基于TensorFlow的信道感知深度学习模型

with tf.name_scope('channel_sensing'):
    # 定义四个可训练的权重矩阵和偏置向量，初始化方式为he_init
    A1 = tf.get_variable('A1', shape=[2*tau, 1024], dtype=tf.float32, initializer=he_init)
        # tf.get_variable: 用于获取一个已存在的变量或者创建一个新的变量。
    A2 = tf.get_variable('A2', shape=[1024, 1024], dtype=tf.float32, initializer=he_init)
    A3 = tf.get_variable('A3', shape=[1024, 1024], dtype=tf.float32, initializer=he_init)
    A4 = tf.get_variable('A4', shape=[1024, 2 * N], dtype=tf.float32, initializer=he_init)

    b1 = tf.get_variable('b1', shape=[1024], dtype=tf.float32, initializer=he_init)
    b2 = tf.get_variable('b2', shape=[1024], dtype=tf.float32, initializer=he_init)
    b3 = tf.get_variable('b3', shape=[1024], dtype=tf.float32, initializer=he_init)
    b4 = tf.get_variable('b4', shape=[2 * N], dtype=tf.float32, initializer=he_init)
    
    # 初始化空列表用于存储权重、后验概率和估计索引
    w_dict = []
    posterior_dict = []
    idx_est_dict = []
    
    # 计算信噪比（SNR）并进行归一化处理
    snr = lay['P'] * tf.ones(shape=[batch_size, 1], dtype=tf.float32)
    snr_dB = tf.log(snr) / np.log(10)
    snr_normal = (snr_dB + 0.5) / np.sqrt(0.75)  # Normalizing for the range -10dB to 30dB

    W_uplink_np = np.random.normal(size=(1,tau, N), loc=0,scale=np.sqrt(0.5))+1j*np.random.normal(size=(1,tau, N), loc=0,scale=np.sqrt(0.5))
    W_uplink_np = W_uplink_np/np.linalg.norm(W_uplink_np,axis=-1,keepdims=True)
    W_uplink = tf.get_variable('W_uplink', dtype=tf.complex64, initializer=W_uplink_np.astype(np.complex64), trainable=False)
    y_noiseless = W_uplink @ tf.reshape(a_phi,(-1,N,1))
    noise = tf.complex(tf.random_normal(tf.shape(y_noiseless), mean=0.0, stddev=noiseSTD_per_dim), \
                         tf.random_normal(tf.shape(y_noiseless), mean=0.0, stddev=noiseSTD_per_dim))
    y_complex = tf.complex(tf.sqrt(lay['P']), 0.0) * y_noiseless + noise
    y_real = tf.concat([tf.real(y_complex), tf.imag(y_complex)], axis=1) / tf.sqrt(lay['P'])
    y_real = tf.reshape(y_real,(-1,tau*2))

    x1 = tf.nn.relu(y_real @ A1 + b1)
    x1 = BatchNormalization()(x1)
    x2 = tf.nn.relu(x1 @ A2 + b2)
    x2 = BatchNormalization()(x2)
    x3 = tf.nn.relu(x2 @ A3 + b3)
    x3 = BatchNormalization()(x3)
    w_her_data = x3 @ A4 + b4
    w_norm_data = tf.reshape(tf.norm(w_her_data, axis=1), (-1, 1))
    w_her_data = tf.divide(w_her_data, w_norm_data)
    w_her_complex_data = tf.complex(w_her_data[:, 0:N], w_her_data[:, N:2 * N])
    y_noiseless_data = tf.reduce_sum(tf.multiply(w_her_complex_data, a_phi), 1, keepdims=True)
    bf_gain = tf.reduce_mean(tf.abs(y_noiseless_data)**2,axis=0)