simet.metrics.roc_auc.mlp

simet.metrics.roc_auc.mlp

RocMLP

RocMLP(in_dim, h1=256, h2=128, p=0.2)

Bases: Module

Two-hidden-layer MLP that outputs a single logit for binary tasks.

Architecture

in_dim → Linear(h1) → ReLU → Dropout(p) → Linear(h2) → ReLU → Dropout(p) → Linear(1) → (logit)

Intended for binary classification / scoring (e.g., ROC-AUC evaluation) using logits (apply torch.nn.BCEWithLogitsLoss or torch.sigmoid externally).

Parameters:

Name	Type	Description	Default
in_dim	int	Input feature dimension D.	required
h1	int	Hidden size of the first layer. Defaults to 256.	256
h2	int	Hidden size of the second layer. Defaults to 128.	128
p	float	Dropout probability applied after each hidden ReLU. Defaults to 0.2.	0.2

Attributes:

Name	Type	Description
net	Sequential	The MLP stack producing a single logit.

Input/Output: - Input: x of shape (N, D) with D == in_dim. - Output: Logits of shape (N,) (after squeeze(-1)).

Example

model = RocMLP(in_dim=2048).eval() x = torch.randn(8, 2048) logits = model(x) logits.shape torch.Size([8])

Source code in simet/metrics/roc_auc/mlp.py

def __init__(self, in_dim: int, h1: int = 256, h2: int = 128, p: float = 0.2):
    super().__init__()
    self.net = nn.Sequential(
        nn.Linear(in_dim, h1),
        nn.ReLU(inplace=True),
        nn.Dropout(p),
        nn.Linear(h1, h2),
        nn.ReLU(inplace=True),
        nn.Dropout(p),
        nn.Linear(h2, 1),  # logits
    )

forward

forward(x)

Compute logits for a batch of feature vectors.

Parameters:

Name	Type	Description	Default
x	Tensor	Input tensor of shape (N, in_dim).	required

Returns:

Type	Description
Tensor	torch.Tensor: Logits of shape (N,). Apply sigmoid for probabilities
Tensor	or use with BCEWithLogitsLoss for numerically stable training.

Source code in simet/metrics/roc_auc/mlp.py

def forward(self, x: torch.Tensor) -> torch.Tensor:
    """Compute logits for a batch of feature vectors.

    Args:
        x (torch.Tensor): Input tensor of shape `(N, in_dim)`.

    Returns:
        torch.Tensor: Logits of shape `(N,)`. Apply `sigmoid` for probabilities
        or use with `BCEWithLogitsLoss` for numerically stable training.
    """
    return self.net(x).squeeze(-1)  # logits