Use tf.int32.min rather than relying on integer overflow (#21077)

wilsbj · web-flow · commit 0cfd2aa18b9e · 2025-03-20T15:08:43.000-07:00
diff --git a/keras/src/backend/tensorflow/numpy.py b/keras/src/backend/tensorflow/numpy.py
@@ -2078,7 +2078,7 @@ def signbit(x):
             tf.bitwise.bitwise_and(
                 tf.bitcast(x, tf.int32),
                 # tf.float32 sign bit
-                tf.constant(0x80000000, dtype=tf.int32),
+                tf.constant(tf.int32.min, dtype=tf.int32),
             ),
             0,
         )

Original file line number	Diff line number	Diff line change
`@@ -2078,7 +2078,7 @@ def signbit(x):`
`2078`	`2078`	`tf.bitwise.bitwise_and(`
`2079`	`2079`	`tf.bitcast(x, tf.int32),`
`2080`	`2080`	`# tf.float32 sign bit`
`2081`		`- tf.constant(0x80000000, dtype=tf.int32),`
	`2081`	`+ tf.constant(tf.int32.min, dtype=tf.int32),`
`2082`	`2082`	`),`
`2083`	`2083`	`0,`
`2084`	`2084`	`)`