[jax2tf] Add more documentation about saving models with custom gradients

jax/experimental/jax2tf/README.md

@@ -95,7 +95,8 @@ is trivial:
 my_model = tf.Module()
 # Save a function that can take scalar inputs.
 my_model.f = tf.function(jax2tf.convert(f_jax), input_signature=[tf.TensorSpec([], tf.float32)])
-tf.saved_model.save(my_model, '/some/directory')
+tf.saved_model.save(my_model, '/some/directory',
+                    options=tf.saved_model.SaveOptions(experimental_custom_gradients=True))
 
 # Restoring (note: the restored model does *not* require JAX to run, just XLA).
 restored_model = tf.saved_model.load('/some/directory')
@@ -113,7 +114,8 @@ SavedModel multiple versions of a function for different input shapes, by
 my_model.f = tf.function(jax2tf.convert(f_jax), autograph=False)
 my_model.f(tf.ones([1, 28, 28]))  # a batch size of 1
 my_model.f(tf.ones([16, 28, 28]))  # a batch size of 16
-tf.saved_model.save(my_model, '/some/directory')
+tf.saved_model.save(my_model, '/some/directory',
+                    options=tf.saved_model.SaveOptions(experimental_custom_gradients=True))
 ```
 
 For examples of how to save a Flax model as a SavedModel see the
@@ -144,6 +146,32 @@ options = tf.saved_model.SaveOptions(experimental_custom_gradients=True)
 tf.saved_model.save(model, path, options=options)
 ```
 
+If you use `with_gradient=True` and forget to use the `experimental_custom_gradients=True` parameter
+to `tf.saved_model.save` when you later load the saved model you will see a warning:
+
+```
+WARNING:absl:Importing a function (__inference_converted_fun_25) with ops with unsaved custom gradients. Will likely fail if a gradient is requested.
+```
+
+and if you do attempt to take a gradient of the loaded model you may get an error:
+
+```
+TypeError: An op outside of the function building code is being passed
+a "Graph" tensor. It is possible to have Graph tensors
+leak out of the function building context by including a
+tf.init_scope in your function building code.
+For example, the following function will fail:
+  @tf.function
+  def has_init_scope():
+    my_constant = tf.constant(1.)
+    with tf.init_scope():
+      added = my_constant * 2
+The graph tensor has name: args_0:0
+```
+
+(We are working with the TF team to give a more explicit error in this case.)
+
+
 ## Shape-polymorphic conversion
 
 **The shape polymorphism support is work in progress. It is meant to be sound,

jax/experimental/jax2tf/examples/saved_model_lib.py

@@ -41,7 +41,7 @@ def convert_and_save_model(
     with_gradient: bool = False,
     enable_xla: bool = True,
     compile_model: bool = True,
-    save_model_options: Optional[tf.saved_model.SaveOptions] = None):
+    saved_model_options: Optional[tf.saved_model.SaveOptions] = None):
   """Convert a JAX function and saves a SavedModel.
 
   This is an example, for serious uses you will likely want to copy and
@@ -89,7 +89,7 @@ def convert_and_save_model(
       `polymorphic_shapes` argument to jax2tf.convert for the second parameter of
       `jax_fn`. In this case, a single `input_signatures` is supported, and
       should have `None` in the polymorphic dimensions.
-    save_model_options: options to pass to savedmodel.save.
+    saved_model_options: options to pass to savedmodel.save.
   """
   if not input_signatures:
     raise ValueError("At least one input_signature must be given")
@@ -124,8 +124,13 @@ def convert_and_save_model(
     # If there are more signatures, trace and cache a TF function for each one
     tf_graph.get_concrete_function(input_signature)
   wrapper = _ReusableSavedModelWrapper(tf_graph, param_vars)
+  if with_gradient:
+    if not saved_model_options:
+      saved_model_options = tf.saved_model.SaveOptions(experimental_custom_gradients=True)
+    else:
+      saved_model_options.experimental_custom_gradients = True
   tf.saved_model.save(wrapper, model_dir, signatures=signatures,
-                      options=save_model_options)
+                      options=saved_model_options)
 
 
 class _ReusableSavedModelWrapper(tf.train.Checkpoint):

jax/experimental/jax2tf/tests/call_tf_test.py

@@ -415,6 +415,48 @@ class CallTfTest(jtu.JaxTestCase):
     self.assertAllClose(g(x), g_rt(x))
     self.assertAllClose(jax.grad(g)(x), jax.grad(g_rt)(x))
 
+  def test_round_trip_without_gradient_saved_model(self):
+    # Explicitly with_gradient=False
+    f_jax = jnp.sum
+
+    x = np.array([0.7, 0.8], dtype=np.float32)
+    f_tf = tf_test_util.SaveAndLoadFunction(
+        jax2tf.convert(f_jax, with_gradient=False),
+        [tf.TensorSpec(x.shape, dtype=x.dtype)])
+    f_rt = jax2tf.call_tf(f_tf)
+
+    self.assertAllClose(f_jax(x), f_rt(x))
+    with self.assertRaisesRegex(Exception,
+                                "Gradient explicitly disabled.*jax2tf-converted function does not support gradients. Use `with_gradient` parameter to enable gradients"):
+      jax.grad(f_rt)(x)
+
+  def test_round_trip_saved_model_no_gradients(self):
+    # Save without gradients
+    f_jax = jnp.sum
+
+    x = np.array([0.7, 0.8], dtype=np.float32)
+    f_tf = tf_test_util.SaveAndLoadFunction(
+        jax2tf.convert(f_jax, with_gradient=True),
+        [tf.TensorSpec(x.shape, dtype=x.dtype)],
+        save_gradients=False)
+    f_rt = jax2tf.call_tf(f_tf)
+
+    self.assertAllClose(f_jax(x), f_rt(x))
+    # TODO: clean this up b/191117111: it should fail with a clear error
+    # The following results in a confusing error:
+    # TypeError: An op outside of the function building code is being passed
+    # a "Graph" tensor. It is possible to have Graph tensors
+    # leak out of the function building context by including a
+    # tf.init_scope in your function building code.
+    # For example, the following function will fail:
+    #   @tf.function
+    #   def has_init_scope():
+    #     my_constant = tf.constant(1.)
+    #     with tf.init_scope():
+    #       added = my_constant * 2
+    # The graph tensor has name: args_0:0
+    # g = jax.grad(f_rt)(x)
+
   def test_module_documentation(self):
     def cos_tf(x):
       return tf.math.cos(x)

jax/experimental/jax2tf/tests/savedmodel_test.py

@@ -42,24 +42,6 @@ class SavedModelTest(tf_test_util.JaxToTfTestCase):
     restored_model = tf_test_util.SaveAndLoadModel(model)
     self.assertAllClose(restored_model.f(x), f_jax(x))
 
-  def test_gradient_disabled(self):
-    f_jax = lambda x: x * x
-
-    model = tf.Module()
-    model.f = tf.function(jax2tf.convert(f_jax, with_gradient=False),
-                          autograph=False,
-                          input_signature=[tf.TensorSpec([], tf.float32)])
-    x = np.array(0.7, dtype=jnp.float32)
-    self.assertAllClose(model.f(x), f_jax(x))
-    restored_model = tf_test_util.SaveAndLoadModel(model)
-    xv = tf.Variable(0.7, dtype=jnp.float32)
-    self.assertAllClose(restored_model.f(x), f_jax(x))
-
-    with self.assertRaisesRegex(LookupError,
-                                "Gradient explicitly disabled.*The jax2tf-converted function does not support gradients"):
-      with tf.GradientTape():
-        _ = restored_model.f(xv)
-
   def test_gradient(self):
     """Save and restore the custom gradient."""
     @jax.custom_jvp
@@ -82,12 +64,95 @@ class SavedModelTest(tf_test_util.JaxToTfTestCase):
     x = np.array(0.7, dtype=jnp.float32)
     self.assertAllClose(model.f(x), f_jax(x))
     restored_model = tf_test_util.SaveAndLoadModel(model)
-    xv = tf.Variable(0.7, dtype=jnp.float32)
+    xv = tf.Variable(x)
     self.assertAllClose(restored_model.f(x), f_jax(x))
     with tf.GradientTape() as tape:
       y = restored_model.f(xv)
     self.assertAllClose(tape.gradient(y, xv).numpy(),
-                        jax.grad(f_jax)(x).astype(np.float32))
+                        jax.grad(f_jax)(x))
+
+  def test_gradient_nested(self):
+    """Save and restore the custom gradient, when combined with other TF code."""
+    @jax.custom_jvp
+    def f_jax(x):
+      return x * x
+
+    @f_jax.defjvp
+    def f_jax_jvp(primals, tangents):
+      # 3 * x * x_t
+      x, = primals
+      x_dot, = tangents
+      primal_out = f_jax(x)
+      tangent_out = x * x_dot * 3.
+      return primal_out, tangent_out
+
+    model = tf.Module()
+    # After conversion, we wrap with some pure TF code
+    model.f = tf.function(lambda x: tf.math.sin(jax2tf.convert(f_jax, with_gradient=True)(x)),
+                          autograph=False,
+                          input_signature=[tf.TensorSpec([], tf.float32)])
+    f_jax_equiv = lambda x: jnp.sin(f_jax(x))
+    x = np.array(0.7, dtype=jnp.float32)
+    self.assertAllClose(model.f(x), f_jax_equiv(x))
+    restored_model = tf_test_util.SaveAndLoadModel(model)
+    xv = tf.Variable(x)
+    self.assertAllClose(restored_model.f(x), f_jax_equiv(x))
+    with tf.GradientTape() as tape:
+      y = restored_model.f(xv)
+    self.assertAllClose(tape.gradient(y, xv).numpy(),
+                        jax.grad(f_jax_equiv)(x))
+
+  def test_gradient_disabled(self):
+    f_jax = lambda x: x * x
+
+    model = tf.Module()
+    model.f = tf.function(jax2tf.convert(f_jax, with_gradient=False),
+                          autograph=False,
+                          input_signature=[tf.TensorSpec([], tf.float32)])
+    x = np.array(0.7, dtype=jnp.float32)
+    self.assertAllClose(model.f(x), f_jax(x))
+    restored_model = tf_test_util.SaveAndLoadModel(model)
+    xv = tf.Variable(0.7, dtype=jnp.float32)
+    self.assertAllClose(restored_model.f(x), f_jax(x))
+
+    with self.assertRaisesRegex(LookupError,
+                                "Gradient explicitly disabled.*The jax2tf-converted function does not support gradients"):
+      with tf.GradientTape():
+        _ = restored_model.f(xv)
+
+  def test_save_without_gradients(self):
+    f_jax = lambda x: x * x
+
+    x = np.array(0.7, dtype=jnp.float32)
+    model = tf.Module()
+    model.f = tf.function(jax2tf.convert(f_jax, with_gradient=True),
+                          autograph=False,
+                          input_signature=[tf.TensorSpec(x.shape, x.dtype)])
+
+    self.assertAllClose(model.f(x), f_jax(x))
+    restored_model = tf_test_util.SaveAndLoadModel(model,
+                                                   save_gradients=False)
+    self.assertAllClose(restored_model.f(x), f_jax(x))
+
+    xv = tf.Variable(x)
+    with tf.GradientTape():
+      _ = restored_model.f(xv)
+      # TODO: clean this up b/191117111: it should fail with a clear error
+      # The following results in a confusing error:
+      # TypeError: An op outside of the function building code is being passed
+      # a "Graph" tensor. It is possible to have Graph tensors
+      # leak out of the function building context by including a
+      # tf.init_scope in your function building code.
+      # For example, the following function will fail:
+      #   @tf.function
+      #   def has_init_scope():
+      #     my_constant = tf.constant(1.)
+      #     with tf.init_scope():
+      #       added = my_constant * 2
+      # The graph tensor has name: args_0:0
+      # g = tape.gradient(res, xv)
+    #self.assertAllClose(g.numpy(), jax.grad(f_jax)(x))
+
 
   def _compare_with_saved_model(self, f_jax, *args):
     # Certain ops are converted to ensure an XLA context, e.g.,

jax/experimental/jax2tf/tests/tf_test_util.py

@@ -75,23 +75,25 @@ class OpMetadataGraph:
   source_line: str
 
 
-def SaveAndLoadModel(model: tf.Module) -> tf.Module:
+def SaveAndLoadModel(model: tf.Module,
+                     save_gradients=True) -> tf.Module:
   # Roundtrip through saved model on disk.
   model_dir = os.path.join(absltest.get_default_test_tmpdir(), str(id(model)))
   tf.saved_model.save(
       model, model_dir,
-      options=tf.saved_model.SaveOptions(experimental_custom_gradients=True))
+      options=tf.saved_model.SaveOptions(experimental_custom_gradients=save_gradients))
   restored_model = tf.saved_model.load(model_dir)
   return restored_model
 
 def SaveAndLoadFunction(f_tf: Callable,
-                        input_signature: Sequence[tf.TensorSpec]) -> Callable:
+                        input_signature: Sequence[tf.TensorSpec],
+                        save_gradients=True) -> Callable:
   # Roundtrip through saved model on disk
   model = tf.Module()
   model.f = tf.function(f_tf,
                         autograph=False,
                         input_signature=input_signature)
-  restored = SaveAndLoadModel(model)
+  restored = SaveAndLoadModel(model, save_gradients=save_gradients)
   return restored.f