Update VITDet to conform to KerasCV scaling standards (#2086)

* Update VITDet to conform to KerasCV scaling standards * dtype fix
keras-team · Sep 22, 2023 · 7712a81 · 7712a81
1 parent e5cd24a
commit 7712a81
Show file tree

Hide file tree

Showing 5 changed files with 10 additions and 4 deletions.
diff --git a/keras_cv/models/backbones/vit_det/data/vitdet_base_out.npz b/keras_cv/models/backbones/vit_det/data/vitdet_base_out.npz
diff --git a/keras_cv/models/backbones/vit_det/vit_det_backbone.py b/keras_cv/models/backbones/vit_det/vit_det_backbone.py
@@ -16,6 +16,7 @@
 
 from keras_cv.api_export import keras_cv_export
 from keras_cv.backend import keras
+from keras_cv.backend import ops
 from keras_cv.layers.vit_det_layers import AddPositionalEmbedding
 from keras_cv.layers.vit_det_layers import ViTDetPatchingAndEmbedding
 from keras_cv.layers.vit_det_layers import WindowedTransformerEncoder
@@ -81,9 +82,9 @@ class ViTDetBackbone(Backbone):
     def __init__(
         self,
         *,
+        include_rescaling,
         input_shape=(1024, 1024, 3),
         input_tensor=None,
-        include_rescaling=False,
         patch_size=16,
         embed_dim=768,
         depth=12,
@@ -123,6 +124,11 @@ def __init__(
             # Use common rescaling strategy across keras_cv
             x = keras.layers.Rescaling(1.0 / 255.0)(x)
 
+        # VITDet scales inputs based on the standard ImageNet mean/stddev.
+        x = (x - ops.array([0.229, 0.224, 0.225], dtype=x.dtype)) / (
+            ops.array([0.485, 0.456, 0.406], dtype=x.dtype)
+        )
+
         x = ViTDetPatchingAndEmbedding(
             kernel_size=(patch_size, patch_size),
             strides=(patch_size, patch_size),

diff --git a/keras_cv/models/backbones/vit_det/vit_det_backbone_presets.py b/keras_cv/models/backbones/vit_det/vit_det_backbone_presets.py
@@ -30,7 +30,7 @@
         "config": {
             "input_shape": (1024, 1024, 3),
             "input_tensor": None,
-            "include_rescaling": False,
+            "include_rescaling": True,
             "patch_size": 16,
             "embed_dim": 768,
             "depth": 12,
@@ -61,7 +61,7 @@
         "config": {
             "input_shape": (1024, 1024, 3),
             "input_tensor": None,
-            "include_rescaling": False,
+            "include_rescaling": True,
             "patch_size": 16,
             "embed_dim": 1024,
             "depth": 24,
@@ -92,7 +92,7 @@
         "config": {
             "input_shape": (1024, 1024, 3),
             "input_tensor": None,
-            "include_rescaling": False,
+            "include_rescaling": True,
             "patch_size": 16,
             "embed_dim": 1280,
             "depth": 32,

diff --git a/keras_cv/models/segmentation/segment_anything/data/sam_base_out_iou_pred.npy b/keras_cv/models/segmentation/segment_anything/data/sam_base_out_iou_pred.npy
diff --git a/keras_cv/models/segmentation/segment_anything/data/sam_base_out_masks.npy b/keras_cv/models/segmentation/segment_anything/data/sam_base_out_masks.npy