fancy-cnn
Multiparadigm Sequential Convolutional Neural Networks for text classification.
First off, this class of sequential convolutional networks is quite GEMM
intensive, and really isn't suited to a CPU. You also really should use CuDNN
when training.
Bleeding edge installations of Theano and Keras are required. The TensorFlow backend has not been tested with the fancier TimeDistributed
architectures. When considering your hardware, note that...
If you want to use CuDNN
, you really should also put
[dnn]
conv.algo_fwd = time_on_shape_change
conv.algo_bwd = time_on_shape_change
to your .theanorc
. This will let Theano (at compilation time) pick the fastest convolution implementation given your input and kernel size.
datasets/aclImdb/
./download.sh
embeddings/
./download.sh
This gets all the embeddings from S3
and sets up the IMDB dataset for processing. Yelp instructions forthcoming.
Look at (and run) prepare_imdb*.py
to prepare your data! Look at this to see how to train the IMDB model.
Look at (and run) prepare_yelp*.py
to prepare your data! Look at this to see how to train a model on Yelp.