Convert the diffusion network's velocity prediction to x0 predidction. velocity: the predicted noise with shape [B, C, H, W]

Convert the diffusion network's velocity prediction to x0 predidction. velocity: the predicted noise with shape [B, C, H, W]

Convert the diffusion network's x0 prediction to noise predidction. x0: the predicted clean data with shape [B, C, H, W] xt:

Create a ScheduleSampler from a library of pre-defined samplers. :param name: the name of the sampler. :param diffusion: the diffusion ob

custom_collate_fn is used to print the index information when the original collate_fn fails

Sample x_{t+1} from the model using DDIM reverse ODE.

Generate samples from the model using DDIM. Same usage as p_sample_loop().

Decode one latent chunk with persistent decoder feature cache. For decoder_causal=False model: decodes each chunk independently with:

A default decoder for webdataset. This handles common file extensions: .txt, .cls, .cls2, .jpg, .png, .json, .npy, .mp, .pt, .pth, .pickl

Input: x: (B, H, W, C) weight: (C, K, K) bias: (C,) Output: y: (B, H, W, C)

The dynamic thresholding method.

The dynamic thresholding method.

Export a model to a platform-specific onnx format. Args: model: a torch.nn.Module object. export_path: export location. s

Get the feature cache map.

Flip a sequence and shift it right by one step. The operation reverses the sequence, drops the last element, and pads the front with ``shift_

Get Inception feature maps Parameters ---------- inp : torch.autograd.Variable Input tensor of shape Bx3xHxW. Val

Forward pass through the RMSNorm layer. Args: x (torch.Tensor): The input tensor. Returns: torch.Te

Fix bfloat16 support for nearest neighbor interpolation.