Fixed bug in depthwise conv when kernel is shared between two nodes (#1384)

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

Author: TomWildenhain-Microsoft

tests/test_backend.py

@@ -620,6 +620,19 @@ def func(x):
         # rtol is a bit high, 2 values have a bit high error. Maybe use different input data.
         self._run_test_case(func, [_OUTPUT], {_INPUT: x_val}, rtol=0.01)
 
+    def test_depthwiseconv_shared_kernel(self):
+        x_shape = [1, 3, 4, 3]
+        kernel_shape = [3, 3, 3, 3]
+        x_val = np.arange(1, 1 + np.prod(x_shape)).astype("float32").reshape(x_shape)
+        kernel_val = np.arange(1, 1 + np.prod(kernel_shape)).astype("float32").reshape(kernel_shape)
+        def func(x, y):
+            kernel = tf.constant(kernel_val, dtype=tf.float32, name='k')
+            conv1 = tf.nn.depthwise_conv2d(x, kernel, strides=[1, 1, 1, 1], padding='VALID')
+            conv2 = tf.nn.depthwise_conv2d(y, kernel, strides=[1, 1, 1, 1], padding='VALID')
+            conv = tf.add(conv1, conv2)
+            return tf.identity(conv, name=_TFOUTPUT)
+        self._run_test_case(func, [_OUTPUT], {_INPUT: x_val, _INPUT1: x_val}, rtol=0.08)
+
     @check_tf_min_version("1.14", "tf depthwise_conv2d dilations")
     @check_opset_min_version(11, "non-const pads")
     def test_depthwiseconv_dilations(self):

tf2onnx/onnx_opset/nn.py

@@ -117,28 +117,22 @@ def conv_convert_inputs(ctx, node, with_kernel=False, new_kernel_shape=None,
     if with_kernel:
         # Some ONNX convolution ops require to reshape the kernel (ie. depthwise_conv2d).
         if new_kernel_shape:
-            if node.inputs[1].is_const():
-                input_node = node.inputs[1]
-                val = input_node.get_tensor_value(as_list=False)
-                val = np.reshape(val, new_kernel_shape)
-                input_node.set_tensor_value(val)
+            kernel_name = node.input[1]
+            if ctx.opset < 5:
+                # Old reshape takes new shape as attribute.
+                reshape = ctx.insert_new_node_on_input(node, "Reshape", kernel_name)
+                reshape.set_attr("shape", new_kernel_shape)
+                reshape.skip_conversion = True
             else:
-                kernel_name = node.input[1]
-                if ctx.opset < 5:
-                    # Old reshape takes new shape as attribute.
-                    reshape = ctx.insert_new_node_on_input(node, "Reshape", kernel_name)
-                    reshape.set_attr("shape", new_kernel_shape)
-                    reshape.skip_conversion = True
-                else:
-                    # New reshape takes new shape as input[1].
-                    shape_name = utils.make_name(node.name)
-                    ctx.make_const(shape_name, np.array(new_kernel_shape, dtype=np.int64))
+                # New reshape takes new shape as input[1].
+                shape_name = utils.make_name(node.name)
+                ctx.make_const(shape_name, np.array(new_kernel_shape, dtype=np.int64))
 
-                    reshape = ctx.make_node("Reshape", [kernel_name, shape_name])
-                    ctx.replace_input(node, kernel_name, reshape.output[0], 1)
+                reshape = ctx.make_node("Reshape", [kernel_name, shape_name])
+                ctx.replace_input(node, kernel_name, reshape.output[0], 1)
 
-                    reshape.skip_conversion = True
-                ctx.set_shape(reshape.output[0], new_kernel_shape)
+                reshape.skip_conversion = True
+            ctx.set_shape(reshape.output[0], new_kernel_shape)
 
         # Get kernel (may have be changed to a reshape above).
         kernel_node = node.inputs[1]