TimesFM Time Series Forecasting (Google AI, Jupyter, and GPUs)

Supercharge your time series forecasting with the TimesFM model from Google Research! In this video, we show you how to harness the power of GPUs and Vertex AI Workbench notebooks to run this cutting-edge model.
Learn how to:
- Set up a GPU-enabled Jupyter notebook on Google Cloud
- Install required dependencies like Hugging Face Hub, PyTorch, and Jax
- Configure the TimesFM model for your forecasting needs
- Prepare your time series data using Pandas
- Generate forecasts and interpret the results
Whether you're a beginner or a pro, this tutorial will level up your time series skills.
Demonstration Code and Diagram: github.com/nod...
Free Trial - Our New Diagram Tool: softwaresim.co... ("KZbin24" for 25% Off)
If you are a cloud, DevOps, or software engineer you’ll probably find our wide range of KZbin tutorials, demonstrations, and walkthroughs useful - please consider subscribing to support the channel.
0:00 Conceptual Overview
1:17 Vertex AI Workbench (Jupyter)
2:40 TimesFM Setup
5:04 Time Series and Model
7:08 Synthetic Data (Pandas and Numpy)
8:29 Forecasting/Prediction

[@surajvardhan8490]

How to do this on audio timeseries dataset?

[@fathimashaniya9805]

Hi, Thank you very much. Your vedio was very beneficial.

[@ziqili2120]

Hi, thanks for your great video! Do you know how to finetune this model on some private dataset? I found the tutorials using Paxml framework are very limited. I've been stuck in the finetuning for weeks.

[@nodematic]

It doesn't look like this is very well supported right now, but the model architecture means this should be relatively easy once it's better documented/supported. We'll try to create a video on this when the fine-tuning experience is improved.

[@soumyaaaaa-z9k]

Hi thanks for your great video! I have a question, is there a way to install the timesFM in my local system in python either using Jupyter or VScode. I have been trying to install it but failing to do so, it fails because of the Paxml library.

[@nodematic]

I can replicate an issue like this, by using an older Python version like 3.9.2. Only Python 3.10+ can install the paxml>=1.4.0 that is required for timesfm. If you're using Python 3.9, can you try 3.10 or newer?

[@fathimashaniya9805]

If I want to consider other features also for training…. What should be my code…??

[@fathimashaniya9805]

Please can you give me example code to use the rest of the features also…. Also..in case there is a date col … while training do we have to include the month col week col etc separately? If so what should be the dtype of them? Also other if there are cols like id, state id , category id,…. Other id columns , do we consider them while training?

[@fathimashaniya9805]

Can I ask u another question? How do I fine tune this model

[@nodematic]

Fine-tuning does not appear to be supported right now, but they may add support for that in the future, since the model architecture lends itself well to fine-tuning.

[@katarzynakuryo197]

Hi, do you know if it is possible to use this model for multivariate time series?

[@nodematic]

Yes, the model is multivariate. Just be sure to include everything in the forecast call "inputs".

[@katarzynakuryo197]

​@@nodematic I want to train my model on both ts at once. Do you know why it is not working? data = pd.DataFrame({ 'ds': train_df.index, 'ts1': train_df["ts1"], 'ts2': train_df["ts2"], 'unique_id': 'sensor_1' }) forecast_df=tfm.forecast_on_df( inputs=data, freq="D", value_name=["ts1", "ts2"] )

[@camillacarpinelli7531]

@@katarzynakuryo197 ts1 and ts2 have to be unique ids, you need to have your dataset like [["ds","sales","unique_id"] that means that you should have date = ds, quantity sold = sales, unique_id = ts1,ts2,ts3 etc. You need to change your dataset in order to have only these three columns

[@camillacarpinelli7531]

of course this means that your date is not unique anymore because in a single day you could have sold multiple products but it doesn't matter.