Merge pull request #336 from mindest/dev_isnan

add op IsNaN and test case test_isnan

Author: nbcsm

tests/test_backend.py

@@ -1702,6 +1702,19 @@ def test_zeros_like(self):
         _ = tf.identity(res1, name=_TFOUTPUT1)
         self._run_test_case([_OUTPUT, _OUTPUT1], {_INPUT: input_val})
 
+    @check_opset_min_version(9, "is_nan")
+    def test_isnan(self):
+        # only compatible with dtype `float32`
+        x_val1 = np.array([1.0, 2.0, -3.0, -4.0], dtype=np.float32).reshape((2, 2))
+        x_val2 = np.array([np.nan, np.nan, np.nan, np.nan], dtype=np.float32).reshape((2, 2))
+        x_val3 = np.array([1.0, np.nan, -3.0, np.nan], dtype=np.float32).reshape((2, 2))
+        for x_val in [x_val1, x_val2, x_val3]:
+            x = tf.placeholder(tf.float32, x_val.shape, name=_TFINPUT)
+            x_ = tf.is_nan(x)
+            _ = tf.identity(x_, name=_TFOUTPUT)
+            self._run_test_case([_OUTPUT], {_INPUT: x_val})
+            tf.reset_default_graph()
+
 
 if __name__ == '__main__':
     unittest_main()

tf2onnx/tfonnx.py

@@ -1894,17 +1894,18 @@ def where_op(ctx, node, name, args):
 }
 
 _OPSET_9 = {
-    "Erf": (direct_op, []),
-    "Fill": (fill_op, []),
-    "Sinh": (direct_op, []),
-    "Cosh": (direct_op, []),
-    "Asinh": (direct_op, []),
     "Acosh": (direct_op, []),
+    "Asinh": (direct_op, []),
     "Atanh": (direct_op, []),
+    "Cosh": (direct_op, []),
+    "Erf": (direct_op, []),
+    "Fill": (fill_op, []),
     "Greater": (logical_compare_op, []),
+    "IsNan": (direct_op, ["IsNaN"]),
     "Less": (logical_compare_op, []),
     "ResizeBilinear": (upsample_op9, ["Upsample", "linear"]),
     "ResizeNearestNeighbor": (upsample_op9, ["Upsample", "nearest"]),
+    "Sinh": (direct_op, []),
     "Where": (where_op, []),
 }
 
@@ -2296,7 +2297,7 @@ def rewrite_incomplete_type_support_rs5(g, ops):
 
 
 def rewrite_incomplete_type_support_rs6(g, ops):
-    return rewrite_incomplete_type_support(g, ops, ["Div", "ReduceSum", "Slice", "Split", "Tile", "Transpose"])
+    return rewrite_incomplete_type_support(g, ops, ["Div", "IsNaN", "ReduceSum", "Slice", "Split", "Tile", "Transpose"])
 
 
 def tensorflow_onnx_mapping(g, continue_on_error, custom_op_handlers):