add yolov6

yolo/config/__init__.py

@@ -4,6 +4,7 @@ from .yolov2_config    import build_yolov2_config
 from .yolov3_config    import build_yolov3_config
 from .yolov5_config    import build_yolov5_config
 from .yolov5_af_config import build_yolov5af_config
+from .yolov6_config    import build_yolov6_config
 from .yolov8_config    import build_yolov8_config
 from .gelan_config     import build_gelan_config
 from .rtdetr_config    import build_rtdetr_config
@@ -23,6 +24,8 @@ def build_config(args):
         cfg = build_yolov5af_config(args)
     elif 'yolov5' in args.model:
         cfg = build_yolov5_config(args)
+    elif 'yolov6' in args.model:
+        cfg = build_yolov6_config(args)
     elif 'yolov8' in args.model:
         cfg = build_yolov8_config(args)
     elif 'gelan' in args.model:

yolo/config/yolov6_config.py

@@ -0,0 +1,184 @@
+# Yolov6 configuration
+
+
+def build_yolov6_config(args):
+    if   args.model == 'yolov6_n':
+        return Yolov6NConfig()
+    elif args.model == 'yolov6_s':
+        return Yolov6SConfig()
+    elif args.model == 'yolov6_m':
+        return Yolov6MConfig()
+    elif args.model == 'yolov6_l':
+        return Yolov6LConfig()
+    else:
+        raise NotImplementedError("No config for model: {}".format(args.model))
+    
+# YOLOv6-Base config
+class Yolov6BaseConfig(object):
+    def __init__(self) -> None:
+        # ---------------- Model config ----------------
+        self.width    = 1.0
+        self.depth    = 1.0
+        self.reg_max  = 16
+        self.out_stride = [8, 16, 32]
+        self.max_stride = 32
+        self.num_levels = 3
+        self.scale      = "b"
+        ## Backbone
+        self.use_pretrained = True
+        ## Neck
+        self.neck_act       = 'silu'
+        self.neck_norm      = 'BN'
+        self.neck_depthwise = False
+        self.neck_expand_ratio = 0.5
+        self.spp_pooling_size  = 5
+        ## FPN
+        self.fpn_act  = 'silu'
+        self.fpn_norm = 'BN'
+        self.fpn_depthwise = False
+        ## Head
+        self.head_act  = 'silu'
+        self.head_norm = 'BN'
+        self.head_depthwise = False
+        self.num_cls_head   = 1
+        self.num_reg_head   = 1
+
+        # ---------------- Post-process config ----------------
+        ## Post process
+        self.val_topk = 1000
+        self.val_conf_thresh = 0.001
+        self.val_nms_thresh  = 0.7
+        self.test_topk = 100
+        self.test_conf_thresh = 0.2
+        self.test_nms_thresh  = 0.5
+
+        # ---------------- Assignment config ----------------
+        ## Matcher
+        self.tal_topk_candidates = 13
+        self.tal_alpha = 1.0
+        self.tal_beta  = 6.0
+        ## Loss weight
+        self.loss_cls = 1.0
+        self.loss_box = 2.5
+
+        # ---------------- ModelEMA config ----------------
+        self.use_ema = True
+        self.ema_decay = 0.9998
+        self.ema_tau   = 2000
+
+        # ---------------- Optimizer config ----------------
+        self.trainer      = 'yolo'
+        self.optimizer    = 'adamw'
+        self.per_image_lr = 0.001 / 64
+        self.base_lr      = None      # base_lr = per_image_lr * batch_size
+        self.min_lr_ratio = 0.01      # min_lr  = base_lr * min_lr_ratio
+        self.momentum     = 0.9
+        self.weight_decay = 0.05
+        self.clip_max_norm   = 35.0
+        self.warmup_bias_lr  = 0.1
+        self.warmup_momentum = 0.8
+
+        # ---------------- Lr Scheduler config ----------------
+        self.warmup_epoch = 3
+        self.lr_scheduler = "cosine"
+        self.max_epoch    = 300
+        self.eval_epoch   = 10
+        self.no_aug_epoch = 20
+
+        # ---------------- Data process config ----------------
+        self.aug_type = 'yolo'
+        self.box_format = 'xyxy'
+        self.normalize_coords = False
+        self.mosaic_prob = 1.0
+        self.mixup_prob  = 0.15
+        self.copy_paste  = 0.0           # approximated by the YOLOX's mixup
+        self.multi_scale = [0.5, 1.25]   # multi scale: [img_size * 0.5, img_size * 1.25]
+        ## Pixel mean & std
+        self.pixel_mean = [0., 0., 0.]
+        self.pixel_std  = [255., 255., 255.]
+        ## Transforms
+        self.train_img_size = 640
+        self.test_img_size  = 640
+        self.use_ablu = True
+        self.affine_params = {
+            'degrees': 0.0,
+            'translate': 0.2,
+            'scale': [0.1, 2.0],
+            'shear': 0.0,
+            'perspective': 0.0,
+            'hsv_h': 0.015,
+            'hsv_s': 0.7,
+            'hsv_v': 0.4,
+        }
+
+    def print_config(self):
+        config_dict = {key: value for key, value in self.__dict__.items() if not key.startswith('__')}
+        for k, v in config_dict.items():
+            print("{} : {}".format(k, v))
+
+# YOLOv6-N
+class Yolov6NConfig(Yolov6BaseConfig):
+    def __init__(self) -> None:
+        super().__init__()
+        # ---------------- Model config ----------------
+        self.width = 0.25
+        self.depth = 0.34
+        self.scale = "n"
+        ## Backbone
+        self.use_pretrained = True
+
+        # ---------------- Data process config ----------------
+        self.mosaic_prob = 1.0
+        self.mixup_prob  = 0.0
+        self.copy_paste  = 0.5
+
+# YOLOv6-S
+class Yolov6SConfig(Yolov6BaseConfig):
+    def __init__(self) -> None:
+        super().__init__()
+        # ---------------- Model config ----------------
+        self.width = 0.50
+        self.depth = 0.34
+        self.scale = "s"
+        ## Backbone
+        self.use_pretrained = True
+
+        # ---------------- Data process config ----------------
+        self.mosaic_prob = 1.0
+        self.mixup_prob  = 0.0
+        self.copy_paste  = 0.5
+
+# YOLOv6-M
+class Yolov6MConfig(Yolov6BaseConfig):
+    def __init__(self) -> None:
+        super().__init__()
+        # ---------------- Model config ----------------
+        self.width = 0.75
+        self.depth = 0.67
+        self.scale = "m"
+        self.bk_csp_expansion = 0.67
+        ## Backbone
+        self.use_pretrained = True
+
+        # ---------------- Data process config ----------------
+        self.mosaic_prob = 1.0
+        self.mixup_prob  = 0.1
+        self.copy_paste  = 0.5
+
+# YOLOv6-L
+class Yolov6LConfig(Yolov6BaseConfig):
+    def __init__(self) -> None:
+        super().__init__()
+        # ---------------- Model config ----------------
+        self.width = 1.0
+        self.depth = 1.0
+        self.scale = "l"
+        self.bk_csp_expansion = 0.5
+        ## Backbone
+        self.use_pretrained = True
+
+        # ---------------- Data process config ----------------
+        self.mosaic_prob = 1.0
+        self.mixup_prob  = 0.1
+        self.copy_paste  = 0.5
+        

yolo/models/__init__.py

@@ -7,6 +7,7 @@ from .yolov2.build    import build_yolov2
 from .yolov3.build    import build_yolov3
 from .yolov5.build    import build_yolov5
 from .yolov5_af.build import build_yolov5af
+from .yolov6.build    import build_yolov6
 from .yolov8.build    import build_yolov8
 from .gelan.build     import build_gelan
 from .rtdetr.build    import build_rtdetr
@@ -30,6 +31,9 @@ def build_model(args, cfg, is_val=False):
     ## Modified YOLOv5
     elif 'yolov5' in args.model:
         model, criterion = build_yolov5(cfg, is_val)
+    ## YOLOv6
+    elif 'yolov6' in args.model:
+        model, criterion = build_yolov6(cfg, is_val)
     ## YOLOv8
     elif 'yolov8' in args.model:
         model, criterion = build_yolov8(cfg, is_val)

yolo/models/yolov6/build.py

@@ -0,0 +1,24 @@
+import torch.nn as nn
+
+from .loss import SetCriterion
+from .yolov6 import Yolov6
+
+
+# build object detector
+def build_yolov6(cfg, is_val=False):
+    # -------------- Build YOLO --------------
+    model = Yolov6(cfg, is_val)
+
+    # -------------- Initialize YOLO --------------
+    for m in model.modules():
+        if isinstance(m, nn.BatchNorm2d):
+            m.eps = 1e-3
+            m.momentum = 0.03    
+            
+    # -------------- Build criterion --------------
+    criterion = None
+    if is_val:
+        # build criterion for training
+        criterion = SetCriterion(cfg)
+        
+    return model, criterion

yolo/models/yolov6/loss.py

@@ -0,0 +1,145 @@
+import torch
+import torch.nn.functional as F
+
+from utils.box_ops import bbox_iou
+from utils.distributed_utils import get_world_size, is_dist_avail_and_initialized
+
+from .matcher import TaskAlignedAssigner
+
+
+class SetCriterion(object):
+    def __init__(self, cfg):
+        # --------------- Basic parameters ---------------
+        self.cfg = cfg
+        self.reg_max = cfg.reg_max
+        self.num_classes = cfg.num_classes
+        # --------------- Loss config ---------------
+        self.loss_cls_weight = cfg.loss_cls
+        self.loss_box_weight = cfg.loss_box
+        # --------------- Matcher config ---------------
+        self.matcher = TaskAlignedAssigner(num_classes     = cfg.num_classes,
+                                           topk_candidates = cfg.tal_topk_candidates,
+                                           alpha           = cfg.tal_alpha,
+                                           beta            = cfg.tal_beta
+                                           )
+
+    def loss_classes(self, pred_cls, gt_score, gt_label):
+        # Compute VFL
+        pred_score = F.sigmoid(pred_cls).detach()
+        target = F.one_hot(gt_label, num_classes=self.num_classes + 1)[..., :-1]
+        weight = 0.75 * pred_score.pow(2.0) * (1 - target) + gt_score
+
+        loss_cls = F.binary_cross_entropy_with_logits(pred_cls, gt_score, weight=weight, reduction='none')
+
+        return loss_cls
+        
+    def loss_bboxes(self, pred_box, gt_box, bbox_weight):
+        # regression loss
+        ious = bbox_iou(pred_box, gt_box, xywh=False, GIoU=True)
+        loss_box = (1.0 - ious.squeeze(-1)) * bbox_weight
+
+        return loss_box
+    
+    def __call__(self, outputs, targets):        
+        """
+            outputs['pred_cls']: List(Tensor) [B, M, C]
+            outputs['pred_reg']: List(Tensor) [B, M, 4*(reg_max+1)]
+            outputs['pred_box']: List(Tensor) [B, M, 4]
+            outputs['anchors']: List(Tensor) [M, 2]
+            outputs['strides']: List(Int) [8, 16, 32] output stride
+            outputs['stride_tensor']: List(Tensor) [M, 1]
+            targets: (List) [dict{'boxes': [...], 
+                                 'labels': [...], 
+                                 'orig_size': ...}, ...]
+        """
+        # preds: [B, M, C]
+        cls_preds = torch.cat(outputs['pred_cls'], dim=1)
+        box_preds = torch.cat(outputs['pred_box'], dim=1)
+        bs, num_anchors = cls_preds.shape[:2]
+        device = cls_preds.device
+        anchors = torch.cat(outputs['anchors'], dim=0)
+        
+        # --------------- label assignment ---------------
+        gt_label_targets = []
+        gt_score_targets = []
+        gt_bbox_targets = []
+        fg_masks = []
+        for batch_idx in range(bs):
+            tgt_labels = targets[batch_idx]["labels"].to(device)     # [Mp,]
+            tgt_boxs = targets[batch_idx]["boxes"].to(device)        # [Mp, 4]
+
+            if self.cfg.normalize_coords:
+                img_h, img_w = outputs['image_size']
+                tgt_boxs[..., [0, 2]] *= img_w
+                tgt_boxs[..., [1, 3]] *= img_h
+            
+            if self.cfg.box_format == 'xywh':
+                tgt_boxs_x1y1 = tgt_boxs[..., :2] - 0.5 * tgt_boxs[..., 2:]
+                tgt_boxs_x2y2 = tgt_boxs[..., :2] + 0.5 * tgt_boxs[..., 2:]
+                tgt_boxs = torch.cat([tgt_boxs_x1y1, tgt_boxs_x2y2], dim=-1)
+
+            # check target
+            if len(tgt_labels) == 0 or tgt_boxs.max().item() == 0.:
+                # There is no valid gt
+                fg_mask  = cls_preds.new_zeros(1, num_anchors).bool()                       # [1, M,]
+                gt_label = cls_preds.new_zeros((1, num_anchors)).long()                     # [1, M,]
+                gt_score = cls_preds.new_zeros((1, num_anchors, self.num_classes)).float()  # [1, M, C]
+                gt_box   = cls_preds.new_zeros((1, num_anchors, 4)).float()                 # [1, M, 4]
+            else:
+                tgt_labels = tgt_labels[None, :, None]      # [1, Mp, 1]
+                tgt_boxs = tgt_boxs[None]                   # [1, Mp, 4]
+                (
+                    gt_label,   # [1, M]
+                    gt_box,     # [1, M, 4]
+                    gt_score,   # [1, M, C]
+                    fg_mask,    # [1, M,]
+                    _
+                ) = self.matcher(
+                    pd_scores = cls_preds[batch_idx:batch_idx+1].detach().sigmoid(), 
+                    pd_bboxes = box_preds[batch_idx:batch_idx+1].detach(),
+                    anc_points = anchors,
+                    gt_labels = tgt_labels,
+                    gt_bboxes = tgt_boxs
+                    )
+            gt_label_targets.append(gt_label)
+            gt_score_targets.append(gt_score)
+            gt_bbox_targets.append(gt_box)
+            fg_masks.append(fg_mask)
+
+        # List[B, 1, M, C] -> Tensor[B, M, C] -> Tensor[BM, C]
+        fg_masks = torch.cat(fg_masks, 0).view(-1)                                    # [BM,]
+        gt_label_targets = torch.cat(gt_label_targets, 0).view(-1)                    # [BM,]
+        gt_score_targets = torch.cat(gt_score_targets, 0).view(-1, self.num_classes)  # [BM, C]
+        gt_bbox_targets = torch.cat(gt_bbox_targets, 0).view(-1, 4)                   # [BM, 4]
+        num_fgs = gt_score_targets.sum()
+        
+        # Average loss normalizer across all the GPUs
+        if is_dist_avail_and_initialized():
+            torch.distributed.all_reduce(num_fgs)
+        num_fgs = (num_fgs / get_world_size()).clamp(1.0)
+
+        # ------------------ Classification loss ------------------
+        cls_preds = cls_preds.view(-1, self.num_classes)
+        loss_cls = self.loss_classes(cls_preds, gt_score_targets, gt_label_targets)
+        loss_cls = loss_cls.sum() / num_fgs
+
+        # ------------------ Regression loss ------------------
+        box_preds_pos = box_preds.view(-1, 4)[fg_masks]
+        box_targets_pos = gt_bbox_targets.view(-1, 4)[fg_masks]
+        bbox_weight = gt_score_targets[fg_masks].sum(-1)
+        loss_box = self.loss_bboxes(box_preds_pos, box_targets_pos, bbox_weight)
+        loss_box = loss_box.sum() / num_fgs
+
+        # total loss
+        losses = loss_cls * self.loss_cls_weight + loss_box * self.loss_box_weight
+        loss_dict = dict(
+                loss_cls = loss_cls,
+                loss_box = loss_box,
+                losses = losses
+        )
+
+        return loss_dict
+    
+
+if __name__ == "__main__":
+    pass

yolo/models/yolov6/matcher.py

@@ -0,0 +1,198 @@
+import torch
+import torch.nn as nn
+from utils.box_ops import bbox_iou
+
+
+# -------------------------- Task Aligned Assigner --------------------------
+class TaskAlignedAssigner(nn.Module):
+    def __init__(self,
+                 num_classes     = 80,
+                 topk_candidates = 10,
+                 alpha           = 0.5,
+                 beta            = 6.0, 
+                 eps             = 1e-9):
+        super(TaskAlignedAssigner, self).__init__()
+        self.topk_candidates = topk_candidates
+        self.num_classes = num_classes
+        self.bg_idx = num_classes
+        self.alpha = alpha
+        self.beta = beta
+        self.eps = eps
+
+    @torch.no_grad()
+    def forward(self,
+                pd_scores,
+                pd_bboxes,
+                anc_points,
+                gt_labels,
+                gt_bboxes):
+        self.bs = pd_scores.size(0)
+        self.n_max_boxes = gt_bboxes.size(1)
+
+        mask_pos, align_metric, overlaps = self.get_pos_mask(
+            pd_scores, pd_bboxes, gt_labels, gt_bboxes, anc_points)
+
+        target_gt_idx, fg_mask, mask_pos = select_highest_overlaps(
+            mask_pos, overlaps, self.n_max_boxes)
+
+        # Assigned target
+        target_labels, target_bboxes, target_scores = self.get_targets(
+            gt_labels, gt_bboxes, target_gt_idx, fg_mask)
+
+        # normalize
+        align_metric *= mask_pos
+        pos_align_metrics = align_metric.amax(axis=-1, keepdim=True)  # b, max_num_obj
+        pos_overlaps = (overlaps * mask_pos).amax(axis=-1, keepdim=True)  # b, max_num_obj
+        norm_align_metric = (align_metric * pos_overlaps / (pos_align_metrics + self.eps)).amax(-2).unsqueeze(-1)
+        target_scores = target_scores * norm_align_metric
+
+        return target_labels, target_bboxes, target_scores, fg_mask.bool(), target_gt_idx
+
+    def get_pos_mask(self, pd_scores, pd_bboxes, gt_labels, gt_bboxes, anc_points):
+        # get in_gts mask, (b, max_num_obj, h*w)
+        mask_in_gts = select_candidates_in_gts(anc_points, gt_bboxes)
+        # get anchor_align metric, (b, max_num_obj, h*w)
+        align_metric, overlaps = self.get_box_metrics(pd_scores, pd_bboxes, gt_labels, gt_bboxes, mask_in_gts)
+        # get topk_metric mask, (b, max_num_obj, h*w)
+        mask_topk = self.select_topk_candidates(align_metric)
+        # merge all mask to a final mask, (b, max_num_obj, h*w)
+        mask_pos = mask_topk * mask_in_gts
+
+        return mask_pos, align_metric, overlaps
+
+    def get_box_metrics(self, pd_scores, pd_bboxes, gt_labels, gt_bboxes, mask_in_gts):
+        """Compute alignment metric given predicted and ground truth bounding boxes."""
+        na = pd_bboxes.shape[-2]
+        mask_in_gts = mask_in_gts.bool()  # b, max_num_obj, h*w
+        overlaps = torch.zeros([self.bs, self.n_max_boxes, na], dtype=pd_bboxes.dtype, device=pd_bboxes.device)
+        bbox_scores = torch.zeros([self.bs, self.n_max_boxes, na], dtype=pd_scores.dtype, device=pd_scores.device)
+
+        ind = torch.zeros([2, self.bs, self.n_max_boxes], dtype=torch.long)  # 2, b, max_num_obj
+        ind[0] = torch.arange(end=self.bs).view(-1, 1).expand(-1, self.n_max_boxes)  # b, max_num_obj
+        ind[1] = gt_labels.squeeze(-1)  # b, max_num_obj
+        # Get the scores of each grid for each gt cls
+        bbox_scores[mask_in_gts] = pd_scores[ind[0], :, ind[1]][mask_in_gts]  # b, max_num_obj, h*w
+
+        # (b, max_num_obj, 1, 4), (b, 1, h*w, 4)
+        pd_boxes = pd_bboxes.unsqueeze(1).expand(-1, self.n_max_boxes, -1, -1)[mask_in_gts]
+        gt_boxes = gt_bboxes.unsqueeze(2).expand(-1, -1, na, -1)[mask_in_gts]
+        overlaps[mask_in_gts] = bbox_iou(gt_boxes, pd_boxes, xywh=False, CIoU=True).squeeze(-1).clamp_(0)
+
+        align_metric = bbox_scores.pow(self.alpha) * overlaps.pow(self.beta)
+        return align_metric, overlaps
+
+    def select_topk_candidates(self, metrics, largest=True):
+        """
+        Args:
+            metrics: (b, max_num_obj, h*w).
+            topk_mask: (b, max_num_obj, topk) or None
+        """
+        # (b, max_num_obj, topk)
+        topk_metrics, topk_idxs = torch.topk(metrics, self.topk_candidates, dim=-1, largest=largest)
+        topk_mask = (topk_metrics.max(-1, keepdim=True)[0] > self.eps).expand_as(topk_idxs)
+        # (b, max_num_obj, topk)
+        topk_idxs.masked_fill_(~topk_mask, 0)
+
+        # (b, max_num_obj, topk, h*w) -> (b, max_num_obj, h*w)
+        count_tensor = torch.zeros(metrics.shape, dtype=torch.int8, device=topk_idxs.device)
+        ones = torch.ones_like(topk_idxs[:, :, :1], dtype=torch.int8, device=topk_idxs.device)
+        for k in range(self.topk_candidates):
+            # Expand topk_idxs for each value of k and add 1 at the specified positions
+            count_tensor.scatter_add_(-1, topk_idxs[:, :, k:k + 1], ones)
+        # count_tensor.scatter_add_(-1, topk_idxs, torch.ones_like(topk_idxs, dtype=torch.int8, device=topk_idxs.device))
+        # Filter invalid bboxes
+        count_tensor.masked_fill_(count_tensor > 1, 0)
+
+        return count_tensor.to(metrics.dtype)
+
+    def get_targets(self, gt_labels, gt_bboxes, target_gt_idx, fg_mask):
+        # Assigned target labels, (b, 1)
+        batch_ind = torch.arange(end=self.bs, dtype=torch.int64, device=gt_labels.device)[..., None]
+        target_gt_idx = target_gt_idx + batch_ind * self.n_max_boxes  # (b, h*w)
+        target_labels = gt_labels.long().flatten()[target_gt_idx]  # (b, h*w)
+
+        # Assigned target boxes, (b, max_num_obj, 4) -> (b, h*w, 4)
+        target_bboxes = gt_bboxes.view(-1, 4)[target_gt_idx]
+
+        # Assigned target scores
+        target_labels.clamp_(0)
+
+        # 10x faster than F.one_hot()
+        target_scores = torch.zeros((target_labels.shape[0], target_labels.shape[1], self.num_classes),
+                                    dtype=torch.int64,
+                                    device=target_labels.device)  # (b, h*w, 80)
+        target_scores.scatter_(2, target_labels.unsqueeze(-1), 1)
+
+        fg_scores_mask = fg_mask[:, :, None].repeat(1, 1, self.num_classes)  # (b, h*w, 80)
+        target_scores = torch.where(fg_scores_mask > 0, target_scores, 0)
+
+        return target_labels, target_bboxes, target_scores
+    
+
+# -------------------------- Basic Functions --------------------------
+def select_candidates_in_gts(xy_centers, gt_bboxes, eps=1e-9):
+    """select the positive anchors's center in gt
+    Args:
+        xy_centers (Tensor): shape(bs*n_max_boxes, num_total_anchors, 4)
+        gt_bboxes (Tensor): shape(bs, n_max_boxes, 4)
+    Return:
+        (Tensor): shape(bs, n_max_boxes, num_total_anchors)
+    """
+    n_anchors = xy_centers.size(0)
+    bs, n_max_boxes, _ = gt_bboxes.size()
+    _gt_bboxes = gt_bboxes.reshape([-1, 4])
+    xy_centers = xy_centers.unsqueeze(0).repeat(bs * n_max_boxes, 1, 1)
+    gt_bboxes_lt = _gt_bboxes[:, 0:2].unsqueeze(1).repeat(1, n_anchors, 1)
+    gt_bboxes_rb = _gt_bboxes[:, 2:4].unsqueeze(1).repeat(1, n_anchors, 1)
+    b_lt = xy_centers - gt_bboxes_lt
+    b_rb = gt_bboxes_rb - xy_centers
+    bbox_deltas = torch.cat([b_lt, b_rb], dim=-1)
+    bbox_deltas = bbox_deltas.reshape([bs, n_max_boxes, n_anchors, -1])
+    return (bbox_deltas.min(axis=-1)[0] > eps).to(gt_bboxes.dtype)
+
+def select_highest_overlaps(mask_pos, overlaps, n_max_boxes):
+    """if an anchor box is assigned to multiple gts,
+        the one with the highest iou will be selected.
+    Args:
+        mask_pos (Tensor): shape(bs, n_max_boxes, num_total_anchors)
+        overlaps (Tensor): shape(bs, n_max_boxes, num_total_anchors)
+    Return:
+        target_gt_idx (Tensor): shape(bs, num_total_anchors)
+        fg_mask (Tensor): shape(bs, num_total_anchors)
+        mask_pos (Tensor): shape(bs, n_max_boxes, num_total_anchors)
+    """
+    fg_mask = mask_pos.sum(-2)
+    if fg_mask.max() > 1:  # one anchor is assigned to multiple gt_bboxes
+        mask_multi_gts = (fg_mask.unsqueeze(1) > 1).expand(-1, n_max_boxes, -1)  # (b, n_max_boxes, h*w)
+        max_overlaps_idx = overlaps.argmax(1)  # (b, h*w)
+
+        is_max_overlaps = torch.zeros(mask_pos.shape, dtype=mask_pos.dtype, device=mask_pos.device)
+        is_max_overlaps.scatter_(1, max_overlaps_idx.unsqueeze(1), 1)
+
+        mask_pos = torch.where(mask_multi_gts, is_max_overlaps, mask_pos).float()  # (b, n_max_boxes, h*w)
+        fg_mask = mask_pos.sum(-2)
+    # Find each grid serve which gt(index)
+    target_gt_idx = mask_pos.argmax(-2)  # (b, h*w)
+
+    return target_gt_idx, fg_mask, mask_pos
+
+def iou_calculator(box1, box2, eps=1e-9):
+    """Calculate iou for batch
+    Args:
+        box1 (Tensor): shape(bs, n_max_boxes, 1, 4)
+        box2 (Tensor): shape(bs, 1, num_total_anchors, 4)
+    Return:
+        (Tensor): shape(bs, n_max_boxes, num_total_anchors)
+    """
+    box1 = box1.unsqueeze(2)  # [N, M1, 4] -> [N, M1, 1, 4]
+    box2 = box2.unsqueeze(1)  # [N, M2, 4] -> [N, 1, M2, 4]
+    px1y1, px2y2 = box1[:, :, :, 0:2], box1[:, :, :, 2:4]
+    gx1y1, gx2y2 = box2[:, :, :, 0:2], box2[:, :, :, 2:4]
+    x1y1 = torch.maximum(px1y1, gx1y1)
+    x2y2 = torch.minimum(px2y2, gx2y2)
+    overlap = (x2y2 - x1y1).clip(0).prod(-1)
+    area1 = (px2y2 - px1y1).clip(0).prod(-1)
+    area2 = (gx2y2 - gx1y1).clip(0).prod(-1)
+    union = area1 + area2 - overlap + eps
+
+    return overlap / union

yolo/models/yolov6/yolov6.py

@@ -0,0 +1,157 @@
+# --------------- Torch components ---------------
+import torch
+import torch.nn as nn
+
+# --------------- Model components ---------------
+from .yolov6_backbone import Yolov6Backbone
+from .yolov6_neck     import SPPF
+from .yolov6_pafpn    import Yolov6PaFPN
+from .yolov6_head     import Yolov6DetHead
+from .yolov6_pred     import Yolov6DetPredLayer
+
+# --------------- External components ---------------
+from utils.misc import multiclass_nms
+
+
+# YOLOv6
+class Yolov6(nn.Module):
+    def __init__(self,
+                 cfg,
+                 is_val = False,
+                 ) -> None:
+        super(Yolov6, self).__init__()
+        # ---------------------- Basic setting ----------------------
+        self.cfg = cfg
+        self.num_classes = cfg.num_classes
+        ## Post-process parameters
+        self.topk_candidates  = cfg.val_topk        if is_val else cfg.test_topk
+        self.conf_thresh      = cfg.val_conf_thresh if is_val else cfg.test_conf_thresh
+        self.nms_thresh       = cfg.val_nms_thresh  if is_val else cfg.test_nms_thresh
+        self.no_multi_labels  = False if is_val else True
+        
+        # ---------------------- Network Parameters ----------------------
+        ## Backbone
+        self.backbone = Yolov6Backbone(cfg)
+        self.pyramid_feat_dims = self.backbone.feat_dims[-3:]
+        ## Neck: SPP
+        self.neck     = SPPF(cfg, self.pyramid_feat_dims[-1], self.pyramid_feat_dims[-1])
+        self.pyramid_feat_dims[-1] = self.neck.out_dim
+        ## Neck: FPN
+        self.fpn      = Yolov6PaFPN(cfg, self.pyramid_feat_dims)
+        ## Head
+        self.head     = Yolov6DetHead(cfg, self.fpn.out_dims)
+        ## Pred
+        self.pred     = Yolov6DetPredLayer(cfg, self.fpn.out_dims)
+
+    def switch_deploy(self,):
+        for m in self.modules():
+            if hasattr(m, "switch_to_deploy"):
+                m.switch_to_deploy()
+
+    def post_process(self, cls_preds, box_preds):
+        """
+        We process predictions at each scale hierarchically
+        Input:
+            cls_preds: List[torch.Tensor] -> [[B, M, C], ...], B=1
+            box_preds: List[torch.Tensor] -> [[B, M, 4], ...], B=1
+        Output:
+            bboxes: np.array -> [N, 4]
+            scores: np.array -> [N,]
+            labels: np.array -> [N,]
+        """
+        all_scores = []
+        all_labels = []
+        all_bboxes = []
+        
+        for cls_pred_i, box_pred_i in zip(cls_preds, box_preds):
+            cls_pred_i = cls_pred_i[0]
+            box_pred_i = box_pred_i[0]
+            if self.no_multi_labels:
+                # [M,]
+                scores, labels = torch.max(cls_pred_i.sigmoid(), dim=1)
+
+                # Keep top k top scoring indices only.
+                num_topk = min(self.topk_candidates, box_pred_i.size(0))
+
+                # topk candidates
+                predicted_prob, topk_idxs = scores.sort(descending=True)
+                topk_scores = predicted_prob[:num_topk]
+                topk_idxs = topk_idxs[:num_topk]
+
+                # filter out the proposals with low confidence score
+                keep_idxs = topk_scores > self.conf_thresh
+                scores = topk_scores[keep_idxs]
+                topk_idxs = topk_idxs[keep_idxs]
+
+                labels = labels[topk_idxs]
+                bboxes = box_pred_i[topk_idxs]
+            else:
+                # [M, C] -> [MC,]
+                scores_i = cls_pred_i.sigmoid().flatten()
+
+                # Keep top k top scoring indices only.
+                num_topk = min(self.topk_candidates, box_pred_i.size(0))
+
+                # torch.sort is actually faster than .topk (at least on GPUs)
+                predicted_prob, topk_idxs = scores_i.sort(descending=True)
+                topk_scores = predicted_prob[:num_topk]
+                topk_idxs = topk_idxs[:num_topk]
+
+                # filter out the proposals with low confidence score
+                keep_idxs = topk_scores > self.conf_thresh
+                scores = topk_scores[keep_idxs]
+                topk_idxs = topk_idxs[keep_idxs]
+
+                anchor_idxs = torch.div(topk_idxs, self.num_classes, rounding_mode='floor')
+                labels = topk_idxs % self.num_classes
+
+                bboxes = box_pred_i[anchor_idxs]
+
+            all_scores.append(scores)
+            all_labels.append(labels)
+            all_bboxes.append(bboxes)
+
+        scores = torch.cat(all_scores, dim=0)
+        labels = torch.cat(all_labels, dim=0)
+        bboxes = torch.cat(all_bboxes, dim=0)
+
+        # to cpu & numpy
+        scores = scores.cpu().numpy()
+        labels = labels.cpu().numpy()
+        bboxes = bboxes.cpu().numpy()
+
+        # nms
+        scores, labels, bboxes = multiclass_nms(
+            scores, labels, bboxes, self.nms_thresh, self.num_classes)
+        
+        return bboxes, scores, labels
+    
+    def forward(self, x):
+        # ---------------- Backbone ----------------
+        pyramid_feats = self.backbone(x)
+        # ---------------- Neck: SPP ----------------
+        pyramid_feats[-1] = self.neck(pyramid_feats[-1])
+
+        # ---------------- Neck: PaFPN ----------------
+        pyramid_feats = self.fpn(pyramid_feats)
+
+        # ---------------- Heads ----------------
+        cls_feats, reg_feats = self.head(pyramid_feats)
+
+        # ---------------- Preds ----------------
+        outputs = self.pred(cls_feats, reg_feats)
+        outputs['image_size'] = [x.shape[2], x.shape[3]]
+
+        if not self.training:
+            all_cls_preds = outputs['pred_cls']
+            all_box_preds = outputs['pred_box']
+
+            # post process
+            bboxes, scores, labels = self.post_process(all_cls_preds, all_box_preds)
+            outputs = {
+                "scores": scores,
+                "labels": labels,
+                "bboxes": bboxes
+            }
+        
+        return outputs 

yolo/models/yolov6/yolov6_backbone.py

@@ -0,0 +1,150 @@
+import torch
+import torch.nn as nn
+
+try:
+    from .yolov6_basic import RepBlock, RepVGGBlock, RepCSPBlock
+except:
+    from  yolov6_basic import RepBlock, RepVGGBlock, RepCSPBlock
+
+# IN1K pretrained weight
+pretrained_urls = {
+    'n': None,
+    's': None,
+    'm': None,
+    'l': None,
+}
+
+
+# --------------------- Yolov3's Backbone -----------------------
+## Modified DarkNet
+class Yolov6Backbone(nn.Module):
+    def __init__(self, cfg):
+        super(Yolov6Backbone, self).__init__()
+        # ------------------ Basic setting ------------------
+        self.cfg = cfg
+        self.model_scale = cfg.scale
+        self.feat_dims = [round(64   * cfg.width),
+                          round(128  * cfg.width),
+                          round(256  * cfg.width),
+                          round(512  * cfg.width),
+                          round(1024 * cfg.width)]
+        
+        # ------------------ Network setting ------------------
+        ## P1/2
+        self.layer_1 = RepVGGBlock(3, self.feat_dims[0],
+                                   kernel_size=3, padding=1, stride=2)
+        # P2/4
+        self.layer_2 = self.make_block(self.feat_dims[0], self.feat_dims[1], round(6*cfg.depth)) 
+        # P3/8
+        self.layer_3 = self.make_block(self.feat_dims[1], self.feat_dims[2], round(12*cfg.depth)) 
+        # P4/16
+        self.layer_4 = self.make_block(self.feat_dims[2], self.feat_dims[3], round(18*cfg.depth)) 
+        # P5/32
+        self.layer_5 = self.make_block(self.feat_dims[3], self.feat_dims[4], round(6*cfg.depth)) 
+
+        # Initialize all layers
+        self.init_weights()
+
+        # Load imagenet pretrained weight
+        if cfg.use_pretrained:
+            self.load_pretrained()
+        
+    def init_weights(self):
+        """Initialize the parameters."""
+        for m in self.modules():
+            if isinstance(m, torch.nn.Conv2d):
+                # In order to be consistent with the source code,
+                # reset the Conv2d initialization parameters
+                m.reset_parameters()
+
+    def load_pretrained(self):
+        url = pretrained_urls[self.model_scale]
+        if url is not None:
+            print('Loading backbone pretrained weight from : {}'.format(url))
+            # checkpoint state dict
+            checkpoint = torch.hub.load_state_dict_from_url(
+                url=url, map_location="cpu", check_hash=True)
+            checkpoint_state_dict = checkpoint.pop("model")
+            # model state dict
+            model_state_dict = self.state_dict()
+            # check
+            for k in list(checkpoint_state_dict.keys()):
+                if k in model_state_dict:
+                    shape_model = tuple(model_state_dict[k].shape)
+                    shape_checkpoint = tuple(checkpoint_state_dict[k].shape)
+                    if shape_model != shape_checkpoint:
+                        checkpoint_state_dict.pop(k)
+                else:
+                    checkpoint_state_dict.pop(k)
+                    print('Unused key: ', k)
+            # load the weight
+            self.load_state_dict(checkpoint_state_dict)
+        else:
+            print('No pretrained weight for model scale: {}.'.format(self.model_scale))
+
+    def make_block(self, in_dim, out_dim, num_blocks=1):
+        if   self.model_scale in ["n", "s"]:
+            block = nn.Sequential(
+                RepVGGBlock(in_dim, out_dim,
+                            kernel_size=3, padding=1, stride=2),
+                RepBlock(in_channels  = out_dim,
+                         out_channels = out_dim,
+                         num_blocks   = num_blocks,
+                         block        = RepVGGBlock)
+                         )
+        elif self.model_scale in ["m", "l"]:
+            block = nn.Sequential(
+                RepVGGBlock(in_dim, out_dim,
+                            kernel_size=3, padding=1, stride=2),
+                RepCSPBlock(in_channels  = out_dim,
+                            out_channels = out_dim,
+                            num_blocks   = num_blocks,
+                            expansion    = self.cfg.bk_csp_expansion)
+                            )
+        else:
+            raise NotImplementedError("Unknown model scale: {}".format(self.model_scale))
+            
+        return block
+
+    def forward(self, x):
+        c1 = self.layer_1(x)
+        c2 = self.layer_2(c1)
+        c3 = self.layer_3(c2)
+        c4 = self.layer_4(c3)
+        c5 = self.layer_5(c4)
+        outputs = [c3, c4, c5]
+
+        return outputs
+
+
+if __name__ == '__main__':
+    import time
+    from thop import profile
+    class BaseConfig(object):
+        def __init__(self) -> None:
+            self.bk_depthwise = False
+            self.width = 0.50
+            self.depth = 0.34
+            self.scale = "s"
+            self.use_pretrained = True
+
+    cfg = BaseConfig()
+    model = Yolov6Backbone(cfg)
+    x = torch.randn(1, 3, 640, 640)
+    t0 = time.time()
+    outputs = model(x)
+    t1 = time.time()
+    print('Time: ', t1 - t0)
+    for out in outputs:
+        print(out.shape)
+    
+    for m in model.modules():
+        if hasattr(m, "switch_to_deploy"):
+            m.switch_to_deploy()
+
+    x = torch.randn(1, 3, 640, 640)
+    print('==============================')
+    flops, params = profile(model, inputs=(x, ), verbose=False)
+    print('==============================')
+    print('GFLOPs : {:.2f}'.format(flops / 1e9 * 2))
+    print('Params : {:.2f} M'.format(params / 1e6))

yolo/models/yolov6/yolov6_basic.py

@@ -0,0 +1,233 @@
+import numpy as np
+import torch
+import torch.nn as nn
+
+
+# --------------------- Basic modules ---------------------
+def get_conv2d(c1, c2, k, p, s, d, g, bias=False):
+    conv = nn.Conv2d(c1, c2, k, stride=s, padding=p, dilation=d, groups=g, bias=bias)
+
+    return conv
+
+def get_activation(act_type=None):
+    if act_type == 'relu':
+        return nn.ReLU(inplace=True)
+    elif act_type == 'lrelu':
+        return nn.LeakyReLU(0.1, inplace=True)
+    elif act_type == 'mish':
+        return nn.Mish(inplace=True)
+    elif act_type == 'silu':
+        return nn.SiLU(inplace=True)
+    elif act_type is None:
+        return nn.Identity()
+    else:
+        raise NotImplementedError
+
+def get_norm(norm_type, dim):
+    if   norm_type == 'BN':
+        return nn.BatchNorm2d(dim)
+    elif norm_type == 'GN':
+        return nn.GroupNorm(num_groups=32, num_channels=dim)
+    elif norm_type is None:
+        return nn.Identity()
+    else:
+        raise NotImplementedError
+
+class BasicConv(nn.Module):
+    def __init__(self, 
+                 in_dim,                   # in channels
+                 out_dim,                  # out channels 
+                 kernel_size=1,            # kernel size 
+                 padding=0,                # padding
+                 stride=1,                 # padding
+                 dilation=1,               # dilation
+                 act_type  :str = 'lrelu', # activation
+                 norm_type :str = 'BN',    # normalization
+                 depthwise :bool = False
+                ):
+        super(BasicConv, self).__init__()
+        self.depthwise = depthwise
+        if not depthwise:
+            self.conv = get_conv2d(in_dim, out_dim, k=kernel_size, p=padding, s=stride, d=dilation, g=1, bias=True)
+            self.norm = get_norm(norm_type, out_dim)
+        else:
+            self.conv1 = get_conv2d(in_dim, in_dim, k=kernel_size, p=padding, s=stride, d=dilation, g=in_dim, bias=True)
+            self.norm1 = get_norm(norm_type, in_dim)
+            self.conv2 = get_conv2d(in_dim, out_dim, k=1, p=0, s=1, d=1, g=1, bias=True)
+            self.norm2 = get_norm(norm_type, out_dim)
+        self.act  = get_activation(act_type)
+
+    def forward(self, x):
+        if not self.depthwise:
+            return self.act(self.norm(self.conv(x)))
+        else:
+            # Depthwise conv
+            x = self.norm1(self.conv1(x))
+            # Pointwise conv
+            x = self.act(self.norm2(self.conv2(x)))
+            return x
+
+class RepVGGBlock(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 stride=1,
+                 padding=1,
+                 dilation=1,
+                 groups=1,
+                 deploy=False,
+                 ):
+        super(RepVGGBlock, self).__init__()
+        assert kernel_size == 3
+        assert padding == 1
+        # --------- Basic parameters ---------
+        self.deploy = deploy
+        self.groups = groups
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        padding_11 = padding - kernel_size // 2
+        # --------- Model parameters ---------
+        if deploy:
+            self.rbr_reparam = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride,
+                                         padding=padding, dilation=dilation, groups=groups, bias=True)
+        else:
+            self.rbr_identity = nn.BatchNorm2d(num_features=in_channels) if out_channels == in_channels and stride == 1 else None
+            self.rbr_dense    = BasicConv(in_channels, out_channels, kernel_size=kernel_size, padding=padding, stride=stride, act_type=None)
+            self.rbr_1x1      = BasicConv(in_channels, out_channels, kernel_size=1, padding=padding_11, stride=stride, act_type=None)
+        self.nonlinearity = nn.ReLU()
+
+    def forward(self, inputs):
+        '''Forward process'''
+        if hasattr(self, 'rbr_reparam'):
+            return self.nonlinearity(self.rbr_reparam(inputs))
+
+        if self.rbr_identity is None:
+            id_out = 0
+        else:
+            id_out = self.rbr_identity(inputs)
+
+        return self.nonlinearity(self.rbr_dense(inputs) + self.rbr_1x1(inputs) + id_out)
+
+    def get_equivalent_kernel_bias(self):
+        kernel3x3, bias3x3 = self._fuse_bn_tensor(self.rbr_dense)
+        kernel1x1, bias1x1 = self._fuse_bn_tensor(self.rbr_1x1)
+        kernelid, biasid = self._fuse_bn_tensor(self.rbr_identity)
+        return kernel3x3 + self._pad_1x1_to_3x3_tensor(kernel1x1) + kernelid, bias3x3 + bias1x1 + biasid
+
+    def _avg_to_3x3_tensor(self, avgp):
+        channels = self.in_channels
+        groups = self.groups
+        kernel_size = avgp.kernel_size
+        input_dim = channels // groups
+        k = torch.zeros((channels, input_dim, kernel_size, kernel_size))
+        k[np.arange(channels), np.tile(np.arange(input_dim), groups), :, :] = 1.0 / kernel_size ** 2
+        return k
+
+    def _pad_1x1_to_3x3_tensor(self, kernel1x1):
+        if kernel1x1 is None:
+            return 0
+        else:
+            return torch.nn.functional.pad(kernel1x1, [1, 1, 1, 1])
+
+    def _fuse_bn_tensor(self, branch):
+        if branch is None:
+            return 0, 0
+        if isinstance(branch, BasicConv):
+            kernel = branch.conv.weight
+            bias   = branch.conv.bias
+            return kernel, bias
+        elif isinstance(branch, nn.BatchNorm2d):
+            if not hasattr(self, 'id_tensor'):
+                input_dim = self.in_channels // self.groups
+                kernel_value = np.zeros((self.in_channels, input_dim, 3, 3), dtype=np.float32)
+                for i in range(self.in_channels):
+                    kernel_value[i, i % input_dim, 1, 1] = 1
+                self.id_tensor = torch.from_numpy(kernel_value).to(branch.weight.device)
+            kernel = self.id_tensor
+            running_mean = branch.running_mean
+            running_var = branch.running_var
+            gamma = branch.weight
+            beta = branch.bias
+            eps = branch.eps
+            std = (running_var + eps).sqrt()
+            t = (gamma / std).reshape(-1, 1, 1, 1)
+            return kernel * t, beta - running_mean * gamma / std
+
+    def switch_to_deploy(self):
+        if hasattr(self, 'rbr_reparam'):
+            return
+        kernel, bias = self.get_equivalent_kernel_bias()
+        self.rbr_reparam = nn.Conv2d(in_channels=self.rbr_dense.conv.in_channels, out_channels=self.rbr_dense.conv.out_channels,
+                                     kernel_size=self.rbr_dense.conv.kernel_size, stride=self.rbr_dense.conv.stride,
+                                     padding=self.rbr_dense.conv.padding, dilation=self.rbr_dense.conv.dilation, groups=self.rbr_dense.conv.groups, bias=True)
+        self.rbr_reparam.weight.data = kernel
+        self.rbr_reparam.bias.data = bias
+        for para in self.parameters():
+            para.detach_()
+        self.__delattr__('rbr_dense')
+        self.__delattr__('rbr_1x1')
+        if hasattr(self, 'rbr_identity'):
+            self.__delattr__('rbr_identity')
+        if hasattr(self, 'id_tensor'):
+            self.__delattr__('id_tensor')
+        self.deploy = True
+
+
+# ---------------------------- Basic Modules ----------------------------
+class RepBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, num_blocks=1, block=RepVGGBlock):
+        super().__init__()
+        self.conv1 = block(in_channels, out_channels)
+        self.block = nn.Sequential(*(block(out_channels, out_channels)
+                                     for _ in range(num_blocks - 1))) if num_blocks > 1 else nn.Identity()
+        if block == BottleRep:
+            self.conv1 = BottleRep(in_channels, out_channels, weight=True)
+            num_blocks = num_blocks // 2
+            self.block = nn.Sequential(*(BottleRep(out_channels, out_channels, weight=True)
+                                         for _ in range(num_blocks - 1))) if num_blocks > 1 else None
+
+    def forward(self, x):
+        x = self.conv1(x)
+        if self.block is not None:
+            x = self.block(x)
+
+        return x
+
+class BottleRep(nn.Module):
+
+    def __init__(self, in_channels, out_channels, weight=False):
+        super().__init__()
+        self.conv1 = RepVGGBlock(in_channels, out_channels, kernel_size=3, padding=1, stride=1)
+        self.conv2 = RepVGGBlock(out_channels, out_channels, kernel_size=3, padding=1, stride=1)
+        if in_channels != out_channels:
+            self.shortcut = False
+        else:
+            self.shortcut = True
+        if weight:
+            self.alpha = nn.Parameter(torch.ones(1))
+        else:
+            self.alpha = 1.0
+
+    def forward(self, x):
+        outputs = self.conv1(x)
+        outputs = self.conv2(outputs)
+
+        return outputs + self.alpha * x if self.shortcut else outputs
+
+class RepCSPBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, num_blocks=1, expansion=0.5):
+        super().__init__()
+        inter_dim = round(out_channels * expansion)  # hidden channels
+        self.cv1 = BasicConv(in_channels, inter_dim, kernel_size=1, act_type='relu')
+        self.cv2 = BasicConv(in_channels, inter_dim, kernel_size=1, act_type='relu')
+        self.cv3 = BasicConv(2 * inter_dim, out_channels, kernel_size=1, act_type='relu')
+
+        self.module = RepBlock(inter_dim, inter_dim, num_blocks, block=BottleRep)
+
+    def forward(self, x):
+        x1 = self.cv1(x)
+        x2 = self.module(self.cv2(x))
+        out = self.cv3(torch.cat((x1, x2), dim=1))
+
+        return out

yolo/models/yolov6/yolov6_head.py

@@ -0,0 +1,169 @@
+import torch
+import torch.nn as nn
+
+try:
+    from .yolov6_basic import BasicConv
+except:
+    from  yolov6_basic import BasicConv
+
+
+## Single-level Detection Head
+class DetHead(nn.Module):
+    def __init__(self,
+                 in_dim       :int  = 256,
+                 cls_head_dim :int  = 256,
+                 reg_head_dim :int  = 256,
+                 num_cls_head :int  = 2,
+                 num_reg_head :int  = 2,
+                 act_type     :str  = "silu",
+                 norm_type    :str  = "BN",
+                 depthwise    :bool = False):
+        super().__init__()
+        # --------- Basic Parameters ----------
+        self.in_dim = in_dim
+        self.num_cls_head = num_cls_head
+        self.num_reg_head = num_reg_head
+        self.act_type = act_type
+        self.norm_type = norm_type
+        self.depthwise = depthwise
+        
+        # --------- Network Parameters ----------
+        ## cls head
+        cls_feats = []
+        self.cls_head_dim = cls_head_dim
+        for i in range(num_cls_head):
+            if i == 0:
+                cls_feats.append(
+                    BasicConv(in_dim, self.cls_head_dim,
+                              kernel_size=3, padding=1, stride=1, 
+                              act_type=act_type,
+                              norm_type=norm_type,
+                              depthwise=depthwise)
+                              )
+            else:
+                cls_feats.append(
+                    BasicConv(self.cls_head_dim, self.cls_head_dim,
+                              kernel_size=3, padding=1, stride=1, 
+                              act_type=act_type,
+                              norm_type=norm_type,
+                              depthwise=depthwise)
+                              )
+        ## reg head
+        reg_feats = []
+        self.reg_head_dim = reg_head_dim
+        for i in range(num_reg_head):
+            if i == 0:
+                reg_feats.append(
+                    BasicConv(in_dim, self.reg_head_dim,
+                              kernel_size=3, padding=1, stride=1, 
+                              act_type=act_type,
+                              norm_type=norm_type,
+                              depthwise=depthwise)
+                              )
+            else:
+                reg_feats.append(
+                    BasicConv(self.reg_head_dim, self.reg_head_dim,
+                              kernel_size=3, padding=1, stride=1, 
+                              act_type=act_type,
+                              norm_type=norm_type,
+                              depthwise=depthwise)
+                              )
+        self.cls_feats = nn.Sequential(*cls_feats)
+        self.reg_feats = nn.Sequential(*reg_feats)
+
+        self.init_weights()
+        
+    def init_weights(self):
+        """Initialize the parameters."""
+        for m in self.modules():
+            if isinstance(m, torch.nn.Conv2d):
+                # In order to be consistent with the source code,
+                # reset the Conv2d initialization parameters
+                m.reset_parameters()
+
+    def forward(self, x):
+        """
+            in_feats: (Tensor) [B, C, H, W]
+        """
+        cls_feats = self.cls_feats(x)
+        reg_feats = self.reg_feats(x)
+
+        return cls_feats, reg_feats
+    
+## Multi-level Detection Head
+class Yolov6DetHead(nn.Module):
+    def __init__(self, cfg, in_dims):
+        super().__init__()
+        ## ----------- Network Parameters -----------
+        self.multi_level_heads = nn.ModuleList(
+            [DetHead(in_dim       = in_dims[level],
+                     cls_head_dim = in_dims[level],
+                     reg_head_dim = in_dims[level],
+                     num_cls_head = cfg.num_cls_head,
+                     num_reg_head = cfg.num_reg_head,
+                     act_type     = cfg.head_act,
+                     norm_type    = cfg.head_norm,
+                     depthwise    = cfg.head_depthwise)
+                     for level in range(cfg.num_levels)
+                     ])
+        # --------- Basic Parameters ----------
+        self.in_dims = in_dims
+
+    def forward(self, feats):
+        """
+            feats: List[(Tensor)] [[B, C, H, W], ...]
+        """
+        cls_feats = []
+        reg_feats = []
+        for feat, head in zip(feats, self.multi_level_heads):
+            # ---------------- Pred ----------------
+            cls_feat, reg_feat = head(feat)
+
+            cls_feats.append(cls_feat)
+            reg_feats.append(reg_feat)
+
+        return cls_feats, reg_feats
+
+
+if __name__=='__main__':
+    import time
+    from thop import profile
+    # Model config
+    
+    # YOLOv3-Base config
+    class Yolov6BaseConfig(object):
+        def __init__(self) -> None:
+            # ---------------- Model config ----------------
+            self.out_stride = 32
+            self.max_stride = 32
+            self.num_levels = 3
+            ## Head
+            self.head_act  = 'lrelu'
+            self.head_norm = 'BN'
+            self.head_depthwise = False
+            self.head_dim  = 256
+            self.num_cls_head   = 2
+            self.num_reg_head   = 2
+
+    cfg = Yolov6BaseConfig()
+    # Build a head
+    pyramid_feats = [torch.randn(1, cfg.head_dim, 80, 80),
+                     torch.randn(1, cfg.head_dim, 40, 40),
+                     torch.randn(1, cfg.head_dim, 20, 20)]
+    head = Yolov6DetHead(cfg, [cfg.head_dim]*3)
+
+
+    # Inference
+    t0 = time.time()
+    cls_feats, reg_feats = head(pyramid_feats)
+    t1 = time.time()
+    print('Time: ', t1 - t0)
+    for cls_f, reg_f in zip(cls_feats, reg_feats):
+        print(cls_f.shape, reg_f.shape)
+
+    print('==============================')
+    flops, params = profile(head, inputs=(pyramid_feats, ), verbose=False)
+    print('==============================')
+    print('GFLOPs : {:.2f}'.format(flops / 1e9 * 2))
+    print('Params : {:.2f} M'.format(params / 1e6))  
+    

yolo/models/yolov6/yolov6_neck.py

@@ -0,0 +1,36 @@
+import torch
+import torch.nn as nn
+
+try:
+    from .yolov6_basic import BasicConv
+except:
+    from  yolov6_basic import BasicConv
+    
+
+# Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
+class SPPF(nn.Module):
+    """
+        This code referenced to https://github.com/ultralytics/yolov5
+    """
+    def __init__(self, cfg, in_dim, out_dim):
+        super().__init__()
+        ## ----------- Basic Parameters -----------
+        inter_dim = round(in_dim * cfg.neck_expand_ratio)
+        self.out_dim = out_dim
+        ## ----------- Network Parameters -----------
+        self.cv1 = BasicConv(in_dim, inter_dim,
+                             kernel_size=1, padding=0, stride=1,
+                             act_type=cfg.neck_act, norm_type=cfg.neck_norm)
+        self.cv2 = BasicConv(inter_dim * 4, out_dim,
+                             kernel_size=1, padding=0, stride=1,
+                             act_type=cfg.neck_act, norm_type=cfg.neck_norm)
+        self.m = nn.MaxPool2d(kernel_size=cfg.spp_pooling_size,
+                              stride=1,
+                              padding=cfg.spp_pooling_size // 2)
+
+    def forward(self, x):
+        x = self.cv1(x)
+        y1 = self.m(x)
+        y2 = self.m(y1)
+
+        return self.cv2(torch.cat((x, y1, y2, self.m(y2)), 1))

yolo/models/yolov6/yolov6_pafpn.py

@@ -0,0 +1,148 @@
+from typing import List
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+try:
+    from .yolov6_basic import BasicConv, RepBlock, RepCSPBlock
+except:
+    from  yolov6_basic import BasicConv, RepBlock, RepCSPBlock
+
+
+# Yolov6FPN
+class Yolov6PaFPN(nn.Module):
+    def __init__(self, cfg, in_dims: List = [256, 512, 1024]):
+        super(Yolov6PaFPN, self).__init__()
+        self.in_dims = in_dims
+        self.model_scale = cfg.scale
+        c3, c4, c5 = in_dims
+
+        # ---------------------- Yolov6's Top down FPN ----------------------
+        ## P5 -> P4
+        self.reduce_layer_1   = BasicConv(c5, round(256*cfg.width),
+                                          kernel_size=1, padding=0, stride=1,
+                                          act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+        self.top_down_layer_1 = self.make_block(in_dim     = c4 + round(256*cfg.width),
+                                                out_dim    = round(256*cfg.width),
+                                                num_blocks = round(12*cfg.depth))
+
+        ## P4 -> P3
+        self.reduce_layer_2   = BasicConv(round(256*cfg.width), round(128*cfg.width),
+                                          kernel_size=1, padding=0, stride=1,
+                                          act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+        self.top_down_layer_2 = self.make_block(in_dim     = c3 + round(128*cfg.width),
+                                                out_dim    = round(128*cfg.width),
+                                                num_blocks = round(12*cfg.depth))
+        
+        # ---------------------- Yolov6's Bottom up PAN ----------------------
+        ## P3 -> P4
+        self.downsample_layer_1 = BasicConv(round(128*cfg.width), round(128*cfg.width),
+                                            kernel_size=3, padding=1, stride=2,
+                                            act_type=cfg.fpn_act, norm_type=cfg.fpn_norm, depthwise=cfg.fpn_depthwise)
+        self.bottom_up_layer_1  = self.make_block(in_dim     = round(128*cfg.width) + round(128*cfg.width),
+                                                  out_dim    = round(256*cfg.width),
+                                                  num_blocks = round(12*cfg.depth))
+
+        ## P4 -> P5
+        self.downsample_layer_2 = BasicConv(round(256*cfg.width), round(256*cfg.width),
+                                            kernel_size=3, padding=1, stride=2,
+                                            act_type=cfg.fpn_act, norm_type=cfg.fpn_norm, depthwise=cfg.fpn_depthwise)
+        self.bottom_up_layer_2  = self.make_block(in_dim     = round(256*cfg.width) + round(256*cfg.width),
+                                                  out_dim    = round(512*cfg.width),
+                                                  num_blocks = round(12*cfg.depth))
+
+        # ---------------------- Yolov6's output projection ----------------------
+        self.out_layers = nn.ModuleList([
+            BasicConv(in_dim, in_dim, kernel_size=1,
+                      act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+                      for in_dim in [round(128*cfg.width), round(256*cfg.width), round(512*cfg.width)]
+                      ])
+        self.out_dims = [round(128*cfg.width), round(256*cfg.width), round(512*cfg.width)]
+
+    def make_block(self, in_dim, out_dim, num_blocks=1):
+        if   self.model_scale in ["n", "s"]:
+            block = RepBlock(in_channels  = in_dim,
+                             out_channels = out_dim,
+                             num_blocks   = num_blocks)
+        elif self.model_scale in ["m", "l"]:
+            block = RepCSPBlock(in_channels  = in_dim,
+                                out_channels = out_dim,
+                                num_blocks   = num_blocks,
+                                expansion    = 0.5)
+        else:
+            raise NotImplementedError("Unknown model scale: {}".format(self.model_scale))
+            
+        return block        
+    def forward(self, features):
+        c3, c4, c5 = features
+        
+        # ------------------ Top down FPN ------------------
+        ## P5 -> P4
+        p5 = self.reduce_layer_1(c5)
+        p5_up = F.interpolate(p5, scale_factor=2.0)
+        p4 = self.top_down_layer_1(torch.cat([c4, p5_up], dim=1))
+
+        ## P4 -> P3
+        p4 = self.reduce_layer_2(p4)
+        p4_up = F.interpolate(p4, scale_factor=2.0)
+        p3 = self.top_down_layer_2(torch.cat([c3, p4_up], dim=1))
+
+        # ------------------ Bottom up PAN ------------------
+        ## P3 -> P4
+        p3_ds = self.downsample_layer_1(p3)
+        p4 = self.bottom_up_layer_1(torch.cat([p4, p3_ds], dim=1))
+
+        ## P4 -> P5
+        p4_ds = self.downsample_layer_2(p4)
+        p5 = self.bottom_up_layer_2(torch.cat([p5, p4_ds], dim=1))
+
+        out_feats = [p3, p4, p5]
+            
+        # output proj layers
+        out_feats_proj = []
+        for feat, layer in zip(out_feats, self.out_layers):
+            out_feats_proj.append(layer(feat))
+            
+        return out_feats_proj
+
+
+if __name__=='__main__':
+    import time
+    from thop import profile
+    # Model config
+    
+    # YOLOv2-Base config
+    class Yolov3BaseConfig(object):
+        def __init__(self) -> None:
+            # ---------------- Model config ----------------
+            self.width    = 0.50
+            self.depth    = 0.34
+            self.out_stride = [8, 16, 32]
+            self.max_stride = 32
+            self.num_levels = 3
+            ## FPN
+            self.fpn_act  = 'silu'
+            self.fpn_norm = 'BN'
+            self.fpn_depthwise = False
+
+    cfg = Yolov3BaseConfig()
+    # Build a head
+    in_dims  = [128, 256, 512]
+    fpn = Yolov6PaFPN(cfg, in_dims)
+
+    # Inference
+    x = [torch.randn(1, in_dims[0], 80, 80),
+         torch.randn(1, in_dims[1], 40, 40),
+         torch.randn(1, in_dims[2], 20, 20)]
+    t0 = time.time()
+    output = fpn(x)
+    t1 = time.time()
+    print('Time: ', t1 - t0)
+    print('====== FPN output ====== ')
+    for level, feat in enumerate(output):
+        print("- Level-{} : ".format(level), feat.shape)
+
+    flops, params = profile(fpn, inputs=(x, ), verbose=False)
+    print('==============================')
+    print('GFLOPs : {:.2f}'.format(flops / 1e9 * 2))
+    print('Params : {:.2f} M'.format(params / 1e6))

yolo/models/yolov6/yolov6_pred.py

@@ -0,0 +1,135 @@
+import torch
+import torch.nn as nn
+
+
+# -------------------- Detection Pred Layer --------------------
+## Single-level pred layer
+class DetPredLayer(nn.Module):
+    def __init__(self,
+                 cls_dim      :int,
+                 reg_dim      :int,
+                 stride       :int,
+                 num_classes  :int,
+                 ):
+        super().__init__()
+        # --------- Basic Parameters ----------
+        self.stride  = stride
+        self.cls_dim = cls_dim
+        self.reg_dim = reg_dim
+        self.num_classes = num_classes
+
+        # --------- Network Parameters ----------
+        self.cls_pred = nn.Conv2d(self.cls_dim, num_classes, kernel_size=1)
+        self.reg_pred = nn.Conv2d(self.reg_dim, 4, kernel_size=1)                
+
+        self.init_bias()
+        
+    def init_bias(self):
+        # Init bias
+        init_prob = 0.01
+        bias_value = -torch.log(torch.tensor((1. - init_prob) / init_prob))
+        # cls pred
+        b = self.cls_pred.bias.view(1, -1)
+        b.data.fill_(bias_value.item())
+        self.cls_pred.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+        # reg pred
+        b = self.reg_pred.bias.view(-1, )
+        b.data.fill_(1.0)
+        self.reg_pred.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+        w = self.reg_pred.weight
+        w.data.fill_(0.)
+        self.reg_pred.weight = torch.nn.Parameter(w, requires_grad=True)
+
+    def generate_anchors(self, fmp_size):
+        """
+            fmp_size: (List) [H, W]
+        """
+        fmp_h, fmp_w = fmp_size
+        anchor_y, anchor_x = torch.meshgrid([torch.arange(fmp_h), torch.arange(fmp_w)])
+
+        # [H, W, 2] -> [HW, 2]
+        anchors = torch.stack([anchor_x, anchor_y], dim=-1).float().view(-1, 2)
+        anchors = anchors + 0.5
+        anchors = anchors * self.stride
+
+        return anchors
+        
+    def forward(self, cls_feat, reg_feat):
+        # 预测层
+        cls_pred = self.cls_pred(cls_feat)
+        reg_pred = self.reg_pred(reg_feat)
+
+        # 生成网格坐标
+        B, _, H, W = cls_pred.size()
+        fmp_size = [H, W]
+        anchors = self.generate_anchors(fmp_size)
+        anchors = anchors.to(cls_pred.device)
+
+        # 对 pred 的size做一些view调整，便于后续的处理
+        # [B, C, H, W] -> [B, H, W, C] -> [B, H*W, C]
+        cls_pred = cls_pred.permute(0, 2, 3, 1).contiguous().view(B, -1, self.num_classes)
+        reg_pred = reg_pred.permute(0, 2, 3, 1).contiguous().view(B, -1, 4)
+        
+        # 解算边界框坐标
+        cxcy_pred = reg_pred[..., :2] * self.stride + anchors
+        bwbh_pred = torch.exp(reg_pred[..., 2:]) * self.stride
+        pred_x1y1 = cxcy_pred - bwbh_pred * 0.5
+        pred_x2y2 = cxcy_pred + bwbh_pred * 0.5
+        box_pred = torch.cat([pred_x1y1, pred_x2y2], dim=-1)
+
+        # output dict
+        outputs = {"pred_cls": cls_pred,       # (torch.Tensor) [B, M, C]
+                   "pred_reg": reg_pred,       # (torch.Tensor) [B, M, 4]
+                   "pred_box": box_pred,       # (torch.Tensor) [B, M, 4]
+                   "anchors" : anchors,        # (torch.Tensor) [M, 2]
+                   "fmp_size": fmp_size,
+                   "stride"  : self.stride,    # (Int)
+                   }
+
+        return outputs
+
+## Multi-level pred layer
+class Yolov6DetPredLayer(nn.Module):
+    def __init__(self, cfg, in_dims):
+        super().__init__()
+        # --------- Basic Parameters ----------
+        self.cfg = cfg
+
+        # ----------- Network Parameters -----------
+        ## pred layers
+        self.multi_level_preds = nn.ModuleList(
+            [DetPredLayer(cls_dim      = in_dims[level],
+                          reg_dim      = in_dims[level],
+                          stride       = cfg.out_stride[level],
+                          num_classes  = cfg.num_classes,)
+                          for level in range(cfg.num_levels)
+                          ])
+
+    def forward(self, cls_feats, reg_feats):
+        all_anchors = []
+        all_fmp_sizes = []
+        all_cls_preds = []
+        all_reg_preds = []
+        all_box_preds = []
+        for level in range(self.cfg.num_levels):
+            # -------------- Single-level prediction --------------
+            outputs = self.multi_level_preds[level](cls_feats[level], reg_feats[level])
+
+            # collect results
+            all_cls_preds.append(outputs["pred_cls"])
+            all_reg_preds.append(outputs["pred_reg"])
+            all_box_preds.append(outputs["pred_box"])
+            all_fmp_sizes.append(outputs["fmp_size"])
+            all_anchors.append(outputs["anchors"])
+        
+        # output dict
+        outputs = {"pred_cls":  all_cls_preds,         # List(Tensor) [B, M, C]
+                   "pred_reg":  all_reg_preds,         # List(Tensor) [B, M, 4*(reg_max)]
+                   "pred_box":  all_box_preds,         # List(Tensor) [B, M, 4]
+                   "fmp_sizes": all_fmp_sizes,         # List(Tensor) [M, 1]
+                   "anchors":   all_anchors,           # List(Tensor) [M, 2]
+                   "strides":   self.cfg.out_stride,   # List(Int) = [8, 16, 32]
+                   }
+
+        return outputs
+