set COARSE_PRED_EACH_LAYER to False for semantic seg

Summary:
Make training much more stable (no NaNs in 7 runs).
Results slightly better (average is around 78.8).

Reviewed By: alexander-kirillov

Differential Revision: D22427114

fbshipit-source-id: 3a3ea9895e5ac955a62454b73cff71f496e0fbfb

Author: ppwwyyxx

projects/PointRend/README.md

@@ -115,9 +115,9 @@ Cityscapes model is trained with ImageNet pretraining.
  <tr><td align="left"><a href="configs/SemanticSegmentation/pointrend_semantic_R_101_FPN_1x_cityscapes.yaml">SemanticFPN + PointRend</a></td>
 <td align="center">R101-FPN</td>
 <td align="center">1024&times;2048</td>
-<td align="center">78.6</td>
-<td align="center">186480235</td>
-<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/PointRend/SemanticSegmentation/pointrend_semantic_R_101_FPN_1x_cityscapes/186480235/model_final_5f3665.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/PointRend/SemanticSegmentation/pointrend_semantic_R_101_FPN_1x_cityscapes/186480235/metrics.json">metrics</a></td>
+<td align="center">78.9</td>
+<td align="center">202576688</td>
+<td align="center"><a href="https://dl.fbaipublicfiles.com/detectron2/PointRend/SemanticSegmentation/pointrend_semantic_R_101_FPN_1x_cityscapes/202576688/model_final_cf6ac1.pkl">model</a>&nbsp;|&nbsp;<a href="https://dl.fbaipublicfiles.com/detectron2/PointRend/SemanticSegmentation/pointrend_semantic_R_101_FPN_1x_cityscapes/202576688/metrics.json">metrics</a></td>
 </tr>
 </tbody></table>
 

projects/PointRend/configs/SemanticSegmentation/Base-PointRend-Semantic-FPN.yaml

@@ -14,6 +14,7 @@ MODEL:
     SUBDIVISION_STEPS: 2
     SUBDIVISION_NUM_POINTS: 8192
     COARSE_SEM_SEG_HEAD_NAME: "SemSegFPNHead"
+    COARSE_PRED_EACH_LAYER: False
 DATASETS:
   TRAIN: ("coco_2017_train_panoptic_stuffonly",)
   TEST: ("coco_2017_val_panoptic_stuffonly",)

projects/PointRend/configs/SemanticSegmentation/pointrend_semantic_R_50_FPN_1x_coco.yaml

@@ -37,7 +37,7 @@ def calculate_uncertainty(sem_seg_logits):
 class PointRendSemSegHead(nn.Module):
     """
     A semantic segmentation head that combines a head set in `POINT_HEAD.COARSE_SEM_SEG_HEAD_NAME`
-        and a point head set in `MODEL.POINT_HEAD.NAME`.
+    and a point head set in `MODEL.POINT_HEAD.NAME`.
     """
 
     def __init__(self, cfg, input_shape: Dict[str, ShapeSpec]):