[Bugfix][Relay][Keras] Fix SeparableConv2D conversion in dilation_rat…

…e attribute (apache#15122)

* Update keras.py

fix the _convert_separable_convolution. The dilation_rate always be the default value (e.g., [1,1]).

* add new test cases to capture bug in separableConv2d

* Update test_forward.py

python/tvm/relay/frontend/keras.py

@@ -569,13 +569,17 @@ def _convert_separable_convolution(inexpr, keras_layer, etab, data_layout, input
         weight0 = weightList[0].transpose([2, 3, 0, 1])
     else:
         weight0 = weightList[0]
+    if isinstance(keras_layer.dilation_rate, (list, tuple)):
+        dilation = [keras_layer.dilation_rate[0], keras_layer.dilation_rate[1]]
+    else:
+        dilation = [keras_layer.dilation_rate, keras_layer.dilation_rate]
     params0 = {
         "weight": etab.new_const(weight0),
         "channels": in_channels * depth_mult,
         "groups": in_channels,
         "kernel_size": [kernel_h, kernel_w],
         "strides": [stride_h, stride_w],
-        "dilation": [1, 1],
+        "dilation": dilation,
         "padding": [0, 0],
         "data_layout": data_layout,
         "kernel_layout": kernel_layout,

tests/python/frontend/keras/test_forward.py

@@ -310,6 +310,8 @@ def test_forward_conv(self, keras_mod):
             keras_mod.layers.DepthwiseConv2D(kernel_size=(3, 3), padding="same"),
             keras_mod.layers.Conv2DTranspose(filters=10, kernel_size=(3, 3), padding="valid"),
             keras_mod.layers.SeparableConv2D(filters=10, kernel_size=(3, 3), padding="same"),
+            keras_mod.layers.SeparableConv2D(filters=10, kernel_size=(3, 3), dilation_rate=(2, 2)),
+            keras_mod.layers.SeparableConv2D(filters=2, kernel_size=(3, 3), dilation_rate=2),
         ]
         for conv_func in conv_funcs:
             x = conv_func(data)