Fix Pad op for dynamic input (opset 11).

Author: jignparm

tf2onnx/onnx_opset/nn.py

@@ -286,7 +286,7 @@ def version_1(cls, ctx, node, **kwargs):
         k_h, k_w, k_input_channels, k_channel_multiplier = kernel_shape
         if k_input_channels < 1:
             raise ValueError("input channel must be positive")
-        k_output_channels = k_input_channels  * k_channel_multiplier
+        k_output_channels = k_input_channels * k_channel_multiplier
 
         node.set_attr("kernel_shape", [k_h, k_w])
         strides = conv_dims_attr(node, "strides")
@@ -448,13 +448,17 @@ def version_11(cls, ctx, node, **kwargs):
         if mode not in [None, "constant", "reflect"]:
             raise ValueError(mode + " pad mode is not supported")
 
-        pads = node.inputs[1].get_tensor_value()
-        pads = np.array(pads).transpose().flatten().astype(np.int64)
-        node.inputs[1].set_tensor_value(pads)
+        # pads must be int64
+        if ctx.get_dtype(node.input[1]) != onnx_pb.TensorProto.INT64:
+            ctx.insert_new_node_on_input(node, "Cast", node.input[1], to=onnx_pb.TensorProto.INT64)
+        ctx.insert_new_node_on_input(node, "Transpose", node.input[1])
+        reshape = ctx.insert_new_node_on_input(node, "Reshape", node.input[1])
+        shape_const = ctx.make_const(utils.make_name(node.name), np.array([-1]).astype(np.int64))
+        reshape.input = [reshape.input[0], shape_const.name]
 
         origin_dtype = ctx.get_dtype(node.output[0])
-        if origin_dtype not in [TensorProto.FLOAT16, TensorProto.FLOAT,
-                                TensorProto.DOUBLE]:
+        if origin_dtype not in [TensorProto.FLOAT, TensorProto.DOUBLE,
+                                TensorProto.INT32, TensorProto.INT64]:
             cast_node = ctx.insert_new_node_on_input(node, "Cast", node.input[0])
             cast_node.set_attr("to", TensorProto.FLOAT)
             ctx.set_dtype(cast_node.output[0], TensorProto.FLOAT)