• PyTorch Implementation of the GRU4REC model.
• Original paper: Session-based Recommendations with Recurrent Neural Networks(ICLR 2016)
• This code is mostly a PyTorch re-implementation of the original Theano code written by the authors of the GRU4REC paper. What I did are as below:
  • Replaced the Theano components with PyTorch
  • Simplifying and Cleaning the session-parallel mini-batch generation code

• It seems that there is a bug in the updated version(05/20/2018) of the code. However, I'm too busy these days, so I won't be able to resolve the issue for a while.(not until July) Sorry for the inconvenience.

• PyTorch 0.4.0 support
  • The code is now compatible with PyTorch >= 0.4.0
• Code cleanup
  • Removed redundant pieces of code thanks to simpler API of PyTorch 0.4.0
  • Improved the readablility of the confusing rnn updating routine
  • Improved the readability of the training/testing routine
• Optimization
  • Testing code is now much faster than before

• PyTorch 0.4.0
• Python 3.6.4
• pandas 0.22.0
• numpy 1.14.0

• Filenames
  • Training set should be named as train.tsv
  • Test set should be named as test.tsv
• File Paths
  • train.tsv, test.tsv should be located under the data directory. i.e. data/train.tsv, data/test.tsv
• Contents
  • train.tsv, test.tsv should be the tsv files that stores the pandas dataframes that satisfy the following requirements(without headers):
    • The 1st column of the tsv file should be the integer Session IDs
    • The 2nd column of the tsv file should be the integer Item IDs
    • The 3rd column of the tsv file should be the Timestamps

See example.ipynb for the full jupyter notebook script that

1. Loads the data
2. Trains & tests a GRU4REC model
3. Loads & tests a pretrained GRU4REC model

• Before using run_train.py, I highly recommend that you to take a look at example.ipynb to see how the implementation works in general.
• Default parameters are the same as the TOP1 loss case in the GRU4REC paper.
• Intermediate models created from each training epochs will be stored to models/, unless specified.
• The log file will be written to logs/train.out.

$ python run_train.py > logs/train.out

Args:
    --loss_type: Loss function type. Should be one of the 'TOP1', 'BPR', 'CrossEntropy'.(Default: 'TOP1')
    --model_name: The prefix for the intermediate models that will be stored during the training.(Default: 'GRU4REC')
    --hidden_size: The dimension of the hidden layer of the GRU.(Default: 100)
    --num_layers: The number of layers for the GRU.(Default: 1)
    --batch_size: Training batch size.(Default: 50)
    --dropout_input: Dropout probability of the input layer of the GRU.(Default: 0)
    --dropout_hidden: Dropout probability of the hidden layer of the GRU.(Default: .5)
    --optimizer_type: Optimizer type. Should be one of the 'Adagrad', 'RMSProp', 'Adadelta', 'Adam', 'SGD'(Default: 'Adagrad')
    --lr: Learning rate for the optimizer.(Default: 0.01)
    --weight_decay: Weight decay for the optimizer.(Default: 0)
    --momentum: Momentum for the optimizer.(Default: 0)
    --eps: eps parameter for the optimizer.(Default: 1e-6)
    --n_epochs: The number of training epochs to run.(Default: 10)
    --time_sort: Whether to sort the sessions in the dataset in a time order.(Default: False)

• The results from this PyTorch Implementation gives a slightly better result compared to the original code that was written in Theano. I guess this comes from the difference between Theano and PyTorch & the fact that dropout has no effect in my single-layered PyTorch GRU implementation.
• The results were reproducible within only 2 or 3 epochs, unlike the original Theano implementation which runs for 10 epochs by default.

$ bash run_train.sh