CNN Classifier

datasets/paired_dataset.py

@@ -0,0 +1,100 @@
+import shelve
+from pathlib import Path
+from typing import Union
+
+import torch
+from torch.utils.data import Dataset, ConcatDataset
+
+from datasets.utils import DatasetMapping
+from lib.modules.model_parts import Generator
+from lib.objects.map import Map
+
+
+class TrajPairDataset(Dataset):
+    @property
+    def map_shape(self):
+        return self.map.as_array.shape
+
+    def __init__(self, data):
+        super(TrajPairDataset, self).__init__()
+        self.alternatives = data['alternatives']
+        self.trajectory = data['trajectory']
+        self.labels = data['labels']
+        self.mapname = data['map']['name'][4:] if data['map']['name'].startswith('map_') else data['map']['name']
+        self.map = data['map']['map']
+
+    def __len__(self):
+        return len(self.alternatives)
+
+    def __getitem__(self, item):
+        return self.trajectory.vertices, self.alternatives[item].vertices, self.labels[item], self.mapname
+
+
+class TrajPairData(object):
+    @property
+    def map_shapes(self):
+        return [dataset.map_shape for dataset in self._dataset.datasets]
+
+    @property
+    def map_shapes_max(self):
+        shapes = self.map_shapes
+        return map(max, zip(*shapes))
+
+    @property
+    def name(self):
+        return self.__class__.__name__
+
+    def __init__(self, data_root, map_root: Union[Path, str] = '', mapname='tate_sw', trajectories=1000, alternatives=10,
+                 train_val_test_split=(0.6, 0.2, 0.2), rebuild=False, equal_samples=True, **_):
+
+        self.rebuild = rebuild
+        self.equal_samples = equal_samples
+        self._alternatives = alternatives
+        self._trajectories = trajectories
+        self.mapname = mapname
+        self.train_split, self.val_split, self.test_split = train_val_test_split
+        self.data_root = Path(data_root)
+        self.maps_root = Path(map_root) if map_root else Path() / 'res' / 'maps'
+        self._dataset, self._train_map, self._val_map, self._test_map = self._load_dataset()
+
+    def _build_data_on_demand(self):
+        map_object = Map(self.mapname).from_image(self.maps_root / f'{self.mapname}.bmp')
+        assert self.maps_root.exists()
+        dataset_file = Path(self.data_root) / f'{self.mapname}.pik'
+        if dataset_file.exists() and self.rebuild:
+            dataset_file.unlink()
+        if not dataset_file.exists():
+            generator = Generator(self.data_root, map_object)
+            generator.generate_n_trajectories_m_alternatives(self._trajectories, self._alternatives,
+                                                             self.mapname, equal_samples=self.equal_samples)
+        return True
+
+    def _load_dataset(self):
+        assert self._build_data_on_demand()
+        with shelve.open(str(self.data_root / f'{self.mapname}.pik')) as d:
+            dataset = ConcatDataset([TrajPairDataset(d[key]) for key in d.keys() if key != 'map'])
+            indices = torch.randperm(len(dataset))
+
+            train_size = int(len(dataset) * self.train_split)
+            val_size = int(len(dataset) * self.val_split)
+            test_size = int(len(dataset) * self.test_split)
+
+            train_map = indices[:train_size]
+            val_map = indices[train_size:val_size]
+            test_map = indices[test_size:]
+        return dataset, train_map, val_map, test_map
+
+    @property
+    def train_dataset(self):
+        return DatasetMapping(self._dataset, self._train_map)
+
+    @property
+    def val_dataset(self):
+        return DatasetMapping(self._dataset, self._val_map)
+
+    @property
+    def test_dataset(self):
+        return DatasetMapping(self._dataset, self._test_map)
+
+    def get_datasets(self):
+        return self.train_dataset, self.val_dataset, self.test_dataset