# PyTorch¶

Computing a counterfactual of a PyTorch model is done by using the ceml.torch.counterfactual.generate_counterfactual() function.

We must provide the PyTorch model within a class that is derived from torch.nn.Module and ceml.model.model.ModelWithLoss. In this class, we must overwrite the predict function and the get_loss function which returns a loss that we want to use - a couple of differentiable loss functions are implemented in ceml.backend.torch.costfunctions.

Besides the model, we must specify the input whose prediction we want to explain and the desired target prediction (prediction of the counterfactual). In addition we can restrict the features that can be used for computing a counterfactual, specify a regularization of the counterfactual and specifying the optimization algorithm used for computing a counterfactual.

A complete example of a softmax regression model using the negative-log-likelihood is given below:

  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 #!/usr/bin/env python3 # -*- coding: utf-8 -*- import torch torch.manual_seed(424242) import numpy as np from sklearn.datasets import load_iris from sklearn.model_selection import train_test_split from sklearn.metrics import accuracy_score from ceml.torch import generate_counterfactual from ceml.backend.torch.costfunctions import NegLogLikelihoodCost from ceml.model import ModelWithLoss # Neural network - Softmax regression class Model(torch.nn.Module, ModelWithLoss): def __init__(self, input_size, num_classes): super(Model, self).__init__() self.linear = torch.nn.Linear(input_size, num_classes) self.softmax = torch.nn.Softmax(dim=0) def forward(self, x): return self.linear(x) # NOTE: Softmax is build into CrossEntropyLoss def predict_proba(self, x): return self.softmax(self.forward(x)) def predict(self, x, dim=1): return torch.argmax(self.forward(x), dim=dim) def get_loss(self, y_target, pred=None): return NegLogLikelihoodCost(input_to_output=self.predict_proba, y_target=y_target) if __name__ == "__main__": # Load data X, y = load_iris(return_X_y=True) X = X.astype(np.dtype(np.float32)) X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=1) # numpy -> torch tensor x = torch.from_numpy(X_train) labels = torch.from_numpy(y_train) x_test = torch.from_numpy(X_test) y_test = torch.from_numpy(y_test) # Create and fit model model = Model(4, 3) learning_rate = 0.001 momentum = 0.9 criterion = torch.nn.CrossEntropyLoss() optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate, momentum=momentum) num_epochs = 800 for epoch in range(num_epochs): optimizer.zero_grad() outputs = model(x) loss = criterion(outputs, labels) loss.backward() optimizer.step() # Evaluation y_pred = model.predict(x_test).detach().numpy() print("Accuracy: {0}".format(accuracy_score(y_test, y_pred))) # Select a data point whose prediction has to be explained x_orig = X_test[1,:] print("Prediction on x: {0}".format(model.predict(torch.from_numpy(np.array([x_orig]))))) # Whitelist of features we can use/change when computing the counterfactual features_whitelist = [0, 2] # Use the first and third feature only # Compute counterfactual print("\nCompute counterfactual ....") print(generate_counterfactual(model, x_orig, y_target=0, features_whitelist=features_whitelist, regularization="l1", C=0.1, optimizer="nelder-mead"))