spacecutter
is a library for implementing ordinal regression models in PyTorch. The library consists of models and loss functions. It is recommended to use skorch to wrap the models to make them compatible with scikit-learn.
pip install spacecutter
Define any PyTorch model you want that generates a single, scalar prediction value. This will be our predictor
model. This model can then be wrapped with spacecutter.models.OrdinalLogisticModel
which will convert the output of the predictor
from a single number to an array of ordinal class probabilities. The following example shows how to do this for a two layer neural network predictor
for a problem with three ordinal classes.
import numpy as np
import torch
from torch import nn
from spacecutter.models import OrdinalLogisticModel
X = np.array([[0.5, 0.1, -0.1],
[1.0, 0.2, 0.6],
[-2.0, 0.4, 0.8]],
dtype=np.float32)
y = np.array([0, 1, 2]).reshape(-1, 1)
num_features = X.shape[1]
num_classes = len(np.unique(y))
predictor = nn.Sequential(
nn.Linear(num_features, num_features),
nn.ReLU(),
nn.Linear(num_features, 1)
)
model = OrdinalLogisticModel(predictor, num_classes)
y_pred = model(torch.as_tensor(X))
print(y_pred)
# tensor([[0.2325, 0.2191, 0.5485],
# [0.2324, 0.2191, 0.5485],
# [0.2607, 0.2287, 0.5106]], grad_fn=<CatBackward>)
It is recommended to use skorch to train spacecutter
models. The following shows how to train the model from the previous section using cumulative link loss with skorch
:
from skorch import NeuralNet
from spacecutter.callbacks import AscensionCallback
from spacecutter.losses import CumulativeLinkLoss
skorch_model = NeuralNet(
module=OrdinalLogisticModel,
module__predictor=predictor,
module__num_classes=num_classes,
criterion=CumulativeLinkLoss,
train_split=None,
callbacks=[
('ascension', AscensionCallback()),
],
)
skorch_model.fit(X, y)
Note that we must add the AscensionCallback
. This ensures that the ordinal cutpoints stay in ascending order. While ideally this constraint would be factored directly into the model optimization, spacecutter
currently hacks an SGD-compatible solution by utilizing a post-backwards-pass callback to clip the cutpoint values.