fix: put_back_round

concrete/quantization/quantized_activations.py

@@ -55,7 +55,9 @@ class QuantizedActivation(ABC):
         assert self.q_out is not None
 
         qoutput_activation = qoutput_activation / self.q_out.scale + self.q_out.zero_point
-        qoutput_activation = (qoutput_activation).clip(0, 2 ** self.q_out.n_bits - 1).astype(int)
+        qoutput_activation = (
+            numpy.rint(qoutput_activation).clip(0, 2 ** self.q_out.n_bits - 1).astype(int)
+        )
 
         # TODO find a better way to do the following (see issue #832)
         q_out = copy.copy(self.q_out)

concrete/quantization/quantized_array.py

@@ -59,7 +59,7 @@ class QuantizedArray:
         qvalues = self.values / scale + zero_point
 
         qvalues = (
-            qvalues.round()
+            numpy.rint(qvalues)
             .clip(-self.offset, 2 ** (self.n_bits) - 1 - self.offset)
             .astype(int)  # Careful this can be very large with high number of bits
         )
@@ -100,8 +100,7 @@ class QuantizedArray:
         """
 
         self.qvalues = (
-            (self.values / self.scale + self.zero_point)
-            .round()
+            numpy.rint(self.values / self.scale + self.zero_point)
             .clip(-self.offset, 2 ** (self.n_bits) - 1 - self.offset)
             .astype(int)
         )

concrete/quantization/quantized_layers.py

@@ -96,7 +96,7 @@ class QuantizedLinear:
         numpy_q_out = m_matmul * numpy_q_out
         numpy_q_out = self.q_out.zero_point + bias_part + numpy_q_out
 
-        numpy_q_out = numpy_q_out.clip(0, 2 ** self.q_out.n_bits - 1).astype(int)
+        numpy_q_out = numpy.rint(numpy_q_out).clip(0, 2 ** self.q_out.n_bits - 1).astype(int)
 
         # TODO find a more intuitive way to do the following (see issue #832)
         # We should be able to reuse q_out quantization parameters