tensorflow · seanpmorgan · Mar 7, 2020 · Feb 23, 2020 · Feb 23, 2020 · Feb 23, 2020
diff --git a/README.md b/README.md
@@ -14,10 +14,10 @@
 
 | Build Type      | Status |
 | ---             | ---    |
-| **MacOS CPU**   | [![Status](https://github.com/tensorflow/addons/workflows/macos-nightly/badge.svg)](https://github.com/tensorflow/addons/actions?query=workflow%3Amacos-nightly) |
-| **Windows CPU**   | [![Status](https://github.com/tensorflow/addons/workflows/windows-nightly/badge.svg)](https://github.com/tensorflow/addons/actions?query=workflow%3Awindows-nightly) |
-| **Ubuntu CPU**   | [![Status](https://github.com/tensorflow/addons/workflows/manylinux-nightly/badge.svg)](https://github.com/tensorflow/addons/actions?query=workflow%3Amanylinux-nightly) |
-| **Ubuntu GPU**   | [![Status](https://storage.googleapis.com/tensorflow-kokoro-build-badges/addons/ubuntu-gpu-py3.svg)](https://storage.googleapis.com/tensorflow-kokoro-build-badges/addons/ubuntu-gpu-py3.html) |
+| **MacOS**   | [![Status](https://github.com/tensorflow/addons/workflows/macos-nightly/badge.svg)](https://github.com/tensorflow/addons/actions?query=workflow%3Amacos-nightly) |
+| **Windows**   | [![Status](https://github.com/tensorflow/addons/workflows/windows-nightly/badge.svg)](https://github.com/tensorflow/addons/actions?query=workflow%3Awindows-nightly) |
+| **Ubuntu**   | [![Status](https://github.com/tensorflow/addons/workflows/manylinux-nightly/badge.svg)](https://github.com/tensorflow/addons/actions?query=workflow%3Amanylinux-nightly) |
+| **Ubuntu custom GPU ops**   | [![Status](https://storage.googleapis.com/tensorflow-kokoro-build-badges/addons/ubuntu-gpu-py3.svg)](https://storage.googleapis.com/tensorflow-kokoro-build-badges/addons/ubuntu-gpu-py3.html) |
 
 **TensorFlow Addons** is a repository of contributions that conform to
 well-established API patterns, but implement new functionality
@@ -105,9 +105,40 @@ TF-Addons. In order to achieve these we require that our additions
 conform to established API patterns seen in core TensorFlow.
 
 #### GPU/CPU Custom-Ops
-A major benefit of TensorFlow Addons is that there are precompiled ops. Should 
-a CUDA 10.1 installation not be found then the op will automatically fall back to 
-a CPU implementation.
+A major benefit of TensorFlow Addons is that there are precompiled ops.
+Custom ops work on GPU only on Ubuntu. This is why on Windows and MacOS, 
+If available, Addons will fallback to a pure TensorFlow Python implementation for custom-ops.
+
+The order of priority in MacOS/Windows:
+1) Pure TensorFlow + Python implementation (work on cpu+gpu)
+2) C++ implementation for CPU
+
+The order of priority for Ubuntu:
+1) CUDA implementation
+2) C++ implementation
+3) Pure TensorFlow + Python implementation (work on cpu+gpu)
+
+If you want to change the default priority, "C++ and CUDA" VS "pure TF Python", 
+you can either set the variable `TF_ADDONS_PY_OPS` from the command line or in 
+your code.
+
+For example, if you're on linux and you have compatibility problems with the compiled ops,
+and you want to give priority to the Python implementation
+you can do:
+
+From the command line:
+```
+export TF_ADDONS_PY_OPS=1
+```
+
+or in your code:
+
+```
+import tensorflow_addons as tfa
+tfa.options.TF_ADDONS_PY_OPS=True
+```
+
+This variable default to `True` on Windows and Mac, and `False` for Linux.
 
 #### Proxy Maintainership
 Addons has been designed to compartmentalize subpackages and submodules so 

diff --git a/tensorflow_addons/BUILD b/tensorflow_addons/BUILD
@@ -11,6 +11,7 @@ py_library(
     name = "tensorflow_addons",
     data = [
         "__init__.py",
+        "options.py",
         "version.py",
     ],
     deps = [

diff --git a/tensorflow_addons/activations/BUILD b/tensorflow_addons/activations/BUILD
@@ -16,6 +16,7 @@ py_library(
         "tanhshrink.py",
     ],
     data = [
+        "//tensorflow_addons:options.py",
         "//tensorflow_addons/custom_ops/activations:_activation_ops.so",
         "//tensorflow_addons/utils",
     ],

@@ -18,6 +18,7 @@
 
 from tensorflow_addons.utils import types
 from tensorflow_addons.utils.resource_loader import LazySO
+from tensorflow_addons import options
 
 _activation_so = LazySO("custom_ops/activations/_activation_ops.so")
 
@@ -40,6 +41,18 @@ def hardshrink(
         A `Tensor`. Has the same type as `x`.
     """
     x = tf.convert_to_tensor(x)
+
+    if not options.TF_ADDONS_PY_OPS:
+        try:
+            return _hardshrink_custom_op(x, lower, upper)
+        except tf.errors.NotFoundError:
+            options.warn_fallback("hardshrink")
+
+    return _hardshrink_py(x, lower, upper)
+
+
+def _hardshrink_custom_op(x, lower=-0.5, upper=0.5):
+    """Alias with lazy loading of the .so file"""
     return _activation_so.ops.addons_hardshrink(x, lower, upper)
 
 
@@ -59,7 +72,6 @@ def _hardshrink_py(
             " not be higher than the value "
             "variable upper, which is {} .".format(lower, upper)
         )
-    x = tf.convert_to_tensor(x)
     mask_lower = x < lower
     mask_upper = upper < x
     mask = tf.logical_or(mask_lower, mask_upper)

diff --git a/tensorflow_addons/activations/hardshrink_test.py b/tensorflow_addons/activations/hardshrink_test.py
@@ -17,16 +17,16 @@
 
 import numpy as np
 import tensorflow as tf
-from tensorflow_addons.activations import hardshrink
-from tensorflow_addons.utils import test_utils
+from tensorflow_addons.activations.hardshrink import _hardshrink_custom_op
 from tensorflow_addons.activations.hardshrink import _hardshrink_py
+from tensorflow_addons.utils import test_utils
 
 
 @test_utils.run_all_in_graph_and_eager_modes
 class HardshrinkTest(tf.test.TestCase, parameterized.TestCase):
     def test_invalid(self):
         with self.assertRaisesOpError("lower must be less than or equal to upper."):
-            y = hardshrink(tf.ones(shape=(1, 2, 3)), lower=2.0, upper=-2.0)
+            y = _hardshrink_custom_op(tf.ones(shape=(1, 2, 3)), lower=2.0, upper=-2.0)
             self.evaluate(y)
 
     @parameterized.named_parameters(
@@ -35,11 +35,11 @@ def test_invalid(self):
     def test_hardshrink(self, dtype):
         x = tf.constant([-2.0, -0.5, 0.0, 0.5, 2.0], dtype=dtype)
         expected_result = tf.constant([-2.0, 0.0, 0.0, 0.0, 2.0], dtype=dtype)
-        self.assertAllCloseAccordingToType(hardshrink(x), expected_result)
+        self.assertAllCloseAccordingToType(_hardshrink_custom_op(x), expected_result)
 
         expected_result = tf.constant([-2.0, 0.0, 0.0, 0.0, 2.0], dtype=dtype)
         self.assertAllCloseAccordingToType(
-            hardshrink(x, lower=-1.0, upper=1.0), expected_result
+            _hardshrink_custom_op(x, lower=-1.0, upper=1.0), expected_result
         )
 
     @parameterized.named_parameters(("float32", np.float32), ("float64", np.float64))
@@ -51,7 +51,7 @@ def test_theoretical_gradients(self, dtype):
         # Avoid these two points to make gradients smooth.
         x = tf.constant([-2.0, -1.5, 0.0, 1.5, 2.0], dtype=dtype)
 
-        theoretical, numerical = tf.test.compute_gradient(hardshrink, [x])
+        theoretical, numerical = tf.test.compute_gradient(_hardshrink_custom_op, [x])
         self.assertAllCloseAccordingToType(theoretical, numerical, atol=1e-4)
 
     @parameterized.named_parameters(("float32", np.float32), ("float64", np.float64))
@@ -68,7 +68,7 @@ def verify_funcs_are_equivalent(self, dtype):
 
         with tf.GradientTape(persistent=True) as t:
             t.watch(x)
-            y_native = hardshrink(x, lower, upper)
+            y_native = _hardshrink_custom_op(x, lower, upper)
             y_py = _hardshrink_py(x, lower, upper)
 
         self.assertAllCloseAccordingToType(y_native, y_py, atol=1e-4)

diff --git a/tensorflow_addons/options.py b/tensorflow_addons/options.py
@@ -0,0 +1,49 @@
+import os
+import platform
+import warnings
+import traceback
+
+try:
+    TF_ADDONS_PY_OPS = bool(int(os.environ["TF_ADDONS_PY_OPS"]))
+except KeyError:
+    if platform.system() == "Linux":
+        TF_ADDONS_PY_OPS = False
+    else:
+        TF_ADDONS_PY_OPS = True
+
+
+FALLBACK_WARNING_TEMPLATE = """{}
+
+The {} C++/CUDA custom op could not be loaded.
+For this reason, Addons will fallback to an implementation written
+in Python with public TensorFlow ops. There worst you might experience with
+this is a moderate slowdown on GPU. There can be multiple
+reason for this loading error, one of them may be an ABI incompatibility between
+the TensorFlow installed on your system and the TensorFlow used to compile
+TensorFlow Addons' custom ops. The stacktrace generated when loading the
+shared object file was displayed above.
+
+If you want this warning to disappear, either make sure the TensorFlow installed
+is compatible with this version of Addons, or tell TensorFlow Addons to
+prefer using Python implementations and not custom C++/CUDA ones. You can do that
+by changing the TF_ADDONS_PY_OPS flag
+either with the environment variable:
+```bash
+TF_ADDONS_PY_OPS=1 python my_script.py
+```
+or in your code, after your imports:
+```python
+import tensorflow_addons as tfa
+import ...
+import ...
+
+tfa.options.TF_ADDONS_PY_OPS = True
+```
+"""
+
+
+def warn_fallback(op_name):
+    warning_msg = FALLBACK_WARNING_TEMPLATE.format(traceback.format_exc(), op_name)
+    warnings.warn(warning_msg, RuntimeWarning)
+    global TF_ADDONS_PY_OPS
+    TF_ADDONS_PY_OPS = True
diff --git a/tools/ci_build/verify/check_typing_info.py b/tools/ci_build/verify/check_typing_info.py
@@ -33,7 +33,7 @@
 modules_list = []
 for attr_name in dir(tensorflow_addons):
     attr = getattr(tensorflow_addons, attr_name)
-    if isinstance(attr, ModuleType):
+    if isinstance(attr, ModuleType) and attr is not tensorflow_addons.options:
         modules_list.append(attr)