loss_functions

Module containing loss functions for training,

MSE_Loss

source

---

wrapper for a simple MSE loss with the same return shape. Forward function takes

Args

yhat, ytrue

Returns

MSE loss mean

Methods:

.forward

source

.forward(
   yhat, ytrue
)

---

GaussNLLLoss

source

---

Wrapper for a Gaussian negative log likelihood loss. Forward:

Args

yhat, ytrue

Returns

NLL loss mean

Methods:

.forward

source

.forward(
   yhat, ytrue
)

---

GaussNLL_VAR0_Loss

source

---

Gaussian negative log likelihood loss with a trainable variance parameter. Forward:

Args

yhat, ytrue

Returns

NLL loss mean

Methods:

.forward

source

.forward(
   yhat, ytrue
)

---

GaussNLL_VAR_Loss

source

---

Gaussian negative log likelihood loss treating logsigma as a second order polynomial expansion similar to the noisemodel in MAVE NN. Forward:

Args

yhat, ytrue

Returns

NLL loss mean

Methods:

.calc_logsigma

source

.calc_logsigma(
   yhat
)

---

function to compute the variance based on a order polinomial expansion

.forward

source

.forward(
   yhat, ytrue
)

---

NoiseLayer

source

NoiseLayer(
   model_params
)

---

Base class for original MAVE-NN noise layers

Args

• model_params (dict) : Dictionary of model parameters.
• key : polynomial_order

Attributes

• poly_order (int) : Order of polynomial expansion for noise model.

Methods:

.compute_nlls

source

.compute_nlls(
   yhat, ytrue
)

---

Compute the negative log-likelihoods for the given predictions and targets defined in the derived noise model classes.

Args

• yhat (torch.Tensor) : Predictions from the model.
• ytrue (torch.Tensor) : Targets for the model.

Returns

• Tensor : The negative log-likelihoods for each sample in the batch.

.forward

source

.forward(
   yhat, ytrue
)

---

GaussianNoise

source

GaussianNoise(
   model_params
)

---

A Gaussian noise distribution for GE regression

Args

• model_params (dict) : Dictionary of model parameters.

Attributes

• poly_order (int) : Order of polynomial expansion for noise model.

---

Methods: compute_nlls: Compute the negative log likelihood using the computed logsigma

Methods:

.compute_params

source

.compute_params(
   yhat, y_true = None
)

---

Compute layer parameters governing p(y|yhat).

.compute_nlls

source

.compute_nlls(
   yhat, ytrue
)

---

Compute negative log likelihood contributions for each datum.

---

KLD_diag_gaussians

source

.KLD_diag_gaussians(
   mu: torch.Tensor, logvar: torch.Tensor, p_mu: torch.Tensor,
   p_logvar: torch.Tensor
)

---

KL divergence between diagonal gaussian with prior diagonal gaussian.

Args

• mu (torch.Tensor) : mean of the posterior
• logvar (torch.Tensor) : log variance of the posterior
• p_mu (torch.Tensor) : mean of the prior
• p_logvar (torch.Tensor) : log variance of the prior

Returns

KL divergence (torch.Tensor)