✨ Add ConvNeXt-XL on mapillary

affromero · Apr 15, 2023 · ff57044 · ff57044
1 parent 7611899
commit ff57044
Show file tree

Hide file tree

Showing 4 changed files with 42 additions and 1 deletion.
diff --git a/README.md b/README.md
@@ -107,6 +107,7 @@ This repo contains the code for our paper **OneFormer: One Transformer to Rule U
 | OneFormer (Mapillary Vistas-Pretrained) | ConvNeXt-L<sup>&dagger;</sup> | 70.1 | 48.7 | 84.6 | 85.2 | 220M | [config](configs/cityscapes/convnext/mapillary_pretrain_oneformer_convnext_large_bs16_90k.yaml) | [model](https://shi-labs.com/projects/oneformer/cityscapes/mapillary_pretrain_250_16_convnext_l_oneformer_cityscapes_90k.pth) &#124; [pretrained](https://shi-labs.com/projects/oneformer/mapillary/mapillary_pretrain_250_16_convnext_l_oneformer_mapillary_300k.pth) |
 | OneFormer | DiNAT-L<sup>&dagger;</sup> | 67.6 | 45.6 | 83.1 | 84.0 | 223M | [config](configs/cityscapes/dinat/oneformer_dinat_large_bs16_90k.yaml) | [model](https://shi-labs.com/projects/oneformer/cityscapes/250_16_dinat_l_oneformer_cityscapes_90k.pth) |
 | OneFormer | ConvNeXt-XL<sup>&dagger;</sup> | 68.4 | 46.7 | 83.6 | 84.6 | 372M | [config](configs/cityscapes/convnext/oneformer_convnext_xlarge_bs16_90k.yaml) | [model](https://shi-labs.com/projects/oneformer/cityscapes/250_16_convnext_xl_oneformer_cityscapes_90k.pth) |
+| OneFormer (Mapillary Vistas-Pretrained) | ConvNeXt-XL<sup>&dagger;</sup> | 69.7 | 48.9 | 84.5 | 85.8 | 372M | [config](configs/cityscapes/convnext/mapillary_pretrain_oneformer_convnext_xlarge_bs16_90k.yaml) | [model](https://shi-labs.com/projects/oneformer/cityscapes/mapillary_pretrain_250_16_convnext_xl_oneformer_cityscapes_90k.pth) &#124; [pretrained](https://shi-labs.com/projects/oneformer/mapillary/mapillary_pretrain_250_16_convnext_xl_oneformer_mapillary_300k.pth) |
 
 ### COCO
 

diff --git a/configs/cityscapes/convnext/mapillary_pretrain_oneformer_convnext_xlarge_bs16_90k.yaml b/configs/cityscapes/convnext/mapillary_pretrain_oneformer_convnext_xlarge_bs16_90k.yaml
@@ -0,0 +1,18 @@
+_BASE_: ../oneformer_R50_bs16_90k.yaml
+MODEL:
+  BACKBONE:
+    NAME: "D2ConvNeXt"
+  CONVNEXT:
+    IN_CHANNELS: 3
+    DEPTHS: [3, 3, 27, 3]
+    DIMS: [256, 512, 1024, 2048]
+    DROP_PATH_RATE: 0.4
+    LSIT: 1.0
+    OUT_INDICES: [0, 1, 2, 3]
+  WEIGHTS: "250_16_convnext_xl_oneformer_mapillary_300k.pth"
+  PIXEL_MEAN: [123.675, 116.280, 103.530]
+  PIXEL_STD: [58.395, 57.120, 57.375]
+  ONE_FORMER:
+    NUM_OBJECT_QUERIES: 250
+TEST:
+  DETECTIONS_PER_IMAGE: 250
diff --git a/configs/coco/Base-COCO-UnifiedSegmentation.yaml b/configs/coco/Base-COCO-UnifiedSegmentation.yaml
@@ -16,7 +16,7 @@ MODEL:
 DATASETS:
   TRAIN: ("coco_2017_train_panoptic_with_sem_seg",)
   TEST_PANOPTIC: ("coco_2017_val_panoptic_with_sem_seg",)  # to evaluate instance and semantic performance as well
-  TEST_INSTANCE: ("coco_2017_val",)
+  TEST_INSTANCE: ("coco_2017_val_panoptic2instance",)
   TEST_SEMANTIC: ("coco_2017_val_panoptic_with_sem_seg",)
 SOLVER:
   IMS_PER_BATCH: 16

diff --git a/...gs/mapillary_vistas/convnext/cityscapes_pretrain_oneformer_convnext_xlarge_bs16_300k.yaml b/...gs/mapillary_vistas/convnext/cityscapes_pretrain_oneformer_convnext_xlarge_bs16_300k.yaml
@@ -0,0 +1,22 @@
+_BASE_: ../oneformer_R50_bs16_300k.yaml
+MODEL:
+  BACKBONE:
+    NAME: "D2ConvNeXt"
+  CONVNEXT:
+    IN_CHANNELS: 3
+    DEPTHS: [3, 3, 27, 3]
+    DIMS: [256, 512, 1024, 2048]
+    DROP_PATH_RATE: 0.4
+    LSIT: 1.0
+    OUT_INDICES: [0, 1, 2, 3]
+  WEIGHTS: "convnext_xlarge_22k_1k_384_ema.pkl"
+  PIXEL_MEAN: [123.675, 116.280, 103.530]
+  PIXEL_STD: [58.395, 57.120, 57.375]
+  ONE_FORMER:
+    NUM_OBJECT_QUERIES: 250
+INPUT:
+  TASK_PROB: 
+    SEMANTIC: 0.33
+    INSTANCE: 0.66
+TEST:
+  DETECTIONS_PER_IMAGE: 250