Zum Inhalt springen

Referenz fĂŒr ultralytics/models/utils/ops.py

Hinweis

Diese Datei ist verfĂŒgbar unter https://github.com/ultralytics/ ultralytics/blob/main/ ultralytics/models/utils/ops .py. Wenn du ein Problem entdeckst, hilf bitte mit, es zu beheben, indem du einen Pull Request đŸ› ïž einreichst. Vielen Dank 🙏!



ultralytics.models.utils.ops.HungarianMatcher

Basen: Module

Ein Modul, das den HungarianMatcher implementiert, der ein differenzierbares Modul ist, um das Zuordnungsproblem in einer Ende-zu-Ende zu lösen.

HungarianMatcher fĂŒhrt eine optimale Zuweisung ĂŒber die vorhergesagten und die wahren Bounding Boxes durch, indem er eine Kostenfunktion Funktion, die die Klassifizierungsergebnisse, die Koordinaten der Boundingboxen und optional die Maskenvorhersagen berĂŒcksichtigt.

Attribute:

Name Typ Beschreibung
cost_gain dict

Wörterbuch der Kostenkoeffizienten: 'class', 'bbox', 'giou', 'mask', und 'dice'.

use_fl bool

Gibt an, ob der Focal Loss fĂŒr die Berechnung der Klassifizierungskosten verwendet werden soll.

with_mask bool

Gibt an, ob das Modell Maskenvorhersagen macht.

num_sample_points int

Die Anzahl der Stichprobenpunkte, die fĂŒr die Berechnung der Maskenkosten verwendet werden.

alpha float

Der Alpha-Faktor bei der Berechnung des Focal Loss.

gamma float

Der Gamma-Faktor in der Focal Loss Berechnung.

Methoden:

Name Beschreibung
forward

Berechnet die Zuordnung zwischen Vorhersagen und Basiswahrheiten fĂŒr einen Stapel.

_cost_mask

Berechnet die Maskenkosten und die WĂŒrfelkosten, wenn Masken vorhergesagt werden.

Quellcode in ultralytics/models/utils/ops.py
class HungarianMatcher(nn.Module):
    """
    A module implementing the HungarianMatcher, which is a differentiable module to solve the assignment problem in an
    end-to-end fashion.

    HungarianMatcher performs optimal assignment over the predicted and ground truth bounding boxes using a cost
    function that considers classification scores, bounding box coordinates, and optionally, mask predictions.

    Attributes:
        cost_gain (dict): Dictionary of cost coefficients: 'class', 'bbox', 'giou', 'mask', and 'dice'.
        use_fl (bool): Indicates whether to use Focal Loss for the classification cost calculation.
        with_mask (bool): Indicates whether the model makes mask predictions.
        num_sample_points (int): The number of sample points used in mask cost calculation.
        alpha (float): The alpha factor in Focal Loss calculation.
        gamma (float): The gamma factor in Focal Loss calculation.

    Methods:
        forward(pred_bboxes, pred_scores, gt_bboxes, gt_cls, gt_groups, masks=None, gt_mask=None): Computes the
            assignment between predictions and ground truths for a batch.
        _cost_mask(bs, num_gts, masks=None, gt_mask=None): Computes the mask cost and dice cost if masks are predicted.
    """

    def __init__(self, cost_gain=None, use_fl=True, with_mask=False, num_sample_points=12544, alpha=0.25, gamma=2.0):
        """Initializes HungarianMatcher with cost coefficients, Focal Loss, mask prediction, sample points, and alpha
        gamma factors.
        """
        super().__init__()
        if cost_gain is None:
            cost_gain = {"class": 1, "bbox": 5, "giou": 2, "mask": 1, "dice": 1}
        self.cost_gain = cost_gain
        self.use_fl = use_fl
        self.with_mask = with_mask
        self.num_sample_points = num_sample_points
        self.alpha = alpha
        self.gamma = gamma

    def forward(self, pred_bboxes, pred_scores, gt_bboxes, gt_cls, gt_groups, masks=None, gt_mask=None):
        """
        Forward pass for HungarianMatcher. This function computes costs based on prediction and ground truth
        (classification cost, L1 cost between boxes and GIoU cost between boxes) and finds the optimal matching between
        predictions and ground truth based on these costs.

        Args:
            pred_bboxes (Tensor): Predicted bounding boxes with shape [batch_size, num_queries, 4].
            pred_scores (Tensor): Predicted scores with shape [batch_size, num_queries, num_classes].
            gt_cls (torch.Tensor): Ground truth classes with shape [num_gts, ].
            gt_bboxes (torch.Tensor): Ground truth bounding boxes with shape [num_gts, 4].
            gt_groups (List[int]): List of length equal to batch size, containing the number of ground truths for
                each image.
            masks (Tensor, optional): Predicted masks with shape [batch_size, num_queries, height, width].
                Defaults to None.
            gt_mask (List[Tensor], optional): List of ground truth masks, each with shape [num_masks, Height, Width].
                Defaults to None.

        Returns:
            (List[Tuple[Tensor, Tensor]]): A list of size batch_size, each element is a tuple (index_i, index_j), where:
                - index_i is the tensor of indices of the selected predictions (in order)
                - index_j is the tensor of indices of the corresponding selected ground truth targets (in order)
                For each batch element, it holds:
                    len(index_i) = len(index_j) = min(num_queries, num_target_boxes)
        """

        bs, nq, nc = pred_scores.shape

        if sum(gt_groups) == 0:
            return [(torch.tensor([], dtype=torch.long), torch.tensor([], dtype=torch.long)) for _ in range(bs)]

        # We flatten to compute the cost matrices in a batch
        # [batch_size * num_queries, num_classes]
        pred_scores = pred_scores.detach().view(-1, nc)
        pred_scores = F.sigmoid(pred_scores) if self.use_fl else F.softmax(pred_scores, dim=-1)
        # [batch_size * num_queries, 4]
        pred_bboxes = pred_bboxes.detach().view(-1, 4)

        # Compute the classification cost
        pred_scores = pred_scores[:, gt_cls]
        if self.use_fl:
            neg_cost_class = (1 - self.alpha) * (pred_scores**self.gamma) * (-(1 - pred_scores + 1e-8).log())
            pos_cost_class = self.alpha * ((1 - pred_scores) ** self.gamma) * (-(pred_scores + 1e-8).log())
            cost_class = pos_cost_class - neg_cost_class
        else:
            cost_class = -pred_scores

        # Compute the L1 cost between boxes
        cost_bbox = (pred_bboxes.unsqueeze(1) - gt_bboxes.unsqueeze(0)).abs().sum(-1)  # (bs*num_queries, num_gt)

        # Compute the GIoU cost between boxes, (bs*num_queries, num_gt)
        cost_giou = 1.0 - bbox_iou(pred_bboxes.unsqueeze(1), gt_bboxes.unsqueeze(0), xywh=True, GIoU=True).squeeze(-1)

        # Final cost matrix
        C = (
            self.cost_gain["class"] * cost_class
            + self.cost_gain["bbox"] * cost_bbox
            + self.cost_gain["giou"] * cost_giou
        )
        # Compute the mask cost and dice cost
        if self.with_mask:
            C += self._cost_mask(bs, gt_groups, masks, gt_mask)

        # Set invalid values (NaNs and infinities) to 0 (fixes ValueError: matrix contains invalid numeric entries)
        C[C.isnan() | C.isinf()] = 0.0

        C = C.view(bs, nq, -1).cpu()
        indices = [linear_sum_assignment(c[i]) for i, c in enumerate(C.split(gt_groups, -1))]
        gt_groups = torch.as_tensor([0, *gt_groups[:-1]]).cumsum_(0)  # (idx for queries, idx for gt)
        return [
            (torch.tensor(i, dtype=torch.long), torch.tensor(j, dtype=torch.long) + gt_groups[k])
            for k, (i, j) in enumerate(indices)
        ]

__init__(cost_gain=None, use_fl=True, with_mask=False, num_sample_points=12544, alpha=0.25, gamma=2.0)

Initialisiert HungarianMatcher mit Kostenkoeffizienten, Focal Loss, Maskenvorhersage, Stichprobenpunkten und Alpha Gamma-Faktoren.

Quellcode in ultralytics/models/utils/ops.py
def __init__(self, cost_gain=None, use_fl=True, with_mask=False, num_sample_points=12544, alpha=0.25, gamma=2.0):
    """Initializes HungarianMatcher with cost coefficients, Focal Loss, mask prediction, sample points, and alpha
    gamma factors.
    """
    super().__init__()
    if cost_gain is None:
        cost_gain = {"class": 1, "bbox": 5, "giou": 2, "mask": 1, "dice": 1}
    self.cost_gain = cost_gain
    self.use_fl = use_fl
    self.with_mask = with_mask
    self.num_sample_points = num_sample_points
    self.alpha = alpha
    self.gamma = gamma

forward(pred_bboxes, pred_scores, gt_bboxes, gt_cls, gt_groups, masks=None, gt_mask=None)

VorwĂ€rtspass fĂŒr HungarianMatcher. Diese Funktion berechnet die Kosten auf der Grundlage der Vorhersage und der Ground Truth (Klassifizierungskosten, L1-Kosten zwischen KĂ€stchen und GIoU-Kosten zwischen KĂ€stchen) und findet auf Basis dieser Kosten die optimale Übereinstimmung zwischen Vorhersagen und der Basiswahrheit auf der Grundlage dieser Kosten.

Parameter:

Name Typ Beschreibung Standard
pred_bboxes Tensor

Vorausgesagte Bounding Boxen mit Form [batch_size, num_queries, 4].

erforderlich
pred_scores Tensor

Vorausgesagte Punktzahlen mit Form [batch_size, num_queries, num_classes].

erforderlich
gt_cls Tensor

Grundwahrheitsklassen mit Form [num_gts, ].

erforderlich
gt_bboxes Tensor

Ground truth bounding boxes mit der Form [num_gts, 4].

erforderlich
gt_groups List[int]

Liste mit der LĂ€nge der StapelgrĂ¶ĂŸe, die die Anzahl der Grundwahrheiten fĂŒr jedes Bildes enthĂ€lt.

erforderlich
masks Tensor

Vorhergesagte Masken mit der Form [batch_size, num_queries, height, width]. Der Standardwert ist Keine.

None
gt_mask List[Tensor]

Liste der Grundwahrheitsmasken, jede mit der Form [num_masks, Height, Width]. Der Standardwert ist Keine.

None

Retouren:

Typ Beschreibung
List[Tuple[Tensor, Tensor]]

Eine Liste der GrĂ¶ĂŸe batch_size, jedes Element ist ein Tupel (index_i, index_j), wobei: - index_i ist die tensor der Indizes der ausgewĂ€hlten Vorhersagen (in Reihenfolge) - index_j ist die tensor der Indizes der entsprechenden ausgewĂ€hlten Bodenwahrheitsziele (in der Reihenfolge) FĂŒr jedes Batch-Element gilt Folgendes: len(index_i) = len(index_j) = min(num_queries, num_target_boxes)

Quellcode in ultralytics/models/utils/ops.py
def forward(self, pred_bboxes, pred_scores, gt_bboxes, gt_cls, gt_groups, masks=None, gt_mask=None):
    """
    Forward pass for HungarianMatcher. This function computes costs based on prediction and ground truth
    (classification cost, L1 cost between boxes and GIoU cost between boxes) and finds the optimal matching between
    predictions and ground truth based on these costs.

    Args:
        pred_bboxes (Tensor): Predicted bounding boxes with shape [batch_size, num_queries, 4].
        pred_scores (Tensor): Predicted scores with shape [batch_size, num_queries, num_classes].
        gt_cls (torch.Tensor): Ground truth classes with shape [num_gts, ].
        gt_bboxes (torch.Tensor): Ground truth bounding boxes with shape [num_gts, 4].
        gt_groups (List[int]): List of length equal to batch size, containing the number of ground truths for
            each image.
        masks (Tensor, optional): Predicted masks with shape [batch_size, num_queries, height, width].
            Defaults to None.
        gt_mask (List[Tensor], optional): List of ground truth masks, each with shape [num_masks, Height, Width].
            Defaults to None.

    Returns:
        (List[Tuple[Tensor, Tensor]]): A list of size batch_size, each element is a tuple (index_i, index_j), where:
            - index_i is the tensor of indices of the selected predictions (in order)
            - index_j is the tensor of indices of the corresponding selected ground truth targets (in order)
            For each batch element, it holds:
                len(index_i) = len(index_j) = min(num_queries, num_target_boxes)
    """

    bs, nq, nc = pred_scores.shape

    if sum(gt_groups) == 0:
        return [(torch.tensor([], dtype=torch.long), torch.tensor([], dtype=torch.long)) for _ in range(bs)]

    # We flatten to compute the cost matrices in a batch
    # [batch_size * num_queries, num_classes]
    pred_scores = pred_scores.detach().view(-1, nc)
    pred_scores = F.sigmoid(pred_scores) if self.use_fl else F.softmax(pred_scores, dim=-1)
    # [batch_size * num_queries, 4]
    pred_bboxes = pred_bboxes.detach().view(-1, 4)

    # Compute the classification cost
    pred_scores = pred_scores[:, gt_cls]
    if self.use_fl:
        neg_cost_class = (1 - self.alpha) * (pred_scores**self.gamma) * (-(1 - pred_scores + 1e-8).log())
        pos_cost_class = self.alpha * ((1 - pred_scores) ** self.gamma) * (-(pred_scores + 1e-8).log())
        cost_class = pos_cost_class - neg_cost_class
    else:
        cost_class = -pred_scores

    # Compute the L1 cost between boxes
    cost_bbox = (pred_bboxes.unsqueeze(1) - gt_bboxes.unsqueeze(0)).abs().sum(-1)  # (bs*num_queries, num_gt)

    # Compute the GIoU cost between boxes, (bs*num_queries, num_gt)
    cost_giou = 1.0 - bbox_iou(pred_bboxes.unsqueeze(1), gt_bboxes.unsqueeze(0), xywh=True, GIoU=True).squeeze(-1)

    # Final cost matrix
    C = (
        self.cost_gain["class"] * cost_class
        + self.cost_gain["bbox"] * cost_bbox
        + self.cost_gain["giou"] * cost_giou
    )
    # Compute the mask cost and dice cost
    if self.with_mask:
        C += self._cost_mask(bs, gt_groups, masks, gt_mask)

    # Set invalid values (NaNs and infinities) to 0 (fixes ValueError: matrix contains invalid numeric entries)
    C[C.isnan() | C.isinf()] = 0.0

    C = C.view(bs, nq, -1).cpu()
    indices = [linear_sum_assignment(c[i]) for i, c in enumerate(C.split(gt_groups, -1))]
    gt_groups = torch.as_tensor([0, *gt_groups[:-1]]).cumsum_(0)  # (idx for queries, idx for gt)
    return [
        (torch.tensor(i, dtype=torch.long), torch.tensor(j, dtype=torch.long) + gt_groups[k])
        for k, (i, j) in enumerate(indices)
    ]



ultralytics.models.utils.ops.get_cdn_group(batch, num_classes, num_queries, class_embed, num_dn=100, cls_noise_ratio=0.5, box_noise_scale=1.0, training=False)

Erhalte eine kontrastive Entrauschungs-Trainingsgruppe. Diese Funktion erstellt eine kontrastive Entrauschungs-Trainingsgruppe mit positiven und negativen Stichproben aus den Ground Truths (gt). Sie wendet Rauschen auf die Klassenbezeichnungen und die Koordinaten der Boundingboxen an, und gibt die geĂ€nderten Bezeichnungen, Bounding Boxes, Aufmerksamkeitsmasken und Metainformationen zurĂŒck.

Parameter:

Name Typ Beschreibung Standard
batch dict

Ein Diktat, das 'gt_cls' (torch.Tensor mit Form [num_gts, ]), 'gt_bboxes' (torch.Tensor mit Form [num_gts, 4]), 'gt_groups' (List(int)), die eine Liste mit der LĂ€nge einer Partie ist die die Anzahl der gts eines jeden Bildes angibt.

erforderlich
num_classes int

Anzahl der Klassen.

erforderlich
num_queries int

Anzahl der Abfragen.

erforderlich
class_embed Tensor

Einbettungsgewichte, um Klassenetiketten auf den Einbettungsraum abzubilden.

erforderlich
num_dn int

Anzahl der Entrauschungen. Der Standardwert ist 100.

100
cls_noise_ratio float

RauschverhĂ€ltnis fĂŒr Klassenlabels. Der Standardwert ist 0,5.

0.5
box_noise_scale float

Rauschskala fĂŒr Bounding-Box-Koordinaten. Der Standardwert ist 1,0.

1.0
training bool

Wenn es sich im Trainingsmodus befindet. Der Standardwert ist False.

False

Retouren:

Typ Beschreibung
Tuple[Optional[Tensor], Optional[Tensor], Optional[Tensor], Optional[Dict]]

Die modifizierten Klasseneinbettungen, Bounding Boxes, Aufmerksamkeitsmaske und Metainformationen fĂŒr die Entrauschung. Wenn sich die Funktion nicht im Trainingsmodus befindet oder 'num_dn' kleiner oder gleich 0 ist, gibt die Funktion fĂŒr alle Elemente des Tupels None zurĂŒck.

Quellcode in ultralytics/models/utils/ops.py
def get_cdn_group(
    batch, num_classes, num_queries, class_embed, num_dn=100, cls_noise_ratio=0.5, box_noise_scale=1.0, training=False
):
    """
    Get contrastive denoising training group. This function creates a contrastive denoising training group with positive
    and negative samples from the ground truths (gt). It applies noise to the class labels and bounding box coordinates,
    and returns the modified labels, bounding boxes, attention mask and meta information.

    Args:
        batch (dict): A dict that includes 'gt_cls' (torch.Tensor with shape [num_gts, ]), 'gt_bboxes'
            (torch.Tensor with shape [num_gts, 4]), 'gt_groups' (List(int)) which is a list of batch size length
            indicating the number of gts of each image.
        num_classes (int): Number of classes.
        num_queries (int): Number of queries.
        class_embed (torch.Tensor): Embedding weights to map class labels to embedding space.
        num_dn (int, optional): Number of denoising. Defaults to 100.
        cls_noise_ratio (float, optional): Noise ratio for class labels. Defaults to 0.5.
        box_noise_scale (float, optional): Noise scale for bounding box coordinates. Defaults to 1.0.
        training (bool, optional): If it's in training mode. Defaults to False.

    Returns:
        (Tuple[Optional[Tensor], Optional[Tensor], Optional[Tensor], Optional[Dict]]): The modified class embeddings,
            bounding boxes, attention mask and meta information for denoising. If not in training mode or 'num_dn'
            is less than or equal to 0, the function returns None for all elements in the tuple.
    """

    if (not training) or num_dn <= 0:
        return None, None, None, None
    gt_groups = batch["gt_groups"]
    total_num = sum(gt_groups)
    max_nums = max(gt_groups)
    if max_nums == 0:
        return None, None, None, None

    num_group = num_dn // max_nums
    num_group = 1 if num_group == 0 else num_group
    # Pad gt to max_num of a batch
    bs = len(gt_groups)
    gt_cls = batch["cls"]  # (bs*num, )
    gt_bbox = batch["bboxes"]  # bs*num, 4
    b_idx = batch["batch_idx"]

    # Each group has positive and negative queries.
    dn_cls = gt_cls.repeat(2 * num_group)  # (2*num_group*bs*num, )
    dn_bbox = gt_bbox.repeat(2 * num_group, 1)  # 2*num_group*bs*num, 4
    dn_b_idx = b_idx.repeat(2 * num_group).view(-1)  # (2*num_group*bs*num, )

    # Positive and negative mask
    # (bs*num*num_group, ), the second total_num*num_group part as negative samples
    neg_idx = torch.arange(total_num * num_group, dtype=torch.long, device=gt_bbox.device) + num_group * total_num

    if cls_noise_ratio > 0:
        # Half of bbox prob
        mask = torch.rand(dn_cls.shape) < (cls_noise_ratio * 0.5)
        idx = torch.nonzero(mask).squeeze(-1)
        # Randomly put a new one here
        new_label = torch.randint_like(idx, 0, num_classes, dtype=dn_cls.dtype, device=dn_cls.device)
        dn_cls[idx] = new_label

    if box_noise_scale > 0:
        known_bbox = xywh2xyxy(dn_bbox)

        diff = (dn_bbox[..., 2:] * 0.5).repeat(1, 2) * box_noise_scale  # 2*num_group*bs*num, 4

        rand_sign = torch.randint_like(dn_bbox, 0, 2) * 2.0 - 1.0
        rand_part = torch.rand_like(dn_bbox)
        rand_part[neg_idx] += 1.0
        rand_part *= rand_sign
        known_bbox += rand_part * diff
        known_bbox.clip_(min=0.0, max=1.0)
        dn_bbox = xyxy2xywh(known_bbox)
        dn_bbox = torch.logit(dn_bbox, eps=1e-6)  # inverse sigmoid

    num_dn = int(max_nums * 2 * num_group)  # total denoising queries
    # class_embed = torch.cat([class_embed, torch.zeros([1, class_embed.shape[-1]], device=class_embed.device)])
    dn_cls_embed = class_embed[dn_cls]  # bs*num * 2 * num_group, 256
    padding_cls = torch.zeros(bs, num_dn, dn_cls_embed.shape[-1], device=gt_cls.device)
    padding_bbox = torch.zeros(bs, num_dn, 4, device=gt_bbox.device)

    map_indices = torch.cat([torch.tensor(range(num), dtype=torch.long) for num in gt_groups])
    pos_idx = torch.stack([map_indices + max_nums * i for i in range(num_group)], dim=0)

    map_indices = torch.cat([map_indices + max_nums * i for i in range(2 * num_group)])
    padding_cls[(dn_b_idx, map_indices)] = dn_cls_embed
    padding_bbox[(dn_b_idx, map_indices)] = dn_bbox

    tgt_size = num_dn + num_queries
    attn_mask = torch.zeros([tgt_size, tgt_size], dtype=torch.bool)
    # Match query cannot see the reconstruct
    attn_mask[num_dn:, :num_dn] = True
    # Reconstruct cannot see each other
    for i in range(num_group):
        if i == 0:
            attn_mask[max_nums * 2 * i : max_nums * 2 * (i + 1), max_nums * 2 * (i + 1) : num_dn] = True
        if i == num_group - 1:
            attn_mask[max_nums * 2 * i : max_nums * 2 * (i + 1), : max_nums * i * 2] = True
        else:
            attn_mask[max_nums * 2 * i : max_nums * 2 * (i + 1), max_nums * 2 * (i + 1) : num_dn] = True
            attn_mask[max_nums * 2 * i : max_nums * 2 * (i + 1), : max_nums * 2 * i] = True
    dn_meta = {
        "dn_pos_idx": [p.reshape(-1) for p in pos_idx.cpu().split(list(gt_groups), dim=1)],
        "dn_num_group": num_group,
        "dn_num_split": [num_dn, num_queries],
    }

    return (
        padding_cls.to(class_embed.device),
        padding_bbox.to(class_embed.device),
        attn_mask.to(class_embed.device),
        dn_meta,
    )





Erstellt 2023-11-12, aktualisiert 2024-06-02
Autoren: glenn-jocher (5), Burhan-Q (1), Laughing-q (1)