PyTensor Scan Dimension Mismatch

Questions v5
,
1

Hi there, I’ve been developing a model for user based repetition wear and restoration effects in online advertising. I’m having some confusion on the dimensions mismatch error within pytensor.scan:

Here’s some of the code excerpt:

coords = {"user_ids": imps_df.index, "weeks": imps_df.columns}

with pm.Model(coords=coords) as model1:
    ## OBSERVED IMPRESSIONS
    imps = pm.MutableData("imps", imps_df, dims={"user_ids", "weeks"})
    wks_since_last_imp = pm.MutableData("wks_since_last_imp", wks_since_last_imp_df, dims={"user_ids", "weeks"})

...
    ###### REPETITION WEAR AND RESTORATION
    #repetition wear
    delta_1 = pm.Uniform("delta_1", lower=0, upper=1)
    #restoration effects - Universal
    rho_1 = pm.Uniform("rho_1", lower=0, upper=1)
    #geometric decay rate
    adstock_decay = pm.Uniform("adstock_decay", lower=.4, upper=.8)
    ## Add restoration and wear to adstock transformation
    R1 = pm.Deterministic("r1", (rho_1*wks_since_last_imp)/(1+rho_1*wks_since_last_imp))
    wearout = pm.Deterministic("wearout", 1 - delta_1**imps)
    eit = pm.Deterministic("eit", advertising_effect*(1 - wearout + R1 * wearout))
    # Recursively Compute Adstock Efficiently via pytensor trace
    adstock0 = pt.zeros(nusers) 

    # def update_adstoc(eit, prior_result, adstock_decay):

    adstock, __ = pytensor.scan(fn=lambda eit, prior_result, adstock_decay: prior_result * adstock_decay + eit,
                                outputs_info=adstock0, #initialize results adstock is equal to eit at wk 1
                                sequences = [eit],
                                non_sequences=[adstock_decay],
                                n_steps=nweeks)

My model graph lines up with what I’m envisioning:

image
image2104×1148 191 KB

But am getting the error:
ValueError: Input dimension mismatch: (input[0].shape[0] = 10000, input[2].shape[0] = 45

I believe this is coming from the prior_value supplied through the scan function. I want to make sure the function is looking back to the adstock from previous weeks not of previous users. Input dataframe was indexed by user_id with columns for successive weeks.

Any thoughts? Do I just need to transpose the input features to the scan function? Appreciate the active community and help moving further along.

1 Like
2

Scan iterates over the leftmost dimensions of sequences

3

Thanks! That makes sense based on the error message I’m seeing where the value error is showing slicing the 0th index of shape:

ValueError: Input dimension mismatch: (input[1].shape[0] = 45, input[4].shape[0] = 10000)

Still hitting some convergence issues but making progress

1 Like