ONNX export: Fully connected operator with no bias

python/mxnet/contrib/onnx/mx2onnx/_op_translations.py

@@ -222,17 +222,45 @@ def convert_fully_connected(node, **kwargs):
     onnx = import_onnx_modules()
     name = node["name"]
     inputs = node["inputs"]
+    attrs = node["attrs"]
+    initializer = kwargs["initializer"]
+
+    no_bias = convert_bool_to_int(attrs, "no_bias")
+
     input_node_id = kwargs["index_lookup"][inputs[0][0]]
     weight_node_id = kwargs["index_lookup"][inputs[1][0]]
-    bias_node_id = kwargs["index_lookup"][inputs[2][0]]
+
     proc_nodes = kwargs["proc_nodes"]
-    input_node = proc_nodes[input_node_id]
-    weights_node = proc_nodes[weight_node_id]
-    bias_node = proc_nodes[bias_node_id]
 
+    input_node = proc_nodes[input_node_id]
     input_name = input_node.name
+
+    weights_node = proc_nodes[weight_node_id]
     weights_name = weights_node.name
-    bias_name = bias_node.name
+
+    fcnode = []
+
+    if no_bias == 0:
+        bias_node_id = kwargs["index_lookup"][inputs[2][0]]
+        bias_node = proc_nodes[bias_node_id]
+        bias_name = bias_node.name
+    else:
+        default_bias = [0]
+        np_arr = np.array(default_bias)
+        data_type = onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[np_arr.dtype]
+        dims = np.shape(np_arr)
+        bias_name = "bias" + str(kwargs["idx"])
+        tensor_node = onnx.helper.make_tensor_value_info(bias_name, data_type, dims)
+        initializer.append(
+            onnx.helper.make_tensor(
+                name=bias_name,
+                data_type=data_type,
+                dims=dims,
+                vals=default_bias,
+                raw=False,
+            )
+        )
+        fcnode.append(tensor_node)
 
     node = onnx.helper.make_node(
         "Gemm",
@@ -245,7 +273,9 @@ def convert_fully_connected(node, **kwargs):
         name=name
     )
 
-    return [node]
+    fcnode.append(node)
+
+    return fcnode
 
 
 @mx_op.register("BatchNorm")