diff --git a/src/finn/transformation/streamline/reorder.py b/src/finn/transformation/streamline/reorder.py
index 3b5d32b1cb6784fcc22cc8dcff04c2842e729d6d..253988185977c6b9d28505d46e00b2ab11c3a76b 100644
--- a/src/finn/transformation/streamline/reorder.py
+++ b/src/finn/transformation/streamline/reorder.py
@@ -29,6 +29,7 @@
import numpy as np
import warnings
from onnx import helper as oh
+from onnx import TensorProto
from finn.transformation import Transformation
import finn.core.data_layout as DataLayout
@@ -676,6 +677,66 @@ class MoveMaxPoolPastMultiThreshold(Transformation):
model = model.transform(InferShapes())
return (model, graph_modified)
+class MoveFlattenPastTopK(Transformation):
+ """Move flatten node past a succeeding topk node, if the "axis" attribute in topk
+ is set to -1 and the data layout before the flatten is NHWC with H=W=1"""
+
+ 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 == "Flatten":
+ consumer = model.find_consumer(n.output[0])
+ if consumer is not None and consumer.op_type == "TopK":
+ axis = get_by_name(consumer.attribute, "axis")
+ if axis is None or axis.i != -1:
+ continue
+ start_name = n.input[0]
+ data_layout = model.get_tensor_layout(start_name)
+ if data_layout != DataLayout.NHWC:
+ warnings.warn(
+ """Transformation can't be applied. The input
+ to flatten has to have DataLayout.NHWC"""
+ )
+ continue
+ (b, h, w, c) = model.get_tensor_shape(start_name)
+ if h != 1 or w != 1:
+ continue
+ # get parameter k from topk
+ k = model.get_tensor_shape(consumer.output[1])[-1]
+
+ # swap conections
+ # new tensor because dims change
+ middle_name = model.make_new_valueinfo_name()
+ topk_indices = oh.make_tensor_value_info(
+ middle_name, TensorProto.INT64, [b, h, w, k]
+ )
+ end_name = consumer.output[1]
+ graph.value_info.append(topk_indices)
+
+ # remove old nodes
+ graph.node.remove(n)
+ graph.node.remove(consumer)
+
+ # set inputs and outputs correctly
+ consumer.input[0] = start_name
+ consumer.output[1] = middle_name
+ model.set_tensor_shape(consumer.output[0], (b, h, w, k))
+
+ n.input[0] = middle_name
+ n.output[0] = end_name
+
+ # insert them back in
+ graph.node.insert(node_ind - 1, consumer)
+ graph.node.insert(node_ind, n)
+
+ graph_modified = True
+
+ model = model.transform(InferShapes())
+ return (model, graph_modified)
+
class MoveFlattenPastAffine(Transformation):
"""Moves a node that implements a (1, -1) reshape past a MatMul, Mul or Add node."""
diff --git a/tests/transformation/test_move_flatten_past_topk.py b/tests/transformation/test_move_flatten_past_topk.py
new file mode 100644
index 0000000000000000000000000000000000000000..65da92c22dbe9f6b1c5a49172ffae59fa6e98607
--- /dev/null
+++ b/tests/transformation/test_move_flatten_past_topk.py
@@ -0,0 +1,89 @@
+# 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.
+import pytest
+
+from onnx import TensorProto, helper
+
+from finn.core.modelwrapper import ModelWrapper
+from finn.core.datatype import DataType
+import finn.core.data_layout as DataLayout
+from finn.util.basic import gen_finn_dt_tensor
+from finn.transformation.insert_topk import InsertTopK
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.transformation.infer_data_layouts import InferDataLayouts
+from finn.transformation.general import GiveUniqueNodeNames, GiveReadableTensorNames
+from finn.transformation.streamline.reorder import MoveFlattenPastTopK
+import finn.core.onnx_exec as oxe
+
+# data layout
+@pytest.mark.parametrize("data_layout", [DataLayout.NHWC, DataLayout.NCHW])
+# batch size
+@pytest.mark.parametrize("batch_size", [1, 2])
+def test_move_flatten_past_affine(data_layout, batch_size):
+ if data_layout == DataLayout.NHWC:
+ ishape = [batch_size, 1, 1, 1024]
+ oshape = [batch_size, 1024]
+ else:
+ ishape = [batch_size, 1024, 1, 1]
+ oshape = [batch_size, 1024]
+
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, ishape)
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, oshape)
+
+ flatten_node = helper.make_node("Flatten", ["inp"], ["outp"])
+
+ graph = helper.make_graph(
+ nodes=[flatten_node], name="move-flatten-graph", inputs=[inp], outputs=[outp],
+ )
+
+ model = helper.make_model(graph, producer_name="move_flatten_model")
+ model = ModelWrapper(model)
+
+ model.set_tensor_datatype("inp", DataType.INT2)
+ model.set_tensor_layout("inp", data_layout)
+ model = model.transform(InsertTopK())
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ model = model.transform(InferDataLayouts())
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(GiveReadableTensorNames())
+
+ # compare execution before and after transformation
+ inp_values = gen_finn_dt_tensor(DataType.INT2, ishape)
+ idict = {model.graph.input[0].name: inp_values}
+ model_transformed = model.transform(MoveFlattenPastTopK())
+ assert oxe.compare_execution(model, model_transformed, idict)
+
+ # depending on data layout check if graph is transformed or not
+ if data_layout == DataLayout.NHWC:
+ # check if nodes have new order in transformed graph
+ assert model.graph != model_transformed.graph
+ assert model_transformed.graph.node[-1].op_type == "Flatten"
+ else:
+ assert model.graph == model_transformed.graph