Lodestar.Embeddings 0.6.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

Every text as its own row of ids, unpadded.

public IReadOnlyList<long[]> EncodeAll(IEnumerable<string> texts, CancellationToken cancellationToken = default)

Parameters — texts are the strings to encode. cancellationToken is observed between texts, because tokenizing a large corpus is not instant.

Returns — one long[] per text, in the order given, each already carrying its template tokens and already truncated. No padding and no rectangle.

Exceptions — OperationCanceledException when cancelled. ArgumentNullException when texts is null.

Example — two texts of different lengths stay different lengths.

using Lodestar.Embeddings.Tokenization;

var vocab = new Dictionary<string, int>(StringComparer.Ordinal)
{
    ["[UNK]"] = 0, ["token"] = 1, ["##ize"] = 2, ["text"] = 3,
    ["[CLS]"] = 4, ["[SEP]"] = 5, ["[PAD]"] = 6,
};
var tokenizer = new WordPieceTokenizer(
    vocab, unkToken: "[UNK]", continuationPrefix: "##", maxCharsPerWord: 100, lowercase: true);

var encoder = new BatchEncoder(tokenizer, new EncodingOptions
{
    Template = SpecialTokenTemplate.Bert,
    MaxLength = 8,
});

IReadOnlyList<long[]> sequences = encoder.EncodeAll(["text", "tokenize text"]);

int shorter = sequences[0].Length;   // => 3
int longer = sequences[1].Length;    // => 5

Remarks — this is the front half of EncodeBatch, and Pad is the back half. EncodeBatch is the two of them in one call, over the whole corpus at once; the pair exists for a caller that has to decide how to group the rows before they are laid out — running a corpus through an inference session a few rows at a time, say, where the widest row in a group is what the group costs.

Knowing the lengths before padding is the point. A group of similar lengths wastes little; a group that mixes 3 and 512 pays 512 for every row in it.

Applies to — net10.0, netstandard2.0.

See also — BatchEncoder.Pad, BatchEncoder.EncodeBatch, EncodedBatch.