VariationalAutoencoder

VariationalAutoencoder draws ManimML’s autoencoder as a NeuralNetwork: the image input, the feed-forward layers of the encoder (neurons per layer, [5, 3] by default), the latent space, the decoder ([3, 5]) and the image output, both size cells square. layers() gives these layers and latent_index() the place of the latent space among them.

• Latent space. An embedding box with a cloud of points points for seed and the 1σ and 2σ rings of its Gaussian in ring_color.
• Code. The image’s point in the latent space, drawn in pulse_color, is μ + L·epsilon (L the Cholesky factor of the covariance, the reparameterisation trick). epsilon is in units of the standard normal distribution, (1.3, -0.7) by default; the point stays inside the box.
• Forward pass. With forward_pass the pulses run from the input to the output, and the point appears as they reach the latent space. Without one it is always shown.

Gallery scenes: VariationalAutoencoder, Vae.

VariationalAutoencoder 5 sPlayUnmuteCopy render command

Vae 5 sPlayUnmuteCopy render command

use fframes_kit::prelude::*;

fn autoencoder<'a>(ui: &mut Ui<'a, '_, '_>) -> Svgr<'a> {
    ui.show(VariationalAutoencoder {
        input: "digit.png",
        output: "digit.png",
        epsilon: (0.4, 1.1),
        forward_pass: Some(ForwardPass { start: 0.9, duration: 3.2 }),
        enter: Some(FadeIn::default()),
        ..ui.defaults()
    })
}

API

Components are plain structs. Set the fields you need and fill the rest with ..ui.defaults(): the defaults below come from the current theme (theme. values are fields of Theme).

VariationalAutoencoder

Field	Type	Default	Description
input	&str	""	File name of the input image in the media folder.
output	&str	""	File name of the reconstructed output image in the media folder.
size	usize	8	Width and height of the images in cells, for the convolution-style maps.
encoder	Vec<usize>	vec![5, 3]	Neurons in each feed-forward layer of the encoder, from the input towards the latent space.
decoder	Vec<usize>	vec![3, 5]	Neurons in each feed-forward layer of the decoder, from the latent space towards the output.
points	usize	120	Number of points in the cloud drawn in the latent box.
seed	u64	1	Seed of the point cloud; the same seed gives the same cloud on every frame and machine.
epsilon	(f64, f64)	(1.3, -0.7)	Noise of this image’s latent point, in units of the standard normal distribution (the point is the mean plus the noise scaled by the Gaussian).
center	(f64, f64)	(0.0, 0.0)	The middle of the network, in units of ui.space.
spacing	f32	130.0	Pixels between neighbouring layers.
node_radius	f32	16.0	Radius in pixels of a neuron.
node_spacing	f32	46.0	Pixels between neighbouring neurons of a layer.
cell	f32	14.0	Size in pixels of one cell of an image.
forward_pass	Option<ForwardPass>	None	Pulses that run from the input to the output; the latent point appears when they reach it. None for none.
node_color	Color	theme.colors.text	Colour of the neuron and box outlines and of the text.
node_fill	Color	theme.colors.surface	Fill colour inside the neurons and boxes.
edge_color	Color	theme.colors.muted	Colour of the connection lines.
pulse_color	Color	theme.colors.highlight	Colour of the pulses and of this image’s latent point.
ring_color	Color	theme.colors.accent	Colour of the Gaussian’s 1σ and 2σ rings in the latent box.
stroke_width	f32	2.5	Width in pixels of the outlines and connection lines.
enter	Option<FadeIn>	None	Fade in at the start; None for none.
exit	Option<FadeOut>	None	Fade out before the end; None for none.

Methods:

• fn layers(&self) -> Vec<Layer>: The layers from the input image to the output image.
• fn latent_index(&self) -> usize: The place of the latent space among layers.

Feature ml · Source