INN.Conv1d

CLASS INN.Conv1d(channels, kernel_size, method='NICE', **args)

A nonlinear 1d-convolutional INN layer.

Common parameters

• channels: number of channels
• kernel_size: kernel size
• method: This can be 'NICE', 'RealNVP' and 'ResFlow'
• activation_fn: Activation function for the coupling function. If the function is given, this argument will be ignored

RealNVP and NICE method

• mask: Mask for splitting input vectors

NICE

• m: Addition function, it should be a neural network maps vector with dimension channels // 2 to channels - channels // 2. If m=None, it will be generated automatically by INN.utils.default_net(dim, k, activation_fn);

RealNVP

• s: Multiplication function. It has the same dimension requirements as m;
• t: Addition function. It has the same dimension requirements as m;
• clip_n: (default: clip_n=1) Clipping the output of s to avoid extreme numbers between [-clip_n, clip_n]. The clipping is using tanh to keep the gradient;
• clip: Clipping the output of s if clip=True

ResFlow

• hidden: Dimension of hidden layers
• lipschitz_constrain: Lipschitz constrain number, it should be lower than 1. Low value may decrease the computation power of the neural network;
• mem_efficient: Using memory-efficient back-propagation if it is True;
• est_steps: Number of iterations for estimating gradients and Jacobians

Methods

forward(x, log_p0=0, log_det_J=0)

Compute the forward pass. The output will be the transformed x. If compute_p=True, logp and logdet will also returned.

inverse(y, **args)

Compute the inverse of y. The **args is a placeholder for consistent format.