[math_model/test] Added "QOp" implementation and basic tests.

math_model.py

@@ -14,11 +14,11 @@ def dequantize(tensor, scale, zero_point):
 
 
 class QuantLinear(nn.Module):
-    def __init__(self, quant_param):
+    def __init__(self, in_ch, out_ch, quant_param):
         super().__init__()
         mul_factor = torch.tensor(quant_param['smoothquant_mul']).view(quant_param['smoothquant_mul_shape'])
         self.register_buffer('mul_factor', mul_factor)
-        self.linear = nn.Linear(128, 128)
+        self.linear = nn.Linear(in_ch, out_ch)
         weight_scale = torch.tensor(quant_param['weight_scale']).view(quant_param['weight_scale_shape'])
         weight_zp = torch.tensor(quant_param['weight_zp']).view(quant_param['weight_zp_shape'])
         input_scale = torch.tensor(quant_param['input_scale']).view(quant_param['input_scale_shape'])
@@ -28,10 +28,9 @@ class QuantLinear(nn.Module):
         self.register_buffer('input_scale', input_scale)
         self.register_buffer('input_zp', input_zp)
 
-    def forward(self, x):
+    # I.e., "fake quantization"
+    def qdq_forward(self, x):
         scaled_x = x * self.mul_factor
-        # With an integer conv kernel, if the weight zero point is not zero,
-        # it is required an extra input channel that is equal to the per-channel zero point of the weights
         quant_weight = quantize(self.linear.weight, self.weight_scale, self.weight_zp, is_asym=True)
         quant_input = quantize(scaled_x, self.input_scale, self.input_zp, is_asym=False)
         dequantized_weight = dequantize(quant_weight, self.weight_scale, self.weight_zp)
@@ -39,12 +38,37 @@ class QuantLinear(nn.Module):
         out = torch.nn.functional.linear(dequantized_input, dequantized_weight, self.linear.bias)
         return out
 
+    # Accelerated version
+    def qop_forward(self, x):
+        # With an integer linear kernel, if the weight zero point is not zero,
+        # A correction term must be calculated to correct the output.
+        # The correction term calculated as follows:
+        #  - sum the input tensor across the dot-product dimentions: (e.g., `torch.sum(quant_input, dim=-1)`)
+        #  - multiply this sum with every weight zero-point (e.g., `torch.sum(quant_input, dim=-1) * self.weight_zp`
+        #  - Subtract from previous output (e.g., `quant_output -= torch.sum(quant_input, dim=-1) * self.weight_zp`)
+        #  - All other code is just to make sure the broadcasting semantics work correctly
+        scaled_x = x * self.mul_factor
+        quant_weight = quantize(self.linear.weight, self.weight_scale, self.weight_zp, is_asym=True).to(torch.uint8)
+        quant_input = quantize(scaled_x, self.input_scale, self.input_zp, is_asym=False).to(torch.int8)
+        quant_output = torch.nn.functional.linear(quant_input.to(torch.float32), quant_weight.to(torch.float32), None).to(torch.int32) # Convert inputs to FP32 to avoid F.linear quantizing the output to int8
+        correction = torch.sum(quant_input, dim=-1).to(torch.int32).unsqueeze(-1) * (-self.weight_zp).to(torch.uint8).view([1]*(quant_input.ndim-1) + [self.weight_zp.nelement()]) # Correct for weight zero-point
+        quant_output = quant_output + correction
+        output = dequantize(quant_output, (self.weight_scale * self.input_scale).view([1]*(quant_output.ndim-1) + [(self.weight_scale * self.input_scale).nelement()]), 0.0)
+        output += self.linear.bias
+        return output
+
+    def forward(self, x, qop=False):
+        if qop:
+            return self.qop_forward(x)
+        else:
+            return self.qdq_forward(x)
+
 class QuantConv2d(nn.Module):
-    def __init__(self, quant_param):
+    def __init__(self, in_ch, out_ch, kernel_size, quant_param):
         super().__init__()
         mul_factor = torch.tensor(quant_param['smoothquant_mul']).view(quant_param['smoothquant_mul_shape'])
         self.register_buffer('mul_factor', mul_factor)
-        self.conv2d = nn.Conv2d(128, 128, 3)
+        self.conv2d = nn.Conv2d(in_ch, out_ch, kernel_size)
         weight_scale = torch.tensor(quant_param['weight_scale']).view(quant_param['weight_scale_shape'])
         weight_zp = torch.tensor(quant_param['weight_zp']).view(quant_param['weight_zp_shape'])
         input_scale = torch.tensor(quant_param['input_scale']).view(quant_param['input_scale_shape'])
@@ -54,13 +78,47 @@ class QuantConv2d(nn.Module):
         self.register_buffer('input_scale', input_scale)
         self.register_buffer('input_zp', input_zp)
 
-    def forward(self, x):
+    # I.e., "fake quantization"
+    def qdq_forward(self, x):
         scaled_x = x * self.mul_factor
-        # With an integer conv kernel, if the weight zero point is not zero,
-        # it is required an extra input channel that is equal to the per-channel zero point of the weights
-        quant_weight = quantize(self.linear.weight, self.weight_scale, self.weight_zp, is_asym=True)
+        quant_weight = quantize(self.conv2d.weight, self.weight_scale, self.weight_zp, is_asym=True)
         quant_input = quantize(scaled_x, self.input_scale, self.input_zp, is_asym=False)
         dequantized_weight = dequantize(quant_weight, self.weight_scale, self.weight_zp)
         dequantized_input = dequantize(quant_input, self.input_scale, self.input_zp)
         out = torch.nn.functional.conv2d(dequantized_input, dequantized_weight, self.conv2d.bias)
         return out
+
+    # Accelerated version
+    def qop_forward(self, x):
+        # With an integer conv2d kernel, if the weight zero point is not zero,
+        # A correction term must be calculated to correct the output.
+        # Conceptually, it's identical to the linear case except that it's difficult
+        # to reduce the input across the dot-product dimension. This leaves us with two obvious options:
+        #  1. Manually compute the reduction via Im2Col -> `torch.sum`
+        #  2. Add an extra _output channel_ to the convolution with a kernel made from all ones (e.g., `torch.ones()`)
+        # In this example, I've used option #2.
+        # The correction term is then calculated as follows:
+        #  - Add an extra output channel to the weight tensor with all values equal to 1 to calculate the sum (e.g., `torch.cat((quant_weight, torch.ones(shape)), dim=0)`)
+        #  - Extract the sum from the output tensor (e.g., `sum = quant_output[:,-1,:,:]`)
+        #  - multiply this sum with every weight zero-point (e.g., `sum * self.weight_zp`
+        #  - Subtract from previous output (e.g., `quant_output -= sum * self.weight_zp`)
+        #  - All other code is just to make sure the broadcasting semantics work correctly
+        scaled_x = x * self.mul_factor
+        quant_weight = quantize(self.conv2d.weight, self.weight_scale, self.weight_zp, is_asym=True).to(torch.uint8)
+        b_shape = list(quant_weight.shape) # Used for weight zero-point correction
+        b_shape[0] = 1 # Used for weight zero-point correction
+        weight_cat = torch.ones((1,1,1,1)).broadcast_to(b_shape).to(torch.uint8) # Used for weight zero-point correction
+        quant_weight = torch.cat((quant_weight,weight_cat),dim=0).to(torch.uint8) # Create extra output channel, used for weight zero-point correction
+        quant_input = quantize(scaled_x, self.input_scale, self.input_zp, is_asym=False).to(torch.int8)
+        quant_output = torch.nn.functional.conv2d(quant_input.to(torch.float32), quant_weight.to(torch.float32), None).to(torch.int32) # Convert inputs to FP32 to avoid F.conv2d quantizing the output to int8
+        correction = quant_output[:,-1,:,:] * (-self.weight_zp).to(torch.uint8).view([1, self.weight_zp.nelement()] +  [1]*(quant_output.ndim-2)) # Correct zero-point for weight
+        quant_output = quant_output[:,:-1,:,:] + correction
+        output = dequantize(quant_output, (self.weight_scale * self.input_scale).view([1, (self.weight_scale * self.input_scale).nelement()] + [1]*(quant_output.ndim-2)), 0.0)
+        output += self.conv2d.bias.view([1, self.conv2d.bias.nelement()] + [1]*(quant_output.ndim-2))
+        return output
+
+    def forward(self, x, qop=False):
+        if qop:
+            return self.qop_forward(x)
+        else:
+            return self.qdq_forward(x)

test_quant_conv2d.py

@@ -0,0 +1,34 @@
+import torch
+from math_model import QuantConv2d
+
+torch.manual_seed(0)
+
+batch_size = 1
+out_ch = 8
+in_ch = 4
+k = 3
+h = 5
+w = 5
+
+quant_params = {
+    'smoothquant_mul': torch.rand((in_ch,)),
+    'smoothquant_mul_shape': (1,in_ch,1,1),
+    'weight_scale': torch.rand((out_ch,)),
+    'weight_scale_shape': (out_ch,1,1,1),
+    'weight_zp': torch.randint(-255, 0, (out_ch,)),
+    'weight_zp_shape': (out_ch,1,1,1),
+    'input_scale': torch.rand((1,)),
+    'input_scale_shape': (1,),
+    'input_zp': torch.zeros((1,)),
+    'input_zp_shape': (1,),
+}
+
+print(quant_params)
+
+l = QuantConv2d(in_ch, out_ch, k, quant_params)
+i = torch.rand((batch_size,in_ch,h,w))
+o_qdq = l(i)
+o_qop = l(i, qop=True)
+print(o_qdq.shape)
+print(o_qop.shape)
+print(o_qdq - o_qop)

test_quant_linear.py

@@ -0,0 +1,31 @@
+import torch
+from math_model import QuantLinear
+
+torch.manual_seed(0)
+
+batch_size = 1
+out_ch = 128
+in_ch = 64
+
+quant_params = {
+    'smoothquant_mul': torch.rand((in_ch,)),
+    'smoothquant_mul_shape': (in_ch,),
+    'weight_scale': torch.rand((out_ch,)),
+    'weight_scale_shape': (out_ch,1),
+    'weight_zp': torch.randint(-255, 0, (out_ch,)),
+    'weight_zp_shape': (out_ch,1),
+    'input_scale': torch.rand((1,)),
+    'input_scale_shape': (1,),
+    'input_zp': torch.zeros((1,)),
+    'input_zp_shape': (1,),
+}
+
+print(quant_params)
+
+l = QuantLinear(in_ch, out_ch, quant_params)
+i = torch.rand((batch_size,in_ch))
+o_qdq = l(i)
+o_qop = l(i, qop=True)
+print(o_qdq.shape)
+print(o_qop.shape)
+print(o_qdq - o_qop)