Stage

DeepAlignmentStage

class DeepAlignmentStage(mean_shape, input_size=112, p_dropout=0.5, is_first=True, patch_size=16, norm_type='batch')

Bases: sphinx.ext.autodoc.importer._MockObject

A single Deep Alignment Stage

Parameters:

• mean_shape (numpy.ndarray or torch.Tensor) – the mean shape
• input_size (int or tuple, optional) – the size of the input images (the default is 112)
• p_dropout (float, optional) – the dropout probability (the default is 0.5)
• is_first (bool) – whether the current stage is the first one or not
• patch_size (int, optional) – the patch size for heatmap generation (the default is 16)
• norm_type (str, optional) – which kind of normalization to apply (the default is “instance”)

forward(input_tensor, prev_lmk=None, prev_hidden=None)

Feeds an input image (and the landmarks and hidden activations if given) through the actual stage

Parameters:

• input_tensor (torch.Tensor) – the input image
• prev_lmk (torch.Tensor or None) – the landmarks of the previous stage, if the current stage is not the first one
• prev_hidden (torch.Tensor or None) – the activations of the previous stage’s hidden layer, if the current stage is not the first one

Raises:

ValueError – If the current stage is not the first one, but no prev_lmk or no prev_hidden are passed

Returns:

• torch.Tensor – the predicted landmarks
• torch.Tensor – the activations of the hidden layer to be used in the next stage