Fix RandomCrop layer called on integer input

We had a tf.cond which called tf.image.resize on one path (returning float32) and
cropping the input dtype on the other, which lead to a error. Instead we cast the
inputs to the layer's compute dtype and do our computation with that type.

Fixes #15681

PiperOrigin-RevId: 417686461

Author: mattdangerw

keras/layers/preprocessing/image_preprocessing.py

@@ -253,7 +253,7 @@ def __init__(self, height, width, seed=None, **kwargs):
   def call(self, inputs, training=True):
     if training is None:
       training = backend.learning_phase()
-    inputs = utils.ensure_tensor(inputs)
+    inputs = utils.ensure_tensor(inputs, dtype=self.compute_dtype)
     input_shape = tf.shape(inputs)
     h_diff = input_shape[H_AXIS] - self.height
     w_diff = input_shape[W_AXIS] - self.width
@@ -266,10 +266,14 @@ def random_crop():
       return tf.image.crop_to_bounding_box(inputs, h_start, w_start,
                                            self.height, self.width)
 
-    outputs = tf.cond(
+    def resize():
+      outputs = smart_resize(inputs, [self.height, self.width])
+      # smart_resize will always output float32, so we need to re-cast.
+      return tf.cast(outputs, self.compute_dtype)
+
+    return tf.cond(
         tf.reduce_all((training, h_diff >= 0, w_diff >= 0)), random_crop,
-        lambda: smart_resize(inputs, [self.height, self.width]))
-    return tf.cast(outputs, self.compute_dtype)
+        resize)
 
   def compute_output_shape(self, input_shape):
     input_shape = tf.TensorShape(input_shape).as_list()

keras/layers/preprocessing/image_preprocessing_test.py

@@ -17,6 +17,7 @@
 import functools
 from absl.testing import parameterized
 
+import keras
 from keras import keras_parameterized
 from keras import testing_utils
 from keras.engine import sequential
@@ -398,6 +399,12 @@ def test_unbatched_image(self):
         actual_output = layer(inp, training=True)
         self.assertAllClose(inp[2:10, 2:10, :], actual_output)
 
+  @testing_utils.run_v2_only
+  def test_uint8_input(self):
+    inputs = keras.Input((128, 128, 3), batch_size=2, dtype=tf.uint8)
+    layer = image_preprocessing.RandomCrop(64, 64)
+    self.assertAllEqual(layer(inputs).dtype, 'float32')
+
   @testing_utils.run_v2_only
   def test_output_dtypes(self):
     inputs = np.array([[[1], [2]], [[3], [4]]], dtype='float64')