Added compatibility check for identity and relu activation conversion.

src/finn/transformation/convert_qonnx_to_finn.py

@@ -198,6 +198,10 @@ class QuantActBaseHandler(ABC):
     def _remove_activation_node(self):
         pass
 
+    @abstractmethod
+    def _check_compatibility(self):
+        pass
+
     def _extract_output_datatype(self):
         dtype = self._model.get_tensor_datatype(self._q_node.output[0]).name
         if "SCALED" in dtype:
@@ -214,6 +218,8 @@ class QuantActBaseHandler(ABC):
         }
 
     def replace_quant_node(self):
+        # Check that we actually support what the user is trying to do
+        self._check_compatibility()
         # Shorten instance variables
         model = self._model
         graph = model.graph
@@ -334,6 +340,16 @@ class QuantReluHandler(QuantActBaseHandler):
     # zero_pt = model.get_initializer(n.input[2])
     # signed = q_inst.get_nodeattr("signed")
 
+    def _check_compatibility(self):
+        q_inst = getCustomOp(self._q_node)
+        narrow = q_inst.get_nodeattr("narrow")
+        signed = q_inst.get_nodeattr("signed")
+        if signed or narrow:
+            raise ValueError(
+                "FINN only supports unsigned and non-narrow Quant nodes "
+                "for Relu activations."
+            )
+
     def _calculate_act_bias(self):
         # No bias allowed for Relu activations, see: https://github.com/Xilinx/
         # brevitas/blob/a5bfd6dc5e030f0047ac1ee47932b60e8e873e17/src/brevitas/
@@ -413,6 +429,15 @@ class QuantIdentityHandler(QuantActBaseHandler):
     # zero_pt = model.get_initializer(n.input[2])
     # signed = q_inst.get_nodeattr("signed")
 
+    def _check_compatibility(self):
+        # Gather parameters to check
+        q_inst = getCustomOp(self._q_node)
+        signed = q_inst.get_nodeattr("signed")
+        if not signed:
+            raise ValueError(
+                "FINN only supports signed Quant nodes for identity activations."
+            )
+
     def _calculate_act_bias(self):
         # Gather parameters
         bit_width = self._model.get_initializer(self._q_node.input[3])