[Bugfix][Relay][keras] Fix the bug about the attribute 'output_paddin…

…g' in Deconv (apache#15060)

* fix deconv about the parameter output_padding

* add test cases to capture the in deconv

* Update keras.py

python/tvm/relay/frontend/keras.py

@@ -421,6 +421,8 @@ def _convert_convolution(inexpr, keras_layer, etab, data_layout, input_shape=Non
         params["groups"] = in_channels
     else:
         params["channels"] = n_filters
+    if is_deconv and keras_layer.output_padding:
+        params["output_padding"] = keras_layer.output_padding
     if keras_layer.padding == "valid":
         pass
     # we insert a separate pad operator
@@ -507,6 +509,8 @@ def _convert_convolution3d(inexpr, keras_layer, etab, data_layout, input_shape=N
         "kernel_layout": kernel_layout,
     }
     params["channels"] = n_filters
+    if is_deconv and keras_layer.output_padding:
+        params["output_padding"] = keras_layer.output_padding
 
     if keras_layer.padding == "valid":
         pass

tests/python/frontend/keras/test_forward.py

@@ -321,6 +321,9 @@ def test_forward_conv_transpose(self, keras_mod):
         data = keras_mod.layers.Input(shape=(32, 32, 128))
         conv_funcs = [
             keras_mod.layers.Conv2DTranspose(filters=64, kernel_size=(2, 2), padding="valid"),
+            keras_mod.layers.Conv2DTranspose(
+                filters=2, kernel_size=(3, 3), strides=(2, 2), output_padding=(1, 1)
+            ),
         ]
         for conv_func in conv_funcs:
             x = conv_func(data)
@@ -619,6 +622,9 @@ def test_forward_conv3d_transpose(self, keras_mod):
                 filters=1, kernel_size=(3, 3, 3), padding="valid", use_bias=False
             ),
             keras_mod.layers.Conv3DTranspose(filters=10, kernel_size=(2, 2, 2), padding="valid"),
+            keras_mod.layers.Conv3DTranspose(
+                filters=2, kernel_size=(3, 3, 3), strides=(2, 2, 2), output_padding=(1, 1, 1)
+            ),
         ]
         for conv_func in conv_funcs:
             x = conv_func(data)