Merge pull request #159 from Xilinx/feature/remove_identity_ops

Feature/remove identity ops

src/finn/transformation/streamline/remove.py

+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+#   list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+#   this list of conditions and the following disclaimer in the documentation
+#   and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+#   contributors may be used to endorse or promote products derived from
+#   this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+
+from finn.transformation import Transformation
+from finn.transformation.infer_shapes import InferShapes
+import numpy as np
+
+class RemoveIdentityOps(Transformation):
+    """Remove identity ops like Add/Sub with zero or Mul/Div with one"""
+
+    def apply(self, model):
+        graph = model.graph
+        node_ind = 0
+        graph_modified = False
+        for n in graph.node:
+            node_ind += 1
+            if (
+                n.op_type in ["Add", "Sub"]
+                and not model.is_fork_node(n)
+                and not model.is_join_node(n)
+            ):
+                A = model.get_initializer(n.input[1])
+                if A is not None and (A == np.zeros_like(A)).all():
+                    producer = model.find_producer(n.input[0])
+                    # remove node and wire output tensor to
+                    # output of producer node
+                    producer.output[0] = n.output[0]
+                    graph.node.remove(n)
+
+            elif (
+                n.op_type in ["Mul", "Div"]
+                and not model.is_fork_node(n)
+                and not model.is_join_node(n)
+            ):
+                A = model.get_initializer(n.input[1])
+                if A is not None and (A == np.ones_like(A)).all():
+                    producer = model.find_producer(n.input[0])
+                    # remove node and wire output tensor to
+                    # output of producer node
+                    producer.output[0] = n.output[0]
+                    graph.node.remove(n)
+        model = model.transform(InferShapes())
+        return (model, graph_modified)

tests/transformation/test_remove_identity_ops.py

+import pytest
+
+import numpy as np
+from onnx import helper, TensorProto
+import finn.core.onnx_exec as oxe
+from finn.core.datatype import DataType
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.streamline.remove import RemoveIdentityOps
+from finn.util.basic import gen_finn_dt_tensor
+
+
+def insert_identity_op(model, op):
+    if op in ["Add", "Sub"]:
+        val = np.asarray([0.0], dtype=np.float32)
+    elif op in ["Mul", "Div"]:
+        val = np.asarray([1.0], dtype=np.float32)
+    else:
+        return
+
+    identity_node = helper.make_node(op, ["div_out", "value"], ["ident_out"])
+    graph = model.graph
+    graph.node.insert(3, identity_node)
+    graph.node[-1].input[0] = "ident_out"
+    model.set_initializer("value", val)
+
+    return model
+
+
+# identity operations to be inserted
+@pytest.mark.parametrize("op", ["Add", "Sub", "Mul", "Div"])
+def test_remove_identity_ops(op):
+
+    # set up onnx model
+    inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, 4, 1, 1])
+    mul = helper.make_tensor_value_info("mul", TensorProto.FLOAT, [])
+    shape = helper.make_tensor_value_info("shape", TensorProto.FLOAT, [2])
+    div = helper.make_tensor_value_info("div", TensorProto.FLOAT, [])
+    matmul = helper.make_tensor_value_info("matmul", TensorProto.FLOAT, [4, 2])
+    outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, 2])
+
+    mul_node = helper.make_node("Mul", ["inp", "mul"], ["mul_out"])
+    reshape_node = helper.make_node("Reshape", ["mul_out", "shape"], ["reshape_out"])
+    div_node = helper.make_node("Div", ["reshape_out", "div"], ["div_out"])
+    matmul_node = helper.make_node("MatMul", ["div_out", "matmul"], ["outp"])
+
+    graph = helper.make_graph(
+        nodes=[mul_node, reshape_node, div_node, matmul_node],
+        name="identity-graph",
+        inputs=[inp],
+        outputs=[outp],
+        value_info=[mul, shape, div, matmul],
+    )
+
+    model = helper.make_model(graph, producer_name="mulpastconv-model")
+    model = ModelWrapper(model)
+    inp_values = gen_finn_dt_tensor(DataType.INT2, [1, 4, 1, 1])
+    mul_values = np.random.uniform(low=0.1, high=0.99, size=(1)).astype(np.float32)
+    shape_values = np.asarray([1, -1], dtype=np.int64)
+    div_values = np.random.uniform(low=0.1, high=0.99, size=(1)).astype(np.float32)
+    matmul_values = gen_finn_dt_tensor(DataType.INT2, [4, 2])
+    model.set_initializer("mul", mul_values)
+    model.set_initializer("shape", shape_values)
+    model.set_initializer("div", div_values)
+    model.set_initializer("matmul", matmul_values)
+    insert_identity_op(model, op)
+    model = model.transform(InferShapes())
+    model = model.transform(InferDataTypes())
+    idict = {"inp": inp_values}
+    odict = oxe.execute_onnx(model, idict)
+    out_before = odict["outp"]
+    num_of_nodes_before = len(model.graph.node)
+
+    model = model.transform(RemoveIdentityOps())
+    num_of_nodes_after = len(model.graph.node)
+    assert num_of_nodes_before - 1 == num_of_nodes_after
+
+    odict = oxe.execute_onnx(model, idict)
+    out_after = odict["outp"]
+    assert (out_before == out_after).all()