Recommendation for Optimal Environment to Run Meridian Model

[gitlingjin]

Hi Meridian team,

I'm currently running the Meridian model in Google Colab under my company Google enterprise account with the Pay-As-You-Go plan. However, I'm finding the training time quite long for typical runs. Eg:

CPU times: user 4h 42min 46s
sys: 3min 52s
total: 4h 46min 38s
Wall time: 3h 21min 6s

Is this typical? I’d love to know if you have any recommendations for a more efficient environment to train the model? My goal is to speed up training time and iterate faster.

Any suggestions would be greatly appreciated!

[sonriks6]

Hi @gitlingjin ,

Under my experience, this is too long, for your reference, running in local machine (laptop CPU 16 cores, no GPU) a national model with 200+ weeks, 16 media channels, 5 controls, and 2 non-media, takes between 5 to 10 minutes.

Computing time really depends on how you configure the sample_posterior params like n_chains, n_keep can influence the time needed, this also consumes a lot of memory, there's a tradeoff here you need to wisely select.

[gousesb]

Hello @gitlingjin ,

Thank you for contacting us!

To accelerate training and iteration, ensure your Colab environment utilizes a GPU by verifying the runtime type and noting the specific GPU. Experiment with MCMC parameters in your Meridian model, particularly n_chains, for improved speed. For a more scalable solution, consider running Meridian on dedicated machine learning compute resources like Vertex AI, instead of Colab. Additionally, start with a simpler Meridian model, such as a basic national model with fewer media channels and controls, to diagnose if extended training times are due to the environment or model complexity, establishing a baseline for reasonable training times in your setup.

To help us pinpoint the exact cause, could you please confirm your Colab runtime is set to GPU, and if so, what type of GPU is being used? Additionally, please provide a brief overview of your current Meridian model's complexity, including the number of media channels and the level of data granularity .

Feel free to reach out for any further queries or feedback regarding Meridian.

Thank you

Google Meridian Support Team