eval_openset.py

"""
Stand alone evaluation script for open set recognition and plotting of different datasets

Uses the same command line parser as main.py

The attributes that need to be specified are the number of variational samples (should be greater than one if prediction
uncertainties are supposed to be calculated and compared), the architecture type and the resume flag pointing to a model
checkpoint file.
Other parameters like open set distance function etc. are optional.

Minimum example usage:
--resume /path/checkpoint.pth.tar --var-samples 100 -a MLP
"""

import collections
from lib.cmdparser import parser
import lib.Datasets.datasets as datasets
import lib.Models.architectures as architectures
from lib.Models.pixelcnn import PixelCNN
from lib.Training.evaluate import eval_dataset as eval_dataset
from lib.Training.evaluate import eval_openset_dataset as eval_openset_dataset
from lib.Utility.visualization import *
from lib.OpenSet.meta_recognition import *


def main():
    # set device
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    # Command line options
    args = parser.parse_args()
    print("Command line options:")
    for arg in vars(args):
        print(arg, getattr(args, arg))

    # Get the dataset which has been trained and the corresponding number of classes
    data_init_method = getattr(datasets, args.dataset)
    dataset = data_init_method(torch.cuda.is_available(), args)
    num_classes = dataset.num_classes
    net_input, _ = next(iter(dataset.train_loader))
    num_colors = net_input.size(1)

    # Split a part of the non-used dataset to use as validation set for determining open set (e.g entropy)
    # rejection thresholds
    split_perc = 0.5
    split_sets = torch.utils.data.random_split(dataset.valset,
                                               [int((1 - split_perc) * len(dataset.valset)),
                                                int(split_perc * len(dataset.valset))])

    # overwrite old set and create new split set to determine thresholds/priors
    dataset.valset = split_sets[0]
    dataset.threshset = split_sets[1]

    # overwrite old data loader and create new loader for thresh set
    is_gpu = torch.cuda.is_available()
    dataset.val_loader = torch.utils.data.DataLoader(dataset.valset, batch_size=args.batch_size, shuffle=False,
                                                     num_workers=args.workers, pin_memory=is_gpu, sampler=None)
    dataset.threshset_loader = torch.utils.data.DataLoader(dataset.threshset, batch_size=args.batch_size, shuffle=False,
                                                           num_workers=args.workers, pin_memory=is_gpu, sampler=None)

    # Load open set datasets
    openset_datasets_names = args.openset_datasets.strip().split(',')
    openset_datasets = []
    for openset_dataset in openset_datasets_names:
        openset_data_init_method = getattr(datasets, openset_dataset)
        openset_datasets.append(openset_data_init_method(torch.cuda.is_available(), args))

    if not args.autoregression:
        args.out_channels = num_colors

    # Initialize empty model
    net_init_method = getattr(architectures, args.architecture)
    model = net_init_method(device, num_classes, num_colors, args)

    # Optional addition of autoregressive decoder portion
    if args.autoregression:
        model.pixelcnn = PixelCNN(device, num_colors, args.out_channels, args.pixel_cnn_channels,
                                  num_layers=args.pixel_cnn_layers, k=args.pixel_cnn_kernel_size,
                                  padding=args.pixel_cnn_kernel_size // 2)

    model = torch.nn.DataParallel(model).to(device)

    # load model (using the resume functionality)
    assert(os.path.isfile(args.resume)), "=> no model checkpoint found at '{}'".format(args.resume)

    # Fill the random model with the parameters of the checkpoint
    print("=> loading checkpoint '{}'".format(args.resume))
    checkpoint = torch.load(args.resume)
    best_prec = checkpoint['best_prec']
    best_loss = checkpoint['best_loss']
    # print the saved model's validation accuracy (as a check to see if the loaded model has really been trained)
    print("Saved model's validation accuracy: ", best_prec)
    print("Saved model's validation loss: ", best_loss)
    model.load_state_dict(checkpoint['state_dict'])
    model.to(device)
    model.eval()

    # set the save path to the directory from which the model has been loaded
    save_path = os.path.dirname(args.resume)

    # start of the model evaluation on the training dataset and fitting
    print("Evaluating original train dataset: " + args.dataset + ". This may take a while...")
    dataset_eval_dict_train = eval_dataset(model, dataset.train_loader, dataset.num_classes, device,
                                           samples=args.var_samples, calc_reconstruction=args.calc_reconstruction,
                                           autoregression=args.autoregression)
    print("Training accuracy: ", dataset_eval_dict_train["accuracy"])

    # Get the mean of z for correctly classified data inputs
    mean_zs = get_means(dataset_eval_dict_train["zs_correct"])

    # visualize the mean z vectors
    mean_zs_tensor = torch.stack(mean_zs, dim=0)
    visualize_means(mean_zs_tensor, dataset.class_to_idx, args.dataset, save_path, "z")

    # calculate each correctly classified example's distance to the mean z
    distances_to_z_means_correct_train = calc_distances_to_means(mean_zs, dataset_eval_dict_train["zs_correct"],
                                                                 args.distance_function)

    # Weibull fitting
    # set tailsize according to command line parameters (according to percentage of dataset size)
    tailsize = int(len(dataset.trainset) * args.openset_weibull_tailsize / num_classes)
    print("Fitting Weibull models with tailsize: " + str(tailsize))
    tailsizes = [tailsize] * num_classes
    weibull_models, valid_weibull = fit_weibull_models(distances_to_z_means_correct_train, tailsizes)
    assert valid_weibull, "Weibull fit is not valid"

    # Determine rejection thresholds/priors on the created split set
    print("Evaluating original threshold split dataset: " + args.dataset + ". This may take a while...")
    threshset_eval_dict = eval_dataset(model, dataset.threshset_loader, num_classes, device, samples=args.var_samples,
                                       calc_reconstruction=args.calc_reconstruction, autoregression=args.autoregression)

    # Again calculate distances to mean z
    print("Split set accuracy: ", threshset_eval_dict["accuracy"])
    distances_to_z_means_threshset = calc_distances_to_means(mean_zs, threshset_eval_dict["zs_correct"],
                                                             args.distance_function)

    outlier_probs_threshset = calc_outlier_probs(weibull_models, distances_to_z_means_threshset)

    threshset_classification = calc_openset_classification(outlier_probs_threshset, num_classes,
                                                           num_outlier_threshs=100)
    max_entropy = np.max(threshset_eval_dict["out_entropy"])
    threshset_entropy_classification = calc_entropy_classification(threshset_eval_dict["out_entropy"],
                                                                   max_entropy,
                                                                   num_outlier_threshs=100)

    # We have added a flag to turn off calculation of the decoder because it is computationally heavy for many samples
    # (repeated calculation of the decoder), whereas latent space sampling and repeated calculation of our latent based
    # EVT approach and even the single layer classifier is cheap.
    if args.calc_reconstruction:
        max_recon_loss = np.max(threshset_eval_dict["recon_loss_mus"])
        threshset_recon_classification = calc_reconstruction_classification(threshset_eval_dict["recon_loss_mus"],
                                                                            max_recon_loss,
                                                                            num_outlier_threshs=1000)

    # determine the index for the corresponding rejection priors/thresholds. Although this should never happen,
    # we also set a default if no threshold satisfies the 95% inlier condition.
    if (np.array(threshset_classification["outlier_percentage"]) <= args.percent_validation_outliers).any() == True:
        EVT_prior_index = np.argwhere(np.array(threshset_classification["outlier_percentage"])
                                      <= 0.05)[0][0]
        EVT_prior = threshset_classification["thresholds"][EVT_prior_index]
    else:
        EVT_prior = 0.5
        EVT_prior_index = 50

    if (np.array(threshset_entropy_classification["entropy_outlier_percentage"]) <=
        args.percent_validation_outliers).any() == True:
        entropy_threshold_index = np.argwhere(np.array(threshset_entropy_classification["entropy_outlier_percentage"])
                                              <= 0.05)[0][0]
        entropy_threshold = threshset_entropy_classification["entropy_thresholds"][entropy_threshold_index]
    else:
        entropy_threshold = np.median(threshset_entropy_classification["entropy_thresholds"])
        entropy_threshold_index = 50

    if args.calc_reconstruction:
        if (np.array(threshset_recon_classification["reconstruction_outlier_percentage"]) <=
            args.percent_validation_outliers).any() == True:
            recon_threshold_index = np.argwhere(
                np.array(threshset_recon_classification["reconstruction_outlier_percentage"]) <= 0.05)[0][0]
            recon_threshold = threshset_recon_classification["reconstruction_thresholds"][recon_threshold_index]
        else:
            recon_threshold = np.median(threshset_recon_classification["reconstruction_thresholds"])
            recon_threshold_index = 500

    print("EVT prior: " + str(EVT_prior) + "; Entropy threshold: " + str(entropy_threshold))
    if args.calc_reconstruction:
        print("Reconstruction loss threshold: " + str(recon_threshold))

    # ------------------------------------------------------------------------------------------
    # Fitting on train dataset complete. Beginning of all testing/open set recognition on validation and unknown sets.
    # ------------------------------------------------------------------------------------------
    # We evaluate the validation set to later evaluate trained dataset's statistical inlier/outlier estimates.
    print("Evaluating original validation dataset: " + args.dataset + ". This may take a while...")
    dataset_eval_dict = eval_dataset(model, dataset.val_loader, num_classes, device, samples=args.var_samples,
                                     calc_reconstruction=args.calc_reconstruction, autoregression=args.autoregression)

    # Again calculate distances to mean z
    print("Validation accuracy: ", dataset_eval_dict["accuracy"])
    distances_to_z_means_correct = calc_distances_to_means(mean_zs, dataset_eval_dict["zs_correct"],
                                                           args.distance_function)

    # Evaluate outlier probability of trained dataset's validation set
    outlier_probs_correct = calc_outlier_probs(weibull_models, distances_to_z_means_correct)

    dataset_classification_correct = calc_openset_classification(outlier_probs_correct, num_classes,
                                                                 num_outlier_threshs=100)
    dataset_entropy_classification_correct = calc_entropy_classification(dataset_eval_dict["out_entropy"],
                                                                         max_entropy,
                                                                         num_outlier_threshs=100)
    if args.calc_reconstruction:
        dataset_recon_classification_correct = calc_reconstruction_classification(dataset_eval_dict["recon_loss_mus"],
                                                                                  max_recon_loss,
                                                                                  num_outlier_threshs=1000)

    print(args.dataset + '(trained) EVT outlier percentage: ' +
          str(dataset_classification_correct["outlier_percentage"][EVT_prior_index]))
    print(args.dataset + '(trained) entropy outlier percentage: ' +
          str(dataset_entropy_classification_correct["entropy_outlier_percentage"][entropy_threshold_index]))
    if args.calc_reconstruction:
        print(args.dataset + '(trained) reconstruction loss outlier percentage: ' +
              str(dataset_recon_classification_correct["reconstruction_outlier_percentage"][recon_threshold_index]))

    # ------------------------------------------------------------------------------------------
    # Repeat process for open set recognition (no fitting, just testing) on all unseen datasets
    # ------------------------------------------------------------------------------------------
    # dicitionaries to hold results
    openset_dataset_eval_dicts = collections.OrderedDict()
    openset_outlier_probs_dict = collections.OrderedDict()
    openset_classification_dict = collections.OrderedDict()
    openset_entropy_classification_dict = collections.OrderedDict()
    if args.calc_reconstruction:
        openset_recon_classification_dict = collections.OrderedDict()

    for od, openset_dataset in enumerate(openset_datasets):
        print("Evaluating openset dataset: " + openset_datasets_names[od] + ". This may take a while...")
        openset_dataset_eval_dict = eval_openset_dataset(model, openset_dataset.val_loader, num_classes, device,
                                                         samples=args.var_samples, autoregression=args.autoregression,
                                                         calc_reconstruction=args.calc_reconstruction)

        openset_distances_to_z_means = calc_distances_to_means(mean_zs, openset_dataset_eval_dict["zs"],
                                                               args.distance_function)

        openset_outlier_probs = calc_outlier_probs(weibull_models, openset_distances_to_z_means)

        # getting outlier classification accuracies across the entire datasets
        openset_classification = calc_openset_classification(openset_outlier_probs, num_classes,
                                                             num_outlier_threshs=100)
        openset_entropy_classification = calc_entropy_classification(openset_dataset_eval_dict["out_entropy"],
                                                                     max_entropy,
                                                                     num_outlier_threshs=100)
        if args.calc_reconstruction:
            openset_recon_classification_correct = calc_reconstruction_classification(
                openset_dataset_eval_dict["recon_loss_mus"], max_recon_loss, num_outlier_threshs=1000)

        openset_dataset_eval_dicts[openset_datasets_names[od]] = openset_dataset_eval_dict
        openset_outlier_probs_dict[openset_datasets_names[od]] = openset_outlier_probs
        openset_classification_dict[openset_datasets_names[od]] = openset_classification
        openset_entropy_classification_dict[openset_datasets_names[od]] = openset_entropy_classification
        if args.calc_reconstruction:
            openset_recon_classification_dict[openset_datasets_names[od]] = openset_recon_classification_correct

    # Print the results
    for other_data_name, other_data_dict in openset_classification_dict.items():
        print(other_data_name + ' EVT outlier percentage: ' +
              str(other_data_dict["outlier_percentage"][entropy_threshold_index]))
    for other_data_name, other_data_dict in openset_entropy_classification_dict.items():
        print(other_data_name + ' entropy outlier percentage: ' +
              str(other_data_dict["entropy_outlier_percentage"][entropy_threshold_index]))
    if args.calc_reconstruction:
        for other_data_name, other_data_dict in openset_recon_classification_dict.items():
            print(other_data_name + ' reconstruction loss outlier percentage: ' +
                  str(other_data_dict["reconstruction_outlier_percentage"][recon_threshold_index]))

    # joint prediction uncertainty plot for all datasets
    if args.var_samples > 1:
        visualize_classification_uncertainty(dataset_eval_dict["out_mus_correct"],
                                             dataset_eval_dict["out_sigmas_correct"],
                                             openset_dataset_eval_dicts,
                                             "out_mus", "out_sigmas",
                                             args.dataset + ' (trained)',
                                             args.var_samples, save_path)

    # visualize the outlier probabilities
    visualize_weibull_outlier_probabilities(outlier_probs_correct, openset_outlier_probs_dict,
                                            args.dataset + ' (trained)', save_path, tailsize)

    # histograms
    visualize_classification_scores(dataset_eval_dict["out_mus_correct"], openset_dataset_eval_dicts, 'out_mus',
                                    args.dataset + ' (trained)', save_path)
    visualize_entropy_histogram(dataset_eval_dict["out_entropy"], openset_dataset_eval_dicts,
                                dataset_entropy_classification_correct["entropy_thresholds"][-1], "out_entropy",
                                args.dataset + ' (trained)', save_path)
    if args.calc_reconstruction:
        visualize_recon_loss_histogram(dataset_eval_dict["recon_loss_mus"], openset_dataset_eval_dicts,
                                       dataset_recon_classification_correct["reconstruction_thresholds"][-1],
                                       "recon_loss_mus", args.dataset + ' (trained)', save_path)

    # joint plot for outlier detection accuracy for both seen and unseen datasets
    visualize_openset_classification(dataset_classification_correct["outlier_percentage"],
                                     openset_classification_dict, "outlier_percentage",
                                     args.dataset + ' (trained)',
                                     dataset_classification_correct["thresholds"], save_path, tailsize)
    visualize_entropy_classification(dataset_entropy_classification_correct["entropy_outlier_percentage"],
                                     openset_entropy_classification_dict, "entropy_outlier_percentage",
                                     args.dataset + ' (trained)',
                                     dataset_entropy_classification_correct["entropy_thresholds"], save_path)
    if args.calc_reconstruction:
        visualize_reconstruction_classification(dataset_recon_classification_correct["reconstruction_outlier_percentage"],
                                                openset_recon_classification_dict, "reconstruction_outlier_percentage",
                                                args.dataset + ' (trained)',
                                                dataset_recon_classification_correct["reconstruction_thresholds"],
                                                save_path, autoregression=args.autoregression)


if __name__ == '__main__':
    main()