transform nn

[yuhanH]

Thanks for answering my issue. #63
Here is an application.
First, I build training umap with precalculated nearest neighbor input and return its model.
Then, I build the test-train nearest neighbor, and put it into umap_transform which avoids annoy_search for X, and use precalculated test-train nearest neighbor as input.
This NN input can give users flexibilities to add/explore their specific kernel into umap and transform_umap, not limited to the features input. Thanks.

This is an example:

devtools::install_github("jlmelville/vizier")
devtools::install_github("jlmelville/snedata")
fashion <- snedata::download_fashion_mnist()
fashion_train <- head(fashion, 60000)
fashion_test <- tail(fashion, 10000)

fashion_train.nn <- uwot:::annoy_nn(X = as.matrix(fashion_train[, 1:784]),
                                    k = 15, 
                                    metric = "cosine",
                                    ret_index = T)
# return umap map with annoy_nn input
set.seed(1337)
fashion_umap <- umap(X = NULL, nn_method = fashion_train.nn, ret_model = T)

# compute the query-reference annoy_nn
query_ref.nn <- annoy_search(X = as.matrix(fashion_test[, 1:784]), 
                             k = 15, 
                             ann = fashion_train.nn$index  )

# use the query-reference annoy_nn to transform query to reference
fashion_umap_test <- umap_transform(X = NULL,  model = fashion_umap, nn_method = query_ref.nn)


vizier::embed_plot(fashion_umap$embedding, fashion, cex = 0.5, title = "Fashion UMAP", alpha_scale = 0.075)
vizier::embed_plot(fashion_umap_test, fashion_test, cex = 0.5, title = "Fashion Test UMAP", alpha_scale = 0.075)

[jlmelville]

I understand now. Thank you for the PR. Happy to merge it if it will be useful to you. I will also add you as a contributor to the DESCRIPTION.

[yuhanH]

Cool. Thanks!