PROJECT NOT UNDER ACTIVE MANAGEMENT

This project will no longer be maintained by Intel.

Intel has ceased development and contributions including, but not limited to, maintenance, bug fixes, new releases, or updates, to this project.

Intel no longer accepts patches to this project.

If you have an ongoing need to use this project, are interested in independently developing it, or would like to maintain patches for the open source software community, please create your own fork of this project.

Contact: [email protected]

Demand Forecasting

Introduction

This reference kit implements an end-to-end (E2E) workflow that, with the help of Intel® Optimization for TensorFlow* and Intel® Neural Compressor, produces a complex time series model for demand forecasting capable of spanning multiple products and stores all at once.

Check out more workflow examples in the Developer Catalog.

Solution Technical Overview

The ability to forecast the demand and thus overcome the problems arising from its variability is one of the top challenges for Supply Chain Managers. When planning the production or the supply of shops, organizations are always managing a short blanket: producing/supplying more to avoid out of stocks implies spending more for production, transportation, and immobilized capital.

To model the complex patterns that may be present when understanding product demand, it is often necessary to capture complicated mechanisms such as extended seasonality and long-term correlations, resulting in significant feature engineering to come to a good solution. Modern artificial Intelligence (AI) solutions, such as Deep Neural Networks can drastically aid in this process by becoming automatic feature extractors where the explicit mechanisms do not need to be written down, but rather the AI is allowed to learn them from historical data.

In this use case, we follow this Deep Learning approach and demonstrate how to train and utilize a Convolutional Neural Network and Long Short-Term Memory Network (CNN-LSTM) time series model, which takes in the last 130 days worth of sales data for a specific item at a specific store, to predict the demand one day ahead. A trained model can then be used to predict the next days demand for every item and every store in a given catalog using only previous purchase data, on a daily basis.

Broadly, we will tackle this problem using the following pipeline:

Historical Purchase Data => CNN-LSTM Training => CNN-LSTM Inference

The solution contained in this repo uses the following Intel® packages:

• Intel® Optimizations for TensorFlow*

  The latest version of Intel® Optimization for TensorFlow* is included as part of the Intel® oneAPI AI Analytics Toolkit (AI Kit). This kit provides a comprehensive and interoperable set of AI software libraries to accelerate end-to-end data science and machine-learning workflows.

• Intel® Neural Compressor*

  Intel® Neural Compressor performs model compression to reduce the model size and increase the speed of deep learning inference for deployment on CPUs or GPUs. This open-source Python* library automates popular model compression technologies, such as quantization, pruning, and knowledge distillation across multiple deep learning frameworks.

  Using this library, you can:

  • Converge quickly on quantized models though automatic accuracy-driven tuning strategies.
  • Prune the least important parameters for large models.
  • Distill knowledge from a larger model to improve the accuracy of a smaller model for deployment.
  • Get started with model compression with one-click analysis and code insertion.

For more details, visit Intel® Optimization for TensorFlow*, Intel® Neural Compressor, the Demand Forecasting GitHub repository.

Solution Technical Details

The reference kit implementation is a reference solution to the described use case that includes:

1. An Optimized E2E architecture to arrive at an AI solution with a CNN-LSTM model implemented in TensorFlow*.
2. The Intel® optimizations enabled for TensorFlow* and Intel® Neural Compressor for Model Quantization.

Optimized E2E Architecture with Intel® oneAPI Components

Dataset

The dataset used for this demo is a synthetic set of daily purchase counts over a period of five years characterized by date, item, store, sales, where each feature corresponds to:

• date: date of purchase
• item: item id
• store: store id
• sales: amount of item purchased at store

There is exactly one row per (date, item, store) tuple.

Treated as a time series, this provides a time view of how purchases change over time for each item at each store.

Expected Data Input-Output

Input	Output
Past Purchases	Predicted Demand at the selected horizon

Example Input	Example Output
Date, Store, Item, Sales 1-1-2010, 1, 1, 5 1-2-2010, 1, 1, 7 1-3-2010, 1, 1, 9 1-4-2010, 1, 1, 11 1-1-2010, 2, 1, 10 1-2-2010, 2, 1, 15 1-3-2010, 2, 1, 20 1-4-2010, 2, 1, 25	Date, Store, Item, Predicted Demand 1-5-2010, 2, 1, 20 1-5-2010, 2,1, 30

Validated Hardware Details

There are workflow-specific hardware and software setup requirements to run this use case.

Recommended Hardware	Precision
CPU: Intel® 2nd Gen Xeon® Platinum 8280 CPU @ 2.70GHz or higher	FP32, INT8
RAM: 187 GB
Recommended Free Disk Space: 20 GB or more

Operating System: Ubuntu* 22.04 LTS.

How it Works

Model Training

Using the synthetic daily dataset described here, we train a Deep Learning model to forecast the next days demand using the previous $n$ days. More specifically, to build a forecasting model for $\text{count}[t]$, representing the sales of a given (item, store) at time $t$, we first transform the time series data to rows of the form:

(item, store, count[t-n], count[t - (n - 1)], ..., count[t-1], count[t])

Using data of this form, the forecasting problem becomes a regression problem that takes in the demand/sales of the previous $n$ time points and predicts the demand at the current time $t$, representing the forecasting assumption as:

$f(\text{count}[t-n], \text{count}[t - (n - 1)], ..., \text{count}[t-1]) = \text{count}[t]$.

Here, $n$ and the forecast horizon are configurable if desired beyond $n$ lags and 1-step ahead. A CNN-LSTM model is used for this application, as opposed to just a LSTM, to allow for the model to learn better long-term dependencies by folding the time series and applying a Convolutional Neural Network (CNN) encoder before modeling the modified sequential dependencies via the Long Short-Term Memory Network (LSTM).

Model Inference

The saved model from the training process can be used to predict demand on new data of the same format. This expects the previous 130 days sales counts. A multi-step ahead forecast can be generated by iterating this process, using predicted values as input and propagating forward though the model.

Get Started

Start by defining an environment variable that will store the workspace path, this can be an existing directory or one to be created in further steps. This ENVVAR will be used for all the commands executed using absolute paths.

export WORKSPACE=$PWD/demand-forecasting

Define DATA_DIR and OUTPUT_DIR.

export DATA_DIR=$WORKSPACE/data
export OUTPUT_DIR=$WORKSPACE/output
export CONFIG_DIR=$WORKSPACE/config

Download the Workflow Repository

Create a working directory for the workflow and clone the Main Repository into your working directory.

mkdir -p $WORKSPACE && cd $WORKSPACE

git clone https://github.com/oneapi-src/demand-forecasting $WORKSPACE

Set Up Conda

To learn more, please visit install anaconda on Linux.

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

Set Up Environment

Install and set the libmamba solver as default solver. Run the following commands:

conda install -n base conda-libmamba-solver -y
conda config --set solver libmamba

The env/intel_env.yml file contains all dependencies to create the Intel® environment.

Packages required in YAML file	Version
python	3.9
intel-aikit-tensorflow	2024.1

Execute next command to create the conda environment.

conda env create -f $WORKSPACE/env/intel_env.yml

During this setup, demand_forecasting_intel conda environment will be created with the dependencies listed in the YAML configuration. Use the following command to activate the environment created above:

conda activate demand_forecasting_intel

Supported Runtime Environment

You can execute this reference pipeline using the following environments:

• Bare Metal

Run Using Bare Metal

Follow these instructions to set up and run this workflow on your own development system.

Set Up System Software

Our examples use the conda package and environment on your local computer. If you don't already have conda installed, go to Set up conda or see the Conda* Linux installation instructions.

Run Workflow

Once we create and activate the demand_forecasting_intel environment, we can run the next steps.

Setting up the data

The benchmarking scripts expects two files to be present in data/demand:

• data/demand/train.csv: training data

• data/demand/test_full.csv: testing data

To setup the data for benchmarking:

1. Use the src/generate_data.py script to generate synthetic data for demand/train.csv and demand/test_full.csv:

   python $WORKSPACE/src/generate_data.py --output_dir $DATA_DIR/demand

Model Building Process

As described in Model Training section, we first transform the data to the regression format expected and feed this data into our CNN-LSTM model. The run_training.py script reads and preprocesses the data, trains the model, and saves the model which can be used for future inference.

The script takes the following arguments:

usage: run_training.py [-l LOGFILE] [-s SAVE_MODEL_DIR] [-b BATCH_SIZE] -i INPUT_CSV

options:
  -l LOGFILE, --logfile LOGFILE
                        log file to output benchmarking results to
  -s SAVE_MODEL_DIR, --save_model_dir SAVE_MODEL_DIR
                        directory to save model to
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        training batch size
  -i INPUT_CSV, --input_csv INPUT_CSV
                        input csv file path

As an example of using this to train a model, run the following command:

python $WORKSPACE/src/run_training.py --save_model_dir $OUTPUT_DIR/saved_models/intel \
  --batch_size 512 --input_csv $DATA_DIR/demand/train.csv -l $OUTPUT_DIR/logs/training_log.txt

Running Inference

The above script will train and save models to the save_model_dir. To use this model to make predictions on new data, a 2-step process is necessary to optimize performance:

1. Convert the saved model from a Keras* saved model to a TensorFlow* frozen graph. To do this, execute the script convert_keras_to_frozen_graph.py which takes the following arguments:

usage: convert_keras_to_frozen_graph.py -s KERAS_SAVED_MODEL_DIR -o OUTPUT_SAVED_DIR

options:
  -s KERAS_SAVED_MODEL_DIR, --keras_saved_model_dir KERAS_SAVED_MODEL_DIR
                        directory with saved keras model
  -o OUTPUT_SAVED_DIR, --output_saved_dir OUTPUT_SAVED_DIR
                        directory to save frozen graph to

For the above saved model, run the command:

python $WORKSPACE/src/convert_keras_to_frozen_graph.py -s $OUTPUT_DIR/saved_models/intel \
  -o $OUTPUT_DIR/saved_models/intel

The convert_keras_to_frozen_graph script takes in the saved Keras* model and outputs a frozen graph in the same directory called saved_models/intel/saved_frozen_model.pb.

2. Once a saved frozen graph is saved, this model can now be used to perform inference using the run_inference.py script which has the following arguments:

usage: run_inference.py [-l LOGFILE] -s SAVED_FROZEN_MODEL [-b BATCH_SIZE]
    --input_file INPUT_FILE [--benchmark_mode] [-n NUM_ITERS]

options:
  -l LOGFILE, --logfile LOGFILE
                        log file to output benchmarking results to
  -s SAVED_FROZEN_MODEL, --saved_frozen_model SAVED_FROZEN_MODEL
                        saved frozen graph
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        batch size to use
  -i INPUT_FILE, --input_file INPUT_FILE
                        input csv data file
  --benchmark_mode      benchmark inference time
  -n NUM_ITERS, --num_iters NUM_ITERS
                        number of iterations to use when benchmarking

Now, run the inference on data file data/demand/test_full.csv:

python $WORKSPACE/src/run_inference.py --input_file $DATA_DIR/demand/test_full.csv \
  --saved_frozen_model $OUTPUT_DIR/saved_models/intel/saved_frozen_model.pb \
  --batch_size 512 --benchmark_mode -l $OUTPUT_DIR/logs/inference_log.txt

On larger sample data set sizes and more complex models, the gains will become more obvious and apparent.

Post Training Optimization with Intel® Neural Compressor

In scenarios where the model or data become very large, such as if there are a huge amount of stores and items, and the model is expanded to capture more complex phenomena, it may be desirable to further optimize the latency and throughput of a model. For these scenarios, one method can utilize model quantization techniques via Intel® Neural Compressor.

Model quantization is the practice of converting the FP32 weights in Deep Neural Networks to a lower precision, such as INT8 in order to accelerate computation time and reduce storage space of trained models. This may be useful if latency and throughput are critical. Intel® offers multiple algorithms and packages for quantizing trained models. This reference implementation includes the run_quantize_inc.py script which can be executed after saving the frozen graph to attempt accuracy-aware quantization on the trained model.

The run_quantize_inc.py script takes the following arguments:

usage: run_quantize_inc.py --saved_frozen_graph SAVED_FROZEN_GRAPH --output_dir OUTPUT_DIR 
    --inc_config_file INC_CONFIG_FILE --input_csv INPUT_CSV

options:
  --saved_frozen_graph SAVED_FROZEN_GRAPH
                        saved pretrained frozen graph to quantize
  --output_dir OUTPUT_DIR
                        directory to save quantized model
  --inc_config_file INC_CONFIG_FILE
                        INC conf yaml
  --input_csv INPUT_CSV
                        input csv dataset file path

Execute model quantization in the saved_frozen_model.pb file as follows:

python $WORKSPACE/src/run_quantize_inc.py \
  --saved_frozen_graph $OUTPUT_DIR/saved_models/intel/saved_frozen_model.pb \
  --output_dir $OUTPUT_DIR/saved_models/intel --inc_config_file $CONFIG_DIR/conf.yaml \
  --input_csv $DATA_DIR/demand/train.csv

The ouput is a quantized model saved in saved_models/intel/saved_frozen_int8_model.pb file. This model is typically smaller at a minor cost to accuracy.

Inference on this newly quantized model can be performed identically as before, pointing the script to the saved quantized graph.

python $WORKSPACE/src/run_inference.py --input_file $DATA_DIR/demand/test_full.csv \
  --saved_frozen_model $OUTPUT_DIR/saved_models/intel/saved_frozen_int8_model.pb \
  --batch_size 512 --benchmark_mode -l $OUTPUT_DIR/logs/quantize_inc_inference_log.txt

Clean Up Bare Metal

Follow these steps to restore your $WORKSPACE directory to an initial step. Please note that all downloaded or created dataset files, conda environment, and logs created by the workflow will be deleted. Before executing next steps back up your important files.

# activate base environment
conda activate base
# delete conda environment created
conda env remove -n demand-forecasting

# remove all data generated
rm -rf $DATA_DIR/demand
# remove all outputs generated
rm -rf $OUTPUT_DIR

Expected Output

The $OUTPUT_DIR/logs/training_log.txt file shows values for Train RMSE, Validation RMSE and Train time:

INFO:root:Starting training on 639009 samples with batch size 512...
INFO:root:======> Train RMSE: 7.5889
INFO:root:======> Validation RMSE: 7.5917
INFO:root:=======> Train time : 467 seconds
INFO:root:Saving model...
INFO:tensorflow:Assets written to: demand-forecasting/output/saved_models/intel/

The $OUTPUT_DIR/logs/inference.txt contains the value for Average Inference Time:

INFO:root:Starting inference on batch size 512 for 100 iterations
INFO:root:=======> Average Inference Time : 0.007982 seconds

Inference on the newly quantized model execution command creates the $OUTPUT_DIR/logs/quantize_inc_inference_log.txt file that contains Average Inference Time value:

INFO:root:Starting inference on batch size 512 for 100 iterations
INFO:root:=======> Average Inference Time : 0.005064 seconds

Generated models will be saved in $OUTPUT_DIR/saved_models/intel/:

assets/
fingerprint.pb
keras_metadata.pb
saved_frozen_int8_model.pb
saved_frozen_model.pb
saved_model.pb
variables/

Summary and Next Steps

Demand Forecasting is a pivotal and ever-present component of many business decisions. The ability for an analyst to quickly train a complex model and forecast into the future, whether to analyze a hypothesis or to make key decisions can heavily impact the day or time to market products. This reference kit implementation provides a performance-optimized guide around demand forecasting using Deep Learning related use cases that be easily scaled across similar use cases.

Training is conducted using Intel® oneAPI Optimizations for TensorFlow* to accelerate performance using oneDNN optimizations.

When it is necessary to forecast demand across a large catalog of items and stores, it is often desirable to predict in large batches to process the workload as quickly as possible. For more complex models, it may also be desirable to reduce the size of the model with minimal accuracy impact to scale the solutions. This use case utilizes model quantization techniques from Intel® Neural Compressor.

Learn More

For more information about or to read about other relevant workflow examples, see these guides and software resources:

• Intel® AI Analytics Toolkit (AI Kit)
• Intel® Distribution for Python*
• Intel® Optimization for TensorFlow*
• Intel® Neural Compressor

Support

If you have questions or issues about this use case, want help with troubleshooting, want to report a bug or submit enhancement requests, please submit a GitHub issue.

Appendix

*Names and brands that may be claimed as the property of others. Trademarks.

Disclaimers

To the extent that any public or non-Intel datasets or models are referenced by or accessed using tools or code on this site those datasets or models are provided by the third party indicated as the content source. Intel does not create the content and does not warrant its accuracy or quality. By accessing the public content, or using materials trained on or with such content, you agree to the terms associated with that content and that your use complies with the applicable license.

Intel expressly disclaims the accuracy, adequacy, or completeness of any such public content, and is not liable for any errors, omissions, or defects in the content, or for any reliance on the content. Intel is not liable for any liability or damages relating to your use of public content.