Fix MaxpoolWithArgmax (#1451)

Signed-off-by: Tom Wildenhain <[email protected]>

Author: TomWildenhain-Microsoft

tests/test_backend.py

@@ -164,12 +164,12 @@ def get_conv_getdata(kind=1):
 
 def get_maxpoolwithargmax_getdata():
     data = [
-        ('SAME', [1, 3, 3, 1], [1, 3, 3, 1], [1, 2, 2, 1]),
-        ('SAME', [1, 5, 5, 1], [1, 4, 4, 1], [1, 2, 2, 1]),
-        ('SAME', [1, 10, 5, 1], [1, 2, 2, 1], [1, 2, 2, 1]),
-        ('SAME', [1, 10, 5, 1], [1, 4, 4, 1], [1, 1, 1, 1]),
-        ('VALID', [1, 3, 3, 1], [1, 3, 3, 1], [1, 2, 2, 1]),
-        ('VALID', [1, 5, 5, 1], [1, 4, 4, 1], [1, 2, 2, 1]),
+        ('SAME', [1, 3, 3, 2], [1, 3, 3, 1], [1, 2, 2, 1]),
+        ('SAME', [2, 5, 5, 3], [1, 4, 4, 1], [1, 2, 2, 1]),
+        ('SAME', [2, 10, 5, 1], [1, 2, 2, 1], [1, 2, 2, 1]),
+        ('SAME', [2, 10, 5, 3], [1, 4, 4, 1], [1, 1, 1, 1]),
+        ('VALID', [2, 3, 3, 3], [1, 3, 3, 1], [1, 2, 2, 1]),
+        ('VALID', [2, 5, 5, 3], [1, 4, 4, 1], [1, 2, 2, 1]),
     ]
     for idx, v in enumerate(data):
         yield (idx,) + v
@@ -3738,13 +3738,41 @@ def func(x):
     def test_maxpoolwithargmax(self):
         for p in get_maxpoolwithargmax_getdata():
             _, padding, x_shape, ksize, strides = p
-            x_val = make_xval(x_shape)
+            x_val = np.random.uniform(0, 10, x_shape)
             def func(x):
                 mp = tf.nn.max_pool_with_argmax(x, ksize, strides, padding=padding)
                 return tf.identity(mp[0], name=_TFOUTPUT), tf.identity(mp[1], name=_TFOUTPUT1)
             self.logger.debug(str(p))
             self._run_test_case(func, [_OUTPUT, _OUTPUT1], {_INPUT: x_val})
 
+    @check_tf_min_version("1.13")
+    @check_opset_min_version(11, "MaxPoolWithArgmax")
+    def test_maxpoolwithargmax_batch_in_index(self):
+        padding = 'SAME'
+        x_shape = [2, 10, 5, 3]
+        ksize = [1, 4, 4, 1]
+        strides = [1, 1, 1, 1]
+        x_val = np.random.uniform(0, 10, x_shape)
+        def func(x):
+            mp = tf.nn.max_pool_with_argmax(x, ksize, strides, padding=padding, include_batch_in_index=True)
+            return tf.identity(mp[0], name=_TFOUTPUT), tf.identity(mp[1], name=_TFOUTPUT1)
+        self._run_test_case(func, [_OUTPUT, _OUTPUT1], {_INPUT: x_val})
+
+    @check_tf_min_version("1.13")
+    @check_opset_min_version(11, "MaxPoolWithArgmax")
+    def test_maxpoolwithargmax_unknown_c(self):
+        padding = 'SAME'
+        x_shape = [2, 10, 5, 1]
+        ksize = [1, 4, 4, 1]
+        strides = [1, 1, 1, 1]
+        x_val = np.random.uniform(0, 10, x_shape)
+        s_val = np.array([2, 10, 5, 4], np.int64)
+        def func(x, s):
+            x = tf.broadcast_to(x, s)
+            mp = tf.nn.max_pool_with_argmax(x, ksize, strides, padding=padding, include_batch_in_index=True)
+            return tf.identity(mp[0], name=_TFOUTPUT), tf.identity(mp[1], name=_TFOUTPUT1)
+        self._run_test_case(func, [_OUTPUT, _OUTPUT1], {_INPUT: x_val, _INPUT1: s_val})
+
     @check_opset_min_version(10, "Selu")
     def test_selu(self):
         x_val = np.random.random_sample([3]).astype(np.float32)

tf2onnx/onnx_opset/nn.py

@@ -694,6 +694,42 @@ def version_8(cls, ctx, node, **kwargs):
         # The input data_format is NHWC for TF MaxPoolWithArgmax
         node.set_attr("data_format", "NHWC")
 
+        # Convert indices from NCHW to NHWC format
+        input_shape = ctx.make_node("Shape", [node.input[0]]).output[0]
+        input_shape_guess = ctx.get_shape(node.input[0])
+        n, h, w, c = ctx.make_node("Split", [input_shape], attr={'axis': 0}, output_count=4).output
+        hw = ctx.make_node("Mul", [h, w]).output[0]
+        chw = ctx.make_node("Mul", [hw, c]).output[0]
+        consumers = ctx.find_output_consumers(node.output[1])
+        if ctx.opset >= 10:
+            xy = ctx.make_node("Mod", [node.output[1], hw]).output[0]
+        else:
+            xy_div = ctx.make_node("Div", [node.output[1], hw]).output[0]
+            xy_mul = ctx.make_node("Mul", [xy_div, hw]).output[0]
+            xy = ctx.make_node("Sub", [node.output[1], xy_mul]).output[0]
+        xy_scale_c = ctx.make_node("Mul", [xy, c]).output[0]
+        const_zero = ctx.make_const(utils.make_name("const_zero"), np.array(0, np.int64)).output[0]
+        const_one = ctx.make_const(utils.make_name("const_one"), np.array(1, np.int64)).output[0]
+        if input_shape_guess is not None and input_shape_guess[3] > 0:
+            c_range_np = np.arange(input_shape_guess[3], dtype=np.int64)
+            c_range = ctx.make_const(utils.make_name("c_range"), c_range_np).output[0]
+        else:
+            utils.make_sure(ctx.opset >= 11, "opset 11 required for MaxPoolWithArgmax with non-const num channels")
+            c_sq = GraphBuilder(ctx).make_squeeze({'data': c, 'axes': [0]})
+            c_range = ctx.make_node("Range", [const_zero, c_sq, const_one]).output[0]
+        xyc = ctx.make_node("Add", [xy_scale_c, c_range]).output[0]
+        single_batch = input_shape_guess is not None and input_shape_guess[0] == 1
+        if node.get_attr_value('include_batch_in_index', False) and not single_batch:
+            utils.make_sure(ctx.opset >= 11, "opset 11 required for MaxPoolWithArgmax with include_batch_in_index")
+            n_sq = GraphBuilder(ctx).make_squeeze({'data': n, 'axes': [0]})
+            n_range = ctx.make_node("Range", [const_zero, n_sq, const_one]).output[0]
+            n_range_unsq = GraphBuilder(ctx).make_unsqueeze({'data': n_range, 'axes': [1, 2, 3]})
+            n_range_scale = ctx.make_node("Mul", [n_range_unsq, chw]).output[0]
+            result = ctx.make_node("Add", [xyc, n_range_scale]).output[0]
+        else:
+            result = xyc
+        ctx.replace_all_inputs(node.output[1], result, ops=consumers)
+
         add_padding(ctx, node, kernel_shape, strides)
         conv_convert_inputs(ctx, node, with_kernel=False, input_indices=[0], output_indices=[0, 1])