Returns the transition matrix q(x_{t+1} | x_t).

Returns the transition matrix q(x_{t+1} | x_t).

Get q(x_t), the n-step posterior. For example, for t = 0, it returns q0 unchanged. Args: q0: an array of floats specifying a

Get q(x_t), the n-step posterior. For example, for t = 0, it returns q0 unchanged. Args: q0: an array of floats specifying a

Returns probabilities from the reverse process p(x_{t-1} | x_t). Args: denoise_fn: the reverse process. Must support embed, call, and attend.

Returns probabilities from the reverse process p(x_{t-1} | x_t). Args: denoise_fn: the reverse process. Must support embed, call, and attend.

Get q(x_{t+1} | x_t), the one-step posterior efficiently. Args: qt_plus_1: an array of floats specifying a distribution over p(x_0).

Get q(x_{t+1} | x_t), the one-step posterior efficiently. Args: qt_plus_1: an array of floats specifying a distribution over p(x_0).

Samples from q(x_{t+1} | x_0), then computes q(x_t | x_{t+1}, x_0). Args: x_0: an array containing x_0 samples. These are expected t

Samples from q(x_{t+1} | x_0), then computes q(x_t | x_{t+1}, x_0). Args: x_0: an array containing x_0 samples. These are expected t

Samples batches of time steps to use.

Remove some keys from ignore list

Initializes values used by the get function.

Initializes values used by the get function.

Returns the KL for one term in the ELBO (time t) (loss L_t). This assumes x_start is a sample from x_0, from which we draw samples from q(x_t

Returns the KL for one term in the ELBO (time t) (loss L_t). This assumes x_start is a sample from x_0, from which we draw samples from q(x_t

Computes KL divergence between q(x_T) and the true distribution.

Computes KL divergence between q(x_T) and the true distribution.

Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols are ignored. This is modifi

We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids. Args: in

Returns q(x_t | x_0), the product of the first t matrices.

Returns q(x_t | x_0), the product of the first t matrices.

Get q(x_t), the n-step posterior. Can do efficiently for masks. For example, for t = 0, it returns q0 unchanged. Args:

Get q(x_t), the n-step posterior. Can do efficiently for masks. For example, for t = 0, it returns q0 unchanged. Args:

Returns the matrix Q = q(x_t | x_0) materialized over all x_0.

Returns the matrix Q = q(x_t | x_0) materialized over all x_0.

This function is only used to evaluate the self-BLEU score in conditional generation tasks to align with the implementation in https://github

A simple scheduler for masking policies. Args: dim: int, the dimensionality of the state space. schedule: a DiffusionSched

A simple scheduler for masking policies. Args: dim: int, the dimensionality of the state space. schedule: a DiffusionSched