Optimization problem with scalers

[Miguel-Adarlo]

Good day. I'm running into an issue when trying to optimize my spend. Right now, my code looks like this:

# Run optimization with the parameters of choice.
solution, kpi_without_optim, previous_media_allocation = optimize_media.find_optimal_budgets(
    n_time_periods=n_time_periods,
    media_mix_model=mmm,
    extra_features=extra_scaler.transform(extra_data_test)[:n_time_periods],
    budget=budget,
    prices=prices,
    bounds_upper_pct = 20, #Needed due to budget being overblown.
#    media_scaler=media_scaler,
#    target_scaler=target_scaler,
    seed=seed)

But when I try to add the scalers, I run into problems. With the media_scaler, I run into this:

ValueError: Expected 2D array, got 1D array instead:
array=[0.12943826 0.29676388 0.08048123 0.06228086 0.09115253 0.04516312
 0.02979681 0.03268982 0.04777034 0.12130852 0.07578947 0.048046
 0.13668245 0.01495924 0.03651248 0.11818695 0.06646596 0.19662156
 0.03996078 0.0097561  0.03880266].
Reshape your data either using array.reshape(-1, 1) if your data has a single feature or array.reshape(1, -1) if it contains a single sample.

And with the target scaler, I run into this.

TracerArrayConversionError: The numpy.ndarray conversion method __array__() was called on traced array with shape float64[2000,12].
The error occurred while tracing the function _objective_function at /Users/miguelantonioadarlo/anaconda3/envs/ROMI/lib/python3.8/site-packages/lightweight_mmm/optimize_media.py:27 for jit. This value became a tracer due to JAX operations on these lines:

  operation a:f64[] = convert_element_type[new_dtype=float64 weak_type=False] b
    from line /var/folders/dd/7df9_d6d2nqdhkq927qz5hn40000gp/T/ipykernel_89368/3525311021.py:2 (<module>)

  operation a:f64[] = convert_element_type[new_dtype=float64 weak_type=False] b
    from line /var/folders/dd/7df9_d6d2nqdhkq927qz5hn40000gp/T/ipykernel_89368/3525311021.py:2 (<module>)

  operation a:f64[] = convert_element_type[new_dtype=float64 weak_type=False] b
    from line /var/folders/dd/7df9_d6d2nqdhkq927qz5hn40000gp/T/ipykernel_89368/3525311021.py:2 (<module>)

  operation a:f64[82,21] = convert_element_type[new_dtype=float64 weak_type=False] b
    from line /var/folders/dd/7df9_d6d2nqdhkq927qz5hn40000gp/T/ipykernel_89368/3525311021.py:2 (<module>)

  operation a:f64[82,31] = convert_element_type[new_dtype=float64 weak_type=False] b
    from line /var/folders/dd/7df9_d6d2nqdhkq927qz5hn40000gp/T/ipykernel_89368/3525311021.py:2 (<module>)

(Additional originating lines are not shown.)
See https://jax.readthedocs.io/en/latest/errors.html#jax.errors.TracerArrayConversionError

It runs without the scalers, but I cannot then get the optimized budget allocation and such properly.

For the record, my scalers are the MaxAbsScaler from sklearn, as the CustomScaler for LightweightMMM produces NaNs.

Help regarding optimization would be appreciated. Thank you!

[DEGI-lucas]

If one of your channels is only zeroes, then you could run into this problem. Also, try checking for null or infinite values.

[Miguel-Adarlo]

I retried the CustomScaler again, this time removing the columns with only zeroes. It worked, but optimization was not able to complete successfully. It instead stated something else.

Positive directional derivative for linesearch    (Exit mode 8)
            Current function value: -17456214035.345318
            Iterations: 64
            Function evaluations: 2580
            Gradient evaluations: 60

Even when I set the max iterations higher, it still quits there.