def plot_all_traversals_grid(model,
latent_dim,
input_size,
input_shape,
num_sample=1,
figsize=(30, 30),
filename=None,
conditional_target=None,
device='cpu'):
"""
Generates a grid of images for all latent dimensions, where each row corresponds to a traversal along a latent
dimension.
:param model: INN used for sampling
:param latent_dim: dimension of the latent space
:param input_size: total number of elements in the input of the INN
:param input_shape: shape of the input for the INN
:param idx: Index of a continuous latent dimension to traverse. If None, no latent is traversed and all latent
dimensions are randomly sampled or kept fixed.
:param num_sample: number of samples to generate
:param figsize: the size of the generated plot
:param filename: file name under which the plot will be saved. (optional)
:return: None
"""
model.to(device)
model.eval()
if len(input_shape) != 3:
raise ValueError("input_shape must be 3-dimensional")
grid = []
for idx in range(latent_dim):
latent_samples = tra.traverse_continous_line(
latent_dim,
input_size,
idx,
num_sample,
False,
lat_img=None,
conditional_target=conditional_target)
latent_samples = latent_samples.view(num_sample, input_shape[0],
input_shape[1], input_shape[2])
generate = model(latent_samples.to(device), rev=True)
grid.append(generate)
grid = torch.cat(grid)
imshow(torchvision.utils.make_grid(grid.detach(), num_sample), figsize,
filename if (filename is not None) else None)
def plot_latent_traversal_line(model, latent_dim, input_size, input_shape, idx, num_sample=1, figsize=(30, 30), filename=None, dataset=None, conditional_target=None, device='cpu'):
"""
Generates an image traversal through a latent dimension.
:param model: INN used for sampling
:param latent_dim: dimension of the latent space
:param input_size: total number of elements in the input of the INN
:param input_shape: shape of the input for the INN
:param idx: Index of a continuous latent dimension to traverse. If None, no latent is traversed and all latent
dimensions are randomly sampled or kept fixed.
:param num_sample: number of samples to generate
:param figsize: the size of the generated plot
:param filename: file name under which the plot will be saved. (optional)
:param dataset: dataset to draw images from for which the latent traversal will be created
:return: None
"""
model.to(device)
model.eval()
if dataset is not None:
loader = dl.get_loader(dataset, num_sample)
img, label = next(iter(loader))
lat_img = model(img)
lat_img = lat_img.view(lat_img.size(0), -1)
else:
lat_img = None
latent_samples = tra.traverse_continous_line(latent_dim, input_size, idx, num_sample, False, lat_img, conditional_target=conditional_target)
if len(input_shape) == 2:
latent_samples = latent_samples.view(num_sample, input_shape[0], input_shape[1])
elif len(input_shape) == 3:
latent_samples = latent_samples.view(num_sample, input_shape[0], input_shape[1], input_shape[2])
else:
raise ValueError("input_shape is neither 2- nor 3-dimensional")
generate = model(latent_samples.to(device), rev=True)
imshow(torchvision.utils.make_grid(generate.detach(), num_sample), figsize, filename if (filename is not None) else None)