Keras 3 API documentation / KerasNLP / Samplers / ContrastiveSampler

ContrastiveSampler

[source]

ContrastiveSampler class

keras_nlp.samplers.ContrastiveSampler(k=5, alpha=0.6, **kwargs)

Contrastive Sampler class.

This sampler implements contrastive search algorithm. In short, the sampler chooses the token having the max "score" as the next token. The "score" is a weighted sum between token's probability and max similarity against previous tokens. By using this joint score, contrastive sampler reduces the behavior of duplicating seen tokens.

Arguments

  • k: int, the k value of top-k. Next token will be chosen from k tokens.
  • alpha: float, the weight of minus max similarity in joint score computation. The larger the value of alpha, the score relies more on the similarity than the token probability.

Call arguments

{{call_args}}

Examples

causal_lm = keras_nlp.models.GPT2CausalLM.from_preset("gpt2_base_en")

# Pass by name to compile.
causal_lm.compile(sampler="contrastive")
causal_lm.generate(["Keras is a"])

# Pass by object to compile.
sampler = keras_nlp.samplers.ContrastiveSampler(k=5)
causal_lm.compile(sampler=sampler)
causal_lm.generate(["Keras is a"])