save_videos#

deepinv.utils.save_videos(vid_list, titles=None, time_dim=2, rescale_mode='min_max', figsize=None, save_fn=None, **plot_kwargs)[source]#

Saves an animation of a list of image sequences.

Plots videos as sequence of side-by-side frames, and saves animation (e.g. GIF) or displays as interactive HTML in notebook. This is useful for e.g. time-varying inverse problems. Individual frames are plotted with deepinv.utils.plot() vid_list can either be a video or a list of them. A video is defined as images of shape (B,C,H,W) augmented with a time dimension specified by time_dim, e.g. of shape (B,C,T,H,W) and time_dim=2. All videos must be same time-length.

Per frame of the videos, this function calls deepinv.utils.plot(), see its params to see how the frames are plotted.

Examples:

Display list of image sequences live in a notebook:

>>> from deepinv.utils import save_videos
>>> x = torch.rand((1, 3, 5, 8, 8)) # B,C,T,H,W image sequence
>>> y = torch.rand((1, 3, 5, 16, 16))
>>> save_videos([x, y], save_fn="vid.gif") # Save video as GIF
Parameters:

  • vid_list (Union[torch.Tensor, List[torch.Tensor]]) – video or list of videos as defined above

  • titles (Union[str, List[str]]) – titles of images in frame, defaults to None

  • time_dim (int) – time dimension of the videos. All videos should have same length in this dimension, or length 1. After indexing this dimension, the resulting images should be of shape (B,C,H,W). Defaults to 2

  • rescale_mode (str) – rescaling mode for deepinv.utils.plot(), defaults to “min_max”

  • figsize (tuple[int], None) – size of the figure. If None, calculated from size of img list.

  • save_fn (str) – if not None, save the animation to this filename. File extension must be provided, note anim_writer might have to be specified. Defaults to None

  • **plot_kwargs – kwargs to pass to deepinv.utils.plot()

Examples using save_videos:#

Building your diffusion posterior sampling method using SDEs

Building your diffusion posterior sampling method using SDEs