Removes label smoothing from CutMix and MixUp (keras-team#70)

* Removes label smoothing from CutMix and MixUp

* Run format

Author: (Ian Stenbit)

keras_cv/layers/preprocessing/cut_mix.py

@@ -28,10 +28,6 @@ class CutMix(layers.Layer):
             distribution.  This controls the shape of the distribution from which the
             smoothing values are sampled.  Defaults 1.0, which is a recommended value
             when training an imagenet1k classification model.
-        label_smoothing: Float in [0, 1]. When > 0, label values are smoothed,
-            meaning the confidence on label values are relaxed. e.g.
-            label_smoothing=0.2 means that we will use a value of 0.1 for label 0 and
-            0.9 for label 1.  Defaults 0.0.
     References:
        [CutMix paper]( https://arxiv.org/abs/1905.04899).
 
@@ -43,11 +39,10 @@ class CutMix(layers.Layer):
     ```
     """
 
-    def __init__(self, rate, label_smoothing=0.0, alpha=1.0, seed=None, **kwargs):
+    def __init__(self, rate, alpha=1.0, seed=None, **kwargs):
         super().__init__(**kwargs)
         self.alpha = alpha
         self.rate = rate
-        self.label_smoothing = label_smoothing
         self.seed = seed
 
     @staticmethod
@@ -85,7 +80,7 @@ def call(self, images, labels, training=True):
         augment_cond = tf.logical_and(rate_cond, training)
         # pylint: disable=g-long-lambda
         cutmix_augment = lambda: self._update_labels(*self._cutmix(images, labels))
-        no_augment = lambda: (images, self._smooth_labels(labels))
+        no_augment = lambda: (images, labels)
         return tf.cond(augment_cond, cutmix_augment, no_augment)
 
     def _cutmix(self, images, labels):
@@ -132,15 +127,8 @@ def _cutmix(self, images, labels):
         return images, labels, lambda_sample, permutation_order
 
     def _update_labels(self, images, labels, lambda_sample, permutation_order):
-        labels_smoothed = self._smooth_labels(labels)
         cutout_labels = tf.gather(labels, permutation_order)
 
         lambda_sample = tf.reshape(lambda_sample, [-1, 1])
-        labels = lambda_sample * labels_smoothed + (1.0 - lambda_sample) * cutout_labels
+        labels = lambda_sample * labels + (1.0 - lambda_sample) * cutout_labels
         return images, labels
-
-    def _smooth_labels(self, labels):
-        label_smoothing = self.label_smoothing or 0.0
-        off_value = label_smoothing / tf.cast(tf.shape(labels)[1], tf.float32)
-        on_value = 1.0 - label_smoothing + off_value
-        return labels * on_value + (1 - labels) * off_value

keras_cv/layers/preprocessing/cut_mix_test.py

@@ -30,21 +30,7 @@ def test_return_shapes(self):
         xs, ys = layer(xs, ys)
 
         self.assertEqual(xs.shape, [2, 512, 512, 3])
-        # one hot smoothed labels
         self.assertEqual(ys.shape, [2, 10])
-        self.assertEqual(len(ys != 0.0), 2)
-
-    def test_label_smoothing(self):
-        xs = tf.ones((2, 512, 512, 3))
-        # randomly sample labels
-        ys = tf.random.categorical(tf.math.log([[0.5, 0.5]]), 2)
-        ys = tf.squeeze(ys)
-        ys = tf.one_hot(ys, NUM_CLASSES)
-
-        layer = CutMix(1.0, label_smoothing=0.2, seed=1)
-        xs, ys = layer(xs, ys)
-        self.assertNotAllClose(ys, 0.0)
-        self.assertAllClose(tf.math.reduce_sum(ys, axis=-1), (1.0, 1.0))
 
     def test_cut_mix_call_results(self):
         xs = tf.cast(
@@ -56,7 +42,7 @@ def test_cut_mix_call_results(self):
         )
         ys = tf.one_hot(tf.constant([0, 1]), 2)
 
-        layer = CutMix(1.0, label_smoothing=0.0, seed=1)
+        layer = CutMix(1.0, seed=1)
         xs, ys = layer(xs, ys)
 
         # At least some pixels should be replaced in the CutMix operation
@@ -78,7 +64,7 @@ def test_cut_mix_call_results_one_channel(self):
         )
         ys = tf.one_hot(tf.constant([0, 1]), 2)
 
-        layer = CutMix(1.0, label_smoothing=0.0, seed=1)
+        layer = CutMix(1.0, seed=1)
         xs, ys = layer(xs, ys)
 
         # At least some pixels should be replaced in the CutMix operation
@@ -97,7 +83,7 @@ def test_in_tf_function(self):
         )
         ys = tf.one_hot(tf.constant([0, 1]), 2)
 
-        layer = CutMix(1.0, label_smoothing=0.0, seed=1)
+        layer = CutMix(1.0, seed=1)
 
         @tf.function
         def augment(x, y):

keras_cv/layers/preprocessing/mix_up.py

@@ -26,10 +26,6 @@ class MixUp(layers.Layer):
             distribution.  This controls the shape of the distribution from which the
             smoothing values are sampled.  Defaults 0.2, which is a recommended value
             when training an imagenet1k classification model.
-        label_smoothing: Float in [0, 1]. When > 0, label values are smoothed,
-            meaning the confidence on label values are relaxed. e.g.
-            label_smoothing=0.2 means that we will use a value of 0.1 for label 0 and
-            0.9 for label 1.  Defaults 0.0.
     References:
         [MixUp paper](https://arxiv.org/abs/1710.09412).
 
@@ -41,11 +37,10 @@ class MixUp(layers.Layer):
     ```
     """
 
-    def __init__(self, rate, label_smoothing=0.0, alpha=0.2, seed=None, **kwargs):
+    def __init__(self, rate, alpha=0.2, seed=None, **kwargs):
         super().__init__(**kwargs)
         self.alpha = alpha
         self.rate = rate
-        self.label_smoothing = label_smoothing
         self.seed = seed
 
     @staticmethod
@@ -83,7 +78,7 @@ def call(self, images, labels, training=True):
         augment_cond = tf.logical_and(rate_cond, training)
         # pylint: disable=g-long-lambda
         mixup_augment = lambda: self._update_labels(*self._mixup(images, labels))
-        no_augment = lambda: (images, self._smooth_labels(labels))
+        no_augment = lambda: (images, labels)
         return tf.cond(augment_cond, mixup_augment, no_augment)
 
     def _mixup(self, images, labels):
@@ -99,18 +94,9 @@ def _mixup(self, images, labels):
         return images, labels, tf.squeeze(lambda_sample), permutation_order
 
     def _update_labels(self, images, labels, lambda_sample, permutation_order):
-        labels_smoothed = self._smooth_labels(labels)
         labels_for_mixup = tf.gather(labels, permutation_order)
 
         lambda_sample = tf.reshape(lambda_sample, [-1, 1])
-        labels = (
-            lambda_sample * labels_smoothed + (1.0 - lambda_sample) * labels_for_mixup
-        )
+        labels = lambda_sample * labels + (1.0 - lambda_sample) * labels_for_mixup
 
         return images, labels
-
-    def _smooth_labels(self, labels):
-        label_smoothing = self.label_smoothing or 0.0
-        off_value = label_smoothing / tf.cast(tf.shape(labels)[1], tf.float32)
-        on_value = 1.0 - label_smoothing + off_value
-        return on_value * labels + (1 - labels) * off_value

keras_cv/layers/preprocessing/mix_up_test.py

@@ -30,21 +30,7 @@ def test_return_shapes(self):
         xs, ys = layer(xs, ys)
 
         self.assertEqual(xs.shape, [2, 512, 512, 3])
-        # one hot smoothed labels
         self.assertEqual(ys.shape, [2, 10])
-        self.assertEqual(len(ys != 0.0), 2)
-
-    def test_label_smoothing(self):
-        xs = tf.ones((2, 512, 512, 3))
-        # randomly sample labels
-        ys = tf.random.categorical(tf.math.log([[0.5, 0.5]]), 2)
-        ys = tf.squeeze(ys)
-        ys = tf.one_hot(ys, NUM_CLASSES)
-
-        layer = MixUp(1.0, label_smoothing=0.2)
-        xs, ys = layer(xs, ys)
-        self.assertNotAllClose(ys, 0.0)
-        self.assertAllClose(tf.math.reduce_sum(ys, axis=-1), (1.0, 1.0))
 
     def test_mix_up_call_results(self):
         xs = tf.cast(
@@ -56,7 +42,7 @@ def test_mix_up_call_results(self):
         )
         ys = tf.one_hot(tf.constant([0, 1]), 2)
 
-        layer = MixUp(1.0, label_smoothing=0.0)
+        layer = MixUp(1.0)
         xs, ys = layer(xs, ys)
 
         # None of the individual values should still be close to 1 or 0
@@ -77,7 +63,7 @@ def test_in_tf_function(self):
         )
         ys = tf.one_hot(tf.constant([0, 1]), 2)
 
-        layer = MixUp(1.0, label_smoothing=0.0)
+        layer = MixUp(1.0)
 
         @tf.function
         def augment(x, y):