CSCI E-89 Ankit Srivastava - Auto Completion of StackOverflow Post Titles
Published 2019-05-12
Final Project: Auto Completion of StackOverflow Post Titles using CharRNN. We train a 2 layer character level LSTM with 512 neurons in each layer. We train on 100 characters in last 100 time steps to predict next character in each iteration. We use minibatch size of 100 and train for 100K iterations to optimize for accuracy and reduce validation loss. We output softmax probabilities for each character in vocabulary. The project leverages Azure Deep Learning VM pre-configured with NVIDIA drivers and CUDA toolkit. We use Tensorflow for training the model and Tensorboard for computational graph and scalar output visualization. We test on various input sequences to find strengths and weaknesses of model. We found model being computationally expensive to train and susceptible to typos. However, after 100K iterations, it was pretty good at synthetic generation of post titles and autocomplete test input sequences. It could auto-complete even when we started from middle of the word or introduced typos in input sequence.
All Comments (2)
