Create the list of output nodes to test automatically

Populate the list of output node names for each operator as part of the
`add_operator!` calls, instead of separately. This will avoid mistakes where a
new op is added in this test but is never actually tested.

src/model.rs

@@ -573,11 +573,15 @@ mod tests {
         let kernel_val = from_data(vec![1, 1, 1, 1], vec![0.5]);
         let kernel = builder.add_float_constant(&kernel_val);
 
-        let add_operator =
+        // Names of all operator output nodes.
+        let mut op_outputs = Vec::new();
+
+        let mut add_operator =
             |builder: &mut ModelBuilder, name: &str, op: OpType, input_nodes: &[Option<u32>]| {
                 let output_name = format!("{}_out", name);
                 let op_output_node = builder.add_value(&output_name);
                 builder.add_operator(name, op, input_nodes, &[op_output_node]);
+                op_outputs.push(output_name);
                 op_output_node
             };
 
@@ -739,51 +743,25 @@ mod tests {
 
         let model = Model::load(&buffer).unwrap();
 
-        // Outputs of ops tested with a 4D input (eg. NCHW image).
-        let outputs = vec![
-            "Add_out",
-            "AveragePool_out",
-            "BatchNormalization_out",
-            "Cast_out",
-            "Clip_out",
-            "Concat_out",
-            "ConstantOfShape_out",
-            "Conv_out",
-            "ConvTranspose_out",
-            "Cos_out",
-            "Div_out",
-            "Equal_out",
-            "Erf_out",
-            "Expand_out",
-            "Identity_out",
-            "Gather_out",
-            "GlobalAveragePool_out",
-            "LeakyRelu_out",
-            "Less_out",
-            "MaxPool_out",
-            "Mul_out",
-            "Pad_out",
-            "Pow_out",
-            "ReduceMean_out",
-            "Relu_out",
-            "Reshape_out",
-            "Resize_out",
-            "Shape_out",
-            "Sigmoid_out",
-            "Slice_out",
-            "Softmax_out",
-            "Sqrt_out",
-            "Squeeze_out",
-            "Sin_out",
-            "Sub_out",
-            "Tanh_out",
-            "Transpose_out",
-            "Unsqueeze_out",
-        ];
+        // Most ops are tested with a 4D input (eg. NCHW image). A few require
+        // different shapes are tested separately.
         let input = from_data(vec![1, 1, 3, 3], vec![1., 2., 3., 4., 5., 6., 7., 8., 9.]);
+        for output in op_outputs {
+            if [
+                "Gemm_out",
+                "MatMul_out",
+                "Split_out_1",
+                "Split_out_2",
+                "Range_out",
+                "Where_out",
+            ]
+            .contains(&output.as_str())
+            {
+                // This op requires special handling. See below.
+                continue;
+            }
 
-        for output in outputs {
-            let output_id = model.find_node(output).unwrap();
+            let output_id = model.find_node(&output).unwrap();
             let result = model
                 .run(
                     &[(input_node as usize, (&input).into())],