Change tf_upgrade_v2.py script to wrap labels argument with tf.stop_gradients in

tf.softmax_cross_entropy_with_logits call.

PiperOrigin-RevId: 228260462

tensorflow/tools/compatibility/BUILD

@@ -69,6 +69,7 @@ py_library(
         ":ast_edits",
         ":renames_v2",
         ":reorders_v2",
+        "@six_archive//:six",
     ],
 )
 

tensorflow/tools/compatibility/reorders_v2.py

@@ -65,6 +65,7 @@ reorders = {
     'tf.nn.moments': ['x', 'axes', 'shift', 'name', 'keep_dims'],
     'tf.nn.pool': ['input', 'window_shape', 'pooling_type', 'padding', 'dilation_rate', 'strides', 'name', 'data_format'],
     'tf.nn.separable_conv2d': ['input', 'depthwise_filter', 'pointwise_filter', 'strides', 'padding', 'rate', 'name', 'data_format'],
+    'tf.nn.softmax_cross_entropy_with_logits': ['_sentinel', 'labels', 'logits', 'dim', 'name'],
     'tf.nn.space_to_batch': ['input', 'paddings', 'block_size', 'name'],
     'tf.nn.space_to_depth': ['input', 'block_size', 'name', 'data_format'],
     'tf.nn.weighted_moments': ['x', 'axes', 'frequency_weights', 'name', 'keep_dims'],

tensorflow/tools/compatibility/testdata/test_file_v1_12.py

@@ -70,6 +70,15 @@ class TestUpgrade(test_util.TensorFlowTestCase):
         [0],
         tf.argmin([[1, 3, 2]], name='abc', dimension=1))
 
+  @test_util.run_v1_only("b/120545219")
+  def testSoftmaxCrossEntropyWithLogits(self):
+    out = tf.nn.softmax_cross_entropy_with_logits(
+        logits=[0.1, 0.8], labels=[0, 1])
+    self.assertAllClose(out, 0.40318608)
+    out = tf.nn.softmax_cross_entropy_with_logits_v2(
+        logits=[0.1, 0.8], labels=[0, 1])
+    self.assertAllClose(out, 0.40318608)
+
 
 if __name__ == "__main__":
   test_lib.main()

tensorflow/tools/compatibility/tf_upgrade_v2.py

@@ -21,6 +21,7 @@ from __future__ import print_function
 import ast
 
 import pasta
+import six
 
 from tensorflow.tools.compatibility import ast_edits
 from tensorflow.tools.compatibility import renames_v2
@@ -94,6 +95,10 @@ class TFAPIChangeSpec(ast_edits.APIChangeSpec):
         "tf.convert_to_tensor": {
             "preferred_dtype": "dtype_hint"
         },
+        "tf.nn.softmax_cross_entropy_with_logits": {
+            "dim": "axis",
+            "_sentinel": None,
+        },
         "tf.nn.softmax_cross_entropy_with_logits_v2": {
             "dim": "axis"
         },
@@ -682,6 +687,10 @@ class TFAPIChangeSpec(ast_edits.APIChangeSpec):
         "tf.norm",
         "tf.reverse_sequence",
         "tf.sparse_split",
+        # tf.nn.softmax_cross_entropy_with_logits *must* be called with
+        # keyword arguments. Add keyword arguments in rare case when they
+        # are not specified.
+        "tf.nn.softmax_cross_entropy_with_logits",
     }
 
     # Functions that were reordered should be changed to the new keyword args
@@ -714,6 +723,8 @@ class TFAPIChangeSpec(ast_edits.APIChangeSpec):
         "tf.to_float": self._cast_transformer,
         "tf.to_int32": self._cast_transformer,
         "tf.to_int64": self._cast_transformer,
+        "tf.nn.softmax_cross_entropy_with_logits":
+            self._softmax_cross_entropy_with_logits_transformer,
     }
 
     decay_function_comment = (
@@ -950,10 +961,6 @@ class TFAPIChangeSpec(ast_edits.APIChangeSpec):
             "'deterministic' arguments. Now it takes a single 'seed' arg. If "
             "'seed' is zero, the execution is random and deterministic "
             "otherwise",
-        "tf.nn.softmax_cross_entropy_with_logits":
-            "tf.nn.softmax_cross_entropy_with_logits behavior has changed. "
-            "'labels' needs to be wrapped with tf.stop_gradient to keep the "
-            "old behavior. Also, 'dim' argument has been renamed to 'axis'.",
         "tf.test.assert_equal_graph_def":
             "tf.assert_equal_graph_def no longer takes 'checkpoint_v2' "
             "argument. 'checkpoint_v2' now defaults to True.",
@@ -1228,6 +1235,36 @@ class TFAPIChangeSpec(ast_edits.APIChangeSpec):
                                                                     dtype_str)))
     return node
 
+  @staticmethod
+  def _softmax_cross_entropy_with_logits_transformer(
+      parent, node, full_name, name, logs, errors):
+    def _wrap_label(parent, old_value):
+      """Wrap labels with tf.stop_gradient."""
+      if six.PY3:
+        new_value = ast.Call(
+            ast.Name(id="tf.stop_gradient", ctx=ast.Load()),
+            [old_value], [])
+      else:
+        new_value = ast.Call(
+            ast.Name(id="tf.stop_gradient", ctx=ast.Load()),
+            [old_value], [], None, None)
+
+      # This copies the prefix and suffix on old_value to new_value.
+      pasta.ast_utils.replace_child(parent, old_value, new_value)
+      ast.copy_location(new_value, old_value)
+
+    # Check if we have a labels keyword arg
+    for karg in node.keywords:
+      if karg.arg == "labels":
+        logs.append((node.lineno, node.col_offset,
+                     "Changing labels arg of "
+                     "tf.nn.softmax_cross_entropy_with_logits to "
+                     "tf.stop_gradient(labels). Please check this "
+                     "transformation.\n"))
+        _wrap_label(karg, karg.value)
+        return node
+    return node
+
   @staticmethod
   def _batch_gather_transformer(parent, node, full_name, name, logs, errors):
     # Check if the call already has a batch_dims argument

tensorflow/tools/compatibility/tf_upgrade_v2_test.py

@@ -684,26 +684,32 @@ bazel-bin/tensorflow/tools/compatibility/update/generate_v2_reorders_map
     self.assertEqual(new_text, expected_text)
 
   def testSoftMaxCrossEntropyWithLogitsV2(self):
-    text = "tf.nn.softmax_cross_entropy_with_logits_v2(labels, logits, dim=2)"
+    text = (
+        "tf.nn.softmax_cross_entropy_with_logits_v2("
+        "labels=labels, logits=logits, dim=2)")
     expected_text = (
-        "tf.nn.softmax_cross_entropy_with_logits(labels, logits, axis=2)")
+        "tf.nn.softmax_cross_entropy_with_logits("
+        "labels=labels, logits=logits, axis=2)")
     _, unused_report, errors, new_text = self._upgrade(text)
     self.assertEqual(new_text, expected_text)
 
     self.assertFalse(errors)
 
   def testSoftMaxCrossEntropyWithLogits(self):
-    text = "tf.nn.softmax_cross_entropy_with_logits(labels, logits, dim=2)"
+    text = ("tf.nn.softmax_cross_entropy_with_logits("
+            "labels=labels, logits=logits, dim=2)")
     expected_text = (
-        "tf.nn.softmax_cross_entropy_with_logits(labels, logits, dim=2)")
+        "tf.nn.softmax_cross_entropy_with_logits("
-    _, report, errors, new_text = self._upgrade(text)
+        "labels=tf.stop_gradient(labels), logits=logits, axis=2)")
+    _, unused_report, unused_errors, new_text = self._upgrade(text)
     self.assertEqual(new_text, expected_text)
-    self.assertIn(
+
-        "tf.nn.softmax_cross_entropy_with_logits requires manual check",
+    text = ("tf.nn.softmax_cross_entropy_with_logits("
-        errors[0])
+            "labels=foo(bar))")
-    self.assertIn(
+    expected_text = ("tf.nn.softmax_cross_entropy_with_logits("
-        "tf.nn.softmax_cross_entropy_with_logits behavior has changed. ",
+                     "labels=tf.stop_gradient(foo(bar)))")
-        report)
+    _, unused_report, unused_errors, new_text = self._upgrade(text)
+    self.assertEqual(expected_text, new_text)
 
   def testSparseMatmul(self):
     text = ("tf.sparse_matmul(a, b, c, d, e, f, g)\n")