webapp/seed/Lib/site-packages/transformers/loss/loss_lw_detr.py

# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import torch
import torch.nn as nn

from ..utils import is_accelerate_available, is_scipy_available, is_vision_available
from .loss_for_object_detection import (
    HungarianMatcher,
    _set_aux_loss,
    box_iou,
    dice_loss,
    generalized_box_iou,
    nested_tensor_from_tensor_list,
    sigmoid_focal_loss,
)


if is_vision_available():
    from transformers.image_transforms import center_to_corners_format


if is_scipy_available():
    from scipy.optimize import linear_sum_assignment

if is_accelerate_available():
    from accelerate import PartialState
    from accelerate.utils import reduce


class LwDetrHungarianMatcher(HungarianMatcher):
    @torch.no_grad()
    def forward(self, outputs, targets, group_detr):
        """
        Differences:
        - out_prob = outputs["logits"].flatten(0, 1).sigmoid() instead of softmax
        - class_cost uses alpha and gamma
        """
        batch_size, num_queries = outputs["logits"].shape[:2]

        # We flatten to compute the cost matrices in a batch
        out_prob = outputs["logits"].flatten(0, 1).sigmoid()  # [batch_size * num_queries, num_classes]
        out_bbox = outputs["pred_boxes"].flatten(0, 1)  # [batch_size * num_queries, 4]

        # Also concat the target labels and boxes
        target_ids = torch.cat([v["class_labels"] for v in targets])
        target_bbox = torch.cat([v["boxes"] for v in targets])

        # Compute the classification cost.
        alpha = 0.25
        gamma = 2.0
        neg_cost_class = (1 - alpha) * (out_prob**gamma) * (-(1 - out_prob + 1e-8).log())
        pos_cost_class = alpha * ((1 - out_prob) ** gamma) * (-(out_prob + 1e-8).log())
        class_cost = pos_cost_class[:, target_ids] - neg_cost_class[:, target_ids]

        # Compute the L1 cost between boxes, cdist only supports float32
        dtype = out_bbox.dtype
        out_bbox = out_bbox.to(torch.float32)
        target_bbox = target_bbox.to(torch.float32)
        bbox_cost = torch.cdist(out_bbox, target_bbox, p=1)
        bbox_cost = bbox_cost.to(dtype)

        # Compute the giou cost between boxes
        giou_cost = -generalized_box_iou(center_to_corners_format(out_bbox), center_to_corners_format(target_bbox))

        # Final cost matrix
        cost_matrix = self.bbox_cost * bbox_cost + self.class_cost * class_cost + self.giou_cost * giou_cost
        cost_matrix = cost_matrix.view(batch_size, num_queries, -1).cpu()

        sizes = [len(v["boxes"]) for v in targets]
        indices = []
        group_num_queries = num_queries // group_detr
        cost_matrix_list = cost_matrix.split(group_num_queries, dim=1)
        for group_id in range(group_detr):
            group_cost_matrix = cost_matrix_list[group_id]
            group_indices = [linear_sum_assignment(c[i]) for i, c in enumerate(group_cost_matrix.split(sizes, -1))]
            if group_id == 0:
                indices = group_indices
            else:
                indices = [
                    (
                        np.concatenate([indice1[0], indice2[0] + group_num_queries * group_id]),
                        np.concatenate([indice1[1], indice2[1]]),
                    )
                    for indice1, indice2 in zip(indices, group_indices)
                ]
        return [(torch.as_tensor(i, dtype=torch.int64), torch.as_tensor(j, dtype=torch.int64)) for i, j in indices]


class LwDetrImageLoss(nn.Module):
    def __init__(self, matcher, num_classes, focal_alpha, losses, group_detr):
        super().__init__()
        self.matcher = matcher
        self.num_classes = num_classes
        self.focal_alpha = focal_alpha
        self.losses = losses
        self.group_detr = group_detr

    # removed logging parameter, which was part of the original implementation
    def loss_labels(self, outputs, targets, indices, num_boxes):
        if "logits" not in outputs:
            raise KeyError("No logits were found in the outputs")
        source_logits = outputs["logits"]

        idx = self._get_source_permutation_idx(indices)
        target_classes_o = torch.cat([t["class_labels"][J] for t, (_, J) in zip(targets, indices)])
        alpha = self.focal_alpha
        gamma = 2
        src_boxes = outputs["pred_boxes"][idx]
        target_boxes = torch.cat([t["boxes"][i] for t, (_, i) in zip(targets, indices)], dim=0)
        iou_targets = torch.diag(
            box_iou(center_to_corners_format(src_boxes.detach()), center_to_corners_format(target_boxes))[0]
        )
        # Convert to the same dtype as the source logits as box_iou upcasts to float32
        iou_targets = iou_targets.to(source_logits.dtype)
        pos_ious = iou_targets.clone().detach()
        prob = source_logits.sigmoid()
        # init positive weights and negative weights
        pos_weights = torch.zeros_like(source_logits)
        neg_weights = prob**gamma

        pos_ind = list(idx)
        pos_ind.append(target_classes_o)

        t = prob[pos_ind].pow(alpha) * pos_ious.pow(1 - alpha)
        t = torch.clamp(t, 0.01).detach()

        pos_weights[pos_ind] = t
        neg_weights[pos_ind] = 1 - t
        loss_ce = -pos_weights * prob.log() - neg_weights * (1 - prob).log()
        loss_ce = loss_ce.sum() / num_boxes
        losses = {"loss_ce": loss_ce}

        return losses

    # Copied from loss.loss_for_object_detection.ImageLoss.loss_cardinality
    @torch.no_grad()
    def loss_cardinality(self, outputs, targets, indices, num_boxes):
        """
        Compute the cardinality error, i.e. the absolute error in the number of predicted non-empty boxes.

        This is not really a loss, it is intended for logging purposes only. It doesn't propagate gradients.
        """
        logits = outputs["logits"]
        device = logits.device
        target_lengths = torch.as_tensor([len(v["class_labels"]) for v in targets], device=device)
        # Count the number of predictions that are NOT "no-object" (which is the last class)
        card_pred = (logits.argmax(-1) != logits.shape[-1] - 1).sum(1)
        card_err = nn.functional.l1_loss(card_pred.float(), target_lengths.float())
        losses = {"cardinality_error": card_err}
        return losses

    # Copied from loss.loss_for_object_detection.ImageLoss.loss_boxes
    def loss_boxes(self, outputs, targets, indices, num_boxes):
        """
        Compute the losses related to the bounding boxes, the L1 regression loss and the GIoU loss.

        Targets dicts must contain the key "boxes" containing a tensor of dim [nb_target_boxes, 4]. The target boxes
        are expected in format (center_x, center_y, w, h), normalized by the image size.
        """
        if "pred_boxes" not in outputs:
            raise KeyError("No predicted boxes found in outputs")
        idx = self._get_source_permutation_idx(indices)
        source_boxes = outputs["pred_boxes"][idx]
        target_boxes = torch.cat([t["boxes"][i] for t, (_, i) in zip(targets, indices)], dim=0)

        loss_bbox = nn.functional.l1_loss(source_boxes, target_boxes, reduction="none")

        losses = {}
        losses["loss_bbox"] = loss_bbox.sum() / num_boxes

        loss_giou = 1 - torch.diag(
            generalized_box_iou(center_to_corners_format(source_boxes), center_to_corners_format(target_boxes))
        )
        losses["loss_giou"] = loss_giou.sum() / num_boxes
        return losses

    # Copied from loss.loss_for_object_detection.ImageLoss.loss_masks
    def loss_masks(self, outputs, targets, indices, num_boxes):
        """
        Compute the losses related to the masks: the focal loss and the dice loss.

        Targets dicts must contain the key "masks" containing a tensor of dim [nb_target_boxes, h, w].
        """
        if "pred_masks" not in outputs:
            raise KeyError("No predicted masks found in outputs")

        source_idx = self._get_source_permutation_idx(indices)
        target_idx = self._get_target_permutation_idx(indices)
        source_masks = outputs["pred_masks"]
        source_masks = source_masks[source_idx]
        masks = [t["masks"] for t in targets]
        # TODO use valid to mask invalid areas due to padding in loss
        target_masks, valid = nested_tensor_from_tensor_list(masks).decompose()
        target_masks = target_masks.to(source_masks)
        target_masks = target_masks[target_idx]

        # upsample predictions to the target size
        source_masks = nn.functional.interpolate(
            source_masks[:, None], size=target_masks.shape[-2:], mode="bilinear", align_corners=False
        )
        source_masks = source_masks[:, 0].flatten(1)

        target_masks = target_masks.flatten(1)
        target_masks = target_masks.view(source_masks.shape)
        losses = {
            "loss_mask": sigmoid_focal_loss(source_masks, target_masks, num_boxes),
            "loss_dice": dice_loss(source_masks, target_masks, num_boxes),
        }
        return losses

    # Copied from loss.loss_for_object_detection.ImageLoss._get_source_permutation_idx
    def _get_source_permutation_idx(self, indices):
        # permute predictions following indices
        batch_idx = torch.cat([torch.full_like(source, i) for i, (source, _) in enumerate(indices)])
        source_idx = torch.cat([source for (source, _) in indices])
        return batch_idx, source_idx

    # Copied from loss.loss_for_object_detection.ImageLoss._get_target_permutation_idx
    def _get_target_permutation_idx(self, indices):
        # permute targets following indices
        batch_idx = torch.cat([torch.full_like(target, i) for i, (_, target) in enumerate(indices)])
        target_idx = torch.cat([target for (_, target) in indices])
        return batch_idx, target_idx

    def get_loss(self, loss, outputs, targets, indices, num_boxes):
        loss_map = {
            "labels": self.loss_labels,
            "cardinality": self.loss_cardinality,
            "boxes": self.loss_boxes,
            "masks": self.loss_masks,
        }
        if loss not in loss_map:
            raise ValueError(f"Loss {loss} not supported")
        return loss_map[loss](outputs, targets, indices, num_boxes)

    def forward(self, outputs, targets):
        """
        This performs the loss computation.

        Args:
             outputs (`dict`, *optional*):
                Dictionary of tensors, see the output specification of the model for the format.
             targets (`list[dict]`, *optional*):
                List of dicts, such that `len(targets) == batch_size`. The expected keys in each dict depends on the
                losses applied, see each loss' doc.
        """
        group_detr = self.group_detr if self.training else 1
        outputs_without_aux_and_enc = {
            k: v for k, v in outputs.items() if k != "enc_outputs" and k != "auxiliary_outputs"
        }

        # Retrieve the matching between the outputs of the last layer and the targets
        indices = self.matcher(outputs_without_aux_and_enc, targets, group_detr)

        # Compute the average number of target boxes across all nodes, for normalization purposes
        num_boxes = sum(len(t["class_labels"]) for t in targets)
        num_boxes = num_boxes * group_detr
        num_boxes = torch.as_tensor([num_boxes], dtype=torch.float, device=next(iter(outputs.values())).device)
        world_size = 1
        if is_accelerate_available():
            if PartialState._shared_state != {}:
                num_boxes = reduce(num_boxes)
                world_size = PartialState().num_processes
        num_boxes = torch.clamp(num_boxes / world_size, min=1).item()

        # Compute all the requested losses
        losses = {}
        for loss in self.losses:
            losses.update(self.get_loss(loss, outputs, targets, indices, num_boxes))

        # In case of auxiliary losses, we repeat this process with the output of each intermediate layer.
        if "auxiliary_outputs" in outputs:
            for i, auxiliary_outputs in enumerate(outputs["auxiliary_outputs"]):
                indices = self.matcher(auxiliary_outputs, targets, group_detr)
                for loss in self.losses:
                    if loss == "masks":
                        # Intermediate masks losses are too costly to compute, we ignore them.
                        continue
                    l_dict = self.get_loss(loss, auxiliary_outputs, targets, indices, num_boxes)
                    l_dict = {k + f"_{i}": v for k, v in l_dict.items()}
                    losses.update(l_dict)

        if "enc_outputs" in outputs:
            enc_outputs = outputs["enc_outputs"]
            indices = self.matcher(enc_outputs, targets, group_detr=group_detr)
            for loss in self.losses:
                l_dict = self.get_loss(loss, enc_outputs, targets, indices, num_boxes)
                l_dict = {k + "_enc": v for k, v in l_dict.items()}
                losses.update(l_dict)

        return losses


def LwDetrForObjectDetectionLoss(
    logits,
    labels,
    device,
    pred_boxes,
    config,
    outputs_class=None,
    outputs_coord=None,
    enc_outputs_class=None,
    enc_outputs_coord=None,
    **kwargs,
):
    # First: create the matcher
    matcher = LwDetrHungarianMatcher(
        class_cost=config.class_cost, bbox_cost=config.bbox_cost, giou_cost=config.giou_cost
    )
    # Second: create the criterion
    losses = ["labels", "boxes", "cardinality"]
    criterion = LwDetrImageLoss(
        matcher=matcher,
        num_classes=config.num_labels,
        focal_alpha=config.focal_alpha,
        losses=losses,
        group_detr=config.group_detr,
    )
    criterion.to(device)
    # Third: compute the losses, based on outputs and labels
    outputs_loss = {}
    auxiliary_outputs = None
    outputs_loss["logits"] = logits
    outputs_loss["pred_boxes"] = pred_boxes
    outputs_loss["enc_outputs"] = {
        "logits": enc_outputs_class,
        "pred_boxes": enc_outputs_coord,
    }
    if config.auxiliary_loss:
        auxiliary_outputs = _set_aux_loss(outputs_class, outputs_coord)
        outputs_loss["auxiliary_outputs"] = auxiliary_outputs
    loss_dict = criterion(outputs_loss, labels)
    # Fourth: compute total loss, as a weighted sum of the various losses
    weight_dict = {"loss_ce": 1, "loss_bbox": config.bbox_loss_coefficient}
    weight_dict["loss_giou"] = config.giou_loss_coefficient
    if config.auxiliary_loss:
        aux_weight_dict = {}
        for i in range(config.decoder_layers - 1):
            aux_weight_dict.update({k + f"_{i}": v for k, v in weight_dict.items()})
        weight_dict.update(aux_weight_dict)
    enc_weight_dict = {k + "_enc": v for k, v in weight_dict.items()}
    weight_dict.update(enc_weight_dict)
    loss = sum(loss_dict[k] * weight_dict[k] for k in loss_dict if k in weight_dict)
    return loss, loss_dict, auxiliary_outputs