A PyTorch implement of Dilated RNN

This code implements the models described in the paper "Dilated Recurrent Neural Networks". We refer to their tensorflow implement.

Requirements

• Python 2.7
• PyTorch 0.3

Setup

1. Download the MNIST data via "torchvision.datasets.MNIST" and make the following configuration at "demo.py".

data_dir = '/home/fox/MNIST_data' # the path of MNIST data
n_steps = 28 # the step of RNN
input_dims = 28 # the input dimension of each RNN cell
n_classes = 10 # MNIST has ten classes for classifying
cell_type = "LSTM" # only support LSTM
hidden_structs = [20, 20] # Give a list of the dimension in each layer
dilations = [1, 2] # Give a list of the dilation in each layer

2. Run "demo.py" and you'll see:

==> Building a dRNN with LSTM cells
Iter 1, Step 100, Avarage Loss: 1.310388
Iter 1, Step 200, Avarage Loss: 0.929560
Iter 1, Step 300, Avarage Loss: 0.943462
Iter 1, Step 400, Avarage Loss: 0.843424
========> Validation Accuarcy: 0.753500
...

Notes

Currently, it only supports CPU and LSTM. Moreover, when the dilation starts not at 1, the output will be fused in the paper. This is not implemented now.