Make DataFrameTextBenchmark script pos_weight optional (#379)

Make `DataFrameTextBenchmark` script `pos_weight` optional as some
datasets may not need this (such as `kick`)

benchmark/data_frame_text_benchmark.py

@@ -61,7 +61,7 @@
     "--idx",
     type=int,
     default=0,
-    help="The index of the dataset within DataFrameBenchmark",
+    help="The index of the dataset within DataFrameTextBenchmark",
 )
 parser.add_argument(
     "--model_type",
@@ -87,6 +87,12 @@
     ],
 )
 parser.add_argument("--finetune", action="store_true")
+parser.add_argument(
+    "--pos_weight",
+    action="store_true",
+    help=("Whether to set `pos_weight` in `BCEWithLogitsLoss` "
+          "for the binary classification task."),
+)
 parser.add_argument('--result_path', type=str, default='')
 parser.add_argument("--api_key", type=str, default=None)
 args = parser.parse_args()
@@ -464,8 +470,12 @@ def main_torch(
 
     if dataset.task_type == TaskType.BINARY_CLASSIFICATION:
         out_channels = 1
-        label_imbalance = sum(train_tensor_frame.y) / len(train_tensor_frame.y)
-        loss_fun = BCEWithLogitsLoss(pos_weight=1 / label_imbalance)
+        if args.pos_weight:
+            label_imbalance = sum(train_tensor_frame.y) / len(
+                train_tensor_frame.y)
+            loss_fun = BCEWithLogitsLoss(pos_weight=1 / label_imbalance)
+        else:
+            loss_fun = BCEWithLogitsLoss()
         metric_computer = AUROC(task='binary').to(device)
         higher_is_better = True
     elif dataset.task_type == TaskType.MULTICLASS_CLASSIFICATION: