Bayesian MMM with Pytensor in place of Theano

Question

I have a piece of code which performs geometric ad stock decay using Theano. It is a old piece of code and I need to update it using the latest version of PyTensor. Can someone please help me convert this?

def adstock_geometric_theano_pymc3(x, theta):
    x = tt.as_tensor_variable(x)
    
    def adstock_geometric_recurrence_theano(index, 
                                            input_x, 
                                            decay_x,   
                                            theta):
        return tt.set_subtensor(decay_x[index], 
               tt.sum(input_x + theta * decay_x[index - 1]))
    len_observed = x.shape[0]
    x_decayed = tt.zeros_like(x)
    x_decayed = tt.set_subtensor(x_decayed[0], x[0])
    output, _ = theano.scan(
        fn = adstock_geometric_recurrence_theano, 
        sequences = [tt.arange(1, len_observed), x[1:len_observed]], 
        outputs_info = x_decayed,
        non_sequences = theta, 
        n_steps = len_observed - 1
    )
    
    return output[-1]

Answer 1

so first i'll share the converted code and then explain how and why everything works:

Suppose x is a sequence of advertising expenditures over time and theta is the decay rate. I'll use a small array of numbers for x and a hypothetical value for theta.

Example Data: x: Advertising expenditures over 10 time periods, e.g., [100, 120, 90, 110, 95, 105, 115, 100, 130, 125] theta: Decay rate, let's say 0.5

import torch

def adstock_geometric_pytensor(x, theta):
    x = torch.tensor(x, dtype=torch.float32)
    theta = torch.tensor(theta, dtype=torch.float32)

    def adstock_geometric_recurrence_pytensor(index, input_x, decay_x, theta):
        decay_x[index] = input_x + theta * decay_x[index - 1]
        return decay_x

    len_observed = x.shape[0]
    x_decayed = torch.zeros_like(x)
    x_decayed[0] = x[0]

    for index in range(1, len_observed):
        x_decayed = adstock_geometric_recurrence_pytensor(index, x[index], x_decayed, theta)

    return x_decayed

# Example usage
x_data = [100, 120, 90, 110, 95, 105, 115, 100, 130, 125] # Advertising expenditures
theta_value = 0.5 # Decay rate
output = adstock_geometric_pytensor(x_data, theta_value)
print(output)

Original Code Issues

Your original Theano code uses theano.scan, a powerful tool for looping over sequences in a way optimized for parallel computation. This is a common approach in Theano for handling recursive operations efficiently. However, when switching to PyTorch (PyTensor), there isn't a direct equivalent of theano.scan. PyTorch tends to favor explicit loops in Python, which are less optimized but more straightforward.

Modified Code Explanation

In the modified PyTorch code, I replaced the theano.scan with a standard Python for-loop. This loop iteratively applies the adstock transformation. This change sacrifices some computational efficiency but maintains the core functionality.