add yolov7_af & remove yolov6 as it is ugly

yolo/config/__init__.py

@@ -4,7 +4,6 @@ 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 .yolov7_af_config import build_yolov7af_config
 from .yolov8_config    import build_yolov8_config
 from .gelan_config     import build_gelan_config
@@ -24,8 +23,6 @@ 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 'yolov7_af' in args.model:
         cfg = build_yolov7af_config(args)
     elif 'yolov8' in args.model:

yolo/config/yolov6_config.py

@@ -1,173 +0,0 @@
-# yolo Config
-
-
-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"
-        ## 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 = 2.0
-
-        # ---------------- 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   = -1.
-        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  = 0.0
-        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"
-
-        # ---------------- 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"
-
-        # ---------------- 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
-
-        # ---------------- 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
-
-        # ---------------- Data process config ----------------
-        self.mosaic_prob = 1.0
-        self.mixup_prob  = 0.1
-        self.copy_paste  = 0.5

yolo/config/yolov7_af_config.py

@@ -124,7 +124,7 @@ class Yolov7AFTConfig(Yolov7AFBaseConfig):
         # ---------------- Model config ----------------
         self.width = 0.50
         self.scale = "t"
-        self.fpn_expansions = [0.5, 1.0]
+        self.fpn_expansions = [0.5, 0.5]
         self.fpn_block_bw = 2
         self.fpn_block_dw = 1
 

yolo/models/__init__.py

@@ -7,7 +7,6 @@ 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 .yolov7_af.build import build_yolov7af
 from .yolov8.build    import build_yolov8
 from .gelan.build     import build_gelan
@@ -31,9 +30,6 @@ 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)
     ## Anchor-free YOLOv7
     elif 'yolov7_af' in args.model:
         model, criterion = build_yolov7af(cfg, is_val)

yolo/models/yolov6/README.md

@@ -1,61 +0,0 @@
-# Redesigned YOLOv6:
-
-- VOC
-
-|   Model  | Batch | Scale | AP<sup>val<br>0.5 | Weight |  Logs  |
-|----------|-------|-------|-------------------|--------|--------|
-| YOLOv6-S | 1xb16 |  640  |               | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v6/releases/download/yolo_tutorial_ckpt/yolov6_s_voc.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v6/releases/download/yolo_tutorial_ckpt/YOLOv6-S-VOC.txt) |
-
-- COCO
-
-|  Model   | Batch | Scale | AP<sup>val<br>0.5:0.95 | AP<sup>val<br>0.5 | FLOPs<br><sup>(G) | Params<br><sup>(M) | Weight |  Logs  |
-|----------|-------|-------|------------------------|-------------------|-------------------|--------------------|--------|--------|
-| YOLOv6-S | 1xb16 |  640  |                    |               |   27.3            |   9.0             |  |  |
-
-- For training, we train redesigned YOLOv6 with 300 epochs on COCO. We also use the gradient accumulation.
-- For data augmentation, we use the RandomAffine, RandomHSV, Mosaic and Mixup augmentation.
-- For optimizer, we use AdamW with weight decay of 0.05 and per image base lr of 0.001 / 64.
-- For learning rate scheduler, we use cosine decay scheduler.
-- For batch size, we set it to 16, and we also use the gradient accumulation to approximate batch size of 256.
-
-
-## Train YOLOv6
-### Single GPU
-Taking training YOLOv6-S on COCO as the example,
-```Shell
-python train.py --cuda -d coco --root path/to/coco -m yolov6_s -bs 16 --fp16 
-```
-
-### Multi GPU
-Taking training YOLOv6-S on COCO as the example,
-```Shell
-python -m torch.distributed.run --nproc_per_node=8 train.py --cuda --distributed -d coco --root path/to/coco -m yolov6_s -bs 16 --fp16 
-```
-
-## Test YOLOv6
-Taking testing YOLOv6-S on COCO-val as the example,
-```Shell
-python test.py --cuda -d coco --root path/to/coco -m yolov6_s --weight path/to/yolov6.pth --show 
-```
-
-## Evaluate YOLOv6
-Taking evaluating YOLOv6-S on COCO-val as the example,
-```Shell
-python eval.py --cuda -d coco --root path/to/coco -m yolov6_s --weight path/to/yolov6.pth 
-```
-
-## Demo
-### Detect with Image
-```Shell
-python demo.py --mode image --path_to_img path/to/image_dirs/ --cuda -m yolov6_s --weight path/to/weight --show
-```
-
-### Detect with Video
-```Shell
-python demo.py --mode video --path_to_vid path/to/video --cuda -m yolov6_s --weight path/to/weight --show --gif
-```
-
-### Detect with Camera
-```Shell
-python demo.py --mode camera --cuda -m yolov6_s --weight path/to/weight --show --gif
-```

yolo/models/yolov6/build.py

@@ -1,24 +0,0 @@
-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

@@ -1,145 +0,0 @@
-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

@@ -1,199 +0,0 @@
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-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

@@ -1,157 +0,0 @@
-# --------------- 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

@@ -1,113 +0,0 @@
-import torch
-import torch.nn as nn
-
-try:
-    from .yolov6_basic import RepBlock, RepVGGBlock, RepCSPBlock
-except:
-    from  yolov6_basic import RepBlock, RepVGGBlock, RepCSPBlock
-
-
-# --------------------- 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()
-    
-    def make_block(self, in_dim, out_dim, num_blocks=1):
-        if self.model_scale in ["s", "t", "n"]:
-            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 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):
-        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

@@ -1,234 +0,0 @@
-import torch
-import torch.nn as nn
-from typing import List
-import numpy as np
-
-# --------------------- 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.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

@@ -1,168 +0,0 @@
-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

@@ -1,33 +0,0 @@
-import torch
-import torch.nn as nn
-
-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

@@ -1,148 +0,0 @@
-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 ["s", "t", "n"]:
-            block = RepBlock(in_channels  = in_dim,
-                             out_channels = out_dim,
-                             num_blocks   = num_blocks)
-        elif self.model_scale in ["m", "l", "x"]:
-            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

@@ -1,131 +0,0 @@
-import torch
-import torch.nn as nn
-from typing import List
-
-# -------------------- 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)
-
-    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