utils.py

import random
import math

import numpy as np
from scipy.stats import t
import tensorflow as tf
from tensorflow.keras import layers, models


def get_random_number():
    mu, sigma = 0, 0.1 # 均值和标准差
    x = []
    while len(x) < 1:
        r = random.uniform(0, 1) # 生成0到1之间的随机数
        y = math.sqrt(-2 * math.log(r)) * math.cos(2 * math.pi * r) # 使用Box-Muller变换生成正态分布随机数
        y = y * sigma + mu # 调整均值和标准差
        if -0.2 <= y <= 0.2: # 限制随机数在-0.1到0.1之间
            x.append(y)
    return x[0]


def get_t_random_number():
    # 设置自由度和均值
    df = 10
    mean = 0

    # 生成一个t分布的随机数
    x = t.rvs(df, loc=mean)

    # 缩放和平移
    x_scaled = (x - mean) / np.sqrt(df / (df - 2))
    x_shifted = x_scaled * 0.1

    # 将结果限制在-0.1到0.1的范围内
    if x_shifted < -0.1:
        x_shifted = -0.1
    elif x_shifted > 0.1:
        x_shifted = 0.1

    return x_shifted


def get_poly_random_number(num_points):
    # 假设有10个点，x和y坐标都相同
    x = np.zeros(num_points)
    y = np.zeros(num_points)

    # 生成随机半径
    r = np.random.uniform(0, 1, size=num_points)

    # 生成随机角度
    theta = np.random.uniform(0, 2 * np.pi, size=num_points)

    # 将极坐标转换为直角坐标
    x = r * np.cos(theta)
    y = r * np.sin(theta)

    # 将x和y坐标平移，使得它们的均值为0
    x = x - np.mean(x)
    y = y - np.mean(y)

    # 将x和y坐标缩放，使得它们的标准差为1
    x = x / np.std(x)
    y = y / np.std(y)

    # 将x和y坐标缩放，使得它们的范围在-0.1到0.1之间
    x = x * 0.1
    y = y * 0.1

    return x, y


def compute_loss_coefficient(step, total_steps, decay_factor):
    # Ensure step is within the range [1, total_steps]
    step = max(1, min(step, total_steps))

    # Calculate the loss coefficient for the current step
    loss_coefficient = decay_factor ** (total_steps - step)

    return loss_coefficient


def residual_block(input_data, filters, stride=1):
    shortcut = input_data

    # 第一个卷积层
    x = layers.Conv2D(filters, 3, strides=stride, padding='same',
                      kernel_regularizer=tf.keras.regularizers.l2(0.01))(input_data)
    x = layers.BatchNormalization()(x)
    x = layers.ReLU()(x)

    # 第二个卷积层
    x = layers.Conv2D(filters, 3, padding='same',
                      kernel_regularizer=tf.keras.regularizers.l2(0.01))(x)
    x = layers.BatchNormalization()(x)
    x = layers.ReLU()(x)

    # 如果输入和输出尺寸不一致，使用1x1卷积调整shortcut的尺寸
    if stride > 1 or input_data.shape[-1] != filters:
        shortcut = layers.Conv2D(filters, 1, strides=stride, padding='same',
                                 kernel_regularizer=tf.keras.regularizers.l2(0.01))(shortcut)
        shortcut = layers.BaStchNormalization()(shortcut)
        shortcut = layers.ReLU()(shortcut)

    # 添加卷积和dropout
    shortcut = layers.Conv2D(filters, 3, padding='same',
                             kernel_regularizer=tf.keras.regularizers.l2(0.01))(shortcut)
    shortcut = layers.BatchNormalization()(shortcut)
    shortcut = layers.Dropout(0.3)(shortcut)
    shortcut = layers.ReLU()(shortcut)
    shortcut = layers.Conv2D(filters, 3, padding='same',
                             kernel_regularizer=tf.keras.regularizers.l2(0.01))(shortcut)
    shortcut = layers.BatchNormalization()(shortcut)
    shortcut = layers.ReLU()(shortcut)

    # 残差连接
    x = layers.Add()([x, shortcut])
    x = layers.ReLU()(x)
    return x


def create_wrn28_10():
    input_data = layers.Input(shape=(28, 28, 1))
    x = layers.Conv2D(16, 3, padding='same',
                      kernel_regularizer=tf.keras.regularizers.l2(0.01))(input_data)
    x = layers.BatchNormalization()(x)
    x = layers.ReLU()(x)

    # 第一个残差块组
    x = residual_block(x, 16)
    for _ in range(9):
        x = residual_block(x, 16)

    # 第二个残差块组
    x = residual_block(x, 32, stride=2)
    for _ in range(9):
        x = residual_block(x, 32)

    # 第三个残差块组
    x = residual_block(x, 64, stride=2)
    for _ in range(9):
        x = residual_block(x, 64)

    # 全局平均池化层
    x = layers.GlobalAveragePooling2D()(x)

    # 输出层
    output = layers.Dense(10, activation='softmax',
                          kernel_regularizer=tf.keras.regularizers.l2(0.01))(x)

    model = models.Model(inputs=input_data, outputs=output)
    return model


if __name__ == "__main__":
    # print(get_poly_random_number(1000))
    model = create_wrn28_10()
    print(model.summary())