fix quant grad function calculation error (#3160)

nni/compression/pytorch/compressor.py

@@ -580,17 +580,55 @@ class QuantType:
     """
     Enum class for quantization type.
     """
-    QUANT_INPUT = 0
-    QUANT_WEIGHT = 1
-    QUANT_OUTPUT = 2
+    QUANT_INPUT = 'input'
+    QUANT_WEIGHT = 'weight'
+    QUANT_OUTPUT = 'output'
 
 
 class QuantGrad(torch.autograd.Function):
     """
     Base class for overriding backward function of quantization operation.
     """
+    @classmethod
+    def _quantize(cls, x, scale, zero_point):
+        """
+        Reference function for quantizing x -- non-clamped.
+        Parameters
+        ----------
+        x : Tensor
+            tensor to be quantized
+        scale : Tensor
+            scale for quantizing x
+        zero_point : Tensor
+            zero_point for quantizing x
+        Returns
+        -------
+        tensor
+            quantized x without clamped
+        """
+        return ((x / scale) + zero_point).round()
+    @classmethod
+    def get_bits_length(cls, config, quant_type):
+        """
+        Get bit for quantize config
+        Parameters
+        ----------
+        config : Dict
+            the configuration for quantization
+        quant_type : str
+            quant type
+        Returns
+        -------
+        int
+            n-bits for quantization configuration
+        """
+        if isinstance(config["quant_bits"], int):
+            return config["quant_bits"]
+        else:
+            return config["quant_bits"].get(quant_type)
+
     @staticmethod
-    def quant_backward(tensor, grad_output, quant_type):
+    def quant_backward(tensor, grad_output, scale, zero_point, qmin, qmax):
         """
         This method should be overrided by subclass to provide customized backward function,
         default implementation is Straight-Through Estimator
@@ -600,32 +638,45 @@ class QuantGrad(torch.autograd.Function):
             input of quantization operation
         grad_output : Tensor
             gradient of the output of quantization operation
-        quant_type : QuantType
+        scale : Tensor
             the type of quantization, it can be `QuantType.QUANT_INPUT`, `QuantType.QUANT_WEIGHT`, `QuantType.QUANT_OUTPUT`,
             you can define different behavior for different types.
+        zero_point : Tensor
+            zero_point for quantizing tensor
+        qmin : Tensor
+            quant_min for quantizing tensor
+        qmax : Tensor
+            quant_max for quantizng tensor
         Returns
         -------
         tensor
             gradient of the input of quantization operation
         """
+        tensor_q = QuantGrad._quantize(tensor, scale, zero_point)
+        mask = (tensor_q < qmin) | (tensor_q > qmax)
+        grad_output[mask] = 0
         return grad_output
 
     @staticmethod
     def forward(ctx, tensor, quant_type, wrapper, **kwargs):
-        ctx.save_for_backward(tensor, torch.Tensor([quant_type]))
         if quant_type == QuantType.QUANT_INPUT:
-            return wrapper.quantizer.quantize_input(tensor, wrapper, **kwargs)
+            output = wrapper.quantizer.quantize_input(tensor, wrapper, **kwargs)
         elif quant_type == QuantType.QUANT_WEIGHT:
-            return wrapper.quantizer.quantize_weight(wrapper, **kwargs)
+            output =  wrapper.quantizer.quantize_weight(wrapper, **kwargs)
         elif quant_type == QuantType.QUANT_OUTPUT:
-            return wrapper.quantizer.quantize_output(tensor, wrapper, **kwargs)
+            output = wrapper.quantizer.quantize_output(tensor, wrapper, **kwargs)
         else:
             raise ValueError("unrecognized QuantType.")
 
+        bits = QuantGrad.get_bits_length(wrapper.config, quant_type)
+        qmin, qmax = torch.Tensor([0], device=tensor.device), torch.Tensor([(1 << bits) - 1], device=tensor.device)
+        ctx.save_for_backward(tensor, wrapper.module.scale, wrapper.module.zero_point, qmin, qmax)
+        return output
+
     @classmethod
     def backward(cls, ctx, grad_output):
-        tensor, quant_type = ctx.saved_variables
-        output = cls.quant_backward(tensor, grad_output, quant_type)
+        tensor, scale, zero_point, qmin, qmax = ctx.saved_variables
+        output = cls.quant_backward(tensor, grad_output, scale, zero_point, qmin, qmax)
         return output, None, None, None
 
 def _check_weight(module):