Merge branch 'fix_torch_v2' into nml_examples2

.github/workflows/ci.yml

@@ -69,41 +69,26 @@ jobs:
           pip list
 
     - name: Install graphviz
-      if: ${{ matrix.runs-on != 'windows-latest' }}
-      run: |
-        if [[  ${{ matrix.runs-on }} == *"macos"* ]]; then brew install graphviz ; fi
-        if [[  ${{ matrix.runs-on }} == *"ubuntu"* ]]; then sudo apt install graphviz ; fi
+      uses: ts-graphviz/setup-graphviz@v1
 
     - name: Test interface ACT-R
-      if: ${{ matrix.python-version != '3.10' || matrix.runs-on != 'windows-latest' }}
       run: |
         python -m pytest -v -m "actr" tests/
 
     - name: Test interface PyTorch
-      if: ${{ matrix.python-version != '3.10' || matrix.runs-on != 'windows-latest' }}
       run: |
         python -m pytest -v -m "pytorch" tests/
 
     - name: Test interface NeuroML
-      if: ${{ matrix.python-version != '3.10' || matrix.runs-on != 'windows-latest' }}
       run: |
         python -m pytest -v -m "neuroml" tests/
 
-    - name: Test interface TensorFlow linux/mac
-      if: ${{ matrix.runs-on != 'windows-latest' }}
-      run: |
-        dot -V
-        python -m pytest -v -m "tensorflow" tests/
-
-    - name: Test interface TensorFlow windows
-      if: ${{ matrix.python-version != '3.10' && matrix.runs-on == 'windows-latest' }}
+    - name: Test interface TensorFlow
       run: |
-        choco install graphviz
         dot -V
         python -m pytest -v -m "tensorflow" tests/
 
     - name: Test interface PsyNeuLink
-      if: ${{ matrix.python-version != '3.10' }}
       run: |
         python -m pip install .[psyneulink]
         python -m pytest -v -m "psyneulink" tests/

docs/sphinx/source/api/Installation.md

@@ -123,6 +123,7 @@ make clean
 make html
 ```
 
+
 ### 7) Change directory into html folder and run the documentation offline
 ```
 # for Windows go into build\html folder and double click on the index.html file, or:

docs/sphinx/source/api/export_format/ONNX/simple_ab.py

@@ -100,7 +100,7 @@ def main():
         for t in test_values:
             print("===================\nEvaluating MDF model with input: %s" % t)
             eg.evaluate(initializer={"input": t})
-            print("Output: %s" % eg.enodes["Mul_3"].evaluable_outputs["_4"].curr_value)
+            print("Output: %s" % eg.enodes["/B/Mul"].evaluable_outputs["_4"].curr_value)
 
 
 if __name__ == "__main__":

examples/ONNX/ab.json

@@ -3,9 +3,9 @@
         "format": "ModECI MDF v0.4",
         "generating_application": "Python modeci-mdf v0.4.7",
         "graphs": {
-            "torch-jit-export": {
+            "torch_jit": {
                 "nodes": {
-                    "Add_1": {
+                    "/A/Add": {
                         "input_ports": {
                             "input": {
                                 "shape": [
@@ -19,7 +19,7 @@
                             "B": {
                                 "value": 1.0
                             },
-                            "Add_1": {
+                            "/A/Add": {
                                 "function": "onnx::Add",
                                 "args": {
                                     "A": "input",
@@ -28,14 +28,14 @@
                             }
                         },
                         "output_ports": {
-                            "onnx_Mul_2": {
-                                "value": "Add_1"
+                            "_A_Add_output_0": {
+                                "value": "/A/Add"
                             }
                         }
                     },
-                    "Mul_3": {
+                    "/B/Mul": {
                         "input_ports": {
-                            "onnx_Mul_2": {
+                            "_A_Add_output_0": {
                                 "shape": [
                                     2,
                                     3
@@ -47,27 +47,27 @@
                             "B": {
                                 "value": 5.0
                             },
-                            "Mul_3": {
+                            "/B/Mul": {
                                 "function": "onnx::Mul",
                                 "args": {
-                                    "A": "onnx_Mul_2",
+                                    "A": "_A_Add_output_0",
                                     "B": "B"
                                 }
                             }
                         },
                         "output_ports": {
                             "_4": {
-                                "value": "Mul_3"
+                                "value": "/B/Mul"
                             }
                         }
                     }
                 },
                 "edges": {
-                    "Add_1.onnx_Mul_2_Mul_3.onnx_Mul_2": {
-                        "sender": "Add_1",
-                        "receiver": "Mul_3",
-                        "sender_port": "onnx_Mul_2",
-                        "receiver_port": "onnx_Mul_2"
+                    "/A/Add._A_Add_output_0_/B/Mul._A_Add_output_0": {
+                        "sender": "/A/Add",
+                        "receiver": "/B/Mul",
+                        "sender_port": "_A_Add_output_0",
+                        "receiver_port": "_A_Add_output_0"
                     }
                 }
             }

examples/ONNX/ab.yaml

@@ -2,9 +2,9 @@ ONNX Model:
     format: ModECI MDF v0.4
     generating_application: Python modeci-mdf v0.4.7
     graphs:
-        torch-jit-export:
+        torch_jit:
             nodes:
-                Add_1:
+                /A/Add:
                     input_ports:
                         input:
                             shape:
@@ -14,35 +14,35 @@ ONNX Model:
                     parameters:
                         B:
                             value: 1.0
-                        Add_1:
+                        /A/Add:
                             function: onnx::Add
                             args:
                                 A: input
                                 B: B
                     output_ports:
-                        onnx_Mul_2:
-                            value: Add_1
-                Mul_3:
+                        _A_Add_output_0:
+                            value: /A/Add
+                /B/Mul:
                     input_ports:
-                        onnx_Mul_2:
+                        _A_Add_output_0:
                             shape:
                             - 2
                             - 3
                             type: float
                     parameters:
                         B:
                             value: 5.0
-                        Mul_3:
+                        /B/Mul:
                             function: onnx::Mul
                             args:
-                                A: onnx_Mul_2
+                                A: _A_Add_output_0
                                 B: B
                     output_ports:
                         _4:
-                            value: Mul_3
+                            value: /B/Mul
             edges:
-                Add_1.onnx_Mul_2_Mul_3.onnx_Mul_2:
-                    sender: Add_1
-                    receiver: Mul_3
-                    sender_port: onnx_Mul_2
-                    receiver_port: onnx_Mul_2
+                /A/Add._A_Add_output_0_/B/Mul._A_Add_output_0:
+                    sender: /A/Add
+                    receiver: /B/Mul
+                    sender_port: _A_Add_output_0
+                    receiver_port: _A_Add_output_0

examples/ONNX/simple_ab.py

@@ -100,7 +100,7 @@ def main():
         for t in test_values:
             print("===================\nEvaluating MDF model with input: %s" % t)
             eg.evaluate(initializer={"input": t})
-            print("Output: %s" % eg.enodes["Mul_3"].evaluable_outputs["_4"].curr_value)
+            print("Output: %s" % eg.enodes["/B/Mul"].evaluable_outputs["_4"].curr_value)
 
 
 if __name__ == "__main__":

examples/TensorFlow/Keras/keras2MDFfunc_cal.py

@@ -24,7 +24,7 @@
 from tensorflow.keras.layers import Input, Dense, Flatten
 
 
-new_model = tf.keras.models.load_model("kr_N_model.h5")
+new_model = tf.keras.models.load_model("kr_N_model.h5", compile=False)
 for i in new_model.layers:
     print(i.name)
 

examples/TensorFlow/Keras/keras_model.py

@@ -4,7 +4,7 @@
 # tf.__version__
 
 from keras.layers import Dense
-from keras.utils.vis_utils import plot_model
+from keras.utils import plot_model
 from keras.models import Sequential
 
 # from keras_visualizer import visualizer

examples/TensorFlow/Keras/keras_to_MDF.json

@@ -1,7 +1,7 @@
 {
     "keras_to_MDF": {
         "format": "ModECI MDF v0.4",
-        "generating_application": "Python modeci-mdf v0.4.4",
+        "generating_application": "Python modeci-mdf v0.4.7",
         "graphs": {
             "Keras_to_MDF_graph": {
                 "nodes": {

examples/TensorFlow/Keras/keras_to_MDF.yaml

@@ -1,6 +1,6 @@
 keras_to_MDF:
     format: ModECI MDF v0.4
-    generating_application: Python modeci-mdf v0.4.4
+    generating_application: Python modeci-mdf v0.4.7
     graphs:
         Keras_to_MDF_graph:
             nodes:

setup.cfg

@@ -40,9 +40,9 @@ install_requires =
     matplotlib
     graphviz
     h5py
-    onnxruntime==1.13.1
-    onnx==1.12.0
-    skl2onnx==1.14.0
+    onnxruntime
+    onnx
+    skl2onnx
     attrs>=21.1.0
     cattrs
     modelspec<0.4,>=0.3.0
@@ -89,7 +89,7 @@ optional =
     torchviz
     netron
     torch>=1.11.0
-    torchvision<=0.12.0
+    torchvision
     h5py
 
 all_except_psyneulink =

src/modeci_mdf/functions/onnx.py

@@ -125,9 +125,12 @@ def run_onnx_op(
     if op_name == "Pad":
         if "constant_value" in inputs:
             cval = inputs["constant_value"]
-            data = list(inputs.values())[0]
-            if cval.dtype != data.dtype:
-                inputs["constant_value"] = cval.astype(data.dtype)
+            if cval is None:
+                inputs.pop("constant_value")
+            else:
+                data = list(inputs.values())[0]
+                if cval.dtype != data.dtype:
+                    inputs["constant_value"] = cval.astype(data.dtype)
 
     # SkLearn ONNX doesn't seem to support ConcatFromSequence, see
     # https://github.com/onnx/sklearn-onnx/issues/710

src/modeci_mdf/interfaces/onnx/exporter.py

@@ -12,6 +12,8 @@
 from onnx import AttributeProto, TensorProto, GraphProto
 from onnx.defs import get_schema
 
+import onnxruntime
+
 from ast import literal_eval
 
 import argparse
@@ -44,7 +46,14 @@ def mdf_to_onnx(mdf_model):
         onnx_graph = generate_onnx_graph(graph, nodenames_in_execution_order)
 
         # Make an onnx model from graph
-        onnx_model = helper.make_model(onnx_graph)
+
+        # Check to see if onnxruntime version is less than 1.15, if so ir_version should
+        # be 8 for now. See: https://github.com/microsoft/onnxruntime/issues/15874
+        make_model_kwargs = {}
+        if onnxruntime.__version__ < "1.15":
+            make_model_kwargs = {"ir_version": 8}
+
+        onnx_model = helper.make_model(onnx_graph, **make_model_kwargs)
 
         # Infer shapes
         onnx_model = shape_inference.infer_shapes(onnx_model)

src/modeci_mdf/interfaces/onnx/importer.py

@@ -21,8 +21,12 @@
 def id_to_port(id: str):
     """Turn unique ONNX output and input value names into valid MDF input and outport names"""
 
+    # Get rid of periods in names
     new_name = str(id).replace(".", "_")
 
+    # Get rid of slashes in names
+    new_name = str(id).replace("/", "_")
+
     # Get rid of double colon in id names, this causes issues with execution engine.
     new_name = new_name.replace("::", "_")
 

src/modeci_mdf/interfaces/pytorch/importer.py

@@ -13,6 +13,18 @@
 
 import torch
 
+# We need to monkey patch the torch._C.Node class to add a __getitem__ method
+# This is for torch 2.0
+# From https://github.com/openai/CLIP/issues/79#issuecomment-1624202950
+def _node_get(node: torch._C.Node, key: str):
+    """Gets attributes of a node which is polymorphic over return type."""
+    sel = node.kindOf(key)
+    return getattr(node, sel)(key)
+
+
+torch._C.Node.__getitem__ = _node_get
+
+
 from modeci_mdf.mdf import Model, Graph, Node, Edge, InputPort, OutputPort, Parameter
 from modeci_mdf.functions.onnx import onnx_opset_version as modeci_onnx_opset_version
 
@@ -293,15 +305,15 @@ def torchnode_to_mdfnode(
     """
     op = node.kind()
 
-    # Lookup the schema. For some reason we cannot just call node.schema(), it returns "(no schema)", huh?
-    # We need to do this the hard way.
-    schema = onnx.defs.get_schema(op.replace("onnx::", ""), modeci_onnx_opset_version)
-
     # Exclude constants (as nodes) from the MDF graph. We will instead insert them as parameters to the nodes that
     # they project to.
     if op in ("prim::Constant", "onnx::Constant"):
         return None
 
+    # Lookup the schema. For some reason we cannot just call node.schema(), it returns "(no schema)", huh?
+    # We need to do this the hard way.
+    schema = onnx.defs.get_schema(op.replace("onnx::", ""), modeci_onnx_opset_version)
+
     # If we are dealing with a loop node, we need to recursively create a sub-graph for the loop body
     if op == "onnx::Loop":
         sub_mdf_graph = Graph(id=f"LoopSubgraph{make_node_id(node)}")
@@ -504,7 +516,7 @@ def pytorch_to_mdf(
             graph = None
 
     if use_onnx_ops:
-        operator_export_type = torch._C._onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK
+        operator_export_type = torch._C._onnx.OperatorExportTypes.ONNX
     else:
         operator_export_type = torch._C._onnx.OperatorExportTypes.RAW
 

tests/interfaces/onnx/test_importer.py

@@ -26,7 +26,7 @@ def test_ab():
     mdf_executable = EvaluableGraph(mdf_model.graphs[0], verbose=False)
 
     mdf_executable.evaluate(initializer={"input": test_input})
-    mdf_output = mdf_executable.enodes["Mul_3"].evaluable_outputs["_4"].curr_value
+    mdf_output = mdf_executable.enodes["/B/Mul"].evaluable_outputs["_4"].curr_value
 
     # Get the translated ONNX model
     onnx_models = mdf_to_onnx(mdf_model)