https://github.com/facebookresearch/DiT/blob/main/diffusion/diffusion_utils.py
calculate warm up beta params: - beta start - beta end - num diffusion timestamps - type = using float 64 - warm up - frac
get the beta schedule need a beta schedule to sample from the diffusion process - beta schedule is a list of beta values - using the linear regression for this schedule * beta start, beta end , num diffusion timestamp
get a predefined beta schedule- this has to be similar to the limit in our diffusion timestamp
create a schedule that discretizes the given alpha_t function - alpha_t function is a function that takes in a float and returns a float - the schedule is a list of tuples where each tuple contains a float and the corresponding alpha_t values - the schedule is created by discretizing the given alpha_t function into num_steps steps - the schedule is then sorted by the float values in ascending order - the schedule is then returned - the schedule is used to create the beta schedule params: - num diffusion timestamps - alpha_t function - using lambda taking the argument t from 0 to 1 and produces a cumulative product of (1-beta) - max-beta - the maximum beta value - lower than 1 to prevent singularities.
Above is our params for Diffusion
this is our maindiffusion.cpp use float 64 for t he beta betas, model mean type, model var type, loss type
Calculate q(x_t | x_{t-1}, t) - this is the forward process
then posterior q(x_{t-1} | x_t, t) - this is the reverse process
log the calculation of the reverse process
first calculate the distribution of q(x_t | x_{t-1}, t)
apply model to get the p(x_{t-1} | x_t) and predict x, x_0 params: - model takes signal(in this case the noise and loss classes) and batch of timesteps as its input - x: the [N x C x ....] at time of t - t: a 1 D vector of timesteps - clip denoised: clip the denoised signal into [-1, 1] - denoised_fn: if none apple x_start preduction before applying the clip denoised - model Kwags : model kwags for the model- dict of extra keywords to pass to the model - return: - mean - variance - log_variance - pred_xstart
predict x_start from eps predict eps from xstart
condition the mean by computing the graient of the log proability of the mean with respect to the mean computes grad(log(p(y|x))) the condition is on y strategy used is Sohl- Dickstein et al (2015) - lets look for something more recient
condition the score: compute the P_mean_variance output
Sample x_{t-1} at given timestep params: - model: model above - x: the current vector at x_{t-1} - t: value of t starting at 0 for first diffusion step - clip denoised- clip the x_start prediction top [-1, 1] - denoised_fn- if none apply x_start prediction before applying the clip denoised - cond_fn - gradient function acts similarly to the model - model _kwargs = extra keywords for conitioning - return: - sample - random _ predict X_start - predict x_0
generate a p_sample loop params: - model: model above - shape: shape of sample (N, C, H, W) - noise: noise schedule- same as sample - clip denoised - denoised_fn - model _kwargs - device: use the models parameters - progress - show a tqdm progress bar return: non differentiable batch of samples
Generate a P_sample _loop_progressibe params: - yield intermediate samples from each timestamp of diffusion - arguments are same as P_sample Loop - returns: - generator over dicts where each one returns p_sample()