bugroup/Eye_Tracking_Drift_Correction

1import torch as t2 3 4def macro_soft_f1(real_vals, predictions, reduction):5    """from https://towardsdatascience.com/the-unknown-benefits-of-using-a-soft-f1-loss-in-classification-systems-753902c0105d"""6    true_positive = (real_vals * predictions).sum(dim=0)7    false_positive = (predictions * (1 - real_vals)).sum(dim=0)8    false_negative = ((1 - predictions) * real_vals).sum(dim=0)9    soft_f1 = 2 * true_positive / (2 * true_positive + false_negative + false_positive + 1e-16)10    if reduction == "mean":11        loss = t.mean(1 - soft_f1)12    else:13        loss = 1 - soft_f114    return loss15 16 17def coral_loss(logits, levels, importance_weights=None, reduction="mean"):18    """Computes the CORAL loss described in19    Cao, Mirjalili, and Raschka (2020)20    *Rank Consistent Ordinal Regression for Neural Networks21       with Application to Age Estimation*22    Pattern Recognition Letters, https://doi.org/10.1016/j.patrec.2020.11.00823    Parameters24    ----------25    logits : torch.tensor, shape(num_examples, num_classes-1)26        Outputs of the CORAL layer.27    levels : torch.tensor, shape(num_examples, num_classes-1)28        True labels represented as extended binary vectors29        (via `coral_pytorch.dataset.levels_from_labelbatch`).30    importance_weights : torch.tensor, shape=(num_classes-1,) (default=None)31        Optional weights for the different labels in levels.32        A tensor of ones, i.e.,33        `torch.ones(num_classes-1, dtype=torch.float32)`34        will result in uniform weights that have the same effect as None.35    reduction : str or None (default='mean')36        If 'mean' or 'sum', returns the averaged or summed loss value across37        all data points (rows) in logits. If None, returns a vector of38        shape (num_examples,)39    Returns40    ----------41        loss : torch.tensor42        A torch.tensor containing a single loss value (if `reduction='mean'` or '`sum'`)43        or a loss value for each data record (if `reduction=None`).44    Examples45    ----------46    >>> import torch47    >>> from coral_pytorch.losses import coral_loss48    >>> levels = torch.tensor(49    ...    [[1., 1., 0., 0.],50    ...     [1., 0., 0., 0.],51    ...    [1., 1., 1., 1.]])52    >>> logits = torch.tensor(53    ...    [[2.1, 1.8, -2.1, -1.8],54    ...     [1.9, -1., -1.5, -1.3],55    ...     [1.9, 1.8, 1.7, 1.6]])56    >>> coral_loss(logits, levels)57    tensor(0.6920)58    https://github.com/Raschka-research-group/coral-pytorch/blob/c6ab93afd555a6eac708c95ae1feafa15f91c5aa/coral_pytorch/losses.py59    """60 61    if not logits.shape == levels.shape:62        raise ValueError(63            "Please ensure that logits (%s) has the same shape as levels (%s). " % (logits.shape, levels.shape)64        )65 66    term1 = t.nn.functional.logsigmoid(logits) * levels + (t.nn.functional.logsigmoid(logits) - logits) * (1 - levels)67 68    if importance_weights is not None:69        term1 *= importance_weights70 71    val = -t.sum(term1, dim=1)72 73    if reduction == "mean":74        loss = t.mean(val)75    elif reduction == "sum":76        loss = t.sum(val)77    elif reduction is None:78        loss = val79    else:80        s = 'Invalid value for `reduction`. Should be "mean", ' '"sum", or None. Got %s' % reduction81        raise ValueError(s)82 83    return loss84 85 86def corn_loss(logits, y_train, num_classes):87    """Computes the CORN loss described in our forthcoming88    'Deep Neural Networks for Rank Consistent Ordinal89    Regression based on Conditional Probabilities'90    manuscript.91    Parameters92    ----------93    logits : torch.tensor, shape=(num_examples, num_classes-1)94        Outputs of the CORN layer.95    y_train : torch.tensor, shape=(num_examples)96        Torch tensor containing the class labels.97    num_classes : int98        Number of unique class labels (class labels should start at 0).99    Returns100    ----------101        loss : torch.tensor102        A torch.tensor containing a single loss value.103    Examples104    ----------105    >>> import torch106    >>> from coral_pytorch.losses import corn_loss107    >>> # Consider 8 training examples108    >>> _  = torch.manual_seed(123)109    >>> X_train = torch.rand(8, 99)110    >>> y_train = torch.tensor([0, 1, 2, 2, 2, 3, 4, 4])111    >>> NUM_CLASSES = 5112    >>> #113    >>> #114    >>> # def __init__(self):115    >>> corn_net = torch.nn.Linear(99, NUM_CLASSES-1)116    >>> #117    >>> #118    >>> # def forward(self, X_train):119    >>> logits = corn_net(X_train)120    >>> logits.shape121    torch.Size([8, 4])122    >>> corn_loss(logits, y_train, NUM_CLASSES)123    tensor(0.7127, grad_fn=<DivBackward0>)124    https://github.com/Raschka-research-group/coral-pytorch/blob/c6ab93afd555a6eac708c95ae1feafa15f91c5aa/coral_pytorch/losses.py125    """126    sets = []127    for i in range(num_classes - 1):128        label_mask = y_train > i - 1129        label_tensor = (y_train[label_mask] > i).to(t.int64)130        sets.append((label_mask, label_tensor))131 132    num_examples = 0133    losses = 0.0134    for task_index, s in enumerate(sets):135        train_examples = s[0]136        train_labels = s[1]137 138        if len(train_labels) < 1:139            continue140 141        num_examples += len(train_labels)142        pred = logits[train_examples, task_index]143 144        loss = -t.sum(145            t.nn.functional.logsigmoid(pred) * train_labels146            + (t.nn.functional.logsigmoid(pred) - pred) * (1 - train_labels)147        )148        losses += loss149 150    return losses / num_examples151 152 153def corn_label_from_logits(logits):154    """155    Returns the predicted rank label from logits for a156    network trained via the CORN loss.157    Parameters158    ----------159    logits : torch.tensor, shape=(n_examples, n_classes)160        Torch tensor consisting of logits returned by the161        neural net.162    Returns163    ----------164    labels : torch.tensor, shape=(n_examples)165        Integer tensor containing the predicted rank (class) labels166    Examples167    ----------168    >>> # 2 training examples, 5 classes169    >>> logits = torch.tensor([[14.152, -6.1942, 0.47710, 0.96850],170    ...                        [65.667, 0.303, 11.500, -4.524]])171    >>> corn_label_from_logits(logits)172    tensor([1, 3])173    https://github.com/Raschka-research-group/coral-pytorch/blob/c6ab93afd555a6eac708c95ae1feafa15f91c5aa/coral_pytorch/dataset.py174    """175    probas = t.sigmoid(logits)176    probas = t.cumprod(probas, dim=1)177    predict_levels = probas > 0.5178    predicted_labels = t.sum(predict_levels, dim=1)179    return predicted_labels180