Add bias quantization in QAT and refactor the code of weight quantization (#2914)

src/sdk/pynni/nni/compression/torch/compressor.py

@@ -481,11 +481,10 @@ class QuantizerModuleWrapper(torch.nn.Module):
                 self)
 
         if 'weight' in self.config['quant_types'] and _check_weight(self.module):
-            new_weight = self.quantizer.quant_grad.apply(
+            self.quantizer.quant_grad.apply(
                 self.module.old_weight,
                 QuantType.QUANT_WEIGHT,
                 self)
-            self.module.weight = new_weight
             result = self.module(*inputs)
         else:
             result = self.module(*inputs)
@@ -617,7 +616,7 @@ class QuantGrad(torch.autograd.Function):
         if quant_type == QuantType.QUANT_INPUT:
             return wrapper.quantizer.quantize_input(tensor, wrapper, **kwargs)
         elif quant_type == QuantType.QUANT_WEIGHT:
-            return wrapper.quantizer.quantize_weight(tensor, wrapper, **kwargs)
+            return wrapper.quantizer.quantize_weight(wrapper, **kwargs)
         elif quant_type == QuantType.QUANT_OUTPUT:
             return wrapper.quantizer.quantize_output(tensor, wrapper, **kwargs)
         else:

src/sdk/pynni/nni/compression/torch/quantization/quantizers.py

@@ -2,6 +2,7 @@
 # Licensed under the MIT license.
 
 import logging
+import copy
 import torch
 from schema import Schema, And, Or, Optional
 from ..utils.config_validation import CompressorSchema
@@ -15,6 +16,7 @@ logger = logging.getLogger(__name__)
 class NaiveQuantizer(Quantizer):
     """quantize weight to 8 bits
     """
+
     def __init__(self, model, config_list, optimizer=None):
         super().__init__(model, config_list, optimizer)
         self.layer_scale = {}
@@ -29,13 +31,15 @@ class NaiveQuantizer(Quantizer):
 
         schema.validate(config_list)
 
-    def quantize_weight(self, weight, wrapper, **kwargs):
+    def quantize_weight(self, wrapper, **kwargs):
+        weight = copy.deepcopy(wrapper.module.old_weight.data)
         new_scale = weight.abs().max() / 127
         scale = max(self.layer_scale.get(wrapper.name, 0), new_scale)
         self.layer_scale[wrapper.name] = scale
         orig_type = weight.type()  # TODO: user layer
-        return weight.div(scale).type(torch.int8).type(orig_type).mul(scale)
-
+        weight = weight.div(scale).type(torch.int8).type(orig_type).mul(scale)
+        wrapper.module.weight = weight
+        return weight
 
 def update_ema(biased_ema, value, decay, step):
     """
@@ -60,6 +64,7 @@ def update_ema(biased_ema, value, decay, step):
     unbiased_ema = biased_ema / (1 - decay ** step)  # Bias correction
     return biased_ema, unbiased_ema
 
+
 def update_quantization_param(bits, rmin, rmax):
     """
     calculate the `zero_point` and `scale`.
@@ -116,6 +121,7 @@ class QAT_Quantizer(Quantizer):
     Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference
     http://openaccess.thecvf.com/content_cvpr_2018/papers/Jacob_Quantization_and_Training_CVPR_2018_paper.pdf
     """
+
     def __init__(self, model, config_list, optimizer=None):
         """
         Parameters
@@ -215,20 +221,35 @@ class QAT_Quantizer(Quantizer):
         real_val = op.scale * (quantized_val - op.zero_point)
         return real_val
 
-    def quantize_weight(self, weight, wrapper, **kwargs):
+    def quantize_weight(self, wrapper, **kwargs):
         config = wrapper.config
         module = wrapper.module
+        weight = copy.deepcopy(wrapper.module.old_weight.data)
         weight_bits = get_bits_length(config, 'weight')
         quant_start_step = config.get('quant_start_step', 0)
         assert weight_bits >= 1, "quant bits length should be at least 1"
 
         if quant_start_step > self.steps:
             return weight
+
+        # if bias exists, quantize bias to uint32
+        if hasattr(wrapper.module, 'bias') and wrapper.module.bias is not None:
+            bias = wrapper.module.bias.data
+            bias_bits = 32
+            rmin, rmax = torch.min(bias), torch.max(bias)
+            module.scale, module.zero_point = update_quantization_param(bias_bits, rmin, rmax)
+            bias = self._quantize(bias_bits, module, bias)
+            bias = self._dequantize(module, bias)
+            wrapper.module.bias.data = bias
+
+
+        # quantize weight
         rmin, rmax = torch.min(weight), torch.max(weight)
         module.scale, module.zero_point = update_quantization_param(weight_bits, rmin, rmax)
-        out = self._quantize(weight_bits, module, weight)
-        out = self._dequantize(module, out)
-        return out
+        weight = self._quantize(weight_bits, module, weight)
+        weight = self._dequantize(module, weight)
+        wrapper.module.weight = weight
+        return weight
 
     def quantize_output(self, output, wrapper, **kwargs):
         config = wrapper.config
@@ -241,8 +262,10 @@ class QAT_Quantizer(Quantizer):
             return output
 
         current_min, current_max = torch.min(output), torch.max(output)
-        module.tracked_min_biased, module.tracked_min = update_ema(module.tracked_min_biased, current_min, module.ema_decay, self.steps)
-        module.tracked_max_biased, module.tracked_max = update_ema(module.tracked_max_biased, current_max, module.ema_decay, self.steps)
+        module.tracked_min_biased, module.tracked_min = update_ema(module.tracked_min_biased, current_min,
+                                                                   module.ema_decay, self.steps)
+        module.tracked_max_biased, module.tracked_max = update_ema(module.tracked_max_biased, current_max,
+                                                                   module.ema_decay, self.steps)
         module.scale, module.zero_point = update_quantization_param(output_bits, module.tracked_min, module.tracked_max)
         out = self._quantize(output_bits, module, output)
         out = self._dequantize(module, out)
@@ -264,6 +287,7 @@ class DoReFaQuantizer(Quantizer):
     Zhou et al., DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients
     (https://arxiv.org/abs/1606.06160)
     """
+
     def __init__(self, model, config_list, optimizer=None):
         super().__init__(model, config_list, optimizer)
 
@@ -287,17 +311,20 @@ class DoReFaQuantizer(Quantizer):
 
         schema.validate(config_list)
 
-    def quantize_weight(self, weight, wrapper, **kwargs):
+    def quantize_weight(self, wrapper, **kwargs):
+        weight = copy.deepcopy(wrapper.module.old_weight.data)
         weight_bits = get_bits_length(wrapper.config, 'weight')
-        out = weight.tanh()
-        out = out / (2 * out.abs().max()) + 0.5
-        out = self.quantize(out, weight_bits)
-        out = 2 * out -1
-        return out
+        weight = weight.tanh()
+        weight = weight / (2 * weight.abs().max()) + 0.5
+        weight = self.quantize(weight, weight_bits)
+        weight = 2 * weight - 1
+        wrapper.module.weight = weight
+        # wrapper.module.weight.data = weight
+        return weight
 
     def quantize(self, input_ri, q_bits):
-        scale = pow(2, q_bits)-1
-        output = torch.round(input_ri*scale)/scale
+        scale = pow(2, q_bits) - 1
+        output = torch.round(input_ri * scale) / scale
         return output
 
 
@@ -314,6 +341,7 @@ class BNNQuantizer(Quantizer):
     Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1
     (https://arxiv.org/abs/1602.02830)
     """
+
     def __init__(self, model, config_list, optimizer=None):
         super().__init__(model, config_list, optimizer)
         self.quant_grad = ClipGrad
@@ -339,11 +367,13 @@ class BNNQuantizer(Quantizer):
 
         schema.validate(config_list)
 
-    def quantize_weight(self, weight, wrapper, **kwargs):
-        out = torch.sign(weight)
+    def quantize_weight(self, wrapper, **kwargs):
+        weight = copy.deepcopy(wrapper.module.old_weight.data)
+        weight = torch.sign(weight)
         # remove zeros
-        out[out == 0] = 1
-        return out
+        weight[weight == 0] = 1
+        wrapper.module.weight = weight
+        return weight
 
     def quantize_output(self, output, wrapper, **kwargs):
         out = torch.sign(output)

src/sdk/pynni/tests/test_compressor_torch.py

@@ -234,20 +234,34 @@ class CompressorTestCase(TestCase):
         model.relu = torch.nn.ReLU()
         quantizer = torch_compressor.QAT_Quantizer(model, config_list)
         quantizer.compress()
+
         # test quantize
         # range not including 0
         eps = 1e-7
         weight = torch.tensor([[1, 2], [3, 5]]).float()
-        quantize_weight = quantizer.quantize_weight(weight, model.conv2)
+        model.conv2.module.old_weight.data = weight
+        quantizer.quantize_weight(model.conv2)
         assert math.isclose(model.conv2.module.scale, 5 / 255, abs_tol=eps)
         assert model.conv2.module.zero_point == 0
         # range including 0
         weight = torch.tensor([[-1, 2], [3, 5]]).float()
-        quantize_weight = quantizer.quantize_weight(weight, model.conv2)
+        model.conv2.module.old_weight.data = weight
+        quantizer.quantize_weight(model.conv2)
         assert math.isclose(model.conv2.module.scale, 6 / 255, abs_tol=eps)
         assert model.conv2.module.zero_point in (42, 43)
+        # test value of weight and bias after quantization
+        weight = torch.tensor([[1.1287, 2.3456], [3.7814, 5.9723]])
+        weight_valid = torch.tensor([[1.1242, 2.3421], [3.7707, 5.9723]])
+        bias = torch.tensor([2.3432, 3.4342, 1.3414, 5.2341])
+        bias_valid = torch.tensor([2.3432, 3.4342, 1.3414, 5.2341])
+        model.conv2.module.old_weight.data = weight
+        model.conv2.module.bias.data = bias
+        quantizer.quantize_weight(model.conv2)
+        assert torch.all(torch.isclose(model.conv2.module.weight.data, weight_valid, rtol=1e-4))
+        assert torch.all(torch.isclose(model.conv2.module.bias.data, bias_valid, rtol=1e-7))
 
         # test ema
+        eps = 1e-7
         x = torch.tensor([[-0.2, 0], [0.1, 0.2]])
         out = model.relu(x)
         assert math.isclose(model.relu.module.tracked_min_biased, 0, abs_tol=eps)