Fixed group attribute in convolution op (#2090)

* Fixed group attribute in convolution op. Also minor change reading shape dimensions so it works for different dimensionalities.
* Check convolution kernel_shape values are valid
Skip adding the optional attribute to Conv nodes if any shape value is negative.
* Fixed convolution kernel dimension checks

Signed-off-by: Javier Dehesa <[email protected]>


Signed-off-by: Javier Dehesa <[email protected]>
Co-authored-by: Jay Zhang <[email protected]>

Author: javidcf

tests/test_backend.py

@@ -6138,6 +6138,20 @@ def func(x):
         x_val = make_xval([2, 3])
         self._run_test_case(func, [_OUTPUT], {_INPUT: x_val})
 
+    @check_opset_min_version(11, "Pad")
+    def test_conv_unknown_kernel_channels(self):
+        x_shape = [2, 10, 3]
+        x_val = make_xval(x_shape)
+        kernel_shape = [4, 3, 5]
+        kernel_val = make_xval(kernel_shape)
+        pad_val = np.array([[0, 0], [0, 0], [0, 0]], np.int64)
+        def func(x, kernel, pad):
+            # Make kernel dimensions unknown
+            kernel = tf.pad(kernel, pad)
+            conv = tf.nn.conv1d(x, kernel, stride=[1], padding='VALID')
+            return tf.identity(conv, name='output')
+        self._run_test_case(func, [_OUTPUT], {_INPUT: x_val, _INPUT1: kernel_val, _INPUT2: pad_val})
+
     @check_tf_min_version("2.3.0")
     @check_opset_min_version(16, "ScatterND")
     @skip_tfjs("not supported in tfjs")

tf2onnx/onnx_opset/nn.py

@@ -311,8 +311,9 @@ def conv_kernel_shape(ctx, node, input_idx, spatial=2):
     # Get spatial part.
     kernel_shape = kernel_shape[:spatial]
 
-    # Set new value and return it.
-    node.set_attr("kernel_shape", kernel_shape)
+    # Set attribute value only if all dimensions are known.
+    if all(d > 0 for d in kernel_shape):
+        node.set_attr("kernel_shape", kernel_shape)
 
     return kernel_shape
 
@@ -379,11 +380,13 @@ def any_version(cls, opset, ctx, node, **kwargs):
         data_format = str(node.attr["data_format"].s, encoding="utf8")
         shape_dim = -1
         if data_format == "NHWC":
-            shape_dim = ctx.get_shape(node.input[0])[3]
+            shape_dim = ctx.get_shape(node.input[0])[-1]
         elif data_format == "NCHW":
             shape_dim = ctx.get_shape(node.input[0])[1]
         if shape_dim != -1:
-            groups = int(shape_dim / ctx.get_shape(node.input[1])[2])
+            filter_in_channels = ctx.get_shape(node.input[1])[-2]
+            if filter_in_channels != -1:
+                groups = shape_dim // filter_in_channels
 
         node.set_attr("group", groups)
 
@@ -649,7 +652,8 @@ def version_1(cls, ctx, node, **kwargs):
             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])
+        if k_h > 0 and k_w > 0:
+            node.set_attr("kernel_shape", [k_h, k_w])
         strides = conv_dims_attr(node, "strides")
         dilations = conv_dims_attr(node, "dilations")
         node.set_attr("group", k_input_channels)