Skip to content
This repository has been archived by the owner on Feb 22, 2024. It is now read-only.

Latest commit

 

History

History
207 lines (150 loc) · 14.9 KB

DEVCATALOG.md

File metadata and controls

207 lines (150 loc) · 14.9 KB
Raw

Time Series Forecasting using BigDL-Chronos - Model development and optimization

Overview

BigDL-Chronos (Chronos in short) is an application framework for building a fast, accurate and scalable time series analysis application, where forecasting is the most popular task. For more detailed information about the framework, go to Chronos Document.

How it works

  • Chronos provides an easy but complete API TSDataset for data loading, preprocessing, feature engineering and sampling, where users could fetch their raw tabular data from local files, distributed file system or TS database like Prometheus and transform them to data that can be fed into DL/ML models. It contains 2 backends, pandas for single node users and spark (with the support of bigdl-orca) for cluster users.

  • Chronos provides Forecaster API for 9+ different ML/DL models' training, tunning, evaluation, model optimization, and exporting for deployment.

  • 9+ different ML/DL model include traditional models cover traditional models like ARIMA, Prophet to SOTA Deep learning models such as Autoformer, TCN, NBeats, and Seq2Seq.

  • During the process, IPEX, intel-tensorflow, onnxruntime, openvino, and intel neural compressor are used as accelerator and low precision tool to accelerate the model during training and inferencing. bigdl-orca is used for cluster level distributed training and tunning.

Chronos-workflow

Get Started

Through PyPI

bigdl-chronos is released on PyPI as a python library and could be installed on nearly any platform user prefer. Here we provide a typical installation method for single node user who prefer a pytorch backend, for more detailed information, users may refer to our installaion page.

# conda environment is highly recommeneded
conda create -n bigdl-chronos-pytorch-env python=3.7 setuptools=58.0.4
conda activate bigdl-chronos-pytorch-env
pip install bigdl-chronos[pytorch]

On colab

Users could also install bigdl-chronos easily on google colab, where they only need to install the library through pip.

Through docker

Some users may prefer a docker installation do seperate the environment. Below setup and how-to-run sessions are for users who want to use the provided docker image.

Pull Docker Image

docker pull intel/ai-workflows:time-series-forecaster

Clone the BigDL repository to the current working directory and checkout the specific tag

git clone https://github.com/intel-analytics/BigDL.git
cd BigDL
git checkout ai-workflow

To run the pipeline, follow the instructions below outside of docker instance.

docker run -it --rm -v ${PWD}:/workspace \
   -w /workspace/python/chronos/colab-notebook --init --net=host \
   intel/ai-workflows:time-series-forecaster \
   sh -c "jupyter nbconvert --to python chronos_nyc_taxi_tsdataset_forecaster.ipynb && \
   sed '26,40d' chronos_nyc_taxi_tsdataset_forecaster.py > chronos_taxi_forecaster.py && \
   python chronos_taxi_forecaster.py"

Output

#1 [internal] load build definition from Dockerfile.chronos
#1 transferring dockerfile: 55B done
#1 DONE 0.0s

#2 [internal] load .dockerignore
#2 transferring context: 2B done
#2 DONE 0.0s

#3 [internal] load metadata for docker.io/library/ubuntu:20.04
#3 DONE 0.0s

#4 [1/5] FROM docker.io/library/ubuntu:20.04
#4 DONE 0.0s

#5 [2/5] RUN apt-get update --fix-missing &&     apt-get install -y apt-utils vim curl nano wget unzip git &&     apt-get install -y gcc g++ make &&     apt-get install -y libsm6 libxext6 libxrender-dev &&     apt-get install -y openjdk-8-jre &&     rm /bin/sh &&     ln -sv /bin/bash /bin/sh &&     echo "auth required pam_wheel.so use_uid" >> /etc/pam.d/su &&     chgrp root /etc/passwd && chmod ug+rw /etc/passwd &&     wget https://repo.anaconda.com/miniconda/Miniconda3-py37_4.12.0-Linux-x86_64.sh &&     chmod +x Miniconda3-py37_4.12.0-Linux-x86_64.sh &&     ./Miniconda3-py37_4.12.0-Linux-x86_64.sh -b -f -p /usr/local &&     rm Miniconda3-py37_4.12.0-Linux-x86_64.sh
#5 CACHED

#6 [4/5] RUN echo "source activate chronos" > ~/.bashrc
#6 CACHED

#7 [3/5] RUN conda create -y -n chronos python=3.7 setuptools=58.0.4 && source activate chronos &&     pip install --no-cache-dir --pre --upgrade bigdl-chronos[pytorch,automl] matplotlib notebook==6.4.12 &&     pip uninstall -y torchtext
#7 CACHED

#8 [5/5] RUN echo "export JAVA_HOME=/usr/lib/jvm/java-8-openjdk-amd64/" >> ~/.bashrc
#8 CACHED

#9 exporting to image
#9 exporting layers done
#9 writing image sha256:329995e99da4001c6d57e243085145acfce61f5bddabd9459aa598b846eae331 done
#9 naming to docker.io/library/time-series-chronos:training-ubuntu-20.04 done
#9 DONE 0.0s
Attaching to training-time-series-chronos-1
training-time-series-chronos-1  | [NbConvertApp] Converting notebook chronos_nyc_taxi_tsdataset_chronos.ipynb to python
training-time-series-chronos-1  | [NbConvertApp] Writing 10692 bytes to chronos_nyc_taxi_tsdataset_chronos.py
training-time-series-chronos-1  | Global seed set to 1
training-time-series-chronos-1  | Global seed set to 1
training-time-series-chronos-1  | /usr/local/envs/chronos/lib/python3.7/site-packages/bigdl/chronos/forecaster/utils.py:157: UserWarning: 'batch_size' cannot be divided with no remainder by 'self.num_processes'. We got 'batch_size' = 32 and 'self.num_processes' = 7
training-time-series-chronos-1  |   format(batch_size, num_processes))
training-time-series-chronos-1  | GPU available: False, used: False
training-time-series-chronos-1  | TPU available: False, using: 0 TPU cores
training-time-series-chronos-1  | IPU available: False, using: 0 IPUs
training-time-series-chronos-1  | HPU available: False, using: 0 HPUs
training-time-series-chronos-1  | Global seed set to 1
training-time-series-chronos-1  | Initializing distributed: GLOBAL_RANK: 2, MEMBER: 3/7

...

 98%|█████████▊| 287/294 [00:09<00:00, 31.55it/s, loss=0.0155]
Epoch 2:  98%|█████████▊| 288/294 [00:09<00:00, 31.55it/s, loss=0.0155]
Epoch 2:  98%|█████████▊| 288/294 [00:09<00:00, 31.55it/s, loss=0.0155]
Epoch 2:  98%|█████████▊| 288/294 [00:09<00:00, 31.55it/s, loss=0.0157]
Epoch 2:  98%|█████████▊| 289/294 [00:09<00:00, 31.56it/s, loss=0.0157]
Epoch 2:  98%|█████████▊| 289/294 [00:09<00:00, 31.56it/s, loss=0.0162]
Epoch 2:  99%|█████████▊| 290/294 [00:09<00:00, 31.57it/s, loss=0.0162]
Epoch 2:  99%|█████████▊| 290/294 [00:09<00:00, 31.57it/s, loss=0.0165]
Epoch 2:  99%|█████████▉| 291/294 [00:09<00:00, 31.58it/s, loss=0.0165]
Epoch 2:  99%|█████████▉| 291/294 [00:09<00:00, 31.58it/s, loss=0.0164]
Epoch 2:  99%|█████████▉| 292/294 [00:09<00:00, 31.59it/s, loss=0.0164]
Epoch 2:  99%|█████████▉| 292/294 [00:09<00:00, 31.59it/s, loss=0.0164]
Epoch 2:  99%|█████████▉| 292/294 [00:09<00:00, 31.59it/s, loss=0.0164]
Epoch 2: 100%|█████████▉| 293/294 [00:09<00:00, 31.58it/s, loss=0.0164]
Epoch 2: 100%|█████████▉| 293/294 [00:09<00:00, 31.58it/s, loss=0.0175]
Epoch 2: 100%|██████████| 294/294 [00:09<00:00, 31.64it/s, loss=0.0175]
Epoch 2: 100%|██████████| 294/294 [00:09<00:00, 31.64it/s, loss=0.017] 
Epoch 2: 100%|██████████| 294/294 [00:09<00:00, 31.64it/s, loss=0.017]
Epoch 2: 100%|██████████| 294/294 [00:09<00:00, 31.60it/s, loss=0.017]
training-time-series-forecaster-1  | Global seed set to 1
training-time-series-forecaster-1  | GPU available: False, used: False
training-time-series-forecaster-1  | TPU available: False, using: 0 TPU cores
training-time-series-forecaster-1  | IPU available: False, using: 0 IPUs
training-time-series-forecaster-1  | HPU available: False, using: 0 HPUs
training-time-series-forecaster-1 exited with code 0

Use-cases

How to guides

How-to guides are bite-sized, executable examples where users could check when meeting with some specific topic during the usage. For friendly experience please visit our how to guide page.

Tutorials/examples

Here is a use-cases list of bigdl-chronos, for more friendly experience please visit our tutorial page.

Use case Format Model Framework
Predict Number of Taxi Passengers with Chronos Forecaster Notebook TCN PyTorch
Tune a Forecasting Task Automatically Notebook TCN PyTorch, Ray
Tune a Customized Time Series Forecasting Model with AutoTSEstimator Notebook Customized PyTorch, Ray
Auto Tune the Prediction of Network Traffic at the Transit Link of WIDE Notebook LSTM PyTorch, Ray
Multivariate Forecasting of Network Traffic at the Transit Link of WIDE Notebook LSTM PyTorch
Multistep Forecasting of Network Traffic at the Transit Link of WIDE Notebook TCN PyTorch
Stock Price Prediction with LSTMForecaster Notebook LSTM PyTorch
Stock Price Prediction with ProphetForecaster and AutoProphet Notebook Prophet Prophet, Ray
Tune a Time Series Forecasting Model with multi-objective hyperparameter optimization Notebook TCN Prophet, Optuna
Auto tuning prophet on nyc taxi dataset Scripts Prophet Prophet, Ray
Use Chronos forecasters in a distributed fashion Scripts Seq2Seq, TCN PyTorch, Ray
Use ONNXRuntime to speed-up forecasters' inferecing Scripts Seq2Seq OnnxRuntime, PyTorch
Quantize Chronos forecasters method to speed-up inference Scripts TCN INC, PyTorch
High dimension time series forecasting with Chronos TCMFForecaster Scripts TCMF Ray, Spark
Penalize underestimation with LinexLoss Notebook TCN PyTorch
Serve Chronos forecaster and predict through TorchServe Scripts TCN TorchServe, PyTorch
Help pytorch-forecasting improve the training speed of DeepAR model Scripts DeepAR Pytorch-forecasting
Help pytorch-forecasting improve the training speed of TFT model Scripts TFT Pytorch-forecasting

Recommended Platform

CPU

  • Intel® Xeon® Scalable Performance processors

  • Intel® Core® processors

OS

  • Ubuntu 16.04/18.04/20.04/22.04

  • Windows (experimentally supported)

  • Mac with Intel Chip (experimentally supported)

Python

  • Python 3.7

  • Python 3.8 (experimentally supported)

Learn More

Chronos Document

Trouble shooting

Chronos tips and know issues

Support Forum

We welcome any questions, bug report or feature request to: