Merge pull request #741 from RandySheriffH/rashuai/FixDepthwiseConv

Rashuai/fix depthwise conv

Author: guschmue

tests/test_backend.py

@@ -431,8 +431,7 @@ def test_depthwiseconv_0(self):
         x = tf.placeholder(tf.float32, shape=x_val.shape, name=_TFINPUT)
         conv = tf.nn.depthwise_conv2d(x, kernel, strides=[1, 1, 1, 1], padding='VALID')
         _ = tf.identity(conv, name=_TFOUTPUT)
-        # rtol is a bit high, 2 values have a bit high error. Maybe use different input data.
-        self._run_test_case([_OUTPUT], {_INPUT: x_val}, rtol=0.08)
+        self._run_test_case([_OUTPUT], {_INPUT: x_val}, rtol=1e-6)
 
     def test_depthwiseconv_1(self):
         x_shape = [1, 112, 112, 32]
@@ -443,8 +442,7 @@ def test_depthwiseconv_1(self):
         x = tf.placeholder(tf.float32, shape=x_val.shape, name=_TFINPUT)
         conv = tf.nn.depthwise_conv2d(x, kernel, strides=_STRIDE1x1, padding='VALID')
         _ = tf.identity(conv, name=_TFOUTPUT)
-        # rtol is a bit high, 2 values have a bit high error. Maybe use different input data.
-        self._run_test_case([_OUTPUT], {_INPUT: x_val}, rtol=0.08)
+        self._run_test_case([_OUTPUT], {_INPUT: x_val}, rtol=1e-6)
 
     def test_dropout(self):
         is_training = tf.placeholder_with_default(False, (), "is_training")

tf2onnx/onnx_opset/nn.py

@@ -72,6 +72,24 @@ def conv_convert_inputs(ctx, node, with_kernel=False, new_kernel_shape=None,
 
     # kernel must to be transposed
     if with_kernel:
+        # some onnx conv ops require the reshape the kernel (ie. depthwise_conv2d)
+        if new_kernel_shape:
+            if ctx.opset < 5:
+                # old reshape takes new shape as attribute
+                input_name = node.input[1]
+                reshape = ctx.insert_new_node_on_input(node, "Reshape", input_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))
+                input_name = node.input[1]
+                reshape = ctx.make_node("Reshape", [input_name, shape_name])
+                ctx.replace_input(node, input_name, reshape.output[0])
+                reshape.skip_conversion = True
+            ctx.set_shape(reshape.output[0], new_kernel_shape)
+
         parent = node.inputs[1]
         need_transpose = True
         if node.inputs[1].is_const():
@@ -91,24 +109,6 @@ def conv_convert_inputs(ctx, node, with_kernel=False, new_kernel_shape=None,
             new_shape = spatial_map(ctx.get_shape(input_name), constants.HWCN_TO_NCHW)
             ctx.set_shape(transpose.output[0], new_shape)
 
-        # some onnx conv ops require the reshape the kernel (ie. depthwise_conv2d)
-        if new_kernel_shape:
-            if ctx.opset < 5:
-                # old reshape takes new shape as attribute
-                input_name = node.input[1]
-                reshape = ctx.insert_new_node_on_input(node, "Reshape", input_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))
-                input_name = node.input[1]
-                reshape = ctx.make_node("Reshape", [input_name, shape_name])
-                ctx.replace_input(node, input_name, reshape.output[0])
-                reshape.skip_conversion = True
-            ctx.set_shape(reshape.output[0], new_kernel_shape)
-
     # transpose outputs if needed
     if node.is_nhwc():
         for idx in output_indices:
@@ -280,24 +280,21 @@ def version_1(cls, ctx, node, **kwargs):
         if len(input_shape) != 4:
             raise ValueError("only Conv2D is supported")
 
-        if node.is_nhwc():
-            i_n, i_h, i_w, i_c = input_shape
-        else:
-            i_n, i_c, i_h, i_w = input_shape
-
         kernel_shape = ctx.get_shape(node.input[1])
         if len(kernel_shape) != 4:
             raise ValueError("only Conv2D is supported")
         k_h, k_w, k_input_channels, k_channel_multiplier = kernel_shape
-        k_output_channels = i_c * k_channel_multiplier
+        if k_input_channels < 1:
+            raise ValueError("input channel must be positive")
+        k_output_channels = k_input_channels  * k_channel_multiplier
 
         node.set_attr("kernel_shape", [k_h, k_w])
         strides = conv_dims_attr(node, "strides")
         conv_dims_attr(node, "dilations")
-        node.set_attr("group", i_c)
+        node.set_attr("group", k_input_channels)
         add_padding(ctx, node, kernel_shape, strides)
 
-        new_kernel_shape = [k_output_channels, 1, k_h, k_w]
+        new_kernel_shape = [k_h, k_w, 1, k_output_channels]
         conv_convert_inputs(ctx, node, with_kernel=True, new_kernel_shape=new_kernel_shape)