1

I have read the paper: Simple Way to Initialize Recurrent Networks of Rectified Linear Units

Where can I download the toy benchmark dataset for RNNs this paper mentions?

I need addition problem benchmark.

Yan-Jen Huang
  • 113
  • 2

1 Answers1

0

In https://arxiv.org/abs/1504.00941 the authors describe the toy dataset:

We generated a training set of 100,000 examples and a test set of 10,000 examples as we varied T.

They give you values, e.g. $T = 400$. The start of $\S~4.1$ said that values in each example are uniformly drawn from the unit interval. Publishing the generated values on an FTP server would be silly. I suppose they might have published code, or the name of a PRNG algorithm together with a seed value, but supporting such code is not what they were focused on. They did supply hyperparameters, so you have enough information to generate similar examples.

The MNIST pixels, billion word corpus, and TIMIT datasets each come from their own sources, but clearly none of those is a "toy" problem as you asked about.

J_H
  • 880
  • 7
  • 11
  • Thank you. As you said, I generated it myself later, and found that the hyperparameters, such as learning rate, gradient clip is very important to train a good RNN model in adding problem. – Yan-Jen Huang Dec 21 '18 at 08:54