Exemplo n.º 1
0
def main():
    parser = argparse.ArgumentParser()

    # Settings
    parser.add_argument('-d',
                        '--dataset',
                        choices=dataset_attributes.keys(),
                        required=True)
    parser.add_argument('-s',
                        '--shift_type',
                        choices=shift_types,
                        required=True)
    # Confounders
    parser.add_argument('-t', '--target_name')
    parser.add_argument('-c', '--confounder_names', nargs='+')
    # Resume?
    parser.add_argument('--resume', default=False, action='store_true')
    # Label shifts
    parser.add_argument('--minority_fraction', type=float)
    parser.add_argument('--imbalance_ratio', type=float)
    # Data
    parser.add_argument('--fraction', type=float, default=1.0)
    parser.add_argument('--root_dir', default=None)
    parser.add_argument('--subsample_to_minority',
                        action='store_true',
                        default=False)
    parser.add_argument('--reweight_groups',
                        action='store_true',
                        default=False)
    parser.add_argument('--augment_data', action='store_true', default=False)
    parser.add_argument('--val_fraction', type=float, default=0.1)
    # Objective
    parser.add_argument('--robust', default=False, action='store_true')
    parser.add_argument('--alpha', type=float, default=0.2)
    parser.add_argument('--generalization_adjustment', default="0.0")
    parser.add_argument('--automatic_adjustment',
                        default=False,
                        action='store_true')
    parser.add_argument('--robust_step_size', default=0.01, type=float)
    parser.add_argument('--use_normalized_loss',
                        default=False,
                        action='store_true')
    parser.add_argument('--btl', default=False, action='store_true')
    parser.add_argument('--hinge', default=False, action='store_true')

    # Model
    parser.add_argument('--model',
                        choices=model_attributes.keys(),
                        default='resnet50')
    parser.add_argument('--train_from_scratch',
                        action='store_true',
                        default=False)
    parser.add_argument('--resnet_width', type=int, default=None)

    # Optimization
    parser.add_argument('--n_epochs', type=int, default=4)
    parser.add_argument('--batch_size', type=int, default=32)
    parser.add_argument('--lr', type=float, default=0.001)
    parser.add_argument('--scheduler', action='store_true', default=False)
    parser.add_argument('--weight_decay', type=float, default=5e-5)
    parser.add_argument('--gamma', type=float, default=0.1)
    parser.add_argument('--minimum_variational_weight', type=float, default=0)
    # Misc
    parser.add_argument('--seed', type=int, default=0)
    parser.add_argument('--show_progress', default=False, action='store_true')
    parser.add_argument('--log_dir', default='./logs')
    parser.add_argument('--log_every', default=50, type=int)
    parser.add_argument('--save_step', type=int, default=10)
    parser.add_argument('--save_best', action='store_true', default=False)
    parser.add_argument('--save_last', action='store_true', default=False)
    parser.add_argument('--model_test', type=str)
    parser.add_argument('--gpu', type=str)

    args = parser.parse_args()
    check_args(args)
    model_test = args.model_test
    gpu = args.gpu
    os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
    os.environ["CUDA_VISIBLE_DEVICES"] = gpu

    # BERT-specific configs copied over from run_glue.py
    if args.model == 'bert':
        args.max_grad_norm = 1.0
        args.adam_epsilon = 1e-8
        args.warmup_steps = 0

    if os.path.exists(args.log_dir) and args.resume:
        resume = True
        mode = 'a'
    else:
        resume = False
        mode = 'w'

    ## Initialize logs
    if not os.path.exists(args.log_dir):
        os.makedirs(args.log_dir)

    logger = Logger(os.path.join(args.log_dir, model_test + '_log.txt'), mode)
    # Record args
    log_args(args, logger)

    set_seed(args.seed)

    # Data
    # Test data for label_shift_step is not implemented yet
    test_data = None
    test_loader = None
    if args.shift_type == 'confounder':
        train_data, val_data, test_data = prepare_data(args, train=True)
    elif args.shift_type == 'label_shift_step':
        train_data, val_data = prepare_data(args, train=True)

    loader_kwargs = {
        'batch_size': args.batch_size,
        'num_workers': 12,
        'pin_memory': True
    }
    train_loader = train_data.get_loader(train=True,
                                         reweight_groups=args.reweight_groups,
                                         **loader_kwargs)
    val_loader = val_data.get_loader(train=False,
                                     reweight_groups=None,
                                     **loader_kwargs)
    if test_data is not None:
        test_loader = test_data.get_loader(train=False,
                                           reweight_groups=None,
                                           **loader_kwargs)

    data = {}
    data['train_loader'] = train_loader
    data['val_loader'] = val_loader
    data['test_loader'] = test_loader
    data['train_data'] = train_data
    data['val_data'] = val_data
    data['test_data'] = test_data
    n_classes = train_data.n_classes

    log_data(data, logger)

    ## Initialize model
    pretrained = not args.train_from_scratch
    if resume:
        model = torch.load(os.path.join(args.log_dir, model_test))
        d = train_data.input_size()[0]
    elif model_attributes[args.model]['feature_type'] in ('precomputed',
                                                          'raw_flattened'):
        assert pretrained
        # Load precomputed features
        d = train_data.input_size()[0]
        model = nn.Linear(d, n_classes)
        model.has_aux_logits = False
    elif args.model == 'resnet50':
        model = torchvision.models.resnet50(pretrained=pretrained)
        d = model.fc.in_features
        model.fc = nn.Linear(d, n_classes)
    elif args.model == 'resnet34':
        model = torchvision.models.resnet34(pretrained=pretrained)
        d = model.fc.in_features
        model.fc = nn.Linear(d, n_classes)
    elif args.model == 'wideresnet50':
        model = torchvision.models.wide_resnet50_2(pretrained=pretrained)
        d = model.fc.in_features
        model.fc = nn.Linear(d, n_classes)
    elif args.model == 'resnet50vw':
        assert not pretrained
        assert args.resnet_width is not None
        model = resnet50vw(args.resnet_width, num_classes=n_classes)
    elif args.model == 'resnet18vw':
        assert not pretrained
        assert args.resnet_width is not None
        model = resnet18vw(args.resnet_width, num_classes=n_classes)
    elif args.model == 'resnet10vw':
        assert not pretrained
        assert args.resnet_width is not None
        model = resnet10vw(args.resnet_width, num_classes=n_classes)
    elif args.model == 'bert':
        assert args.dataset == 'MultiNLI'

        from pytorch_transformers import BertConfig, BertForSequenceClassification
        config_class = BertConfig
        model_class = BertForSequenceClassification

        config = config_class.from_pretrained('bert-base-uncased',
                                              num_labels=3,
                                              finetuning_task='mnli')
        model = model_class.from_pretrained('bert-base-uncased',
                                            from_tf=False,
                                            config=config)
    else:
        raise ValueError('Model not recognized.')

    logger.flush()

    ## Define the objective
    if args.hinge:
        assert args.dataset in ['CelebA', 'CUB']  # Only supports binary

        def hinge_loss(yhat, y):
            # The torch loss takes in three arguments so we need to split yhat
            # It also expects classes in {+1.0, -1.0} whereas by default we give them in {0, 1}
            # Furthermore, if y = 1 it expects the first input to be higher instead of the second,
            # so we need to swap yhat[:, 0] and yhat[:, 1]...
            torch_loss = torch.nn.MarginRankingLoss(margin=1.0,
                                                    reduction='none')
            y = (y.float() * 2.0) - 1.0
            return torch_loss(yhat[:, 1], yhat[:, 0], y)

        criterion = hinge_loss
    else:
        criterion = torch.nn.CrossEntropyLoss(reduction='none')

    if False:
        df = pd.read_csv(os.path.join(args.log_dir, 'test.csv'))
        epoch_offset = df.loc[len(df) - 1, 'epoch'] + 1
        logger.write(f'starting from epoch {epoch_offset}')
    else:
        epoch_offset = 0
    train_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'train.csv'),
                                      train_data.n_groups,
                                      mode=mode)
    val_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'val.csv'),
                                    train_data.n_groups,
                                    mode=mode)
    test_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'test.csv'),
                                     train_data.n_groups,
                                     mode=mode)

    train(model,
          criterion,
          data,
          logger,
          train_csv_logger,
          val_csv_logger,
          test_csv_logger,
          args,
          epoch_offset=epoch_offset)

    train_csv_logger.close()
    val_csv_logger.close()
    test_csv_logger.close()
Exemplo n.º 2
0
def main(args):
    if args.wandb:
        wandb.init(project=f"{args.project_name}_{args.dataset}")
        wandb.config.update(args)

    # BERT-specific configs copied over from run_glue.py
    if (args.model.startswith("bert") and args.use_bert_params): 
        args.max_grad_norm = 1.0
        args.adam_epsilon = 1e-8
        args.warmup_steps = 0

    if os.path.exists(args.log_dir) and args.resume:
        resume = True
        mode = "a"
    else:
        resume = False
        mode = "w"

    ## Initialize logs
    if not os.path.exists(args.log_dir):
        os.makedirs(args.log_dir)

    logger = Logger(os.path.join(args.log_dir, "log.txt"), mode)
    # Record args
    log_args(args, logger)

    set_seed(args.seed)

    # Data
    # Test data for label_shift_step is not implemented yet
    test_data = None
    test_loader = None
    if args.shift_type == "confounder":
        train_data, val_data, test_data = prepare_data(
            args,
            train=True,
        )

    elif args.shift_type == "label_shift_step":
        raise NotImplementedError
        train_data, val_data = prepare_data(args, train=True)

    #########################################################################
    ###################### Prepare data for our method ######################
    #########################################################################

    # Should probably not be upweighting if folds are specified.
    assert not args.fold or not args.up_weight

    # Fold passed. Use it as train and valid.
    if args.fold:
        train_data, val_data = folds.get_fold(
            train_data,
            args.fold,
            cross_validation_ratio=(1 / args.num_folds_per_sweep),
            num_valid_per_point=args.num_sweeps,
            seed=args.seed,
        )

    if args.up_weight != 0:
        assert args.aug_col is not None
        # Get points that should be upsampled
        metadata_df = pd.read_csv(args.metadata_path)
        if args.dataset == "jigsaw":
            train_col = metadata_df[metadata_df["split"] == "train"]
        else:
            train_col = metadata_df[metadata_df["split"] == 0]
        aug_indices = np.where(train_col[args.aug_col] == 1)[0]
        print("len", len(train_col), len(aug_indices))
        if args.up_weight == -1:
            up_weight_factor = int(
                (len(train_col) - len(aug_indices)) / len(aug_indices)) - 1
        else:
            up_weight_factor = args.up_weight

        print(f"Up-weight factor: {up_weight_factor}")
        upsampled_points = Subset(train_data,
                                  list(aug_indices) * up_weight_factor)
        # Convert to DRODataset
        train_data = dro_dataset.DRODataset(
            ConcatDataset([train_data, upsampled_points]),
            process_item_fn=None,
            n_groups=train_data.n_groups,
            n_classes=train_data.n_classes,
            group_str_fn=train_data.group_str,
        )
    elif args.aug_col is not None:
        print("\n"*2 + "WARNING: aug_col is not being used." + "\n"*2)

    #########################################################################
    #########################################################################
    #########################################################################

    loader_kwargs = {
        "batch_size": args.batch_size,
        "num_workers": 4,
        "pin_memory": True,
    }
    train_loader = dro_dataset.get_loader(train_data,
                                          train=True,
                                          reweight_groups=args.reweight_groups,
                                          **loader_kwargs)

    val_loader = dro_dataset.get_loader(val_data,
                                        train=False,
                                        reweight_groups=None,
                                        **loader_kwargs)

    if test_data is not None:
        test_loader = dro_dataset.get_loader(test_data,
                                             train=False,
                                             reweight_groups=None,
                                             **loader_kwargs)

    data = {}
    data["train_loader"] = train_loader
    data["val_loader"] = val_loader
    data["test_loader"] = test_loader
    data["train_data"] = train_data
    data["val_data"] = val_data
    data["test_data"] = test_data

    n_classes = train_data.n_classes

    log_data(data, logger)

    ## Initialize model
    model = get_model(
        model=args.model,
        pretrained=not args.train_from_scratch,
        resume=resume,
        n_classes=train_data.n_classes,
        dataset=args.dataset,
        log_dir=args.log_dir,
    )
    if args.wandb:
        wandb.watch(model)

    logger.flush()

    ## Define the objective
    if args.hinge:
        assert args.dataset in ["CelebA", "CUB"]  # Only supports binary
        criterion = hinge_loss
    else:
        criterion = torch.nn.CrossEntropyLoss(reduction="none")

    if resume:
        raise NotImplementedError  # Check this implementation.
        df = pd.read_csv(os.path.join(args.log_dir, "test.csv"))
        epoch_offset = df.loc[len(df) - 1, "epoch"] + 1
        logger.write(f"starting from epoch {epoch_offset}")
    else:
        epoch_offset = 0

    
    train_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, f"train.csv"),
                                      train_data.n_groups,
                                      mode=mode)
    val_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, f"val.csv"),
                                    val_data.n_groups,
                                    mode=mode)
    test_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, f"test.csv"),
                                     test_data.n_groups,
                                     mode=mode)
    train(
        model,
        criterion,
        data,
        logger,
        train_csv_logger,
        val_csv_logger,
        test_csv_logger,
        args,
        epoch_offset=epoch_offset,
        csv_name=args.fold,
        wandb=wandb if args.wandb else None,
    )

    train_csv_logger.close()
    val_csv_logger.close()
    test_csv_logger.close()
Exemplo n.º 3
0
def main():
    parser = argparse.ArgumentParser()

    # Settings
    parser.add_argument('-d',
                        '--dataset',
                        choices=dataset_attributes.keys(),
                        required=True)
    parser.add_argument('-s',
                        '--shift_type',
                        choices=shift_types,
                        required=True)
    # Confounders
    parser.add_argument('-t', '--target_name')
    parser.add_argument('-c', '--confounder_names', nargs='+')
    # Resume?
    parser.add_argument('--resume', default=False, action='store_true')
    # Label shifts
    parser.add_argument('--minority_fraction', type=float)
    parser.add_argument('--imbalance_ratio', type=float)
    # Data
    parser.add_argument('--fraction', type=float, default=1.0)
    parser.add_argument('--root_dir', default=None)
    parser.add_argument('--subsample_to_minority',
                        action='store_true',
                        default=False)
    parser.add_argument('--reweight_groups',
                        action='store_true',
                        default=False)
    parser.add_argument('--augment_data', action='store_true', default=False)
    parser.add_argument('--val_fraction', type=float, default=0.1)
    # Objective
    parser.add_argument('--robust', default=False, action='store_true')
    parser.add_argument('--alpha', type=float, default=0.2)
    parser.add_argument('--generalization_adjustment', default="0.0")
    parser.add_argument('--automatic_adjustment',
                        default=False,
                        action='store_true')
    parser.add_argument('--robust_step_size', default=0.01, type=float)
    parser.add_argument('--use_normalized_loss',
                        default=False,
                        action='store_true')
    parser.add_argument('--btl', default=False, action='store_true')
    parser.add_argument('--hinge', default=False, action='store_true')

    # Model
    parser.add_argument('--model',
                        choices=model_attributes.keys(),
                        default='resnet50')
    parser.add_argument('--train_from_scratch',
                        action='store_true',
                        default=False)
    parser.add_argument('--resnet_width', type=int, default=None)

    # Optimization
    parser.add_argument('--n_epochs', type=int, default=4)
    parser.add_argument('--batch_size', type=int, default=32)
    parser.add_argument('--lr', type=float, default=0.001)
    parser.add_argument('--scheduler', action='store_true', default=False)
    parser.add_argument('--weight_decay', type=float, default=5e-5)
    parser.add_argument('--gamma', type=float, default=0.1)
    parser.add_argument('--minimum_variational_weight', type=float, default=0)
    # Misc
    parser.add_argument('--seed', type=int, default=0)
    parser.add_argument('--show_progress', default=False, action='store_true')
    parser.add_argument('--log_dir', default='./logs')
    parser.add_argument('--log_every', default=50, type=int)
    parser.add_argument('--save_step', type=int, default=10)
    parser.add_argument('--save_best', action='store_true', default=False)
    parser.add_argument('--save_last', action='store_true', default=True)
    parser.add_argument('--student_width', type=int)
    parser.add_argument('--teacher_dir', type=str)
    parser.add_argument('--teacher_width', type=int)
    parser.add_argument('--gpu', type=str)
    parser.add_argument('--temp', type=str)

    args = parser.parse_args()
    gpu = args.gpu
    temp = args.temp
    check_args(args)
    teacher_dir = args.teacher_dir
    student_width = args.student_width
    teacher_width = args.teacher_width
    os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
    os.environ["CUDA_VISIBLE_DEVICES"] = gpu

    def DistillationLoss(temperature):
        cross_entropy = torch.nn.CrossEntropyLoss()

        def loss(student_logits, teacher_logits, target):
            last_dim = len(student_logits.shape) - 1
            p_t = nn.functional.softmax(teacher_logits / temperature,
                                        dim=last_dim)
            log_p_s = nn.functional.log_softmax(student_logits / temperature,
                                                dim=last_dim)
            return cross_entropy(student_logits, target) - (p_t * log_p_s).sum(
                dim=last_dim).mean() * temperature**2

        return loss

    # BERT-specific configs copied over from run_glue.py
    if args.model == 'bert':
        args.max_grad_norm = 1.0
        args.adam_epsilon = 1e-8
        args.warmup_steps = 0

    if os.path.exists(args.log_dir) and args.resume:
        resume = True
        mode = 'a'
    else:
        resume = False
        mode = 'w'

    ## Initialize logs
    if not os.path.exists(args.log_dir):
        os.makedirs(args.log_dir)

    logger = Logger(os.path.join(args.log_dir, 'log.txt'), mode)
    # Record args
    log_args(args, logger)

    set_seed(args.seed)
    print("starting prep")
    # Data
    # Test data for label_shift_step is not implemented yet
    test_data = None
    test_loader = None
    if args.shift_type == 'confounder':
        train_data, val_data, test_data = prepare_data(args, train=True)
    elif args.shift_type == 'label_shift_step':
        train_data, val_data = prepare_data(args, train=True)
    print("done prep")
    loader_kwargs = {
        'batch_size': args.batch_size,
        'num_workers': 16,
        'pin_memory': True
    }
    train_loader = train_data.get_loader(train=True,
                                         reweight_groups=args.reweight_groups,
                                         **loader_kwargs)
    val_loader = val_data.get_loader(train=False,
                                     reweight_groups=None,
                                     **loader_kwargs)
    if test_data is not None:
        test_loader = test_data.get_loader(train=False,
                                           reweight_groups=None,
                                           **loader_kwargs)

    data = {}
    data['train_loader'] = train_loader
    data['val_loader'] = val_loader
    data['test_loader'] = test_loader
    data['train_data'] = train_data
    data['val_data'] = val_data
    data['test_data'] = test_data
    n_classes = train_data.n_classes

    log_data(data, logger)
    logger.flush()

    ## Define the objective
    if args.hinge:
        assert args.dataset in ['CelebA', 'CUB']  # Only supports binary

        def hinge_loss(yhat, y):
            # The torch loss takes in three arguments so we need to split yhat
            # It also expects classes in {+1.0, -1.0} whereas by default we give them in {0, 1}
            # Furthermore, if y = 1 it expects the first input to be higher instead of the second,
            # so we need to swap yhat[:, 0] and yhat[:, 1]...
            torch_loss = torch.nn.MarginRankingLoss(margin=1.0,
                                                    reduction='none')
            y = (y.float() * 2.0) - 1.0
            return torch_loss(yhat[:, 1], yhat[:, 0], y)

        criterion = hinge_loss
    else:
        criterion = torch.nn.CrossEntropyLoss(reduction='none')

    if resume:
        df = pd.read_csv(os.path.join(args.log_dir, 'test.csv'))
        epoch_offset = df.loc[len(df) - 1, 'epoch'] + 1
        logger.write(f'starting from epoch {epoch_offset}')
    else:
        epoch_offset = 0

    train_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'train.csv'),
                                      train_data.n_groups,
                                      mode=mode)
    val_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'val.csv'),
                                    train_data.n_groups,
                                    mode=mode)
    test_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'test.csv'),
                                     train_data.n_groups,
                                     mode=mode)
    strain_csv_logger = CSVBatchLogger(os.path.join(args.log_dir,
                                                    'strain.csv'),
                                       train_data.n_groups,
                                       mode=mode)
    sval_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'sval.csv'),
                                     train_data.n_groups,
                                     mode=mode)
    stest_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'stest.csv'),
                                      train_data.n_groups,
                                      mode=mode)

    teacher = resnet10vw(teacher_width, num_classes=n_classes)
    teacher_old = torch.load(teacher_dir + "/10_model.pth")
    for k, m in teacher_old.named_modules():
        m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatability
    teacher.load_state_dict(teacher_old.state_dict())
    teacher = teacher.to('cuda')
    #    def DistillationLoss(temperature):
    #        cross_entropy = torch.nn.CrossEntropyLoss()
    #
    #        def loss(student_logits, teacher_logits, target):
    #            last_dim = len(student_logits.shape) - 1
    #
    #            p_t = nn.functional.softmax(teacher_logits/temperature, dim=last_dim)
    #            log_p_s = nn.functional.log_softmax(student_logits/temperature, dim=last_dim)
    #
    #            return cross_entropy(student_logits, target) - (p_t * log_p_s).sum(dim=last_dim).mean()
    #
    #        return loss

    distill_criterion = DistillationLoss(float(temp))
    student = resnet10vw(int(student_width), num_classes=n_classes).to('cuda')

    #student.to(device)
    train(teacher,
          student,
          criterion,
          distill_criterion,
          data,
          logger,
          train_csv_logger,
          val_csv_logger,
          test_csv_logger,
          strain_csv_logger,
          sval_csv_logger,
          test_csv_logger,
          args,
          epoch_offset=epoch_offset)
    train_csv_logger.close()
    val_csv_logger.close()
    test_csv_logger.close()
    strain_csv_logger.close()
    sval_csv_logger.close()
    stest_csv_logger.close()
Exemplo n.º 4
0
def main():
    print("Loading and checking args...")
    args = parse_args()
    check_args(args)
    # BERT-specific configs copied over from run_glue.py
    if args.model.startswith('bert'):
        args.max_grad_norm = 1.0
        args.adam_epsilon = 1e-8
        args.warmup_steps = 0

    #Write for logging; assumes no existing logs.
    mode = 'w'

    ## Initialize logs
    if not os.path.exists(args.log_dir):
        os.makedirs(args.log_dir)

    logger = Logger(os.path.join(args.log_dir, 'log.txt'), mode)
    # Record args
    log_args(args, logger)
    set_seed(args.seed)

    # Data
    print("Preparing data")
    train_data, val_data, test_data = prepare_data(args, train=True)

    print("Setting up loader")
    loader_kwargs = {
        'batch_size': args.batch_size,
        'num_workers': 4,
        'pin_memory': True
    }
    train_loader = train_data.get_loader(train=True,
                                         reweight_groups=args.reweight_groups,
                                         **loader_kwargs)
    val_loader = val_data.get_loader(train=False,
                                     reweight_groups=None,
                                     **loader_kwargs)
    test_loader = test_data.get_loader(train=False,
                                       reweight_groups=None,
                                       **loader_kwargs)

    data = {}
    data['train_loader'] = train_loader
    data['val_loader'] = val_loader
    data['test_loader'] = test_loader
    data['train_data'] = train_data
    data['val_data'] = val_data
    data['test_data'] = test_data
    n_classes = train_data.n_classes

    log_data(data, logger)

    ## Initialize model
    if args.model == 'resnet50':
        model = torchvision.models.resnet50(pretrained=True)
        d = model.fc.in_features
        model.fc = nn.Linear(d, n_classes)
        if args.mc_dropout:
            model = add_dropout(model, 'fc')
    elif args.model == 'densenet121':
        model = torchvision.models.densenet121(pretrained=True)
        d = model.classifier.in_features
        model.classifier = nn.Linear(d, n_classes)
        if args.mc_dropout:
            model = add_dropout(model, 'classifier')
    elif args.model == 'bert-base-uncased':
        print("Loading bert")
        model = BertForSequenceClassification.from_pretrained(
            args.model, num_labels=n_classes)
    else:
        raise ValueError('Model not recognized.')

    logger.flush()
    criterion = torch.nn.CrossEntropyLoss(reduction='none')
    print("Getting loggers")
    train_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'train.csv'),
                                      train_data.n_groups,
                                      mode=mode)
    val_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'val.csv'),
                                    train_data.n_groups,
                                    mode=mode)
    test_csv_logger = CSVBatchLogger(os.path.join(args.log_dir, 'test.csv'),
                                     train_data.n_groups,
                                     mode=mode)

    print("Starting to train...")
    train(model,
          criterion,
          data,
          logger,
          train_csv_logger,
          val_csv_logger,
          test_csv_logger,
          args,
          epoch_offset=0,
          train=True)

    train_csv_logger.close()
    val_csv_logger.close()
    test_csv_logger.close()

    if args.save_preds:
        save_preds(model, data, args)
        return