TCMFForecaster


In this guide, we will show you how to use Chronos TCMFForecaster for high dimension time series forecasting.

Refer to TCMFForecaster example for demonstration of using TCMFForecaster for distributed training and inference.

Refer to TCMFForecaster API Guide for more details of AutoTS APIs.


Step 0: Prepare environment

a. We recommend conda to set up your environment. Note that conda environment is required to run onq yarn, but not strictly necessary for running on local.

conda create -n zoo python=3.7
conda activate zoo

b. If you want to enable TCMFForecaster distributed training, it requires pre-install pytorch and horovod. You can follow the horovod document to install the horovod and pytorch with Gloo support. And here are the commands that work on for us on ubuntu 16.04. The exact steps may vary from different machines.

conda install -y pytorch==1.4.0 torchvision==0.5.0 cpuonly -c pytorch
conda install -y cmake==3.16.0 -c conda-forge
conda install cxx-compiler==1.0 -c conda-forge
conda install openmpi
HOROVOD_WITH_PYTORCH=1; HOROVOD_WITH_GLOO=1; pip install --no-cache-dir horovod==0.19.1
pip install analytics_zoo-${VERSION}-${TIMESTAMP}-py2.py3-none-${OS}_x86_64.whl[ray]

If you don't need distributed training. You only need to install pytorch in your environment.

pip install torch==1.4.0 torchvision==0.5.0

c. Download and install nightly build analytics zoo whl by following instructions (here).

pip install analytics_zoo-${VERSION}-${TIMESTAMP}-py2.py3-none-${OS}_x86_64.whl[ray]

d. Install other packages

pip install scikit-learn==0.22
pip install pandas==1.0
pip install requests

Step 1: Init Orca Context

You need to init an orca context with init_ray_on_spark=True before distributed training, and stop it after training is completed. Note orca context is not needed if you don't want to enable distributed training.

from zoo.orca import init_orca_context, stop_orca_context

# run in local mode
init_orca_context(cluster_mode="local", cores=4, memory='2g', num_nodes=1, init_ray_on_spark=True)

# run in yarn client mode
init_orca_context(cluster_mode="yarn-client", 
                  num_nodes=2, cores=2, 
                  driver_memory="6g", driver_cores=4, 
                  conda_name='zoo', 
                  extra_memory_for_ray="10g", 
                  object_store_memory='5g')

Step 2: Create a TCMFForecaster

from zoo.chronos.forecaster.tcmf_forecaster import TCMFForecaster
model = TCMFForecaster(
        vbsize=128,
        hbsize=256,
        num_channels_X=[32, 32, 32, 32, 32, 1],
        num_channels_Y=[16, 16, 16, 16, 16, 1],
        kernel_size=7,
        dropout=0.1,
        rank=64,
        kernel_size_Y=7,
        learning_rate=0.0005,
        normalize=False,
        use_time=True,
        svd=True,)

Step 3: Use TCMFForecaster

Fit with TCMFForecaster

model.fit(
        x,
        val_len=24,
        start_date="2020-4-1",
        freq="1H",
        covariates=None,
        dti=None,
        period=24,
        y_iters=10,
        init_FX_epoch=100,
        max_FX_epoch=300,
        max_TCN_epoch=300,
        alt_iters=10,
        num_workers=num_workers_for_fit)

Evaluate with TCMFForecaster

You can either directly call model.evaluate as

model.evaluate(target_value,
               metric=['mae'],
               target_covariates=None,
               target_dti=None,
               num_workers=num_workers_for_predict,
               )

Or you could predict first and then evaluate with metric name.

yhat = model.predict(horizon,
                     future_covariates=None,
                     future_dti=None,
                     num_workers=num_workers_for_predict)

from zoo.orca.automl.metrics import Evaluator
evaluate_mse = Evaluator.evaluate("mse", target_data, yhat)

Incremental fit TCMFForecaster

model.fit_incremental(x_incr, covariates_incr=None, dti_incr=None)

Save and Load

model.save(dirname)
loaded_model = TCMFForecaster.load(dirname)