Skip to content

Latest commit

 

History

History
75 lines (50 loc) · 2.65 KB

README.md

File metadata and controls

75 lines (50 loc) · 2.65 KB

Clockwork RNN

This project is an implementation of the Clockwork RNN (see paper).

See this section for dependencies.

People : Paul Mustière, Pandav2 aka David Panou

Organization : UPMC - Master Data Science

Clockwork-RNN

The model can be found under models/clockwork_rnn.py.

The current main.py replicates the sequence generation task described in the paper, but the implementation should be able to handle other tasks.

To monitor the training, you can use TensorBoard:

tensorboard --reload_interval 2 --logdir log

Results

We ran different sizes of Clockwork RNN as well as LSTMs to compare performance with similar numbers of parameters.

The following table summarizes the results:

Number of parameters Clockwork RNN (MSE) LSTM (MSE)
~70 5.0e-2 1.4e-1
~480 2.8e-3 8.2e-2
~800 1.7e-3 8.8-2

They were obtained with a learning rate of 0.01 for Clockwork RNN, and 0.001 for LSTM (which was more unstable). Please beware due to hardward constraints, those results are not average over several runs.

This table shows the dimension of the hidden state used and the resulting exact number of parameters:

Number of parameters Clockwork RNN (Hidden dimension / Total parameters) LSTM (Hidden dimension / Total parameters)
~70 9 / 73 3 / 64
~480 27 / 469 10 / 491
~800 36 / 802 13 / 794

The following graph shows the MSE loss during training, scaled by epochs:

Graph of loss during training

The next graph shows the same MSE loss, without the scaling by epochs. It shows that Clockwork RNN are indeed faster to train since they require less computations.

Graph of loss during training

The following plot shows the generated signal (and the target) for the best performing Clockwork RNN:

Plot of generated signal

Generated signals for other models can be found under results/

Dependencies

  • NumPy
  • TensorFlow (tested with 0.12.1)
  • Matplotlib

For generating sequences from WAV files:

  • SciPy