fix subgraph bug (#260)

* add functions to resort Signed-off-by: xiaowuhu <[email protected]> * change code and add test case Signed-off-by: xiaowuhu <[email protected]> * Update float16.py Signed-off-by: xiaowuhu <[email protected]> * Delete float16_old.py Signed-off-by: xiaowuhu <[email protected]> * Update float16.py Signed-off-by: xiaowuhu <[email protected]> * fix flake 8 Signed-off-by: xiaowuhu <[email protected]> * Update perfstats.py Signed-off-by: xiaowuhu <[email protected]> * Update test_float16.py Signed-off-by: xiaowuhu <[email protected]> * Update __init__.py Signed-off-by: xiaowuhu <[email protected]> * Create pyproject.toml * add more info * update Signed-off-by: xiaowuhu <[email protected]> * Delete setup.py * Update float16.py Signed-off-by: xiaowuhu <[email protected]> * disable python Signed-off-by: xiaowuhu <[email protected]> * Revert "Delete setup.py" This reverts commit 2da0727. --------- Signed-off-by: xiaowuhu <[email protected]>
microsoft · Aug 21, 2023 · 3873b29 · 3873b29
1 parent 8c81a0e
commit 3873b29
Show file tree

Hide file tree

Showing 10 changed files with 115 additions and 12 deletions.
diff --git a/.azure-pipelines/OneBranchPipeline-Official.yml b/.azure-pipelines/OneBranchPipeline-Official.yml
@@ -33,7 +33,7 @@ extends:
         break: true # always break the build on policheck issues. You can disable it by setting to 'false'
       codeql:
         python:
-          enabled: true
+          enabled: false
 
     stages:
     - stage: firststage
@@ -57,7 +57,7 @@ extends:
             pip install -r requirements.txt
             pip install onnxruntime
             pip install onnxmltools
-            pip install onnx==1.12.0
+            pip install onnx
             pip install pytest
             pip install -e .
           displayName: 'Install dependencies'

diff --git a/.azure-pipelines/OneBranchPipeline-PullRequest.yml b/.azure-pipelines/OneBranchPipeline-PullRequest.yml
@@ -34,7 +34,7 @@ extends:
         break: true # always break the build on policheck issues.
       codeql:
         python:
-          enabled: true
+          enabled: false
 
     stages:
     - stage: firststage
@@ -55,7 +55,7 @@ extends:
             pip install -r requirements.txt
             pip install onnxruntime
             pip install onnxmltools
-            pip install onnx==1.12.0
+            pip install onnx
             pip install pytest
             pip install -e .
           displayName: 'Install dependencies'

diff --git a/MANIFEST.in b/MANIFEST.in
@@ -1 +1,2 @@
 include LICENSE
+include requirements.txt
diff --git a/onnxconverter_common/__init__.py b/onnxconverter_common/__init__.py
@@ -8,7 +8,7 @@
 This framework performs optimization for ONNX models and
 includes common utilities for ONNX converters.
 """
-__version__ = "1.13.0"
+__version__ = "1.14.0"
 __author__ = "Microsoft"
 __producer__ = "OnnxMLTools"
 __producer_version__ = __version__

diff --git a/onnxconverter_common/data_types.py b/onnxconverter_common/data_types.py
@@ -248,9 +248,9 @@ def find_type_conversion(source_type, target_type):
     """
     Find the operator name for converting source_type into target_type
     """
-    if type(source_type) == type(target_type):
+    if isinstance(source_type, target_type):
         return 'identity'
-    if type(target_type) == FloatTensorType:
+    if isinstance(target_type, FloatTensorType):
         return 'imageToFloatTensor'
     raise ValueError('Unsupported type conversion from %s to %s' % (
                      source_type, target_type))

diff --git a/onnxconverter_common/float16.py b/onnxconverter_common/float16.py
@@ -157,6 +157,8 @@ def convert_float_to_float16(model, min_positive_val=1e-7, max_finite_val=1e4,
     queue = []
     value_info_list = []
     node_list = []
+    # key = node, value = graph, used to distinguish global with sub-graph
+    node_dict = {}
     # type inference on input model
     if func_infer_shape is not None:
         model = func_infer_shape(model)
@@ -222,6 +224,7 @@ def convert_float_to_float16(model, min_positive_val=1e-7, max_finite_val=1e4,
                     # so it will not be converted to float16
                     if n.op_type in op_block_list or n.name in node_block_list:
                         node_list.append(n)
+                        node_dict[n.name] = q
                     else:
                         if n.op_type == 'Cast':
                             for attr in n.attribute:
@@ -264,15 +267,16 @@ def convert_float_to_float16(model, min_positive_val=1e-7, max_finite_val=1e4,
             for value_info in value_info_list:
                 if input == value_info.name:
                     # create new value_info for current node's new input name
-                    new_value_info = model.graph.value_info.add()
+                    graph = node_dict[node.name]  # get the correct graph instead of the global graph
+                    new_value_info = graph.value_info.add()
                     new_value_info.CopyFrom(value_info)
                     output_name = node.name + '_input_cast_' + str(i)
                     new_value_info.name = output_name
                     new_value_info.type.tensor_type.elem_type = onnx_proto.TensorProto.FLOAT
                     # add Cast node (from tensor(float16) to tensor(float) before current node
                     node_name = node.name + '_input_cast' + str(i)
                     new_node = [helper.make_node('Cast', [input], [output_name], to=1, name=node_name)]
-                    model.graph.node.extend(new_node)
+                    graph.node.extend(new_node)
                     # change current node's input name
                     node.input[i] = output_name
                     break
@@ -283,19 +287,21 @@ def convert_float_to_float16(model, min_positive_val=1e-7, max_finite_val=1e4,
             for value_info in value_info_list:
                 if output == value_info.name:
                     # create new value_info for current node's new output
-                    new_value_info = model.graph.value_info.add()
+                    graph = node_dict[node.name]  # get the correct graph instead of the global graph
+                    new_value_info = graph.value_info.add()
                     new_value_info.CopyFrom(value_info)
                     input_name = node.name + '_output_cast_' + str(i)
                     new_value_info.name = input_name
                     new_value_info.type.tensor_type.elem_type = onnx_proto.TensorProto.FLOAT
                     # add Cast node (from tensor(float) to tensor(float16) after current node
                     node_name = node.name + '_output_cast' + str(i)
                     new_node = [helper.make_node('Cast', [input_name], [output], to=10, name=node_name)]
-                    model.graph.node.extend(new_node)
+                    graph.node.extend(new_node)
                     # change current node's input name
                     node.output[i] = input_name
                     break
 
+    sort_topology(model.graph)
     return model
 
 
@@ -332,3 +338,53 @@ def convert_float_to_float16_model_path(model_path, min_positive_val=1e-7, max_f
     if not disable_shape_infer:
         model = onnx.load(model_path)
     return convert_float_to_float16(model, min_positive_val, max_finite_val, keep_io_types, disable_shape_infer)
+
+
+def sort_graph_node(graph_proto):
+    # find the "first" node in Nodes that its input is not any node's output
+    def find_first_node(output2node_dict):
+        for node in org_nodes:
+            is_not_first_node = any(item in output2node_dict for item in node.input)
+            if not is_not_first_node:
+                return node
+        return None
+
+    # remove the node from output2node_dict using output as key
+    def remove_first_node_from_dict2(first_node):
+        for output in first_node.output:
+            if output in output2node_dict:
+                del output2node_dict[output]
+
+    org_nodes = graph_proto.node
+    # create a dict to store output as key and node as value
+    output2node_dict = {}
+    for node in org_nodes:
+        for output in node.output:
+            output2node_dict[output] = node
+
+    # save the final node after sorted
+    sorted_node = []
+    # traverse the Nodes to find the first node
+    while (len(output2node_dict) > 0):
+        first_node = find_first_node(output2node_dict)
+        sorted_node.append(first_node)
+        remove_first_node_from_dict2(first_node)
+        # del node from original nodes list to avoid duplicate traverse
+        org_nodes.remove(first_node)
+
+    for new_node in sorted_node:
+        graph_proto.node.extend([new_node])
+
+
+# The input graph should be mode.graph
+# Recursevly sort the topology for each sub-graph
+def sort_topology(graph_proto):
+    assert (isinstance(graph_proto, onnx_proto.GraphProto))
+    sort_graph_node(graph_proto)  # sort global graph
+    for node in graph_proto.node:
+        for attr in node.attribute:
+            if isinstance(attr.g, onnx_proto.GraphProto) and len(attr.g.node) > 0:
+                sort_topology(attr.g)  # sort sub-graph
+            for g in attr.graphs:
+                if isinstance(g, onnx_proto.GraphProto):
+                    sort_topology(g)  # sort sub-graph
diff --git a/onnxconverter_common/perfstats.py b/onnxconverter_common/perfstats.py
@@ -114,7 +114,7 @@ def compute_op_type_entries(raw_entries):
 def read_raw_entries(profile_path):
     with open(profile_path, "r") as f:
         data = json.load(f)
-    if type(data) == dict:
+    if isinstance(data, dict):
         data = data['traceEvents']
     entries = []
     for item in data:

diff --git a/pyproject.toml b/pyproject.toml
@@ -0,0 +1,30 @@
+[build-system]
+requires = ["setuptools>=61.0.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "onnxconverter-common"
+dynamic = ["version"]
+description="ONNX Converter and Optimization Tools"
+authors = [{ name = "Microsoft Corporation", email = "[email protected]" }]
+urls =  { "Repository" = "https://github.com/microsoft/onnxconverter-common" }
+readme = "README.md"
+requires-python = ">=3.8"
+license = { file = "LICENSE" }
+classifiers = [
+  "Development Status :: 5 - Production/Stable",
+  "Environment :: Console",
+  "Intended Audience :: Developers",
+  "Operating System :: MacOS :: MacOS X",
+  "Operating System :: Microsoft :: Windows",
+  "Programming Language :: Python",
+  "Programming Language :: Python :: 3.8",
+  "Programming Language :: Python :: 3.9",
+  "Programming Language :: Python :: 3.10",
+  "Programming Language :: Python :: 3.11",
+  "License :: OSI Approved :: MIT License",
+]
+dependencies = ["numpy", "onnx", "packaging", "protobuf==3.20.2"]
+
+[tool.setuptools.dynamic]
+version = {attr = "onnxconverter_common.__version__"}
diff --git a/tests/data/test_subgraph.onnx b/tests/data/test_subgraph.onnx
diff --git a/tests/test_float16.py b/tests/test_float16.py
@@ -108,6 +108,22 @@ def test_convert_to_float16_with_truncated(self):
         np_array = np.array([1e-10, -2.0, 15, -1e-9, 65536.1, -100000])
         convert_np_to_float16(np_array)
 
+    @unittest.skipIf(pv.Version(onnx.__version__) == pv.Version('1.9.0'), "ONNX 1.9 has different Optype behavior for Max operator")
+    def test_convert_to_float16_with_subgraph(self):
+        model32_name = "test_subgraph.onnx"
+        working_path = os.path.abspath(os.path.dirname(__file__))
+        data_path = os.path.join(working_path, 'data')
+        model_path = os.path.join(data_path, model32_name)
+        onnx_model32 = onnxmltools.utils.load_model(model_path)
+        x = np.array([1.0], dtype=np.float32)
+        y = np.array([2.0], dtype=np.float32)
+        output_32 = _ort_inference(onnx_model32, {"x":x, "y":y})
+
+        onnx_model16 = convert_float_to_float16(onnx_model32, keep_io_types=True)
+        output_16 = _ort_inference(onnx_model16, {"x":x, "y":y})
+        self.assertTrue(np.allclose(output_16, output_32, atol=1e-2))
+
+
 
 if __name__ == '__main__':
     suite = unittest.TestLoader().loadTestsFromTestCase(ONNXFloat16Test)