Merge pull request #202 from st-tech/superset_matching

add n_comb options

Author: wildsnowman

benchmarks/set_matching_pytorch/test.py

@@ -31,7 +31,7 @@ def get_test_loader(
 
     iqon_outfits = IQONOutfits(root=root, split=split)
 
-    test_examples = iqon_outfits.get_fitb_data(label_dir_name)
+    test_examples = iqon_outfits.get_fitb_data(label_dir_name, n_comb=args.n_comb)
     feature_dir = iqon_outfits.feature_dir
     dataset = FINBsDataset(
         test_examples,
@@ -113,6 +113,7 @@ def main(args):
     parser.add_argument("--valid_year", type=int)
     parser.add_argument("--split", type=int, choices=[0, 1, 2])
     parser.add_argument("--model_dir", "-d", type=str)
+    parser.add_argument("--n_comb", type=int, default=1)
     args = parser.parse_args()
 
     main(args)

benchmarks/set_matching_pytorch/train_sm.py

@@ -24,6 +24,7 @@ def get_train_val_loader(
     valid_year: Union[str, int],
     split: int,
     batch_size: int,
+    n_comb: int,
     root: str = C.ROOT,
     num_workers: Optional[int] = None,
 ) -> Tuple[Any, Any]:
@@ -33,8 +34,12 @@ def get_train_val_loader(
 
     train, valid = iqon_outfits.get_trainval_data(label_dir_name)
     feature_dir = iqon_outfits.feature_dir
-    train_dataset = MultisetSplitDataset(train, feature_dir, n_sets=1, n_drops=None)
-    valid_dataset = MultisetSplitDataset(valid, feature_dir, n_sets=1, n_drops=None)
+    train_dataset = MultisetSplitDataset(
+        train, feature_dir, n_comb=n_comb, n_drops=None
+    )
+    valid_dataset = MultisetSplitDataset(
+        valid, feature_dir, n_comb=n_comb, n_drops=None
+    )
     return (
         get_loader(train_dataset, batch_size, num_workers=num_workers, is_train=True),
         get_loader(valid_dataset, batch_size, num_workers=num_workers, is_train=False),
@@ -70,7 +75,11 @@ def main(args):
 
     # dataset
     train_loader, valid_loader = get_train_val_loader(
-        args.train_year, args.valid_year, args.split, args.batchsize
+        train_year=args.train_year,
+        valid_year=args.valid_year,
+        split=args.split,
+        batch_size=args.batchsize,
+        n_comb=args.n_comb,
     )
 
     # logger
@@ -222,6 +231,7 @@ def eval_process(engine, batch):
     parser.add_argument("--valid_year", type=int)
     parser.add_argument("--split", type=int, choices=[0, 1, 2])
     parser.add_argument("--weight_path", "-w", type=str, default=None)
+    parser.add_argument("--n_comb", type=int, default=1)
 
     args = parser.parse_args()
 

shift15m/datasets/outfitfeature.py

@@ -36,18 +36,18 @@ def __init__(
         self,
         sets: List,
         root: pathlib.Path,
-        n_sets: int,
+        n_comb: int,
         n_drops: Optional[int] = None,
         max_elementnum_per_set: Optional[int] = 8,
     ):
         self.sets = sets
         self.feat_dir = root
-        self.n_sets = n_sets
+        self.n_comb = n_comb
         self.n_drops = n_drops
         if n_drops is None:
             n_drops = max_elementnum_per_set // 2
-        setX_size = (max_elementnum_per_set - n_drops) * n_sets
-        setY_size = n_drops * n_sets
+        setX_size = (max_elementnum_per_set - n_drops) * n_comb
+        setY_size = n_drops * n_comb
         self.transform_x = FeatureListTransform(
             max_set_size=setX_size, apply_shuffle=True, apply_padding=True
         )
@@ -59,9 +59,9 @@ def __len__(self):
         return len(self.sets)
 
     def __getitem__(self, i):
-        if self.n_sets > 1:  # you can conduct "superset matching" by using n_sets > 1
+        if self.n_comb > 1:  # you can conduct "superset matching" by using n_comb > 1
             indices = np.delete(np.arange(len(self.sets)), i)
-            indices = np.random.choice(indices, self.n_sets - 1, replace=False)
+            indices = np.random.choice(indices, self.n_comb - 1, replace=False)
             indices = [i] + list(indices)
         else:
             indices = [i]
@@ -306,7 +306,7 @@ def _make_test_examples(
         n_cands: int = 8,
         seed: int = 0,
     ):
-        print("Make test dataset.")
+        print("Making test dataset.")
         np.random.seed(seed)
 
         test_sets = json.load(open(path / "test.json"))