Added optimization for dequantized constant folding (#1316)

Signed-off-by: Tom Wildenhain <tomwi@microsoft.com>

tests/test_optimizers.py

@@ -9,7 +9,7 @@
 
 import unittest
 import numpy as np
-from onnx import helper, TensorProto, OperatorSetIdProto
+from onnx import helper, numpy_helper, TensorProto, OperatorSetIdProto
 from backend_test_base import Tf2OnnxBackendTestBase
 from common import unittest_main, group_nodes_by_type, check_opset_min_version, check_opset_max_version, get_test_config
 from tf2onnx import utils, constants
@@ -1241,6 +1241,30 @@ def test_const_fold_cast_with_const(self):
 
     # Const Fold Optimizer Tests End
 
+    # Const Dequantize Optimizer Tests Start
+
+    @check_opset_min_version(10, "DequantizeLinear")
+    def test_const_dequantize_reshape(self):
+        inputval = numpy_helper.from_array(np.random.randint(0, 100, (2, 3, 4, 5), np.uint8), name='X')
+        scale = numpy_helper.from_array(np.array(0.75, dtype=np.float32), name='scale')
+        zero_point = numpy_helper.from_array(np.array(3, dtype=np.uint8), name='zero_point')
+        shape = numpy_helper.from_array(np.array([6, 20], dtype=np.int64), name='shape')
+        node1 = helper.make_node("DequantizeLinear", ["X", "scale", "zero_point"], ["Y"], name="dequantize")
+        node2 = helper.make_node("Reshape", ["Y", "shape"], ["Z"], name="reshape")
+
+        graph = helper.make_graph(
+            [node1, node2],
+            "const-dequantize-test",
+            [],
+            [helper.make_tensor_value_info("Z", TensorProto.FLOAT, (6, 20))],
+            [inputval, scale, zero_point, shape]
+        )
+
+        model_proto = self.make_model(graph, producer_name="onnx-tests")
+        self.run_and_compare(["Z"], {}, model_proto, "Reshape", 0)
+
+    # Const Dequantize Optimizer Tests End
+
     def test_transpose_back_to_back_non_const(self):
 
         node0 = helper.make_node("Transpose", ["u"], ["v"], perm=[0, 2, 3, 1], name="trans_0")

tf2onnx/optimizer/__init__.py

@@ -16,13 +16,15 @@
 from .loop_optimizer import LoopOptimizer
 from .back_to_back_optimizer import BackToBackOptimizer
 from .upsample_optimizer import UpsampleOptimizer
+from .const_dequantize_optimizer import ConstDequantizeOptimizer
 from .. import logging
 
 # optimizer sequence need to be considered carefully
 _optimizers = OrderedDict([
     ("optimize_transpose", TransposeOptimizer),
     ("remove_redundant_upsample", UpsampleOptimizer),
     ("fold_constants", ConstFoldOptimizer),
+    ("const_dequantize_optimizer", ConstDequantizeOptimizer),
     ("loop_optimizer", LoopOptimizer),
     # merge_duplication should be used after optimize_transpose
     # for optimize_transpose may have some trans nodes that can be merge

tf2onnx/optimizer/const_dequantize_optimizer.py

@@ -0,0 +1,67 @@
+# SPDX-License-Identifier: Apache-2.0
+
+
+"""const dequantize Optimizer.
+   if a dequantize op's inputs are const we may be able to fold it through the next op
+"""
+
+from .optimizer_base import GraphOptimizerBase
+from .const_fold_optimizer import ConstFoldOptimizer
+
+# pylint: disable=logging-not-lazy,unused-argument,missing-docstring
+
+
+class ConstDequantizeOptimizer(GraphOptimizerBase):
+
+    def __init__(self):  # pylint: disable=useless-super-delegation
+        super(ConstDequantizeOptimizer, self).__init__()
+
+    def _optimize(self, graph):
+        return self._apply_optimization(graph, self._optimize_at_current_graph_level)
+
+    def _optimize_at_current_graph_level(self, graph):
+        graph_changed = True
+        while graph_changed:
+            graph_changed = False
+            ops = graph.get_nodes()
+            for op in ops:
+                if self._fold_node(op, graph):
+                    graph_changed = True
+                    self.graph_been_opt = True
+        return graph
+
+    def _fold_node(self, node, graph):
+        """ if a dequantize op's inputs are const and it is fed into a tensor reshaping op, we can apply the op
+            directly to the quantized inputs.  Returns True if the graph is changed.
+        """
+        if node.type not in ["Transpose", "Reshape", "Unsqueeze"]:
+            return False
+        dequant_node = node.inputs[0]
+        if dequant_node.type != "DequantizeLinear":
+            return False
+        if len(graph.find_output_consumers(dequant_node.output[0])) > 1:
+            return False
+        if not self._all_inputs_are_const(node.inputs[1:]) or self._is_graph_output(node, graph):
+            return False
+        if not self._all_inputs_are_const(dequant_node.inputs):
+            return False
+        graph.replace_input(node, node.input[0], dequant_node.input[0], 0)
+        const_outputs = ConstFoldOptimizer.compute_const_folding(node, graph)
+        graph.replace_all_inputs(node.output[0], dequant_node.output[0])
+        graph.remove_node(node.name)
+        dequant_const = dequant_node.inputs[0]
+        if len(graph.find_output_consumers(dequant_const.output[0])) > 1:
+            dequant_const = graph.copy_const(dequant_const)
+            graph.replace_input(dequant_node, dequant_node.input[0], dequant_const.output[0], 0)
+        dequant_const.set_tensor_value(const_outputs[0])
+        return True
+
+    @staticmethod
+    def _all_inputs_are_const(nodes):
+        return all(node.is_const() for node in nodes if node)
+
+    @staticmethod
+    def _is_graph_output(node, graph):
+        node_out_set = set(node.output)
+        graph_out_set = set(graph.outputs)
+        return node_out_set.intersection(graph_out_set)

tf2onnx/optimizer/const_fold_optimizer.py

@@ -54,7 +54,7 @@ def _should_skip(node):
         if node.is_const() or node.is_graph_input():
             return True
 
-        skip_type = ["Identity"]
+        skip_type = ["Identity", "DequantizeLinear"]
         if node.type in skip_type:
             return True
 
@@ -73,6 +73,10 @@ def _fold_node(self, node, graph):
             self.logger.debug("need to add function to fold op %s whose op_type is %s", node.name, node.type)
         return False
 
+    @staticmethod
+    def compute_const_folding(node, graph):
+        return _func_map[node.type](node, graph)
+
     @staticmethod
     def _all_inputs_are_const(nodes):
         return all(node.is_const() for node in nodes if node)