Time series prediction with RNN

We train a RNN on the Rossmann Store Sales dataset in order to predict the future store sells.

Training Logs

To see the logs from the training with tensorboard use:

tensorboard --logdir=${RESULTS_DIR}/logs

or within a jupyter notebook:

%reload_ext tensorboard
%tensorboard --logdir <logs directory>

The dashed vertical line at t=708 separates the training and validation dataset of the stores with id=4 and id=8.

Setup & Training

conda create -n rnn python=3.6
conda activate rnn
pip install -r requirements.txt
python setup.py install

Up to now only the CPU execution is tested, the data preprocessing has to be optimized. To run a training job execute:

# remove --n_stores to use the entire dataset
python exec/train_03.py \
    --data_dir ${DATA_DIR} \
    --model_dir ${RESULTS_DIR}/models_rossmann_03 \
    --log_dir ${RESULTS_DIR}/logs_rossmann_03 \
    --n_stores 4 \
    --num_epochs 201 \
    --store_id_embedding_dim 4 \
    --rnn_hidden_size 32 \
    --rnn_num_layers 4 \
    --batch_size 20 \
    --t_train 708

Some notes:

• The first 708 days are used for training, the remaining days until 2015-07-31 are used for validation.
• The model has a logic that automatically predicts sales=0 if the store is closed (open=0).
• A mask makes sure that the predictions for the first 100 days in which a store is open are not taken into account in the loss function.