Merge pull request #1238 from onnx/tom/multiscanout

Add support for multiple scan outputs

tests/test_loops.py

@@ -188,6 +188,40 @@ def b(i, out_ta):
         output_names_with_port = ["i:0", "output_ta:0"]
         self.run_test_case(func, feed_dict, input_names_with_port, output_names_with_port, rtol=1e-06)
 
+    def test_while_loop_with_multi_scan_outputs(self):
+        def func(i, inputs1, inputs2):
+            inputs1_ = tf.identity(inputs1)
+            inputs2_ = tf.identity(inputs2)
+            input_ta = tf.TensorArray(dtype=tf.float32, size=0, dynamic_size=True).unstack(inputs1_)
+            input_ta2 = tf.TensorArray(dtype=tf.float32, size=0, dynamic_size=True).unstack(inputs2_)
+            output_ta = tf.TensorArray(dtype=tf.float32, size=0, dynamic_size=True)
+            output_ta2 = tf.TensorArray(dtype=tf.float32, size=0, dynamic_size=True)
+
+            c = lambda i, *_: tf.logical_and(tf.less(i, 10), i >= 0)
+
+            def b(i, out_ta, out_ta2):
+                new_i = tf.add(i, 1)
+                x = input_ta.read(i)
+                y = input_ta2.read(i)
+                z = x + 3 + y
+                p = x * y * 2
+                out_ta_new = out_ta.write(i, z)
+                out_ta_new2 = out_ta2.write(i, p)
+                return new_i, out_ta_new, out_ta_new2
+
+            i_final, out_final, out_final2 = tf.while_loop(c, b, [i, output_ta, output_ta2])
+            i_final_ = tf.identity(i_final, name="i")
+            out_final_ = tf.identity(out_final.stack(), name="output_ta")
+            out_final2_ = tf.identity(out_final2.stack(), name="output_ta2")
+            return i_final_, out_final_, out_final2_
+
+        input_names_with_port = ["input_1:0", "input_2:0", "input_3:0"]
+        feed_dict = {"input_1:0": np.array(0, dtype=np.int32),
+                     "input_2:0": np.array([2.0, 16.0, 5.0, 1.6, 5.0, 6.0, 7.0, 8.0, 9.0, 10.], dtype=np.float32),
+                     "input_3:0": np.array([1.0, 2.0, 3.0, 4.0, 5.0, 16.0, 7.0, 8.0, 9.0, 10.], dtype=np.float32)}
+        output_names_with_port = ["i:0", "output_ta:0", "output_ta2:0"]
+        self.run_test_case(func, feed_dict, input_names_with_port, output_names_with_port, rtol=1e-06)
+
     @check_onnxruntime_min_version(
         "0.5.0",
         "disable this case due to onnxruntime loop issue: https://github.com/microsoft/onnxruntime/issues/1272"

tf2onnx/onnx_opset/controlflow.py

@@ -421,7 +421,7 @@ def version_7(cls, ctx, node, **kwargs):
             del output_names[idx]
             del body.outputs[idx]
 
-        removed_scan_outputs = {}
+        scan_output_names = []
         # remove tensor array that are passed in to the loop
         for idx, n in reversed(to_remove):
             ctx.remove_node(n.name)
@@ -430,19 +430,15 @@ def version_7(cls, ctx, node, **kwargs):
             del body.func_inputs[idx]
             del cond_graph.func_inputs[idx]
             del tf_while_inputs[idx]
-            # save the index of the scan output
-            removed_scan_outputs[body.outputs[idx]] = idx
+            scan_output_names.append(body.outputs[idx])
             del body.outputs[idx]
-            # FIXME: Output shapes may be in wrong order if there are multiple scan outputs
             output_shapes.append(output_shapes[idx])
             output_dtypes.append(output_dtypes[idx])
             output_names.append(output_names[idx])
             del output_shapes[idx]
             del output_dtypes[idx]
             del output_names[idx]
 
-        utils.make_sure(len(removed_scan_outputs) <= 1, "converter only supports while loops with a single scan output")
-
         ctx.remove_node(node.name)
 
         # In onnx 'cond' is a variable, not a function. We need to inject the subgraph into the main graph
@@ -467,7 +463,7 @@ def version_7(cls, ctx, node, **kwargs):
             ctx.replace_all_inputs(k, v)  # ops=ctx.get_nodes()
 
         wire_while_body(ctx, body, loop_node.inputs, body_input_to_state_var, cond_input_to_state_var, output_shapes,
-                        output_dtypes, body_name, node.name, cond_graph, tf_while_inputs, removed_scan_outputs)
+                        output_dtypes, body_name, node.name, cond_graph, tf_while_inputs, scan_output_names)
 
         # if there was a tensorflow variant type, bind in a real type here
         # FIXME: I don't think this is needed anymore
@@ -477,7 +473,7 @@ def version_7(cls, ctx, node, **kwargs):
 
 
 def wire_while_body(parent_g, g, loop_node_inputs, body_input_to_state_var, cond_input_to_state_var, output_shapes,
-                    output_dtypes, scope, parent, cond_graph, tf_while_inputs, removed_scan_outputs):
+                    output_dtypes, scope, parent, cond_graph, tf_while_inputs, scan_output_names):
     """Wire subgraph graph into main."""
     remove_parents = []
     to_remove = []
@@ -521,9 +517,10 @@ def wire_while_body(parent_g, g, loop_node_inputs, body_input_to_state_var, cond
             g.replace_inputs(node, [node.input[2]])
             scan_outputs.append(node.output[0])
 
-    if len(scan_outputs) != len(removed_scan_outputs):
+    if len(scan_outputs) != len(scan_output_names):
         raise ValueError("While loop couldn't find scan output index for nodes")
 
+    names_to_scan_outputs = {}
     for output in scan_outputs:
         last_output = output
         consumers = g.find_output_consumers(last_output)
@@ -533,10 +530,12 @@ def wire_while_body(parent_g, g, loop_node_inputs, body_input_to_state_var, cond
                 raise ValueError("While loop couldn't find scan output index for node " + node.name)
             last_output = node.output[0]
             consumers = g.find_output_consumers(last_output)
-        if last_output not in removed_scan_outputs:
+        if last_output not in scan_output_names:
             raise ValueError("While loop couldn't find scan output index for node " + node.name)
-        # TODO: store index to ensure scan outputs are in correct order for multiple outputs
-        # initial_output_index = removed_scan_outputs[last_output]
+        names_to_scan_outputs[last_output] = output
+
+    # Reorder scan outputs
+    scan_outputs = [names_to_scan_outputs[name] for name in scan_output_names]
 
     # remove all nodes feeding to TensorListSetItem's reserved tensor
     while remove_parents: