torch.distributions.transforms.TanhTransform

class TanhTransform extends Transform

new TanhTransform(options?: TransformOptions)

readonlydomain(unknown)
readonlycodomain(unknown)
readonlybijective(unknown)
readonlysign(unknown)

Tanh transform: y = tanh(x) (maps ℝ → (-1, 1)).

Transform via the mapping y = tanh(x). This transform maps real numbers to the interval (-1, 1). The inverse is arctanh: x = 0.5 * log((1 + y) / (1 - y)).

\begin{aligned} \text{Forward: } y = \tanh(x) = \frac{e^x - e^{-x}}{e^x + e^{-x}} \\ \text{Inverse: } x = \text{arctanh}(y) = \frac{1}{2}\ln\left(\frac{1+y}{1-y}\right) \\ \text{Log-det Jacobian: } \log\left|\frac{dy}{dx}\right| = \ln(1 - y^2) \\ \text{Derivative: } \frac{dy}{dx} = 1 - \tanh^2(x) = 1 - y^2 \\ \text{Domain: } \mathbb{R}, \text{Codomain: } (-1, 1) \end{aligned}

torch.distributions.transforms.TanhTransform

readonlydomain(unknown)

readonlycodomain(unknown)

readonlybijective(unknown)

readonlysign(unknown)

Tanh transform: y = tanh(x) (maps ℝ → (-1, 1)).

Transform via the mapping y = tanh(x). This transform maps real numbers to the interval (-1, 1). The inverse is arctanh: x = 0.5 * log((1 + y) / (1 - y)).

\begin{aligned} \text{Forward: } y = \tanh(x) = \frac{e^x - e^{-x}}{e^x + e^{-x}} \\ \text{Inverse: } x = \text{arctanh}(y) = \frac{1}{2}\ln\left(\frac{1+y}{1-y}\right) \\ \text{Log-det Jacobian: } \log\left|\frac{dy}{dx}\right| = \ln(1 - y^2) \\ \text{Derivative: } \frac{dy}{dx} = 1 - \tanh^2(x) = 1 - y^2 \\ \text{Domain: } \mathbb{R}, \text{Codomain: } (-1, 1) \end{aligned}