use empty COCO to avoid loading empty result via coco api (#270)

maskrcnn_benchmark/data/datasets/evaluation/coco/coco_eval.py

@@ -277,9 +277,11 @@ def evaluate_predictions_on_coco(
     with open(json_result_file, "w") as f:
         json.dump(coco_results, f)
 
+    from pycocotools.coco import COCO
     from pycocotools.cocoeval import COCOeval
 
-    coco_dt = coco_gt.loadRes(str(json_result_file))
+    coco_dt = coco_gt.loadRes(str(json_result_file)) if coco_results else COCO()
+
     # coco_dt = coco_gt.loadRes(coco_results)
     coco_eval = COCOeval(coco_gt, coco_dt, iou_type)
     coco_eval.evaluate()