Update mask_generator_with_label.py

CarlHuangNuc · Jul 6, 2023 · 3aa8a70 · 3aa8a70
1 parent 20247b7
commit 3aa8a70
Showing 1 changed file with 23 additions and 3 deletions.
diff --git a/configs/common/models/mask_generator_with_label.py b/configs/common/models/mask_generator_with_label.py
@@ -1,3 +1,4 @@
+
 # ------------------------------------------------------------------------------
 # Copyright (c) 2022-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
@@ -25,10 +26,15 @@
 from mask2former.modeling.matcher import HungarianMatcher
 from mask2former.modeling.pixel_decoder.msdeformattn import MSDeformAttnPixelDecoder
 
+
 model = L(CategoryODISE)(
     sem_seg_head=L(MaskFormerHead)(
         ignore_value=255,
-        num_classes=133,
+        #ignore_value=2555,
+        #num_classes=133,
+        #num_classes=1203,
+        num_classes=150,
+        #num_classes=847,
         pixel_decoder=L(MSDeformAttnPixelDecoder)(
             conv_dim=256,
             mask_dim=256,
@@ -83,14 +89,28 @@
     ),
     category_head=L(CategoryEmbed)(
         clip_model_name="ViT-L-14-336",
-        labels=L(get_openseg_labels)(dataset="coco_panoptic", prompt_engineered=True),
+        #labels=L(get_openseg_labels)(dataset="coco_panoptic", prompt_engineered=True),
+        labels=L(get_openseg_labels)(dataset="ade20k_150", prompt_engineered=True),
+        #labels=L(get_openseg_labels)(dataset="ade20k_847", prompt_engineered=True),
+
+
+#        ade20k_847   
+        #labels=L(get_openseg_labels)(dataset="lvis_1203", prompt_engineered=True),
+
         projection_dim="${..sem_seg_head.transformer_predictor.post_mask_embed.projection_dim}",
     ),
     clip_head=L(PoolingCLIPHead)(),
     num_queries=100,
     object_mask_threshold=0.0,
     overlap_threshold=0.8,
-    metadata=L(MetadataCatalog.get)(name="coco_2017_train_panoptic_with_sem_seg"),
+
+    #metadata=L(MetadataCatalog.get)(name="coco_2017_train_panoptic_with_sem_seg"),
+
+    #### carl ...change ....
+    metadata=L(MetadataCatalog.get)(name="ade20k_panoptic_train"),
+    #metadata=L(MetadataCatalog.get)(name="ade20k_full_panoptic_val"),
+
+    #metadata=L(MetadataCatalog.get)(name="lvis_v1_train_with_sem_seg"),
     size_divisibility=64,
     sem_seg_postprocess_before_inference=True,
     # normalize to [0, 1]