handy-nn 0.0.5

Delightful and useful neural networks models, including OrdinalRegressionLoss, etc.

handy-nn

Delightful and useful neural networks models, including OrdinalRegressionLoss, etc.

Install

$ pip install handy-nn

Usage

from handy_nn import OrdinalRegressionLoss

# Initialize the loss function
num_classes = 5
criterion = OrdinalRegressionLoss(num_classes)

# For training
logits = model(inputs)  # Shape: (batch_size, 1)
loss = criterion(logits, targets)
loss.backward()  # shape: torch.Size([])

# To get class probabilities
probas = criterion.predict_probas(logits)  # Shape: (batch_size, num_classes)

Shapes

Variable	Shape
logits	(batch_size, 1)
targets	(batch_size,)
loss	torch.Size([])
probas	(batch_size, num_classes)

APIs

OrdinalRegressionLoss(num_classes, learn_thresholds=True, init_scale=2.0)

• num_classes int: Number of ordinal classes (ranks)
• learn_thresholds bool=True: Whether to learn threshold parameters or use fixed ones, defaults to True.
• init_scale float=2.0: Scale for initializing thresholds, defaults to 2.0

Creates the loss function for ordinal regression.

The goal of ordinal regression is to model the relationship between one or more independent variables and an ordinal dependent variable. It predicts the probability that an observation falls into a specific ordinal category or a category higher than a certain threshold. This is particularly useful in fields like social sciences, medicine, and customer surveys where outcomes are often ordinal.

TrendAwareLoss()

criterion = TrendAwareLoss()
loss = criterion(logits, targets)
loss.backward()

• logits torch.Tensor of shape (batch_size, num_classes), the logits of your model output
• targets torch.Tensor of either (batch_size,) (label classifications) or (batch_size, num_classes) (one-hot tensors)

TrendAwareLoss penalizes "too-early / too-late" misclassification inside a label segment more heavily by multiplying per-sample cross-entropy with the segment remaining-length weight.

This loss function is useful for those situations where misclassification leads to an indirect loss, such as financial trading, etc.

License

MIT