请求将基于Theano的几何广告衰减代码转换为PyTensor版本

import pytensor.tensor as pt

def adstock_geometric_pytensor(x, theta):
    x = pt.as_tensor_variable(x)
    
    def adstock_geometric_recurrence_pytensor(index, input_x, decay_x, theta):
        # 计算当前步的衰减值并更新到decay_x对应位置
        updated_decay = pt.set_subtensor(decay_x[index], input_x + theta * decay_x[index - 1])
        return updated_decay
    
    len_observed = x.shape[0]
    x_decayed = pt.zeros_like(x)
    # 初始化第一个时间步的衰减值
    x_decayed = pt.set_subtensor(x_decayed[0], x[0])
    
    output, _ = pt.scan(
        fn=adstock_geometric_recurrence_pytensor,
        sequences=[pt.arange(1, len_observed), x[1:len_observed]],
        outputs_info=x_decayed,
        non_sequences=theta,
        n_steps=len_observed - 1
    )
    
    return output[-1]

关键修改说明 #

• 替换Theano的tensor模块为PyTensor的pytensor.tensor（导入后用pt别名）
• 将theano.scan替换为pt.scan，PyTensor的scan接口与Theano基本兼容，参数无需大幅调整
• 移除了原代码中多余的tt.sum——因为input_x和decay_x[index-1]都是单时间步的标量，相加后本身就是标量，无需求和
• 递归函数名称调整为匹配PyTensor上下文，其余变量命名保持与原代码一致

内容的提问来源于stack exchange，提问作者Shuvayan Das