects_type] = {i: k for i, k in enumerate(keys)}
object_dict['inv_mapping_dict'][objects_type] = {k: i for i, k in enumerate(keys)}
return object_dict
# def _read_objects_Hague(self, dataset_config):
# objects_path_dict = read_object_path_dict(dataset_config)
# object_dict = defaultdict(dict)
# for objects_type, objects_path in objects_path_dict.items():
# print(f"Reading {objects_type} objects")
# file_list = [f for f in os.listdir(objects_path) if f.endswith('.json')]
# for file_ind, file_name in enumerate(file_list):
# print(f"File number {file_ind + 1} out of {len(file_list)}")
# file_path = ''.join([objects_path, file_name])
# with open(file_path, 'r') as f:
# data = json.load(f)
# vertices = data['vertices']
# for obj_key in data['CityObjects'].keys():
# try:
# new_obj_key = self.standardize_obj_key(obj_key, objects_type)
# polygon_mesh_data = self._get_polygon_mesh(data, obj_key, vertices)
# if polygon_mesh_data is not None:
# object_dict[objects_type][new_obj_key] = polygon_mesh_data
# except:
# continue
# intersection_keys = set(object_dict['cands'].keys()).intersection(set(object_dict['index'].keys()))
# object_dict['cands'] = {obj_key: object_dict['cands'][obj_key] for obj_key in intersection_keys}
# print(f"Number of overlapping objects: {len(intersection_keys)}")
# print(f"Number of cands: {len(object_dict['cands'])}")
# print(f"Number of index: {len(object_dict['index'])}")
# for objects_type in objects_path_dict.keys():
# object_dict['mapping_dict'][objects_type] = {ind: obj_key for ind, obj_key in
# enumerate(object_dict[objects_type].keys())}
# object_dict['inv_mapping_dict'][objects_type] = {obj_key: ind for ind, obj_key in
# enumerate(object_dict[objects_type].keys())}
# return object_dict
@staticmethod
def standardize_obj_key(obj_key, object_type):
if object_type == 'cands':
return obj_key.split('bag_')[1]
elif object_type == 'index':
return obj_key.split('NL.IMBAG.Pand.')[1].split('-0')[0]
else:
raise ValueError('Invalid source')
def _insert_polygon_mesh(self, object_dict, obj_type, obj_data, obj_ind=None):
vertices = obj_data['vertices']
obj_key = list(obj_data['CityObjects'].keys())[0]
polygon_mesh = self._get_polygon_mesh(obj_data, obj_key, vertices)
if polygon_mesh is not None:
object_dict[obj_type][obj_ind] = polygon_mesh
return object_dict
def _get_polygon_mesh(self, obj_data, obj_key, vertices):
boundaries = obj_data['CityObjects'][obj_key]['geometry'][0]['boundaries'][0]
if len(boundaries) < self.min_surfaces_num:
return None
polygon_mesh = []
for surface in boundaries:
polygon_mesh.append([vertices[i] for sub_surface_list in surface for i in sub_surface_list])
vertices = self._get_vertices(polygon_mesh)
centroid = self._compute_object_centroid(vertices)
return {'polygon_mesh': polygon_mesh, 'vertices': vertices, 'centroid': centroid}
@staticmethod
def _compute_object_centroid(vertices):
unique_vertices = np.array(vertices)
return unique_vertices.mean(axis=0)
@staticmethod
def _get_vertices(polygon_mesh):
return np.unique(np.array([coord for surface in polygon_mesh for coord in surface]), axis=0)
def _save_dataset_partition_dict(self, dataset_partition_dict):
if not os.path.exists(config.FilePaths.dataset_partition_path):
os.makedirs(config.FilePaths.dataset_partition_path)
path = f"{config.FilePaths.dataset_partition_path}{self.dataset_name}_seed{self.seed}.pkl"
pkl.dump(dataset_partition_dict, open(path, 'wb'))
print(f"Saved the dataset partition dict to {path}")
return
def generate_partition_dicts(args):
partition_dict_obj = DataPartitionGenerator(args)
for seed in range(1, args.seeds_num + 1):
print(f"Creating dataset partition dict for seed {seed}")
start_time = time()
partition_dict_obj.create_dataset_partition_dict(seed)
end_time = time()
print(f"Elapsed time for seed {seed}: {end_time - start_time}")
print(3 * '--------------------------')
print("Done!")
def get_potnetial_neg_pairs(dataset_size_version, bkafi_dim, train_or_test, seed):
file_name = get_file_name()
file_name.replace('concatenation', 'division')
# if train_or_test == 'train':
# file_name = file_name.replace('Operator', 'Train_Operator')
blocking_results_dir = config.FilePaths.results_path + 'blocking_output/'
blocking_results_path = (f"{blocking_results_dir}{file_name}_"
f"{dataset_size_version}_neg_samples_num2_vector_normalization_True_sdr_factor_False_"
f"bkafi_criterion=feature_importance_seed={seed}.joblib")
print(f"\nLoading blocking results from {blocking_results_path}")
blocking_dict = joblib.load(blocking_results_path)
print(f"Loaded blocking results from {blocking_results_path}")
neg_pairs = blocking_dict['neg_pairs'][bkafi_dim]
return neg_pairs
def process_blocking_based_pairs(seed, neg_samples_num, cands_with_match_ids, potential_neg_pairs):
np.random.seed(seed)
pos_pairs = [(cand_id, cand_id) for cand_id in cands_with_match_ids]
neg_pairs = potential_neg_pairs[neg_samples_num + 1]
all_pairs = pos_pairs + neg_pairs
np.random.shuffle(all_pairs)
return all_pairs
def get_blocking_based_pairs(args, seed, train_or_test, dataset_partition_dict):
local_test_ids_dict = {}
neg_samples_list = args.train_neg_samples_list if train_or_test == 'train' else args.test_negative_samples_list
# Only inject pairs for the dataset_size we actually ran blocking for.
# If dataset_size_version is not set on args, fall back to all sizes.
sizes_to_process = (
[args.dataset_size_version]
if hasattr(args, 'dataset_size_version') and args.dataset_size_version
else ['small', 'medium', 'large']
)
for set_size in sizes_to_process:
local_test_ids_dict[set_size] = {}
if train_or_test == 'train':
negative_sampling_pair_set = dataset_partition_dict['train']['negative_sampling'][set_size][2]
else:
negative_sampling_pair_set = dataset_partition_dict['test']['matching']['negative_sampling'][set_size][2]
cands_with_match_ids = set([pair[0] for pair in negative_sampling_pair_set if pair[0] == pair[1]])
potential_neg_pairs = get_potnetial_neg_pairs(set_size, args.bkafi_dim, train_or_test, seed)
for neg_samples_num in neg_samples_list:
local_test_ids_dict[set_size][neg_samples_num] = process_blocking_based_pairs(seed, neg_samples_num,
cands_with_match_ids,
potential_neg_pairs)
return local_test_ids_dict
def add_blocking_based_mode_pairs(args):
for seed in range(1, args.seeds_num + 1):
# read the existing dataset partition dict
dataset_partition_dict = pkl.load(open(f"data/dataset_partitions/{args.dataset_name}_seed{seed}.pkl", 'rb'))
print(f"Loaded dataset partition dict for seed {seed} with blocking-based pairs")
dataset_partition_dict['train']['blocking-based'] = get_blocking_based_pairs(args, seed, 'train',
dataset_partition_dict)
dataset_partition_dict['test']['matching']['blocking-based'] = get_blocking_based_pairs(args, seed, 'test',
dataset_partition_dict)
# save the updated dataset partition dict
pkl.dump(dataset_partition_dict, open(f"data/dataset_partitions/{args.dataset_name}_seed{seed}.pkl", 'wb'))
print(f"Updated dataset partition dict for seed {seed} with blocking-based pairs")
print(3 * '--------------------------')
return
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--dataset_name', type=str, default=config.Constants.dataset_name)
parser.add_argument('--blocking_based_mode', type=bool, default=True)
parser.add_argument('--seeds_num', type=int, default=config.Constants.seeds_num)
parser.add_argument('--train_neg_samples_list', type=list, default=[2, 5])
parser.add_argument('--test_negative_samples_list', type=list, default=[2, 5])
parser.add_argument('--train_size_ratio_list', type=dict, default={"small": 0.1, "medium": 0.4, "large": 0.6})
parser.add_argument('--test_size_ratio_list', type=dict, default={"small": 0.1, "medium": 0.5, "large": 1.0})
parser.add_argument('--bkafi_dim', type=int, default=3)
args = parser.parse_args()
if args.blocking_based_mode: # load existing partitions and add blocking-based partitions for the matching mode
add_blocking_based_mode_pairs(args)
else:
generate_partition_dicts(args)