DALL-E autoregressive image-token scoring (Geron Ch 16, morie.fn grdal)

Teacher-forced log-likelihood of the prefix plus the next-token distribution after temperature and optional top-k truncation. logits_fn receives the 0-based context as a numeric vector.

Usage

morie_geron_dalle_autoregressive_token(
  text_tokens,
  image_tokens_prefix,
  logits_fn,
  temperature = 1,
  top_k = NULL
)

Arguments

text_tokens, image_tokens_prefix

0-based token vectors.

logits_fn

Function(context) -> numeric logits of length V.

temperature

Positive temperature for the next-token step only.

top_k

Optional truncation.

Value

List with log_likelihood, token_logprobs, perplexity, next_token_probs, next_token, vocab_size.

Examples

skew <- function(ctx) c(0, 1, 2) + 0.1 * length(ctx)
morie_geron_dalle_autoregressive_token(c(0, 1), c(2, 0), skew,
    temperature = 0.5)
#> $log_likelihood
#> [1] -2.815212
#> 
#> $token_logprobs
#> [1] -0.407606 -2.407606
#> 
#> $perplexity
#> [1] 4.086161
#> 
#> $next_token_probs
#> [1] 0.01587624 0.11731043 0.86681333
#> 
#> $next_token
#> [1] 2
#> 
#> $vocab_size
#> [1] 3
#> 
#> $context_length
#> [1] 4
#> 
#> $estimate
#> [1] -2.815212
#> 
#> $n
#> [1] 2
#> 
#> $method
#> [1] "Autoregressive image-token likelihood"
#>