Working with datasets: Env, FrameEnv, SeqEnv, UserDataset

Introduction

Although this project is primarily built for the ML20M dataset, it comes with native support for any similar dataset that more or less looks like the latter. I named the name class environment (similar to any of the RL library), it takes two string locations as main arguments. In most of the snippets, it is defined as:

# fixed-length frame
env = recnn.env.FrameEnv('../../data/embeddings/ml20_pca128.pkl',
                         '../../data/ml-20m/ratings.csv', frame_size, batch_size)

# OR

# dynamical length sequential
state_encoder = nn.LSTM(129, 256, batch_first=True).to(cuda)
env = recnn.env.SeqEnv('../../data/embeddings/ml20_pca128.pkl',
                         '../../data/ml-20m/ratings.csv', batch_size, state_encoder, cuda)

Some cool features and things this library handles by itself:

1. Label encoding items (movies). In the original ml20m dataset, the index of movies is not linear, and cannot be fit into int32 by default. I label encode the movies for efficient indexing and better memory usage.

2. Sorting and clustering. Items by default are sorted by timestamp, alternatively, you can provide different argument to sort it.

Making in work with your own data

A quick reminder of how the ml20 looks like:

   user  movie  rating   timestamp
0    55     19       2  1568310847
1    39      5       5  1568318423
2    88      8       2  1568323421
3    79     16       7  1568334534
4    98     19       5  1568343643

You can make it work with your data by providing data_cols argument. This dataset is a little different

data_cols = {user_id='user', rating='rating', timestamp='timestamp', item='movie', sort_users=False}
env = recnn.env.FrameEnv('your embeddings',
                         'your dataset', frame_size, batch_size, data_cols=data_cols)

Sequential state representation:

As in the example above, sequential state representation is defined as:

# dynamical length sequential
state_encoder = nn.LSTM(129, 256, batch_first=True).to(cuda)
env = recnn.env.SeqEnv('../../data/embeddings/ml20_pca128.pkl',
                         '../../data/ml-20m/ratings.csv', batch_size, state_encoder, cuda)

You can use different state encoders, varying from basic LSTM/RNN/GRU to more efficient models such as Chaos Free Networks and Temporal Convolution.

The important thing is to include the state rep model in the optimizer with other networks!

pm1 = list(policy_net.parameters()) + list(state_encoder.parameters())
pm2 = list(value_net.parameters()) + list(state_encoder.parameters())
value_optimizer = recnn.optim.RAdam(pm1, ...)
policy_optimizer = recnn.optim.RAdam(pm2, ...)