Hello everyone on the forum
In my process of learning PyTorch quantization, I implemented convolution in two ways. Given quantized input and weight tensors, along with the scale and zero_point for the input, weight, and output tensors:
The first approach involved converting the quantized input and weight tensors back to floating-point numbers, performing 2D convolution using the floating-point function nn.Conv2d, and then quantizing the result back to integers.
The second approach directly used the quantized.Conv2d function to perform quantized convolution.
After comparing the results from both methods, I noticed a considerable discrepancy. What could be the reason for this?
import torch.nn as nn
import torch.quantization
# 量化参数设置 (示例值，根据实际需求修改)
input_scale = 0.012728
input_zero_point = 33
weight_scale = 0.00292
weight_zero_point = -11
activation_scale = 0.02091#0.02086#0.020389
activation_zero_point = 0
# 读取32x32输入张量 (假设CSV是32行32列)#
input_tensor = torch.tensor([[[[200.,200.,200.,200.,200.],
[200.,200.,200.,200.,200.],
[200.,200.,200.,200.,200.],
[200.,200.,200.,200.,200.],
[200.,200.,200.,200.,200.]]]],dtype=torch.float32)
# 读取5x5权重矩阵 (假设CSV是5行5列)
weight_matrix =torch.tensor( [[[[ 9, -32, 17, 3, -67],
[ -97, -41, 47, 84, -56],
[-128, -125, 11, 63, 54],
[ -15, -58, -31, 50, 48],
[ 23, -3, 64, -87, -91]]]],dtype=torch.int8)
bias = torch.tensor(0, dtype=torch.float32)
bias = bias.reshape(-1)
print(input_tensor)
print(weight_matrix)
#===========浮点模型计算=================
# 创建浮点模型
float_model = nn.Conv2d(
in_channels=1,
out_channels=1,
kernel_size=5,
stride=1,
padding=0,
bias=True
)
float_weight_matrix = weight_scale*(weight_matrix - weight_zero_point)
# 手动设置权重
with torch.no_grad():
float_model.weight = nn.Parameter(float_weight_matrix)
float_model.bias = nn.Parameter(bias)
float_input_tensor = input_scale*(input_tensor - input_zero_point)
print(float_model.weight)
print(float_input_tensor)
# 执行量化卷积计算
float_output = float_model(float_input_tensor)
q_output = torch.quantize_per_tensor(
float_output,
scale= activation_scale,
zero_point=activation_zero_point,
dtype=torch.quint8
)
q_output_int = q_output.int_repr().squeeze().numpy()
#=======================量化版本=========================
#直接使用量化Conv2d
quant_conv = torch.ao.nn.quantized.Conv2d(1, 1, 5, bias = False)
quant_conv.scale = activation_scale
quant_conv.zero_point = activation_zero_point
# 量化权重张量
quant_weight = torch.quantize_per_tensor(
float_weight_matrix,
scale=weight_scale,
zero_point=weight_zero_point,
dtype=torch.qint8
)
# 设置权重（使用浮点权重）
quant_conv.set_weight_bias(quant_weight,None )
# 量化输入张量
quant_input = torch.quantize_per_tensor(
float_input_tensor,
scale=input_scale,
zero_point=input_zero_point,
dtype=torch.quint8
)
print("Actual quantized weight:", quant_conv.weight)
print("Integer representation:", quant_conv.weight().int_repr())
# 执行量化卷积
quant_output = quant_conv(quant_input)
print("float_output float is:", float_output)
print("float_output int is:", q_output_int)
print("quant_output float is:",quant_output)
print("quant_output int is:",quant_output.int_repr())
print("compute done!")
the result is
