Keras implementation of the Crepe character-level convolutional neural net.
Switch branches/tags
Nothing to show
Clone or download
johnb30 Merge pull request #10 from liufuyang/master
Update to Keras v2.1.5 and use tensorflow
Latest commit 48eadde Aug 12, 2018


This is an re-implementation of the Crepe character-level convolutional neural net model described in this paper. The re-implementation is done using the Python library Keras(v2.1.5), using the Tensorflow-gpu(v1.6) backend. Keras sits on top of Tensorflow so can make use of cuDNN for faster models.


The model implemented here follows the one described in the paper rather closely. This means that the vocabulary is the same, fixed vocabulary described in the paper, as well as using stochastic gradient descent as the optimizer. Running the model for 9 epochs returns results in line with those described in the paper.

It's relatively easy to modify the code to use a different vocabulary, e.g., one learned from the training data, and to use different optimizers such as adam.



The usual command to run the model is (for tensorflow-gpu backend):



With current code it reaches similar level of accuracy stated on the paper. We get test accuracy above 0.88 after 10 epoch of running (with Adam(lr=0.001) optimizer)


We had to specify the kernel_initializer for all the Convolution1D layers as RandomNormal(mean=0.0, stddev=0.05, seed=None) as with the default initializer for Keras(v2.1.5) the model cannot converge. I don't exactly know the reason behind this but the lesson learned is that: initialization matters!