diff --git a/src/finn/core/modelwrapper.py b/src/finn/core/modelwrapper.py
index 3ddcaa03c3d62daaf1ffd9f5ae6c3857460994fc..e99a6ef4cd40d6323d77354d3c9b4be341d7649c 100644
--- a/src/finn/core/modelwrapper.py
+++ b/src/finn/core/modelwrapper.py
@@ -253,14 +253,12 @@ class ModelWrapper:
return None
def find_producer(self, tensor_name):
- """Finds and returns the node that produces the tensor with given name.
- Currently only works for linear graphs."""
- all_outputs = [x.output[0] for x in self._model_proto.graph.node]
- try:
- producer_ind = all_outputs.index(tensor_name)
- return self._model_proto.graph.node[producer_ind]
- except ValueError:
- return None
+ """Finds and returns the node that produces the tensor with given name."""
+ ret = None
+ for x in self._model_proto.graph.node:
+ if tensor_name in x.output:
+ ret = x
+ return ret
def find_upstream(self, tensor_name, finder_fxn):
"""Follow the producer chain upstream, calling finder_fxn on each upstream
diff --git a/src/finn/core/onnx_exec.py b/src/finn/core/onnx_exec.py
index 0f47a9104e3d2ef3ee06ef908e302344d78e0b17..172ba25b223fd087df134add460a42d0a9935e0e 100644
--- a/src/finn/core/onnx_exec.py
+++ b/src/finn/core/onnx_exec.py
@@ -61,6 +61,10 @@ def execute_node(node, context, graph):
# onnxruntime unfortunately does not implement run_node as defined by ONNX,
# it can only execute entire models -- so we create a model which solely
# consists of our current node.
+ # note: ensure that the same ValueInfo does not appear both in
+ # graph.value_info as well as graph.output or graph.input
+ # nodes with multiple outputs that are a mix of value_info and
+ # input/outputs may get them reordered below
node_inputs = list(filter(lambda x: x.name in node.input, graph.input))
node_inputs += list(
filter(lambda x: x.name in node.input, graph.value_info)
@@ -84,17 +88,25 @@ def execute_node(node, context, graph):
output_list = sess.run(None, input_dict)
for output_ind in range(len(node.output)):
+ # get the name of the target buffer from node.output
outp = node.output[output_ind]
- if output_list[output_ind].shape != context[outp].shape:
+
+ # retrieve the index of that name in node_outputs
+ for i in range(len(node_outputs)):
+ if outp == node_outputs[i].name:
+ list_ind = i
+
+ # use that index to index output_list
+ if output_list[list_ind].shape != context[outp].shape:
raise Exception(
"""Output shapes disagree after node execution:
found %s vs expected %s"""
% (
- str(output_list[output_ind].shape.shape),
+ str(output_list[list_ind].shape.shape),
str(context[outp].shape),
)
)
- context[outp] = output_list[output_ind]
+ context[outp] = output_list[list_ind]
def execute_onnx(model, input_dict, return_full_exec_context=False):
diff --git a/src/finn/transformation/insert_topk.py b/src/finn/transformation/insert_topk.py
new file mode 100644
index 0000000000000000000000000000000000000000..213d2cedf92c0276e33fcf2b50e6966aeee8c847
--- /dev/null
+++ b/src/finn/transformation/insert_topk.py
@@ -0,0 +1,96 @@
+# 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 numpy as np
+
+from onnx import TensorProto
+from onnx import helper as oh
+
+from finn.transformation import Transformation
+from finn.core.datatype import DataType
+
+
+class InsertTopK(Transformation):
+ """Add TopK node at the network output and replace the graph output with
+ the TopK indices."""
+
+ def __init__(self, k=5, axis=-1, largest=1, sorted=1):
+ super().__init__()
+ self.k = k
+ self.axis = axis
+ self.largest = largest
+ self.sorted = sorted
+
+ def apply(self, model):
+ # get name of output tensor
+ graph_out_name = model.graph.output[0].name
+ # find final node
+ final_node = model.find_producer(graph_out_name)
+ # if a top-select op is already present, do nothing
+ if final_node.op_type == "TopK":
+ return (model, False)
+ else:
+ out_shape = model.get_tensor_shape(graph_out_name)
+ out_dtype = model.get_tensor_datatype(graph_out_name)
+ # adjust shape
+ out_shape[self.axis] = self.k
+ # make new buffer
+ k_tensor = np.array([self.k]).astype(np.int64)
+ k_value = oh.make_tensor_value_info(
+ model.make_new_valueinfo_name(), TensorProto.INT64, [1]
+ )
+ topk_values = oh.make_tensor_value_info(
+ model.make_new_valueinfo_name(), TensorProto.FLOAT, out_shape
+ )
+ topk_indices = oh.make_tensor_value_info(
+ model.make_new_valueinfo_name(), TensorProto.INT64, out_shape
+ )
+ model.graph.value_info.append(k_value)
+ model.set_tensor_datatype(k_value.name, out_dtype) # TODO set to int64
+ model.graph.value_info.append(topk_values)
+ model.set_tensor_datatype(topk_values.name, out_dtype)
+ # create and append topk node
+ model.set_initializer(k_value.name, k_tensor)
+ topk_node = oh.make_node(
+ "TopK",
+ inputs=[graph_out_name, k_value.name],
+ outputs=[topk_values.name, topk_indices.name],
+ axis=self.axis,
+ largest=self.largest,
+ sorted=self.sorted,
+ )
+ model.graph.node.append(topk_node)
+ # replace the existing output definition with topk indices
+ model.graph.output.insert(0, topk_indices)
+ model.graph.output.pop(1)
+ # set quantization annotation for indices
+ # minimal output dtype for TopK indices dependens on num. classes
+ # assuming UINT32 is large enough for now (FINN has currently no
+ # DataType.INT64)
+ model.set_tensor_datatype(topk_indices.name, DataType.UINT32)
+ return (model, True)
diff --git a/tests/transformation/test_topk_insert.py b/tests/transformation/test_topk_insert.py
new file mode 100644
index 0000000000000000000000000000000000000000..ac32c30edbbf466b2b441bcc92975a7d50f42bda
--- /dev/null
+++ b/tests/transformation/test_topk_insert.py
@@ -0,0 +1,58 @@
+import onnx
+from finn.util.test import get_test_model_trained
+import brevitas.onnx as bo
+import numpy as np
+import onnx.numpy_helper as nph
+import torch
+
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.general import GiveReadableTensorNames, GiveUniqueNodeNames
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.transformation.fold_constants import FoldConstants
+from finn.transformation.insert_topk import InsertTopK
+
+import finn.core.onnx_exec as oxe
+from pkgutil import get_data
+
+import pytest
+
+export_onnx_path = "test_output_lfc.onnx"
+
+
+@pytest.mark.parametrize("k", [1, 5, 10])
+def test_topk_insert(k):
+ tfc = get_test_model_trained("TFC", 1, 1)
+ bo.export_finn_onnx(tfc, (1, 1, 28, 28), export_onnx_path)
+ model = ModelWrapper(export_onnx_path)
+
+ # do transformations (no topk)
+ model = model.transform(InferShapes())
+ model = model.transform(FoldConstants())
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(GiveReadableTensorNames())
+ model = model.transform(InferDataTypes())
+
+ # verification: generate random input, run through net, streamline,
+ # run again, check that output is top-k
+ raw_i = get_data("finn", "data/onnx/mnist-conv/test_data_set_0/input_0.pb")
+ input_tensor = onnx.load_tensor_from_string(raw_i)
+ input_brevitas = torch.from_numpy(nph.to_array(input_tensor)).float()
+ output_golden = tfc.forward(input_brevitas).detach().numpy()
+ output_golden_topk = np.flip(output_golden.flatten().argsort())[:k]
+ output_golden_topk = output_golden_topk.flatten()
+
+ input_dict = {"global_in": nph.to_array(input_tensor)}
+
+ # insert top-k
+ model = model.transform(InsertTopK(k))
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(GiveReadableTensorNames())
+ model = model.transform(InferShapes())
+
+ # verify output of top-k
+ output_dict_topk = oxe.execute_onnx(model, input_dict)
+ output_pysim_topk = output_dict_topk[list(output_dict_topk.keys())[0]]
+ output_pysim_topk = output_pysim_topk.astype(np.int).flatten()
+
+ assert np.array_equal(output_golden_topk, output_pysim_topk)