Use consistent dtypes for augmentation layers.

markmcd · copybara-github · commit 8da9fb2cb478 · 2022-02-22T10:19:27.000-08:00
We were calling the `data_augmentation()` layer interactively with `uint8` images, and then during training using `float32` images. This change ensures we use consistent dtypes in both instances.

PiperOrigin-RevId: 430239809
diff --git a/site/en/tutorials/images/data_augmentation.ipynb b/site/en/tutorials/images/data_augmentation.ipynb
@@ -318,7 +318,7 @@
       "outputs": [],
       "source": [
         "# Add the image to a batch.\n",
-        "image = tf.expand_dims(image, 0)"
+        "image = tf.cast(tf.expand_dims(image, 0), tf.float32)"
       ]
     },
     {

Original file line number	Diff line number	Diff line change
`@@ -318,7 +318,7 @@`
`318`	`318`	`"outputs": [],`
`319`	`319`	`"source": [`
`320`	`320`	`"# Add the image to a batch.\n",`
`321`		`- "image = tf.expand_dims(image, 0)"`
	`321`	`+ "image = tf.cast(tf.expand_dims(image, 0), tf.float32)"`
`322`	`322`	`]`
`323`	`323`	`},`
`324`	`324`	`{`