Merge branch 'fastmachinelearning:main' into dev

.github/workflows/test.yml

@@ -15,7 +15,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ['3.8', '3.9', '3.10', '3.11']
+        python-version: ['3.8', '3.9', '3.10']
 
     steps:
       - uses: actions/checkout@v2

docs/qonnx-custom-ops/quant_op.md

@@ -19,7 +19,7 @@ This operator is not part of the ONNX standard and is not currently versioned.
 <dt><tt>narrow</tt> : int (default is 0)</dt>
 <dd>Defines if the value range should be interpreted as narrow, when signed=1. E.g. at 8b regular=[-128, 127] vs narrow=[-127, 127].</dd>
 <dt><tt>rounding_mode</tt> : string (default is "ROUND")</dt>
-<dd>Defines how rounding should be applied during quantization. Currently available modes are: "ROUND", "CEIL" and "FLOOR". Here "ROUND" implies a round-to-even operation.</dd>
+<dd>Defines how rounding should be applied during quantization. Currently available modes are: "ROUND", "CEIL" and "FLOOR". Here "ROUND" implies a round-to-even operation. Lowercase variants for the rounding mode string are also supported: "round", "ceil", "floor".</dd>
 </dl>
 
 #### Inputs

docs/qonnx-custom-ops/trunc_op.md

@@ -12,7 +12,7 @@ This operator is not part of the ONNX standard and is not currently versioned.
 
 <dl>
 <dt><tt>rounding_mode</tt> : string (default is "FLOOR")</dt>
-<dd>Defines how rounding should be applied during truncation. Currently available modes are: "ROUND", "CEIL" and "FLOOR". Here "ROUND" implies a round-to-even operation.</dd>
+<dd>Defines how rounding should be applied during truncation. Currently available modes are: "ROUND", "CEIL" and "FLOOR". Here "ROUND" implies a round-to-even operation. Lowercase variants for the rounding mode string are also supported: "round", "ceil", "floor".</dd>
 </dl>
 
 #### Inputs

src/qonnx/custom_op/general/maxpoolnhwc.py

@@ -97,7 +97,10 @@ def execute_node(self, context, graph):
         inp_vi = helper.make_tensor_value_info(inp_name, TensorProto.FLOAT, inp.shape)
         out_vi = helper.make_tensor_value_info(out_name, TensorProto.FLOAT, dummy_out.shape)
         tmp_graph = helper.make_graph(nodes=[node], name="tmp_graph", inputs=[inp_vi], outputs=[out_vi])
-        tmp_model = qonnx_make_model(tmp_graph, producer_name="finn")
+        opset_version = self.onnx_opset_version
+        opset_imports = [helper.make_opsetid("", opset_version)]
+        onnx_kwargs = {"opset_imports": opset_imports}
+        tmp_model = qonnx_make_model(tmp_graph, producer_name="finn", **onnx_kwargs)
         tmp_model = ModelWrapper(tmp_model)
         new_ctx = {inp_name: inp}
         from qonnx.core.onnx_exec import execute_onnx

src/qonnx/custom_op/general/quant.py

@@ -134,14 +134,15 @@ def quant(inp_tensor, scale, zeropt, bitwidth, signed, narrow, rounding_mode):
 def resolve_rounding_mode(mode_string):
     """Resolve the rounding mode string of Quant and Trunc ops
     to the corresponding numpy functions."""
-    if mode_string == "ROUND":
+    normalized_mode_string = mode_string.upper()
+    if normalized_mode_string == "ROUND":
         return np.round
-    elif mode_string == "CEIL":
+    elif normalized_mode_string == "CEIL":
         return np.ceil
-    elif mode_string == "FLOOR":
+    elif normalized_mode_string == "FLOOR":
         return np.floor
     else:
-        raise ValueError(f"Could not resolve rounding mode called: {mode_string}")
+        raise ValueError(f"Could not resolve rounding mode called: {normalized_mode_string}")
 
 
 class Quant(CustomOp):

src/qonnx/custom_op/general/quantavgpool2d.py

@@ -131,7 +131,11 @@ def execute_node(self, context, graph):
             inputs=[inp],
             outputs=[outp],
         )
-        model_avgpool = qonnx_make_model(graph_avgpool)
+
+        opset_version = self.onnx_opset_version
+        opset_imports = [helper.make_opsetid("", opset_version)]
+        onnx_kwargs = {"opset_imports": opset_imports}
+        model_avgpool = qonnx_make_model(graph_avgpool, **onnx_kwargs)
         idict = {node.input[0]: inp_values}
         sess = rt.InferenceSession(model_avgpool.SerializeToString())
         result_temp = sess.run(None, idict)

src/qonnx/transformation/merge_onnx_models.py

@@ -63,6 +63,14 @@ def apply(self, model):
         # to avoid mix-ups, start by giving all tensors random names
         pre_model = pre_model.transform(GiveRandomTensorNames())
         post_model = post_model.transform(GiveRandomTensorNames())
+        pre_model_opset = pre_model.model.opset_import[0].version
+        post_model_opset = post_model.model.opset_import[0].version
+        merged_model_opset = max(pre_model_opset, post_model_opset)
+        if pre_model_opset != post_model_opset:
+            warnings.warn(
+                "[MergeONNXModels] opsets for models to merge differ: %d vs %d, output model will use opset %d"
+                % (pre_model_opset, post_model_opset, merged_model_opset)
+            )
 
         # check for dynamic outputs of pre model
         dyn_outp = []
@@ -143,7 +151,9 @@ def apply(self, model):
             value_info=vi_new,
         )
 
-        new_model = qonnx_make_model(new_graph, producer_name="fuse_model")
+        new_model = qonnx_make_model(
+            new_graph, producer_name="fuse_model", opset_imports=[helper.make_opsetid("", merged_model_opset)]
+        )
         new_model = ModelWrapper(new_model)
 
         for i in init_new:

tests/transformation/test_merge_onnx_models.py

@@ -73,7 +73,8 @@ def test_merge_onnx_models():
         value_info=[a0, a1],
     )
 
-    model2 = qonnx_make_model(graph, producer_name="model2")
+    exp_opset_id = 13
+    model2 = qonnx_make_model(graph, producer_name="model2", opset_imports=[helper.make_opsetid("", exp_opset_id)])
     model2 = ModelWrapper(model2)
     # initialize model2
     a0_value = np.random.uniform(low=0, high=1, size=(1)).astype(np.float32)
@@ -122,3 +123,5 @@ def test_merge_onnx_models():
 
     # check if finn datatype of graph.input[0] is still set to UINT8
     assert model_transformed.get_tensor_datatype("global_in") == DataType["UINT8"]
+    # check that the merged model uses the greater of the two input opsets
+    assert model_transformed.model.opset_import[0].version == exp_opset_id