Creates and pretrains both classifier and adversarial. The user is also able to provide its own architecture of both models through clf_model and adv_model parameters. Moreover the classifier can be also trained, but then one has to change trained to TRUE and provide the optimizer from this training. It is also possible to create both models from our interface witch is described in create_model documentation.
pretrain( clf_model = NULL, adv_model = NULL, clf_optimizer = NULL, trained = FALSE, train_x = NULL, train_y = NULL, sensitive_train, sensitive_test, batch_size = 50, partition = 0.7, neurons_clf = c(32, 32, 32), neurons_adv = c(32, 32, 32), dimension_clf = 2, dimension_adv = 1, learning_rate_clf = 0.001, learning_rate_adv = 0.001, n_ep_preclf = 5, n_ep_preadv = 10, dsl, dev, verbose = TRUE, monitor = TRUE, seed = 7 )
| clf_model | optional value, net, nn_module, provide the pretrain with your own classification neural network. Default: NULL |
|---|---|
| adv_model | optional value, net, nn_module, provide the pretrain with your own adversarial neural network. Default: NULL |
| clf_optimizer | optional value, provide the optimizer of classifier if you decided to provide your own pre trained classifier. Default: NULL |
| trained | 0 if the classificator is untrained, 1 if the classificator is already pretrained. Default: 0 |
| train_x | numeric, scaled matrix of predictors used for training |
| train_y | numeric, scaled vector of target used for training |
| sensitive_train | integer, vector of sensitive values used for training |
| sensitive_test | integer, vector of sensitive values used for testing |
| batch_size | integer indicating a batch size used in dataloader. Default: 50 |
| partition | float from [0,1] range setting the size of train vector (test size equals 1-partition). Default = 0.7. |
| neurons_clf | integer vector describing a neural architecture of classifier network. Default: c(32,32,32). This notation means that the network has 3 layers with 32 neurons each. |
| neurons_adv | integer vector describing a neural architecture of adversarial network. Default: c(32,32,32). This notation means that the network has 3 layers with 32 neurons each. |
| dimension_clf | integer from [1,2] setting nnf_softmax dimension for classifier. Default: 2 (suggested to use 2 for classifier and 1 for adversarial) |
| dimension_adv | integer from [1,2] setting nnf_softmax dimension for adversarial. Default: 1 (suggested to use 2 for classifier and 1 for adversarial) |
| learning_rate_clf | float from [0,1] setting learning rate for classifier. Default: 0.001 |
| learning_rate_adv | float from [0,1] setting learning rate for classifier. Default: 0.001 |
| n_ep_preclf | integer setting number of epochs for preclassifiers training. Default: 5 |
| n_ep_preadv | integer setting number of epochs for preadversarials training. Default : 10 |
| dsl | dataset_loader object from pretrain |
| dev | device used to calculations (cpu or gpu) |
| verbose | logical indicating if we want to print monitored outputs or not |
| monitor | logical indicating if we want to monitor the learning process or not (monitoring tends to slow down the training process, but provides some useful info to adjust parameters and training process) |
| seed | integer, seed for initial weights, set NULL for none. Default: 7. |
list of two objects : clf_model and adv_model which are pretrained neural networks (net, nn_module).
if (FALSE) { adult <- fairmodels::adult processed <- preprocess( adult, "salary", "sex", "Male", "Female", c("race"), sample = 0.05, train_size = 0.65, test_size = 0.35, validation_size = 0, seed = 7 ) dev <- "cpu" dsl <- dataset_loader(processed$train_x, processed$train_y, processed$test_x, processed$test_y, batch_size = 5, dev = dev) # Both models created with our package models <- pretrain( train_x = processed$train_x, train_y = processed$train_y, sensitive_train = processed$sensitive_train, sensitive_test = processed$sensitive_test, batch_size = 5, partition = 0.65, neurons_clf = c(32, 32, 32), neurons_adv = c(32, 32, 32), dimension_clf = 2, dimension_adv = 1, learning_rate_clf = 0.001, learning_rate_adv = 0.001, n_ep_preclf = 1, n_ep_preadv = 1, dsl = dsl, dev = dev, verbose = FALSE, monitor = FALSE ) # presaved models and states of the optimizers clf <- torch_load(system.file("extdata","clf2",package="fairpan")) clf_optimizer_state <- torch_load(system.file("extdata","clf_optimizer2",package="fairpan")) clf_optimizer <- optim_adam(clf$parameters, lr = 0.001) acc2 <- eval_accuracy(clf, dsl$test_ds, dev) clf_optimizer$load_state_dict(clf_optimizer_state) # Clf provided and pretrained models2 <- pretrain( clf_model = clf, clf_optimizer = clf_optimizer, trained = TRUE, train_x = processed$train_x, train_y = processed$train_y, sensitive_train = processed$sensitive_train, sensitive_test = processed$sensitive_test, batch_size = 5, partition = 0.65, neurons_clf = c(32, 32, 32), neurons_adv = c(32, 32, 32), dimension_clf = 2, dimension_adv = 1, learning_rate_clf = 0.001, learning_rate_adv = 0.001, n_ep_preclf = 1, n_ep_preadv = 1, dsl = dsl, dev = dev, verbose = FALSE, monitor = FALSE ) clf2 <- create_model(processed$train_x, processed$train_y, c(4, 4), 2) # Clf provided, but not pretrained models3 <- pretrain( clf_model = clf2, trained = FALSE, train_x = processed$train_x, train_y = processed$train_y, sensitive_train = processed$sensitive_train, sensitive_test = processed$sensitive_test, batch_size = 5, partition = 0.65, neurons_clf = c(32, 32, 32), neurons_adv = c(32, 32, 32), dimension_clf = 2, dimension_adv = 1, learning_rate_clf = 0.001, learning_rate_adv = 0.001, n_ep_preclf = 1, n_ep_preadv = 1, dsl = dsl, dev = dev, verbose = FALSE, monitor = FALSE ) }