如何让HMM在Tensorflow中处理实值数据

pip install tensorflow numpy scipy

import tensorflow as tf
import tensorflow_probability as tfp

tfd = tfp.distributions

class ContinuousHMM:
    def __init__(self, num_states, num_steps, obs_dim):
        self.num_states = num_states
        self.num_steps = num_steps
        self.obs_dim = obs_dim
        
        # 初始状态分布
        self.initial_distribution = tfd.Categorical(probs=tf.ones(num_states) / num_states)
        
        # 状态转移矩阵
        self.transition_distribution = tfd.Categorical(probs=tf.ones((num_states, num_states)) / num_states)
        
        # 观测分布（使用高斯混合模型）
        self.observation_distribution = tfd.MixtureSameFamily(
            mixture_distribution=tfd.Categorical(probs=tf.ones(num_states) / num_states),
            components_distribution=tfd.MultivariateNormalDiag(
                loc=tf.Variable(tf.random.normal([num_states, obs_dim])),
                scale_diag=tf.Variable(tf.ones([num_states, obs_dim]))
            )
        )

    def log_prob(self, observations):
        return tf.reduce_sum(self.observation_distribution.log_prob(observations))

# 示例参数
num_states = 3
num_steps = 10
obs_dim = 2

# 创建HMM模型实例
hmm_model = ContinuousHMM(num_states, num_steps, obs_dim)

# 生成模拟数据
observations = tf.random.normal([num_steps, obs_dim])

# 计算对数概率
log_prob = hmm_model.log_prob(observations)
print("Log Probability:", log_prob.numpy())