Skip to content

Extract layers of hidden states

Source: R/1_1_textEmbed.R
textEmbedRawLayers.Rd

textEmbedRawLayers extracts layers of hidden states (word embeddings) for all character variables in a given dataframe.

Usage

textEmbedRawLayers(
  texts,
  model = "bert-base-uncased",
  layers = -2,
  return_tokens = TRUE,
  word_type_embeddings = FALSE,
  decontextualize = FALSE,
  keep_token_embeddings = TRUE,
  device = "cpu",
  tokenizer_parallelism = FALSE,
  model_max_length = NULL,
  max_token_to_sentence = 4,
  hg_gated = FALSE,
  hg_token = Sys.getenv("HUGGINGFACE_TOKEN", unset = ""),
  trust_remote_code = FALSE,
  logging_level = "error",
  sort = TRUE
)

Arguments

texts

A character variable or a tibble with at least one character variable.

model

(character) Character string specifying pre-trained language model (default = 'bert-base-uncased'). For full list of options see pretrained models at HuggingFace. For example use "bert-base-multilingual-cased", "openai-gpt", "gpt2", "ctrl", "transfo-xl-wt103", "xlnet-base-cased", "xlm-mlm-enfr-1024", "distilbert-base-cased", "roberta-base", or "xlm-roberta-base". Only load models that you trust from HuggingFace; loading a malicious model can execute arbitrary code on your computer).

layers

(character or numeric) Specify the layers that should be extracted (default -2, which give the second to last layer). It is more efficient to only extract the layers that you need (e.g., 11). You can also extract several (e.g., 11:12), or all by setting this parameter to "all". Layer 0 is the decontextualized input layer (i.e., not comprising hidden states) and thus should normally not be used. These layers can then be aggregated in the textEmbedLayerAggregation function.

return_tokens

(boolean) If TRUE, provide the tokens used in the specified transformer model. (default = TRUE)

word_type_embeddings

(boolean) Wether to provide embeddings for each word/token type. (default = FALSE)

decontextualize

(boolean) Wether to dectonextualise embeddings (i.e., embedding one word at a time). (default = TRUE)

keep_token_embeddings

(boolean) Whether to keep token level embeddings in the output (when using word_types aggregation). (default= TRUE)

device

(character) Name of device to use: 'cpu', 'gpu', 'gpu:k' or 'mps'/'mps:k' for MacOS, where k is a specific device number. (default = "cpu")

tokenizer_parallelism

(boolean) If TRUE this will turn on tokenizer parallelism. (default = FALSE).

model_max_length

The maximum length (in number of tokens) for the inputs to the transformer model (default the value stored for the associated model).

max_token_to_sentence

(numeric) Maximum number of tokens in a string to handle before switching to embedding text sentence by sentence. (default= 4)

hg_gated

Set to TRUE if the accessed model is gated.

hg_token

The token needed to access the gated model. Create a token from the ['Settings' page](https://huggingface.co/settings/tokens) of the Hugging Face website. An an environment variable HUGGINGFACE_TOKEN can be set to avoid the need to enter the token each time.

trust_remote_code

use a model with custom code on the Huggingface Hub

logging_level

(character) Set the logging level. (default ="error") Options (ordered from less logging to more logging): critical, error, warning, info, debug

sort

(boolean) If TRUE sort the output to tidy format. (default = TRUE)

Value

The textEmbedRawLayers() takes text as input, and returns the hidden states for each token of the text, including the [CLS] and the [SEP]. Note that layer 0 is the input embedding to the transformer, and should normally not be used.

See also

See textEmbedLayerAggregation and textEmbed.

Examples

# Get hidden states of layer 11 and 12 for "I am fine".
if (FALSE) { # \dontrun{
imf_embeddings_11_12 <- textEmbedRawLayers(
  "I am fine",
  layers = 11:12
)

# Show hidden states of layer 11 and 12.
imf_embeddings_11_12
} # }

GitHub