Merge pull request #1060 from xadupre/input2

Use the same syntax to replace an node input (2) + optimize replace_all_inputs

Author: xadupre

tests/test_internals.py

@@ -139,7 +139,7 @@ def test_rewrite_subgraph(self):
             op_name = utils.make_name("ReplacedOp")
             out_name = utils.port_name(op_name)
             new_node = g.make_node("Sub", inputs=input_node.input, outputs=[out_name], name=op_name)
-            g.replace_all_inputs(ops, output_node.output[0], new_node.output[0])
+            g.replace_all_inputs(output_node.output[0], new_node.output[0])  # ops=ops
             for n in set(match.get_nodes()):
                 g.remove_node(n.name)
         g.topological_sort(ops)

tf2onnx/graph.py

@@ -456,6 +456,8 @@ def __init__(self, nodes, output_shapes=None, dtypes=None, target=None, opset=No
         self._nodes = []
         self._nodes_by_name = {}
         self._output_to_node_name = {}
+        self._output_to_consumers = {}
+        self._input_to_graph = {}
         self.shapes = {}
         self.graph_name = graph_name or "tf2onnx"
         self._is_subgraph = is_subgraph
@@ -502,7 +504,7 @@ def __init__(self, nodes, output_shapes=None, dtypes=None, target=None, opset=No
                         body_graph.parent_graph = self
                         new_node.set_body_graph_as_attr(attr_name, body_graph)
 
-                self.replace_all_inputs(self.get_nodes(), o, new_output_name)
+                self.replace_all_inputs(o, new_output_name, ops=self.get_nodes())
                 self.make_node("Identity", [new_output_name], outputs=[o], op_name_scope=n.name + "_" + "graph_outputs")
                 self.copy_shape(new_output_name, o)
                 self.copy_dtype(new_output_name, o)
@@ -607,6 +609,9 @@ def make_node(self, op_type, inputs, attr=None, output_count=1, outputs=None, sk
 
         onnx_node = helper.make_node(op_type, inputs, outputs, name=name, domain=domain, **raw_attr)
 
+        for name2 in onnx_node.input:
+            self._register_input_name(name2, onnx_node)
+
         if op_type in ["If", "Loop", "Scan"]:
             # we force the op containing inner graphs not skipped during conversion.
             skip_conversion = False
@@ -635,6 +640,7 @@ def make_node(self, op_type, inputs, attr=None, output_count=1, outputs=None, sk
         return node
 
     def append_node(self, node):
+        "Add a node to the graph."
         output_shapes = node.output_shapes
         output_dtypes = node.output_dtypes
         node.graph = self
@@ -644,6 +650,8 @@ def append_node(self, node):
             self._output_to_node_name[name] = node.name
             self.set_dtype(name, output_dtypes[i])
             self.set_shape(name, output_shapes[i])
+        for name in node.input:
+            self._register_input_name(name, node)
 
     def remove_node(self, node_name):
         """Remove node in current graph."""
@@ -664,6 +672,12 @@ def remove_node(self, node_name):
             if op_output in self._dtypes:
                 del self._dtypes[op_output]
 
+        for op_input in node.input:
+            utils.make_sure(
+                op_input in self._output_to_consumers,
+                "Input %r of node %r not found.", op_input, node_name)
+            self._unregister_input_name(op_input, node)
+
         self._nodes.remove(node)
         node.graph = None
 
@@ -687,9 +701,13 @@ def reset_nodes(self, ops):
         self.contained_graphs = remained_sub_graphs
         self._nodes_by_name = {op.name: op for op in ops}
         self._output_to_node_name = {}
+        self._output_to_consumers = {}
         for op in ops:
             for op_output in op.output:
                 self._output_to_node_name[op_output] = op.name
+            inps = op.input
+            for op_input in inps:
+                self._register_input_name(op_input, op)
 
         for n in self._order_sensitive_inputs:
             if n not in ops:
@@ -823,6 +841,8 @@ def set_node_by_name(self, node):
         self._nodes_by_name[node.name] = node
         for op_output in node.output:
             self._output_to_node_name[op_output] = node.name
+        for name in node.input:
+            self._register_input_name(name, node)
 
     def change_node_name(self, node, new_name):
         """Remove node in current graph."""
@@ -838,7 +858,7 @@ def change_node_name(self, node, new_name):
                 if k == old_output:
                     self.outputs[j] = new_output
                     break
-            self.replace_all_inputs(self.get_nodes(), old_output, new_output)
+            self.replace_all_inputs(old_output, new_output, ops=self.get_nodes())
         return new_node
 
     def add_graph_input(self, name, dtype=None, shape=None):
@@ -1164,13 +1184,12 @@ def dump_node_statistics(self):
             op_cnt[n.type] += 1
             body_graphs = n.get_body_graphs()
             if body_graphs:
-                for _, b_g in body_graphs.items():
+                for b_g in body_graphs.values():
                     op_cnt += b_g.dump_node_statistics()
 
         return op_cnt
 
-    @staticmethod
-    def remove_input(node, to_be_removed, input_index=None):
+    def remove_input(self, node, to_be_removed, input_index=None):
         """Remove input from Node.
         Args:
             node: the node we expect the input on
@@ -1182,15 +1201,24 @@ def remove_input(node, to_be_removed, input_index=None):
         assert isinstance(node, Node) and isinstance(to_be_removed, six.text_type)
         if input_index is not None:
             assert node.input[input_index] == to_be_removed
+            if node.input[input_index] in self._output_to_consumers:
+                to_ops = self._output_to_consumers[node.input[input_index]]
+                if node.name in to_ops:
+                    to_ops.remove(node.name)
             del node.input[input_index]
-            return True
+            return
 
         for i, name in enumerate(node.input):
             if name == to_be_removed:
+                utils.make_sure(
+                    node.input.count(node.input[i]) <= 1,
+                    "Node %r takes multiple times the same input %r. This case is not handled.",
+                    node.name, node.input[i])
+                self._unregister_input_name(node.input[i], node)
                 del node.input[i]
                 break
+
         # don't remove output from parent since others might depend on it
-        return True
 
     def insert_new_node_on_input(self, node, op_type, input_name, name=None, domain=None, **kwargs):
         """Create and insert a new node into the graph.
@@ -1238,43 +1266,93 @@ def insert_new_node_on_output(self, op_type, output_name, name, domain=None, **k
         new_output = port_name(name)
         new_node = self.make_node(op_type, [output_name], attr=kwargs, outputs=[new_output], name=name, domain=domain)
 
-        to_replace = [n for n in self.get_nodes() if n != new_node]
-        self.replace_all_inputs(to_replace, output_name, new_output)
+        to_replace = [self.get_node_by_name(n) for n in self._output_to_consumers[output_name]]
+        to_replace = [n for n in to_replace if n != new_node]
+        self.replace_all_inputs(output_name, new_output, ops=to_replace)
         return new_node
 
     def find_output_consumers(self, output_name):
         """Find all nodes consuming a given output."""
+        if output_name in self._output_to_consumers:
+            ops = self._output_to_consumers[output_name]
+            ops = [self.get_node_by_name(n) for n in ops]
+        else:
+            ops = []  # self.get_nodes()
         nodes = []
-        for node in self.get_nodes():
+        for node in ops:
+            if node is None:
+                continue
             if output_name in node.input:
                 nodes.append(node)
 
-            # find consumers in sub graphs
-            body_graphs = node.get_body_graphs()
-            if body_graphs:
-                for g in body_graphs.values():
-                    nodes.extend(g.find_output_consumers(output_name))
+        # find consumers in sub graphs
+        if output_name in self._input_to_graph:
+            for g in self._input_to_graph[output_name].values():
+                nodes.extend(g.find_output_consumers(output_name))
         return nodes
 
-    @staticmethod
-    def replace_all_inputs(ops, old_input, new_input):
-        """Replace all inputs pointing to old_input with new_input."""
+    def _register_input_name(self, input_name, node, only_graph=False):
+        "Register node taking a specific input."
+        if not only_graph:
+            if input_name not in self._output_to_consumers:
+                self._output_to_consumers[input_name] = set()
+            self._output_to_consumers[input_name].add(node.name)
+        if self.parent_graph is not None:
+            if input_name not in self.parent_graph._input_to_graph:
+                self.parent_graph._input_to_graph[input_name] = {}
+            self.parent_graph._input_to_graph[input_name][id(self)] = self
+            self.parent_graph._register_input_name(input_name, node, only_graph=True)
+
+    def _unregister_input_name(self, input_name, node, only_graph=False):
+        "Unregister node taking a specific input."
+        node_name = node.name
+        if not only_graph:
+            if input_name in self._output_to_consumers[input_name]:
+                if node_name in self._output_to_consumers[input_name]:
+                    self._output_to_consumers[input_name].remove(node_name)
+        if (self.parent_graph is not None and
+                input_name in self.parent_graph._input_to_graph and
+                id(self) in self.parent_graph._input_to_graph[input_name]):
+            del self.parent_graph._input_to_graph[input_name][id(self)]
+            self.parent_graph._unregister_input_name(input_name, node, only_graph=True)
+
+    def replace_all_inputs(self, old_input, new_input, ops=None):
+        """
+        Replace all inputs pointing to old_input with new_input.
+        *ops* is used if defined, otherwise `_output_to_consumers`
+        is used to determine the impacted nodes.
+        """
         if old_input == new_input:
             return
+        if new_input not in self._output_to_consumers:
+            self._output_to_consumers[new_input] = set()
+
+        if ops is not None:
+            keep_ops = True
+        elif old_input in self._output_to_consumers:
+            ops = list(
+                filter(lambda a: a is not None,
+                       map(self.get_node_by_name, self._output_to_consumers[old_input])))
+            keep_ops = False
+        else:
+            ops = []
+            keep_ops = False
 
         for node in ops:
+            assert node is not None
             if old_input in node.input and new_input in node.output:
                 raise RuntimeError("creating a circle in the graph is not allowed: " + node.name)
+            self._register_input_name(new_input, node)
 
             for i, input_name in enumerate(node.input):
                 if input_name == old_input:
-                    node.input[i] = new_input
+                    self.replace_input(node, node.input[i], new_input, i)
 
-            # modify references in sub graphs
-            body_graphs = node.get_body_graphs()
-            if body_graphs:
-                for g in body_graphs.values():
-                    g.replace_all_inputs(g.get_nodes(), old_input, new_input)
+        # modify references in sub graphs
+        if old_input in self._input_to_graph:
+            for g in self._input_to_graph[old_input].values():
+                g.replace_all_inputs(old_input, new_input,
+                                     ops=g.get_nodes() if keep_ops else None)
 
     def replace_input(self, node, old_input, new_input, input_index=None):
         """
@@ -1294,11 +1372,31 @@ def replace_input(self, node, old_input, new_input, input_index=None):
             is_replaced = True
         else:
             raise RuntimeError("Unable to replace input %r into %r for node %r." % (old_input, new_input, node.name))
+
+        to_ops = self._output_to_consumers.get(old_input, None)
+        if to_ops is not None:
+            if node.name in to_ops:
+                # A node may take twice the same entry.
+                to_ops.remove(node.name)
+
+        self._register_input_name(new_input, node)
         return is_replaced
 
     def replace_inputs(self, node, new_inputs):
         """Replace node inputs."""
         assert isinstance(node, Node) and isinstance(new_inputs, list)
+
+        for old_input in node.input:
+            to_ops = self._output_to_consumers.get(old_input, None)
+            if to_ops is not None and old_input in to_ops:
+                # To avoid issues when a node
+                # takes twice the same entry.
+                to_ops.remove(old_input)
+
+        for input_name in new_inputs:
+            assert isinstance(input_name, six.text_type)
+            self._register_input_name(input_name, node)
+
         node.input = new_inputs
         return True
 
@@ -1374,7 +1472,7 @@ def delete_unused_nodes(self, outputs_name):
         for node in related_nodes:
             attr_body_graphs = node.get_body_graphs()
             if attr_body_graphs:
-                for _, body_graph in attr_body_graphs.items():
+                for body_graph in attr_body_graphs.values():
                     body_graph.delete_unused_nodes(body_graph.outputs)
         self.reset_nodes(related_nodes)
 

tf2onnx/onnx_opset/controlflow.py

@@ -492,7 +492,7 @@ class TensorListStack:
     def version_7(cls, ctx, node, **kwargs):
         if node.inputs[0].is_while():
             ctx.remove_node(node.name)
-            ctx.replace_all_inputs(ctx.get_nodes(), node.output[0], node.input[0])
+            ctx.replace_all_inputs(node.output[0], node.input[0])  # ops=ctx.get_nodes()
 
 
 @tf_op(["While", "StatelessWhile"])
@@ -582,7 +582,7 @@ def version_7(cls, ctx, node, **kwargs):
         for idx, n in reversed(to_remove):
             ctx.remove_node(n.name)
             # make the node output bad
-            ctx.replace_all_inputs(ctx.get_nodes(), n.output[0], "@@ALLOC")
+            ctx.replace_all_inputs(n.output[0], "@@ALLOC")  # ops=ctx.get_nodes()
             del body.func_inputs[idx]
             del cond_graph.func_inputs[idx]
             del tf_while_inputs[idx]
@@ -618,7 +618,7 @@ def version_7(cls, ctx, node, **kwargs):
 
         # shift output consumers
         for k, v in output_map.items():
-            ctx.replace_all_inputs(ctx.get_nodes(), k, v)
+            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)
@@ -813,7 +813,7 @@ def prefix_graph(g, scope):
                 if old_output == oname:
                     g.outputs[i] = new_output
                     break
-            g.replace_all_inputs(ops, old_output, new_output)
+            g.replace_all_inputs(old_output, new_output, ops=ops)
         to_remove.append(node)
     for node in to_remove:
         g.remove_node(node.name)

tf2onnx/onnx_opset/math.py

@@ -695,4 +695,4 @@ def atan2(y, x):
             "Add", inputs=[atan_node.output[0], pi_part.output[0]],
             op_name_scope=node.name + 'all',
             shapes=[shape], dtypes=[onnx_dtype])
-        ctx.replace_all_inputs(ctx.get_nodes(), node.output[0], last_node.output[0])
+        ctx.replace_all_inputs(node.output[0], last_node.output[0])  # ops=ctx.get_nodes()

tf2onnx/onnx_opset/misc.py

@@ -30,7 +30,7 @@ def version_1(cls, ctx, node, **kwargs):
             # if identity has a const as input, remove it
             input_name = node.input[0]
             output_name = node.output[0]
-            ctx.replace_all_inputs(ctx.get_nodes(), output_name, input_name)
+            ctx.replace_all_inputs(output_name, input_name)  # ops=ctx.get_nodes()
             ctx.remove_node(node.name)
 
 

tf2onnx/onnx_opset/nn.py

@@ -451,7 +451,7 @@ def version_1(cls, ctx, node, **kwargs):
             downstream_nodes = ctx.find_output_consumers(node.output[0])
             downstream_nodes.remove(output_shape)
             downstream_nodes.remove(slice_node)
-            ctx.replace_all_inputs(downstream_nodes, node.output[0], slice_node.output[0])
+            ctx.replace_all_inputs(node.output[0], slice_node.output[0], ops=downstream_nodes)
 
         conv_dims_attr(node, "strides", spatial=spatial)
         conv_dims_attr(node, "dilations", spatial=spatial)

tf2onnx/onnx_opset/quantize.py

@@ -78,4 +78,4 @@ def version_10(cls, ctx, node, **kwargs):
             "DequantizeLinear", [new_node.output[0], pb_scale.name, zero_point.name],
             op_name_scope=node.name, attr={"axis": axis},
             shapes=[shape], dtypes=[dtype])
-        ctx.replace_all_inputs(ctx.get_nodes(), node.output[0], last_node.output[0])
+        ctx.replace_all_inputs(node.output[0], last_node.output[0])  # ops=ctx.get_nodes()

tf2onnx/onnx_opset/rnn.py

@@ -153,7 +153,7 @@ def make_sigmoid(i, w, b):
         h_node = ctx.make_node("Mul", [co_node.output[0], o])
 
         def replace_output(old_output, new_output):
-            ctx.replace_all_inputs(ctx.get_nodes(), old_output, new_output)
+            ctx.replace_all_inputs(old_output, new_output)  # ops=ctx.get_nodes()
             ctx.copy_dtype(old_output, new_output)
             ctx.copy_shape(old_output, new_output)
 

tf2onnx/onnx_opset/tensor.py

@@ -115,7 +115,7 @@ def version_1(cls, ctx, node, **kwargs):
             # if identity has a const as input, remove it
             input_name = node.input[0]
             output_name = node.output[0]
-            ctx.replace_all_inputs(ctx.get_nodes(), output_name, input_name)
+            ctx.replace_all_inputs(output_name, input_name)  # ops=ctx.get_nodes()
             ctx.remove_node(node.name)
 
 
@@ -125,7 +125,7 @@ class IdentityN:
     def version_1(cls, ctx, node, **kwargs):
         ctx.remove_node(node.name)
         for input_name, output_name in zip(node.input, node.output):
-            ctx.replace_all_inputs(ctx.get_nodes(), output_name, input_name)
+            ctx.replace_all_inputs(output_name, input_name)  # ops=ctx.get_nodes()
 
 
 @tf_op("Reshape")
@@ -1050,7 +1050,7 @@ def version_1(cls, ctx, node, **kwargs):
         # concat all unqueezes
         concat = ctx.make_node("Concat", inputs, op_name_scope=node.name, attr={"axis": axis},
                                shapes=shapes, dtypes=dtypes)
-        ctx.replace_all_inputs(ctx.get_nodes(), node.output[0], concat.output[0])
+        ctx.replace_all_inputs(node.output[0], concat.output[0])  # ops=ctx.get_nodes()
 
 
 @tf_op("Unpack")