pass parallelization_factor, rename rule change

src/HGQ/proxy/convert.py

@@ -127,8 +127,6 @@ def _get_out_tensor(
     if len(inps) == 1:
         inps = inps[0]
 
-    name = namer.next_name(layer.name)
-    layer._name = name
     satisfied[node] = apply_layer(layer, inps, namer=namer, layer_xformer=layer_xformer)
     return satisfied[node]
 
@@ -149,6 +147,7 @@ def apply_layer(
 
     layer_xf = layer_xformer(layer)
     n = 0
+    namer = namer or Namer()
     while layer_xf is not layer:
         if isinstance(layer_xf, keras.Model):
             # layer_transformer may return a keras.Model from a layer.
@@ -157,6 +156,10 @@ def apply_layer(
         layer = layer_xf
         layer_xf = layer_xformer(layer)
         n += 1
+
+    name = namer.next_name(layer_xf.name)
+    layer_xf._name = name
+
     if layer_xf is not None:
         # Remove all inbound and outbound nodes to clean up the graph.
         layer_xf = copy(layer_xf)
@@ -306,6 +309,11 @@ def _(self, layer: HLayerBase):
         else:
             inputs = [keras.layers.Input(shape=shape[1:]) for shape in input_shape]
 
+        overrides = {'layers': {}}
+        if hasattr(layer, 'parallel_factor'):
+            parallel_factor = layer.parallel_factor
+            overrides = {'layers': {name: {'parallelization_factor': int(parallel_factor)}}}
+
         if not isinstance(klayer, keras.layers.Activation):
             if layer._relu_act:  # a non-activation layer can only have relu or liner as activation
                 k, i, f = layer.paq.get_bits_exact(pos_only=False)
@@ -315,11 +323,11 @@ def _(self, layer: HLayerBase):
 
                 f_add_bits = 0 if R == 'TRN' else 1 if R == 'RND' else 2
                 fq1 = FixedPointQuantizer(k, b + f_add_bits, i, SAT=S, RND='TRN', name=f'{name}_quantizer')
-                fq2 = FixedPointQuantizer(rk, rb, ri, SAT=S, RND=R, name=f'{name}_relu_quantizer')
+                fq2 = FixedPointQuantizer(rk, rb, ri, SAT=S, RND=R, name=f'{name}_relu_quantizer', overrides=overrides)
                 return keras.Model(inputs, fq2(keras.layers.ReLU()(fq1(klayer(inputs)))))
 
         k, b, i, R, S = self.get_kbiRS(layer)
-        q = FixedPointQuantizer(k, b, i, R, SAT=S, name=f'{name}_quantizer')
+        q = FixedPointQuantizer(k, b, i, R, SAT=S, name=f'{name}_quantizer', overrides=overrides)
         return keras.Model(inputs, q(klayer(inputs)))
 
     @__call__.register