credit.diffusion ================ .. py:module:: credit.diffusion Classes ------- .. autoapisummary:: credit.diffusion.ModelPrediction credit.diffusion.GaussianDiffusion credit.diffusion.ModifiedGaussianDiffusion Functions --------- .. autoapisummary:: credit.diffusion.extract credit.diffusion.standard_normal_noise credit.diffusion.log_uniform_noise credit.diffusion.linear_beta_schedule credit.diffusion.cosine_beta_schedule credit.diffusion.sigmoid_beta_schedule Module Contents --------------- .. py:class:: ModelPrediction Bases: :py:obj:`tuple` .. py:attribute:: pred_noise .. py:attribute:: pred_x_start .. py:function:: extract(a, t, x_shape) .. py:function:: standard_normal_noise(shape=None, tensor=None, device=None) Generate standard normal noise (mean=0, std=1), either using a shape or a reference tensor. :param shape: Shape for the output tensor. :type shape: tuple, optional :param tensor: Reference tensor to match shape and device. :type tensor: torch.Tensor, optional :param device: Device for output. Optional if tensor is provided. :type device: torch.device or str, optional :returns: Noise sampled from N(0, 1). :rtype: torch.Tensor .. py:function:: log_uniform_noise(shape=None, tensor=None, sigma_min=0.02, sigma_max=200.0, device=None) Sample noise from a log-uniform distribution over standard deviations. :param shape: Shape of the output noise tensor. Required if `tensor` is not provided. :type shape: tuple, optional :param tensor: Reference tensor to match shape. Used if `shape` is not provided. :type tensor: torch.Tensor, optional :param sigma_min: Minimum standard deviation. :type sigma_min: float :param sigma_max: Maximum standard deviation. :type sigma_max: float :param device: Device for the output. If None, inferred from `tensor` or defaults to 'cpu'. :type device: torch.device or str, optional :returns: Noise tensor sampled from N(0, σ^2), with σ ~ log-uniform. :rtype: torch.Tensor .. py:function:: linear_beta_schedule(timesteps) linear schedule, proposed in original ddpm paper .. py:function:: cosine_beta_schedule(timesteps, s=0.008) cosine schedule as proposed in https://openreview.net/forum?id=-NEXDKk8gZ .. py:function:: sigmoid_beta_schedule(timesteps, start=-3, end=3, tau=1, clamp_min=1e-05) sigmoid schedule proposed in https://arxiv.org/abs/2212.11972 - Figure 8 better for images > 64x64, when used during training .. py:class:: GaussianDiffusion(model, *, image_size, timesteps=1000, sampling_timesteps=None, objective='pred_v', beta_schedule='sigmoid', noise_type='normal', schedule_fn_kwargs=dict(), ddim_sampling_eta=0.0, auto_normalize=True, offset_noise_strength=0.0, min_snr_loss_weight=False, min_snr_gamma=5, immiscible=False) Bases: :py:obj:`torch.nn.Module` Base class for all neural network modules. Your models should also subclass this class. Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes:: import torch.nn as nn import torch.nn.functional as F class Model(nn.Module): def __init__(self) -> None: super().__init__() self.conv1 = nn.Conv2d(1, 20, 5) self.conv2 = nn.Conv2d(20, 20, 5) def forward(self, x): x = F.relu(self.conv1(x)) return F.relu(self.conv2(x)) Submodules assigned in this way will be registered, and will also have their parameters converted when you call :meth:`to`, etc. .. note:: As per the example above, an ``__init__()`` call to the parent class must be made before assignment on the child. :ivar training: Boolean represents whether this module is in training or evaluation mode. :vartype training: bool .. py:attribute:: model .. py:attribute:: channels .. py:attribute:: self_condition .. py:attribute:: condition .. py:attribute:: image_size .. py:attribute:: objective :value: 'pred_v' .. py:attribute:: num_timesteps .. py:attribute:: sampling_timesteps :value: None .. py:attribute:: is_ddim_sampling .. py:attribute:: ddim_sampling_eta :value: 0.0 .. py:attribute:: immiscible :value: False .. py:attribute:: offset_noise_strength :value: 0.0 .. py:attribute:: normalize .. py:attribute:: unnormalize .. py:property:: device .. py:method:: predict_start_from_noise(x_t, t, noise) .. py:method:: predict_noise_from_start(x_t, t, x0) .. py:method:: predict_v(x_start, t, noise) .. py:method:: predict_start_from_v(x_t, t, v) .. py:method:: q_posterior(x_start, x_t, t) .. py:method:: model_predictions(x, t, x_self_cond=None, x_cond=None, clip_x_start=False, rederive_pred_noise=False) .. py:method:: p_mean_variance(x, t, x_self_cond=None, x_cond=None, clip_denoised=True) .. py:method:: p_sample(x, t: int, x_self_cond=None, x_cond=None) .. py:method:: p_sample_loop(shape, x_cond, return_all_timesteps=False) .. py:method:: ddim_sample(shape, x_cond, return_all_timesteps=False, disable_tqdm=True) .. py:method:: sample(x_cond, batch_size=16, return_all_timesteps=False) .. py:method:: interpolate(x1, x2, x_cond=None, t=None, lam=0.5) .. py:method:: noise_assignment(x_start, noise) .. py:method:: q_sample(x_start, t, noise=None) .. py:method:: p_losses(x_start, t, x_cond, noise=None, offset_noise_strength=None) .. py:method:: forward(img, x_cond, *args, **kwargs) .. py:class:: ModifiedGaussianDiffusion(*args, **kwargs) Bases: :py:obj:`GaussianDiffusion` Base class for all neural network modules. Your models should also subclass this class. Modules can also contain other Modules, allowing them to be nested in a tree structure. You can assign the submodules as regular attributes:: import torch.nn as nn import torch.nn.functional as F class Model(nn.Module): def __init__(self) -> None: super().__init__() self.conv1 = nn.Conv2d(1, 20, 5) self.conv2 = nn.Conv2d(20, 20, 5) def forward(self, x): x = F.relu(self.conv1(x)) return F.relu(self.conv2(x)) Submodules assigned in this way will be registered, and will also have their parameters converted when you call :meth:`to`, etc. .. note:: As per the example above, an ``__init__()`` call to the parent class must be made before assignment on the child. :ivar training: Boolean represents whether this module is in training or evaluation mode. :vartype training: bool .. py:attribute:: channels .. py:attribute:: history_len .. py:attribute:: criterion :value: None .. py:method:: load_loss(criterion) .. py:method:: forward(img, x_cond=None, *args, **kwargs) .. py:method:: p_losses(x_start, t, x_cond, noise=None, offset_noise_strength=None) .. py:method:: model_predictions(x, t, x_self_cond=None, x_cond=None, clip_x_start=False, rederive_pred_noise=False)