rm redundant

Author: Jiang-Stan

sparsebit/quantization/observers/aciq.py

@@ -61,8 +61,6 @@ def __init__(self, config, qdesc):
             8: 11.16,
         }
         self.gaus_const = (0.5 * 0.35) * (1 + (math.pi * math.log(4)) ** 0.5)
-        self.distribution = config.OBSERVER.ACIQ.DISTRIBUTION.lower()
-        assert self.distribution in ["gaus", "laplace"]
 
     def calc_laplace_minmax(self):
         if self.is_perchannel:

sparsebit/quantization/observers/mse.py

@@ -17,11 +17,9 @@ def __init__(self, config, qdesc):
 
     def calc_minmax(self, data_c_first):
         if self.is_perchannel:
-            data_c_first = self.get_calibration_data(c_first=True)
             max_val = data_c_first.max(axis=1).values
             min_val = data_c_first.min(axis=1).values
         else:
-            data_c_first = self.get_calibration_data(c_first=False)
             min_val, max_val = data_c_first.min(), data_c_first.max()
         self.min_val = min_val.to(self.device)
         self.max_val = max_val.to(self.device)

sparsebit/quantization/observers/percentile.py

@@ -14,6 +14,7 @@ def __init__(self, config, qdesc):
         self.alpha = config.OBSERVER.PERCENTILE.ALPHA
 
     def calc_minmax(self):
+
         if self.is_perchannel:
             data = self.data_cache.get_data_for_calibration(Granularity.CHANNELWISE)
         else:

sparsebit/quantization/quant_model.py

@@ -246,6 +246,7 @@ def export_onnx(
         extra_info=False,
     ):
         self.eval()
+        self.set_quant(w_quant=True, a_quant=True)  # quant must prepared before export
         for n, m in self.model.named_modules():
             if isinstance(m, Quantizer):
                 m.enable_export_onnx()

sparsebit/quantization/quantizers/lsq.py

@@ -42,10 +42,8 @@ def calc_qparams(self):
                 )
                 self.qdesc.set_symmetric(True)
             if self.is_perchannel:
-                x_oc = self.observer.get_calibration_data(c_first=True)
                 scale = 2 * x_oc.abs().mean(axis=1) / math.sqrt(self.qdesc.qmax)
             else:
-                x_oc = self.observer.get_calibration_data(c_first=False)
                 scale = 2 * x_oc.abs().mean() / math.sqrt(self.qdesc.qmax)
             self.scale = nn.Parameter(self._broadcast_qparams(scale.to(self.device)))
             self.zero_point = self._broadcast_qparams(torch.zeros_like(self.scale))