"Where" Op Is Not Implemented

[GuanLuo]

Hi all, I created a model containing "Where" operator but failed to load it via Onnx Runtime C API:

Could not find an implementation for the node Where(9)

However, the version matrix says it supports Onnx opset version 10.

The Python script to generate the model:

import onnx
onnx_input_str = onnx.helper.make_tensor_value_info("INPUT0", onnx.TensorProto.STRING, [1])
onnx_input_int = onnx.helper.make_tensor_value_info("INPUT1", onnx.TensorProto.INT32, [1])
onnx_flag = onnx.helper.make_tensor_value_info("READY", onnx.TensorProto.INT32, [1])
onnx_output = onnx.helper.make_tensor_value_info("OUTPUT", onnx.TensorProto.STRING, [1])

internal_input = onnx.helper.make_node("Cast", ["INPUT0"], ["_INPUT"], to=onnx.TensorProto.INT32)
add = onnx.helper.make_node("Add", ["_INPUT", "INPUT1"], ["add"])
zeros = onnx.helper.make_node("Sub", ["READY", "READY"], ["zeros"])
equal = onnx.helper.make_node("Equal", ["READY", "zeros"], ["equal"])
where = onnx.helper.make_node("Where", ["equal", "zeros", "add"], ["CAST"])
cast = onnx.helper.make_node("Cast", ["CAST"], ["OUTPUT"], to=onnx.TensorProto.STRING)

onnx_nodes = [internal_input, add, zeros, equal, where, cast]
onnx_inputs = [onnx_input_str, onnx_input_int, onnx_flag]
onnx_outputs = [onnx_output]

graph_proto = onnx.helper.make_graph(onnx_nodes, "conditional_and_cast", onnx_inputs, onnx_outputs)
model_def = onnx.helper.make_model(graph_proto)
onnx.save(model_def, "model.onnx")

[skottmckay]

Could not find an implementation for the node Where(9)

This suggests you model has opset set to 9. What version of ONNX are you using?

[GuanLuo]

@skottmckay The onnx version is '1.4.1', and it is using opset 9 as you said (versioning doc). So it is not backward capatible?

[skottmckay]

Sorry - I thought Where was new in opset 10. What version of onnxruntime are you using as our implementation of Where was added a few months ago (and is valid for opset 9).

https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/tensor/where_op.cc

[GuanLuo]

I am using Onnx Runtime 0.4.0 built from source. Now I looked at the where_op.cc at branch rel-0.4.0 and I noticed that the types int32_t and bool is commented out, so I suppose that "where" op is implemented in Onnx Runtime but it is limited to certain data types?

[skottmckay]

The current implementation does limit the supported types, I believe to try and keep the overall binary size small. Support for int32_t was enabled last week though, so if you pulled the latest master it would be in that.

31cbb5d#diff-3276925863124d4b46cb2f4e29e49b9f

[GuanLuo]

I see. I wonder if there is a option / mode that I can enable to build the library with full Onnx support? Right now it seems to be inconvenient as I need to do something similar to 31cbb5d and to rebuild to support some data types.

[GuanLuo]

Opened a new issue #1122 for the feature requested in followup comment.