Fix overflow of CohenKappa (#675)

* fix overflow of int32

Author: WindQAQ

tensorflow_addons/metrics/cohens_kappa.py

@@ -37,6 +37,7 @@ class CohenKappa(Metric):
     while calculating the Cohen's Kappa score.
 
     Usage:
+
     ```python
     actuals = np.array([4, 4, 3, 4, 2, 4, 1, 1], dtype=np.int32)
     preds = np.array([4, 4, 3, 4, 4, 2, 1, 1], dtype=np.int32)
@@ -51,103 +52,102 @@ class CohenKappa(Metric):
     m.update_state(actuals, preds, sample_weight=weights)
     print('Final result: ', m.result().numpy()) # Result: 0.37209308
     ```
+
     Usage with tf.keras API:
+
     ```python
-    model = keras.models.Model(inputs, outputs)
+    model = tf.keras.models.Model(inputs, outputs)
     model.add_metric(tfa.metrics.CohenKappa(num_classes=5)(outputs))
     model.compile('sgd', loss='mse')
     ```
-
-    Args:
-      num_classes : Number of unique classes in your dataset
-      weightage   : Weighting to be considered for calculating
-                    kappa statistics. A valid value is one of
-                    [None, 'linear', 'quadratic']. Defaults to None.
-
-    Returns:
-      kappa_score : float
-        The kappa statistic, which is a number between -1 and 1. The maximum
-        value means complete agreement; zero or lower means chance agreement.
-
-    Raises:
-      ValueError: If the value passed for `weightage` is invalid
-        i.e. not any one of [None, 'linear', 'quadratic']
     """
 
     def __init__(self,
                  num_classes,
                  name='cohen_kappa',
                  weightage=None,
-                 dtype=tf.float32):
+                 dtype=None):
+        """Creates a `CohenKappa` instance.
+
+        Args:
+          num_classes: Number of unique classes in your dataset.
+          name: (Optional) String name of the metric instance.
+          weightage: (Optional) Weighting to be considered for calculating
+            kappa statistics. A valid value is one of
+            [None, 'linear', 'quadratic']. Defaults to `None`.
+          dtype: (Optional) Data type of the metric result.
+            Defaults to `None`.
+
+        Raises:
+          ValueError: If the value passed for `weightage` is invalid
+            i.e. not any one of [None, 'linear', 'quadratic']
+        """
         super(CohenKappa, self).__init__(name=name, dtype=dtype)
 
         if weightage not in (None, 'linear', 'quadratic'):
             raise ValueError("Unknown kappa weighting type.")
-        else:
-            self.weightage = weightage
 
+        self.weightage = weightage
         self.num_classes = num_classes
         self.conf_mtx = self.add_weight(
             'conf_mtx',
             shape=(self.num_classes, self.num_classes),
             initializer=tf.keras.initializers.zeros,
-            dtype=tf.int32)
+            dtype=tf.int64)
 
     def update_state(self, y_true, y_pred, sample_weight=None):
         """Accumulates the confusion matrix condition statistics.
 
         Args:
-          y_true : array, shape = [n_samples]
-                   Labels assigned by the first annotator.
-          y_pred : array, shape = [n_samples]
-                   Labels assigned by the second annotator. The kappa statistic
-                   is symmetric, so swapping ``y_true`` and ``y_pred`` doesn't
-                   change the value.
-          sample_weight(optional) : for weighting labels in confusion matrix
-                   Default is None. The dtype for weights should be the same
-                   as the dtype for confusion matrix. For more details,
-                   please check tf.math.confusion_matrix.
-
+          y_true: Labels assigned by the first annotator with shape
+            `[num_samples,]`.
+          y_pred: Labels assigned by the second annotator with shape
+            `[num_samples,]`. The kappa statistic is symmetric,
+            so swapping `y_true` and `y_pred` doesn't change the value.
+          sample_weight (optional): for weighting labels in confusion matrix
+            Defaults to `None`. The dtype for weights should be the same
+            as the dtype for confusion matrix. For more details,
+            please check `tf.math.confusion_matrix`.
 
         Returns:
           Update op.
         """
-        y_true = tf.cast(y_true, dtype=tf.int32)
-        y_pred = tf.cast(y_pred, dtype=tf.int32)
+        y_true = tf.cast(y_true, dtype=tf.int64)
+        y_pred = tf.cast(y_pred, dtype=tf.int64)
 
         if y_true.shape != y_pred.shape:
             raise ValueError(
-                "Number of samples in y_true and y_pred are different")
+                "Number of samples in `y_true` and `y_pred` are different")
 
         # compute the new values of the confusion matrix
         new_conf_mtx = tf.math.confusion_matrix(
             labels=y_true,
             predictions=y_pred,
             num_classes=self.num_classes,
-            weights=sample_weight)
+            weights=sample_weight,
+            dtype=tf.int64)
 
         # update the values in the original confusion matrix
         return self.conf_mtx.assign_add(new_conf_mtx)
 
     def result(self):
         nb_ratings = tf.shape(self.conf_mtx)[0]
-        weight_mtx = tf.ones([nb_ratings, nb_ratings], dtype=tf.int32)
+        weight_mtx = tf.ones([nb_ratings, nb_ratings], dtype=tf.int64)
 
         # 2. Create a weight matrix
         if self.weightage is None:
-            diagonal = tf.zeros([nb_ratings], dtype=tf.int32)
+            diagonal = tf.zeros([nb_ratings], dtype=tf.int64)
             weight_mtx = tf.linalg.set_diag(weight_mtx, diagonal=diagonal)
-            weight_mtx = tf.cast(weight_mtx, dtype=tf.float32)
-
         else:
-            weight_mtx += tf.range(nb_ratings, dtype=tf.int32)
-            weight_mtx = tf.cast(weight_mtx, dtype=tf.float32)
+            weight_mtx += tf.cast(tf.range(nb_ratings), dtype=tf.int64)
+            weight_mtx = tf.cast(weight_mtx, dtype=self.dtype)
 
             if self.weightage == 'linear':
                 weight_mtx = tf.abs(weight_mtx - tf.transpose(weight_mtx))
             else:
                 weight_mtx = tf.pow((weight_mtx - tf.transpose(weight_mtx)), 2)
-            weight_mtx = tf.cast(weight_mtx, dtype=tf.float32)
+
+        weight_mtx = tf.cast(weight_mtx, dtype=self.dtype)
 
         # 3. Get counts
         actual_ratings_hist = tf.reduce_sum(self.conf_mtx, axis=1)
@@ -161,8 +161,8 @@ def result(self):
         conf_mtx = self.conf_mtx / tf.reduce_sum(self.conf_mtx)
         out_prod = out_prod / tf.reduce_sum(out_prod)
 
-        conf_mtx = tf.cast(conf_mtx, dtype=tf.float32)
-        out_prod = tf.cast(out_prod, dtype=tf.float32)
+        conf_mtx = tf.cast(conf_mtx, dtype=self.dtype)
+        out_prod = tf.cast(out_prod, dtype=self.dtype)
 
         # 6. Calculate Kappa score
         numerator = tf.reduce_sum(conf_mtx * weight_mtx)
@@ -185,4 +185,6 @@ def reset_states(self):
 
         for v in self.variables:
             K.set_value(
-                v, np.zeros((self.num_classes, self.num_classes), np.int32))
+                v,
+                np.zeros((self.num_classes, self.num_classes),
+                         v.dtype.as_numpy_dtype))

tensorflow_addons/metrics/cohens_kappa_test.py

@@ -130,6 +130,16 @@ def test_kappa_with_sample_weights(self):
         self.check_results([kp_obj1, kp_obj2, kp_obj3],
                            [-0.25473321, -0.38992332, -0.60695344])
 
+    def test_large_values(self):
+        y_true = [1] * 10000 + [0] * 20000 + [1] * 20000
+        y_pred = [0] * 20000 + [1] * 30000
+
+        obj = CohenKappa(num_classes=2)
+        self.evaluate(tf.compat.v1.variables_initializer(obj.variables))
+
+        self.evaluate(obj.update_state(y_true, y_pred))
+        self.assertAllClose(0.166666666, obj.result())
+
 
 if __name__ == '__main__':
     tf.test.main()