search

anomal / wordsimilarity

0 stars 0 forks source link

Performance issue: response time is 35+ seconds #1

Open anomal opened 5 years ago

anomal commented 5 years ago

Response time is 35+ seconds due to the reduction of n-dimensional coordinates to 2-dimensional coordinates through t-distributed stochastic neighbor embedding (t-SNE).

anomal commented 5 years ago

This is due to using BarnesHutTsne.java with theta=0.0, which uses Tsne.java instead. The Barnes-Hut approximation method is supposed to be much faster than the default Tsne.java, but due to a bug in BarnesHutTsne.java that both produces poor results and makes it much slower than Tsne.java, Tsne.java is used instead. The next release of deeplearning4j has BarnesHutTsne.java rewritten to match the reference implementation, which should solve both the performance and accuracy issues.