filter_layer_embeddings: Filter and Combine Embedding Vectors
In jonathanbratt/RBERTviz: Visualization Tools for RBERT

Description Usage Arguments Value Examples

Applies a filter + summarize to the given table of embeddings. Can be used to filter to specific layer outputs, or to average the outputs from multiple layers.

1	filter_layer_embeddings(embedding_df, layer_indices = 12L, sum_fun = mean)

`embedding_df`	A tbl_df of embedding vectors; the "output" component of `extract_features`.
`layer_indices`	Integer vector; which layers embeddings to keep.
`sum_fun`	A summarizing function to apply to the embedding vector components of the retained layers. Will be passed as the `.funs` argument to `summarize_at`. By default, `mean` is used, but other functions are possible. If this parameter is given a NULL value, no summarizing will be done, and only the filtered tbl_df will be returned.

The input tbl_df of embedding vectors, with the specified filtering and summarizing applied.

## Not run: 
# assuming something like the following has been run:
# feats <- RBERT::extract_features(...) # See RBERT documentation
# Then:
embeddings <- feats$output
embeddings_layer1 <- embeddings %>%
    filter_layer_embeddings(layer_indices = 1L)
# To keep the layer_index column, use sum_fun = NULL:
embeddings_layer1b <- embeddings %>%
    filter_layer_embeddings(layer_indices = 1L, sum_fun = NULL)
# Take the component-wise mean of the embeddings across layers 10-12.
embeddings_layer_mean <- embeddings %>%
    filter_layer_embeddings(layer_indices = 10:12)
# Take the component-wise maximum of the embeddings across layers 8-12.
# (Why would we do this? I don't know.)
embeddings_layer_max <- embeddings %>%
    filter_layer_embeddings(layer_indices = 8:12, sum_fun = max)

## End(Not run)