Sales forecasting with Keras and Tensorflow

The contents will allow you to play with a Keras model forecasting sales for the data from: https://www.kaggle.com/c/rossmann-store-sales

You will be able to run your code locally and on google cloud platform for scalability.

Description

This beginner-level workshop will guide you through implementing a sales forecasting neural network model based on an openly available dataset. We will go through some basic neural network theory and apply a systematic model improvement approach to get from a simple model to a more complex one, trying out various tricks along the way.

data: train and store coming from kaggle Rossman case:
- train1, eval1: data filtered for just one store (date ascending)
output - where the checkpoint models are stored
gcp-output - where the checkpoint models from the cloud are downloaded
notebooks - useful notebooks
scripts - scripts (mainly gcloud) for dealing with google cloud
trainer - main model

Requirements

python: beginner level
a laptop with some software configured (see setup instructions)
pandas, numpy, sklearn: nice-to-have

Setup instructions

You will need:

a clone of: https://github.com/gregaw/sales-forecasting-with-nn.git
anaconda (https://conda.io/docs/user-guide/install/index.html)
[optional] google cloud ml-engine (there's free tier available) and shell (https://cloud.google.com/sdk/)

Once you get those installed (and have this git repo cloned locally), run:

conda create -n mlengine python=2.7 anaconda

source activate mlengine

pip install -r requirements.txt

Note : (you can also set it up with python=3.6, but you might have problems running this on GCP ML-engine) Note2 : check the setup.txt with some dumps of the environment correctly setup.

Windows setup

tensorflow is available with python 3.x only, so use:

conda create -n mlengine python=3.5 anaconda

import model in task.py becomes:

from trainer import model

print 'text' becomes:

print('text')

source activate mlengine becomes:

activate mlengine

Running locally

Use run_next_local.py, which will create a new job_name (with sequential numbers)

python run_next_local.py

Running on GCP ML-engine

Use scripts under scripts

set your arguments in scripts/config.txt

Note gcloud job name has to be unique in your space!

run the job

scripts/job-cloud.sh

download the results locally

scripts/download.sh

explore the results with local TensorBoard

gregaw/sales-forecasting-with-nn