likelihood="bernoulli")
#train mu network
run_auto_lr_range(train_loader, bae_model)
bae_model.fit(train_loader, num_epochs=10)
#for each model, evaluate and plot:
bae_models = [bae_model]
id_data_test = test_loader
ood_data_list = [ood_loader]
train_set_name = "CIFAR"
#run evaluation test of model on ood data set
run_test_model(bae_models=bae_models,
id_data_test=test_loader,
ood_data_list=ood_data_list,
id_data_name=train_set_name,
output_reshape_size=(32, 32, 3))
#experimental here
x_test = get_sample_dataloader(test_loader)[0].cuda()
x_ood = get_sample_dataloader(ood_loader)[0].cuda()
test_latent = bae_model.predict_latent(x_test, transform_pca=False)
ood_latent = bae_model.predict_latent(x_ood, transform_pca=False)
plt.figure()
plt.boxplot([test_latent[0].mean(1), ood_latent[0].mean(1)])
plt.figure()
plt.boxplot([test_latent[1].mean(1), ood_latent[1].mean(1)])
last_activation="sigmoid") #symmetrical to encoder
#combine them into autoencoder
autoencoder = Autoencoder(encoder, decoder_mu)
#convert into BAE-MCDropout
bae_mcdropout = BAE_MCDropout(autoencoder=autoencoder,
dropout_p=0.2,
num_train_samples=5,
num_samples=50,
use_cuda=use_cuda,
denoising_factor=noise_factor)
#train mu network
run_auto_lr_range(train_loader, bae_mcdropout)
bae_mcdropout.fit(train_loader, num_epochs=1)
#for each model, evaluate and plot:
bae_models = [bae_mcdropout]
id_data_test = test_loader
ood_data_list = [ood_loader, noisy_loader]
train_set_name = "FashionMNIST"
#run evaluation test of model on ood data set
run_test_model(bae_models=bae_models,
id_data_test=test_loader,
ood_data_names=["OOD", "NOISY"],
ood_data_list=ood_data_list,
id_data_name=train_set_name,
output_reshape_size=(28, 28))