Merge pull request #50 from onnx/gs/onnx-1.2

work around missing onnx broadcast support for min/max

Author: guschmue

tests/test_backend.py

@@ -173,6 +173,17 @@ def test_trig_ops(self):
 
     @unittest.skipIf(BACKEND in ["caffe2", "onnxmsrt"], "not supported correctly in caffe2")
     def test_multinomial(self):
+        x_val = np.array([[10., 10.]], dtype=np.float32)
+        x = tf.placeholder(tf.float32, shape=x_val.shape, name=_TFINPUT)
+        op = tf.multinomial(tf.log(x), 5, output_dtype=tf.int64)
+        output = tf.identity(op, name=_TFOUTPUT)
+        actual, expected = self._run(output, {x: x_val}, {_INPUT: x_val})
+        # since returned indexes are random we can only check type and shape
+        self.assertEqual(expected.dtype, actual.dtype)
+        self.assertEqual(expected.shape, actual.shape)
+
+    @unittest.skipIf(BACKEND in ["caffe2", "onnxmsrt"], "not supported correctly in caffe2")
+    def test_multinomial1(self):
         shape = [2, 10]
         x_val = np.ones(np.prod(shape)).astype("float32").reshape(shape)
         x = tf.placeholder(tf.float32, shape=x_val.shape, name=_TFINPUT)
@@ -181,7 +192,7 @@ def test_multinomial(self):
         actual, expected = self._run(output, {x: x_val}, {_INPUT: x_val})
         # since returned indexes are random we can only check type and shape
         self.assertEqual(expected.dtype, actual.dtype)
-        self.assertAllClose(expected.shape, actual.shape)
+        self.assertEqual(expected.shape, actual.shape)
 
     def test_maxppol(self):
         x_val = make_xval((1, 4, 4, 1))
@@ -462,13 +473,25 @@ def test_square(self):
     def test_min(self):
         x_val1 = np.array([4.0, 16.0, 4.0, 1.6], dtype=np.float32).reshape((2, 2))
         x_val2 = np.array([4.0, 4.0, 4.0, 4.0], dtype=np.float32).reshape((2, 2))
-        x1 = tf.placeholder(tf.float32, [2, 2], name=_TFINPUT)
-        x2 = tf.placeholder(tf.float32, [2, 2], name=_TFINPUT1)
+        x1 = tf.placeholder(tf.float32, x_val1.shape, name=_TFINPUT)
+        x2 = tf.placeholder(tf.float32, x_val2.shape, name=_TFINPUT1)
         mi = tf.minimum(x1, x2)
         output = tf.identity(mi, name=_TFOUTPUT)
         actual, expected = self._run(output, {x1: x_val1, x2: x_val2}, {_INPUT: x_val1, _INPUT1: x_val2, })
         self.assertAllClose(expected, actual)
 
+    @unittest.skipIf(BACKEND in ["caffe2", "onnxmsrt"], "issue with broadcastnig scalar")
+    def test_min_broadcast(self):
+        # tests if the broadcast for min/max is working
+        x_val1 = np.array([2.0, 16.0, 5.0, 1.6], dtype=np.float32).reshape((2, 2))
+        x_val2 = np.array([4.0], dtype=np.float32)
+        x1 = tf.placeholder(tf.float32, x_val1.shape, name=_TFINPUT)
+        x2 = tf.constant(x_val2, dtype=tf.float32, name='x2')
+        mi = tf.minimum(x1, x2)
+        output = tf.identity(mi, name=_TFOUTPUT)
+        actual, expected = self._run(output, {x1: x_val1}, {_INPUT: x_val1})
+        self.assertAllClose(expected, actual)
+
     def test_logicaland(self):
         x_val1 = np.array([1, 0, 1, 1], dtype=np.bool).reshape((2, 2))
         x_val2 = np.array([0, 1, 1, 1], dtype=np.bool).reshape((2, 2))
@@ -712,7 +735,6 @@ def test_pad(self):
 
     @unittest.skipIf(BACKEND in ["caffe2", "onnxmsrt"], "not supported correctly in caffe2")
     def test_randomuniform(self):
-        # not supported by onnxmsrt or caffe2
         shape = tf.constant([2, 3], name="shape")
         x_ = tf.random_uniform(shape, name="rand", dtype=tf.float32)
         x_ = tf.identity(x_, name="output1")
@@ -722,6 +744,17 @@ def test_randomuniform(self):
         # since results are random, compare the shapes only
         self.assertAllClose(expected.shape, actual.shape)
 
+    @unittest.skip
+    def test_randomuniform_int(self):
+        shape = tf.constant([2, 3], name="shape")
+        x_ = tf.random_uniform(shape, name="rand", dtype=tf.int32, maxval=10)
+        x_ = tf.identity(x_, name="output1")
+        x_ = tf.identity(x_, name="output2")
+        output = tf.identity(x_, name=_TFOUTPUT)
+        actual, expected = self._run(output, {}, {})
+        # since results are random, compare the shapes only
+        self.assertAllClose(expected.shape, actual.shape)
+
     @unittest.skip
     def test_argminmax(self):
         # TODO: fails on onnxmsrt caffe2

tests/unity.yaml

@@ -1,7 +1,5 @@
 3DBall:
-  # caffe2: needs RandomNormal
-  # onnxmsrtnext: fails on missing onnx min/max broadcast
-  disabled: true
+  check_only_shape: true
   url: https://github.com/Unity-Technologies/ml-agents/raw/master/unity-environment/Assets/ML-Agents/Examples/3DBall/TFModels/3DBall.bytes
   model: 3DBall.bytes
   input_get: get_random
@@ -57,9 +55,7 @@ Basic:
     - value_estimate:0
 
 Bouncer:
-  # caffe2: needs RandomNormal
-  # onnxmsrtnext: fails on missing onnx min/max broadcast
-  disabled: true
+  check_only_shape: true
   url: https://github.com/Unity-Technologies/ml-agents/raw/master/unity-environment/Assets/ML-Agents/Examples/Bouncer/TFModels/Bouncer.bytes
   model: Bouncer.bytes
   input_get: get_random
@@ -70,9 +66,7 @@ Bouncer:
     - value_estimate:0
 
 crawler:
-  # caffe2: needs RandomNormal
-  # onnxmsrtnext: fails on missing onnx min/max broadcast
-  disabled: true
+  check_only_shape: true
   url: https://github.com/Unity-Technologies/ml-agents/raw/master/unity-environment/Assets/ML-Agents/Examples/Crawler/TFModels/crawler.bytes
   model: crawler.bytes
   input_get: get_random
@@ -134,9 +128,7 @@ PushBlock:
     - action:0
 
 Reacher:
-  # caffe2: needs RandomNormal
-  # onnxmsrtnext: fails on missing onnx min/max broadcast
-  disabled: true
+  check_only_shape: true
   url: https://github.com/Unity-Technologies/ml-agents/raw/master/unity-environment/Assets/ML-Agents/Examples/Reacher/TFModels/Reacher.bytes
   model: Reacher.bytes
   input_get: get_random

tf2onnx/tfonnx.py

@@ -57,6 +57,7 @@ def tensorflow_to_onnx(graph):
                 shape = out.get_shape().as_list()
             except Exception as ex:
                 shape = []
+            dtypes[out.name] = utils.map_tf_dtype(out.dtype)
             output_shapes[out.name] = shape
 
     # minimal conversion of attributes
@@ -805,6 +806,32 @@ def spacetodepth_op(ctx, node, name, args):
     return nodes
 
 
+def minmax_op(ctx, node, name, args):
+    # tensorflow minimum/maximum support broadcast. Onnx <= opset 7 does not.
+    # inject a add(0) as 'broadcast' operator if needed.
+    shapeo = ctx.get_shape(node.output[0])
+    needs_broadcast_op = []
+    for i, name in enumerate(node.input):
+        if ctx.get_shape(name) != shapeo:
+            needs_broadcast_op.append(i)
+    if needs_broadcast_op:
+        new_nodes = []
+        for i in needs_broadcast_op:
+            input_node = node.inputs[i]
+            dtype = ctx.dtypes[node.input[i]]
+            zero_name = utils.make_name(input_node.name)
+            zero_node = ctx.make_const(zero_name, "Const", np.zeros(shapeo, dtype=utils.ONNX_TO_NUMPY_DTYPE[dtype]))
+            op_name = utils.make_name(input_node.name)
+            output_name = op_name + ":0"
+            add_node = Node(helper.make_node("Add", [input_node.output[0], zero_name],
+                                             [output_name], name=op_name), ctx)
+            node.input[i] = output_name
+            new_nodes.append(add_node)
+        new_nodes.append(node)
+        return new_nodes
+    return node
+
+
 # pylint: enable=W0613,C0111,W0612
 
 # map tensorflow ops to onnx ops. The format below is
@@ -847,12 +874,12 @@ def spacetodepth_op(ctx, node, name, args):
     "LogicalOr": (broadcast_op, ["Or"]),
     "Max": (reduce_op, ["ReduceMax"]),
     "MatMul": (direct_op, ["MatMul"]),
-    "Maximum": (direct_op, ["Max"]),
+    "Maximum": (minmax_op, ["Max"]),
     "MaxPool": (pool_op, ["MaxPool"]),
     "MaxPoolV2": (pool_op, ["MaxPool"]),
     "Mean": (reduce_op, ["ReduceMean"]),
     "Min": (reduce_op, ["ReduceMin"]),
-    "Minimum": (direct_op, ["Min"]),
+    "Minimum": (minmax_op, ["Min"]),
     "Mul": (broadcast_op, []),
     "Neg": (direct_op, []),
     "NoOp": (no_op, []),