Revert "Fix dtype bug in image converter (#2147)" (#2180)

Author: mattdangerw

keras_hub/src/layers/preprocessing/image_converter.py

@@ -16,6 +16,7 @@
 from keras_hub.src.utils.preset_utils import get_preset_saver
 from keras_hub.src.utils.python_utils import classproperty
 from keras_hub.src.utils.tensor_utils import check_bounding_box_support
+from keras_hub.src.utils.tensor_utils import in_tf_function
 from keras_hub.src.utils.tensor_utils import preprocessing_function
 
 
@@ -270,36 +271,45 @@ def call(self, inputs):
         else:
             x = inputs
         if self.scale is not None:
-            x = x * self._expand_non_channel_dims(self.scale, x)
+            # If we are scaling always cast to the compute dtype. We can't
+            # leave things as an int type if we are scaling to [0, 1].
+            scale = self._expand_non_channel_dims(self.scale, x)
+            x, scale = self._convert_types(x, scale, self.compute_dtype)
+            x = x * scale
         if self.offset is not None:
-            x = x + self._expand_non_channel_dims(self.offset, x)
+            offset = self._expand_non_channel_dims(self.offset, x)
+            x, offset = self._convert_types(x, offset, x.dtype)
+            x = x + offset
         if isinstance(inputs, dict):
             inputs["images"] = x
         else:
             inputs = x
         return inputs
 
     def _expand_non_channel_dims(self, value, inputs):
-        input_dtype = keras.backend.standardize_dtype(inputs.dtype)
-
+        """Expand non channel dims so value is broadcastable with inputs."""
         unbatched = len(ops.shape(inputs)) == 3
         channels_first = self.data_format == "channels_first"
         if unbatched:
             broadcast_dims = (1, 2) if channels_first else (0, 1)
         else:
             broadcast_dims = (0, 2, 3) if channels_first else (0, 1, 2)
-        # If inputs are not a tensor type, return a numpy array.
-        # This might happen when running under tf.data.
-        if ops.is_tensor(inputs):
-            # preprocessing decorator moves tensors to cpu in torch backend and
-            # processed on CPU, and then converted back to the appropriate
-            # device (potentially GPU) after preprocessing.
-            if keras.backend.backend() == "torch" and self.image_size is None:
-                return ops.expand_dims(value, broadcast_dims).cpu()
-            expanded = ops.expand_dims(value, broadcast_dims)
-            return ops.cast(expanded, input_dtype)
-        else:
-            return np.expand_dims(value, broadcast_dims).astype(input_dtype)
+        # An numpy value will work backend native ops or with tf.data.
+        return np.expand_dims(value, broadcast_dims)
+
+    def _convert_types(self, x, y, dtype):
+        """Make sure x and y have the same dtype and are on ths same device."""
+        if in_tf_function():
+            # This could happen on any backend if we are running in tf.data.
+            import tensorflow as tf
+
+            return tf.cast(x, dtype), tf.cast(y, dtype)
+        x = ops.cast(x, dtype)
+        y = ops.cast(y, dtype)
+        if keras.backend.backend() == "torch":
+            # Place on the same device as x (the image).
+            y = y.to(x.device)
+        return x, y
 
     def get_config(self):
         config = super().get_config()

keras_hub/src/layers/preprocessing/image_converter_test.py

@@ -4,6 +4,7 @@
 import keras
 import numpy as np
 import pytest
+import tensorflow as tf
 from absl.testing import parameterized
 from keras import ops
 
@@ -22,6 +23,12 @@ def test_resize_simple(self):
         outputs = converter(inputs)
         self.assertAllClose(outputs, ops.ones((4, 4, 3)))
 
+    def test_resize_dataset(self):
+        converter = ImageConverter(image_size=(4, 4), scale=1 / 255.0)
+        ds = tf.data.Dataset.from_tensor_slices(tf.zeros((8, 10, 10, 3)))
+        batch = ds.batch(2).map(converter).take(1).get_single_element()
+        self.assertAllClose(batch, tf.zeros((2, 4, 4, 3)))
+
     def test_unbatched(self):
         converter = ImageConverter(
             image_size=(4, 4),
@@ -35,20 +42,21 @@ def test_unbatched(self):
         self.assertAllClose(outputs[:, :, 1], np.ones((4, 4)) * 0.301569)
         self.assertAllClose(outputs[:, :, 2], np.ones((4, 4)) * 0.852353)
 
-    def test_bfloat16_input(self):
+    def test_dtypes(self):
+        converter = ImageConverter(image_size=(4, 4), scale=1.0 / 255.0)
+        int_image = ops.ones((10, 10, 3), dtype="uint8") * 255
+        float_image = ops.ones((10, 10, 3), dtype="float64") * 255
+        self.assertDTypeEqual(converter(int_image), "float32")
+        self.assertDTypeEqual(converter(float_image), "float32")
+        self.assertAllClose(converter(int_image), np.ones((4, 4, 3)))
+        self.assertAllClose(converter(float_image), np.ones((4, 4, 3)))
         converter = ImageConverter(
-            image_size=(4, 4),
-            scale=(1.0 / 255.0, 0.8 / 255.0, 1.2 / 255.0),
-            offset=(0.2, -0.1, 0.25),
-            dtype="bfloat16",
+            image_size=(4, 4), scale=1.0 / 255.0, dtype="bfloat16"
         )
-        inputs = ops.ones((10, 10, 3)) * 128
-        inputs = ops.cast(inputs, "bfloat16")
-        outputs = converter(inputs)
-        self.assertEqual(ops.shape(outputs), (4, 4, 3))
-        self.assertAllClose(outputs[:, :, 0], np.ones((4, 4)) * 0.703125)
-        self.assertAllClose(outputs[:, :, 1], np.ones((4, 4)) * 0.302734)
-        self.assertAllClose(outputs[:, :, 2], np.ones((4, 4)) * 0.851562)
+        self.assertDTypeEqual(converter(int_image), "bfloat16")
+        self.assertDTypeEqual(converter(float_image), "bfloat16")
+        self.assertAllClose(converter(int_image), np.ones((4, 4, 3)))
+        self.assertAllClose(converter(float_image), np.ones((4, 4, 3)))
 
     @parameterized.parameters(
         (True, False),

keras_hub/src/models/vit/vit_image_converter.py

@@ -53,12 +53,17 @@ def __init__(
 
     @preprocessing_function
     def call(self, inputs):
+        # TODO: Remove this whole function. Why can just use scale and offset
+        # in the base class.
         x = super().call(inputs)
-        # By default normalize using imagenet mean and std
         if self.norm_mean:
-            x = x - self._expand_non_channel_dims(self.norm_mean, x)
+            norm_mean = self._expand_non_channel_dims(self.norm_mean, x)
+            x, norm_mean = self._convert_types(x, norm_mean, self.compute_dtype)
+            x = x - norm_mean
         if self.norm_std:
-            x = x / self._expand_non_channel_dims(self.norm_std, x)
+            norm_std = self._expand_non_channel_dims(self.norm_std, x)
+            x, norm_std = self._convert_types(x, norm_std, x.dtype)
+            x = x / norm_std
 
         return x
 

keras_hub/src/utils/tensor_utils.py

@@ -28,6 +28,12 @@ def no_convert_scope():
         NO_CONVERT_COUNTER.count = getattr(NO_CONVERT_COUNTER, "count", 0) - 1
 
 
+def in_tf_function():
+    if tf is None:
+        return False
+    return not tf.executing_eagerly()
+
+
 def in_no_convert_scope():
     return getattr(NO_CONVERT_COUNTER, "count", 0) > 0