YOLOX2 is a test model

yolo/config/__init__.py

@@ -9,6 +9,8 @@ from .yolov8_config    import build_yolov8_config
 from .gelan_config     import build_gelan_config
 from .rtdetr_config    import build_rtdetr_config
 
+from .yolox2_config    import build_yolox2_config
+
 def build_config(args):
     print('==============================')
     print('Model: {} ...'.format(args.model.upper()))
@@ -32,6 +34,10 @@ def build_config(args):
     # ----------- RT-DETR -----------
     elif 'rtdetr' in args.model:
         cfg = build_rtdetr_config(args)
+
+    elif 'yolox2' in args.model:
+        cfg = build_yolox2_config(args)
+
     else:
         raise NotImplementedError("Unknown model config: {}".format(args.model))
     

yolo/config/yolox2_config.py

@@ -0,0 +1,191 @@
+# yolo Config
+
+
+def build_yolox2_config(args):
+    if   args.model == 'yolox2_n':
+        return Yolox2NConfig()
+    elif args.model == 'yolox2_s':
+        return Yolox2SConfig()
+    elif args.model == 'yolox2_m':
+        return Yolox2MConfig()
+    elif args.model == 'yolox2_l':
+        return Yolox2LConfig()
+    elif args.model == 'yolox2_x':
+        return Yolox2XConfig()
+    else:
+        raise NotImplementedError("No config for model: {}".format(args.model))
+    
+# YOLOx2-Base config
+class Yolox2BaseConfig(object):
+    def __init__(self) -> None:
+        # ---------------- Model config ----------------
+        self.width    = 1.0
+        self.depth    = 1.0
+        self.out_stride = [8, 16, 32]
+        self.max_stride = 32
+        self.num_levels = 3
+        self.scale      = "b"
+        ## Backbone
+        self.bk_act   = 'silu'
+        self.bk_norm  = 'BN'
+        self.bk_depthwise = False
+        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.head_dim       = 256
+        self.num_cls_head   = 2
+        self.num_reg_head   = 2
+
+        # ---------------- 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 = 0.3
+        self.test_nms_thresh  = 0.5
+
+        # ---------------- Assignment config ----------------
+        ## Matcher
+        self.ota_soft_center_radius = 3.0
+        self.ota_topk_candidates = 13
+        ## 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   = 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  = 0.0
+        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))
+
+# YOLOx2-N
+class Yolox2NConfig(Yolox2BaseConfig):
+    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
+
+# YOLOx2-S
+class Yolox2SConfig(Yolox2BaseConfig):
+    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
+
+# YOLOx2-M
+class Yolox2MConfig(Yolox2BaseConfig):
+    def __init__(self) -> None:
+        super().__init__()
+        # ---------------- Model config ----------------
+        self.width = 0.75
+        self.depth = 0.67
+        self.scale = "m"
+
+        # ---------------- Data process config ----------------
+        self.mosaic_prob = 1.0
+        self.mixup_prob  = 0.1
+        self.copy_paste  = 0.5
+
+# YOLOx2-L
+class Yolox2LConfig(Yolox2BaseConfig):
+    def __init__(self) -> None:
+        super().__init__()
+        # ---------------- Model config ----------------
+        self.width = 1.0
+        self.depth = 1.0
+        self.scale = "l"
+
+        # ---------------- Data process config ----------------
+        self.mosaic_prob = 1.0
+        self.mixup_prob  = 0.1
+        self.copy_paste  = 0.5
+
+# YOLOx2-X
+class Yolox2XConfig(Yolox2BaseConfig):
+    def __init__(self) -> None:
+        super().__init__()
+        # ---------------- Model config ----------------
+        self.width = 1.25
+        self.depth = 1.34
+        self.scale = "x"
+
+        # ---------------- Data process config ----------------
+        self.mosaic_prob = 1.0
+        self.mixup_prob  = 0.1
+        self.copy_paste  = 0.5

yolo/engine.py

@@ -62,7 +62,7 @@ class YoloTrainer(object):
         self.scaler = torch.cuda.amp.GradScaler(enabled=args.fp16)
 
         # ---------------------------- Build Optimizer ----------------------------
-        self.grad_accumulate = max(64 // args.batch_size, 1)
+        self.grad_accumulate = max(256 // args.batch_size, 1)
         cfg.base_lr = cfg.per_image_lr * args.batch_size * self.grad_accumulate
         cfg.min_lr  = cfg.base_lr * cfg.min_lr_ratio
         self.optimizer, self.start_epoch = build_yolo_optimizer(cfg, model, args.resume)

yolo/models/__init__.py

@@ -12,6 +12,8 @@ from .yolov8.build    import build_yolov8
 from .gelan.build     import build_gelan
 from .rtdetr.build    import build_rtdetr
 
+from .yolox2.build import build_yolox2
+
 # build object detector
 def build_model(args, cfg, is_val=False):
     # ------------ build object detector ------------
@@ -43,6 +45,9 @@ def build_model(args, cfg, is_val=False):
     elif 'rtdetr' in args.model:
         model, criterion = build_rtdetr(cfg, is_val)
 
+    elif 'yolox2' in args.model:
+        model, criterion = build_yolox2(cfg, is_val)
+
     if is_val:
         # ------------ Load pretrained weight ------------
         if args.pretrained is not None:

yolo/models/yolox2/README.md

@@ -0,0 +1,61 @@
+# Anchor-free YOLOv5:
+
+- VOC
+
+|     Model   | Batch | Scale | AP<sup>val<br>0.5 | Weight |  Logs  |
+|-------------|-------|-------|-------------------|--------|--------|
+| YOLOv5-AF-S | 1xb16 |  640  |       82.4        | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v5/releases/download/yolo_tutorial_ckpt/yolov5_af_s_voc.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v5/releases/download/yolo_tutorial_ckpt/YOLOv5-AF-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  |
+|-------------|-------|-------|------------------------|-------------------|-------------------|--------------------|--------|--------|
+| YOLOv5-AF-S | 1xb16 |  640  |                    |               |   26.9            |   8.9             |  |  |
+
+- For training, we train redesigned YOLOv5-AF with 300 epochs on COCO. We also use the gradient accumulation.
+- For data augmentation, we use the RandomAffine, RandomHSV, Mosaic and YOLOX's 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 YOLOv5-AF
+### Single GPU
+Taking training YOLOv5-AF-S on COCO as the example,
+```Shell
+python train.py --cuda -d coco --root path/to/coco -m yolov5_af_s -bs 16 --fp16 
+```
+
+### Multi GPU
+Taking training YOLOv5-AF-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 yolov5_af_s -bs 16 --fp16 
+```
+
+## Test YOLOv5-AF
+Taking testing YOLOv5-AF-S on COCO-val as the example,
+```Shell
+python test.py --cuda -d coco --root path/to/coco -m yolov5_af_s --weight path/to/yolov5.pth --show 
+```
+
+## Evaluate YOLOv5-AF
+Taking evaluating YOLOv5-AF-S on COCO-val as the example,
+```Shell
+python eval.py --cuda -d coco --root path/to/coco -m yolov5_af_s --weight path/to/yolov5.pth 
+```
+
+## Demo
+### Detect with Image
+```Shell
+python demo.py --mode image --path_to_img path/to/image_dirs/ --cuda -m yolov5_af_s --weight path/to/weight --show
+```
+
+### Detect with Video
+```Shell
+python demo.py --mode video --path_to_vid path/to/video --cuda -m yolov5_af_s --weight path/to/weight --show --gif
+```
+
+### Detect with Camera
+```Shell
+python demo.py --mode camera --cuda -m yolov5_af_s --weight path/to/weight --show --gif
+```

yolo/models/yolox2/build.py

@@ -0,0 +1,24 @@
+import torch.nn as nn
+
+from .loss import SetCriterion
+from .yolox2 import Yolox2
+
+
+# build object detector
+def build_yolox2(cfg, is_val=False):
+    # -------------- Build YOLO --------------
+    model = Yolox2(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/yolox2/loss.py

@@ -0,0 +1,128 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from utils.box_ops import get_ious, bbox2dist
+from utils.distributed_utils import get_world_size, is_dist_avail_and_initialized
+
+from .matcher import AlignedSimOTA
+
+
+class SetCriterion(object):
+    def __init__(self, cfg):
+        self.cfg = cfg
+        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 = AlignedSimOTA(soft_center_radius = cfg.ota_soft_center_radius,
+                                     topk_candidates    = cfg.ota_topk_candidates,
+                                     num_classes        = cfg.num_classes,
+                                     )
+
+    def loss_classes(self, pred_cls, target, beta=2.0):
+        # Quality FocalLoss
+        """
+            pred_cls: (torch.Tensor): [N, C]。
+            target:   (tuple([torch.Tensor], [torch.Tensor])): label -> (N,), score -> (N)
+        """
+        label, score = target
+        pred_sigmoid = pred_cls.sigmoid()
+        scale_factor = pred_sigmoid
+        zerolabel = scale_factor.new_zeros(pred_cls.shape)
+
+        ce_loss = F.binary_cross_entropy_with_logits(
+            pred_cls, zerolabel, reduction='none') * scale_factor.pow(beta)
+        
+        bg_class_ind = pred_cls.shape[-1]
+        pos = ((label >= 0) & (label < bg_class_ind)).nonzero().squeeze(1)
+        if pos.shape[0] > 0:
+            pos_label = label[pos].long()
+
+            scale_factor = score[pos] - pred_sigmoid[pos, pos_label]
+
+            ce_loss[pos, pos_label] = F.binary_cross_entropy_with_logits(
+                pred_cls[pos, pos_label], score[pos],
+                reduction='none') * scale_factor.abs().pow(beta)
+
+        return ce_loss
+    
+    def loss_bboxes(self, pred_box, gt_box):
+        ious = get_ious(pred_box, gt_box, box_mode="xyxy", iou_type='giou')
+        loss_box = 1.0 - ious
+
+        return loss_box
+
+    def __call__(self, outputs, targets):        
+        """
+            outputs['pred_cls']: List(Tensor) [B, M, C]
+            outputs['pred_box']: List(Tensor) [B, M, 4]
+            outputs['pred_box']: List(Tensor) [B, M, 4]
+            outputs['strides']: List(Int) [8, 16, 32] output stride
+            targets: (List) [dict{'boxes': [...], 
+                                 'labels': [...], 
+                                 'orig_size': ...}, ...]
+        """
+        bs          = outputs['pred_cls'][0].shape[0]
+        device      = outputs['pred_cls'][0].device
+        fpn_strides = outputs['strides']
+        anchors     = outputs['anchors']
+        # preds: [B, M, C]
+        cls_preds = torch.cat(outputs['pred_cls'], dim=1)
+        box_preds = torch.cat(outputs['pred_box'], dim=1)
+        
+        # --------------- label assignment ---------------
+        cls_targets = []
+        box_targets = []
+        assign_metrics = []
+        for batch_idx in range(bs):
+            tgt_labels = targets[batch_idx]["labels"].to(device)  # [N,]
+            tgt_bboxes = targets[batch_idx]["boxes"].to(device)   # [N, 4]
+            assigned_result = self.matcher(fpn_strides=fpn_strides,
+                                           anchors=anchors,
+                                           pred_cls=cls_preds[batch_idx].detach(),
+                                           pred_box=box_preds[batch_idx].detach(),
+                                           gt_labels=tgt_labels,
+                                           gt_bboxes=tgt_bboxes
+                                           )
+            cls_targets.append(assigned_result['assigned_labels'])
+            box_targets.append(assigned_result['assigned_bboxes'])
+            assign_metrics.append(assigned_result['assign_metrics'])
+
+        # List[B, M, C] -> Tensor[BM, C]
+        cls_targets = torch.cat(cls_targets, dim=0)
+        box_targets = torch.cat(box_targets, dim=0)
+        assign_metrics = torch.cat(assign_metrics, dim=0)
+
+        # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+        bg_class_ind = self.num_classes
+        pos_inds = ((cls_targets >= 0) & (cls_targets < bg_class_ind)).nonzero().squeeze(1)
+        num_fgs = assign_metrics.sum()
+
+        if is_dist_avail_and_initialized():
+            torch.distributed.all_reduce(num_fgs)
+        num_fgs = (num_fgs / get_world_size()).clamp(1.0).item()
+
+        # ------------------ Classification loss ------------------
+        cls_preds = cls_preds.view(-1, self.num_classes)
+        loss_cls = self.loss_classes(cls_preds, (cls_targets, assign_metrics))
+        loss_cls = loss_cls.sum() / num_fgs
+
+        # ------------------ Regression loss ------------------
+        box_preds_pos = box_preds.view(-1, 4)[pos_inds]
+        box_targets_pos = box_targets[pos_inds]
+        loss_box = self.loss_bboxes(box_preds_pos, box_targets_pos)
+        loss_box = loss_box.sum() / num_fgs
+
+        # total loss
+        losses = self.loss_cls_weight * loss_cls + \
+                 self.loss_box_weight * loss_box
+        loss_dict = dict(
+                loss_cls = loss_cls,
+                loss_box = loss_box,
+                losses = losses
+        )
+
+        return loss_dict
+    

yolo/models/yolox2/matcher.py

@@ -0,0 +1,160 @@
+# ------------------------------------------------------------------------------------------
+# This code referenced to https://github.com/open-mmlab/mmyolo/models/task_modules/assigners/batch_dsl_assigner.py
+# ------------------------------------------------------------------------------------------
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from utils.box_ops import box_iou
+
+
+# -------------------------- Aligned SimOTA assigner --------------------------
+class AlignedSimOTA(object):
+    def __init__(self, num_classes, soft_center_radius=3.0, topk_candidates=13):
+        self.num_classes = num_classes
+        self.soft_center_radius = soft_center_radius
+        self.topk_candidates = topk_candidates
+
+    @torch.no_grad()
+    def __call__(self, 
+                 fpn_strides, 
+                 anchors, 
+                 pred_cls, 
+                 pred_box, 
+                 gt_labels,
+                 gt_bboxes):
+        # [M,]
+        strides = torch.cat([torch.ones_like(anchor_i[:, 0]) * stride_i
+                                for stride_i, anchor_i in zip(fpn_strides, anchors)], dim=-1)
+        # List[F, M, 2] -> [M, 2]
+        num_gt = len(gt_labels)
+        anchors = torch.cat(anchors, dim=0)
+
+        # check gt
+        if num_gt == 0 or gt_bboxes.max().item() == 0.:
+            return {
+                'assigned_labels': gt_labels.new_full(pred_cls[..., 0].shape, self.num_classes).long(),
+                'assigned_bboxes': gt_bboxes.new_full(pred_box.shape, 0).float(),
+                'assign_metrics':  gt_bboxes.new_full(pred_cls[..., 0].shape, 0).float(),
+            }
+        
+        # get inside points: [N, M]
+        is_in_gt = self.find_inside_points(gt_bboxes, anchors)
+        valid_mask = is_in_gt.sum(dim=0) > 0  # [M,]
+
+        # ----------------------------------- soft center prior -----------------------------------
+        gt_center = (gt_bboxes[..., :2] + gt_bboxes[..., 2:]) / 2.0
+        distance = (anchors.unsqueeze(0) - gt_center.unsqueeze(1)
+                    ).pow(2).sum(-1).sqrt() / strides.unsqueeze(0)  # [N, M]
+        distance = distance * valid_mask.unsqueeze(0)
+        soft_center_prior = torch.pow(10, distance - self.soft_center_radius)
+
+        # ----------------------------------- regression cost -----------------------------------
+        pair_wise_ious, _ = box_iou(gt_bboxes, pred_box)  # [N, M]
+        pair_wise_ious_loss = -torch.log(pair_wise_ious + 1e-8) * 3.0
+
+        # ----------------------------------- classification cost -----------------------------------
+        ## select the predicted scores corresponded to the gt_labels
+        pairwise_pred_scores = pred_cls.permute(1, 0)  # [M, C] -> [C, M]
+        pairwise_pred_scores = pairwise_pred_scores[gt_labels.long(), :].float()   # [N, M]
+        ## scale factor
+        scale_factor = (pair_wise_ious - pairwise_pred_scores.sigmoid()).abs().pow(2.0)
+        ## cls cost
+        pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
+            pairwise_pred_scores, pair_wise_ious,
+            reduction="none") * scale_factor # [N, M]
+            
+        del pairwise_pred_scores
+
+        ## foreground cost matrix
+        cost_matrix = pair_wise_cls_loss + pair_wise_ious_loss + soft_center_prior
+        max_pad_value = torch.ones_like(cost_matrix) * 1e9
+        cost_matrix = torch.where(valid_mask[None].repeat(num_gt, 1),   # [N, M]
+                                  cost_matrix, max_pad_value)
+
+        # ----------------------------------- dynamic label assignment -----------------------------------
+        matched_pred_ious, matched_gt_inds, fg_mask_inboxes = self.dynamic_k_matching(
+            cost_matrix, pair_wise_ious, num_gt)
+        del pair_wise_cls_loss, cost_matrix, pair_wise_ious, pair_wise_ious_loss
+
+        # -----------------------------------process assigned labels -----------------------------------
+        assigned_labels = gt_labels.new_full(pred_cls[..., 0].shape,
+                                             self.num_classes)  # [M,]
+        assigned_labels[fg_mask_inboxes] = gt_labels[matched_gt_inds].squeeze(-1)
+        assigned_labels = assigned_labels.long()  # [M,]
+
+        assigned_bboxes = gt_bboxes.new_full(pred_box.shape, 0)        # [M, 4]
+        assigned_bboxes[fg_mask_inboxes] = gt_bboxes[matched_gt_inds]  # [M, 4]
+
+        assign_metrics = gt_bboxes.new_full(pred_cls[..., 0].shape, 0) # [M, 4]
+        assign_metrics[fg_mask_inboxes] = matched_pred_ious            # [M, 4]
+
+        assigned_dict = dict(
+            assigned_labels=assigned_labels,
+            assigned_bboxes=assigned_bboxes,
+            assign_metrics=assign_metrics
+            )
+        
+        return assigned_dict
+
+    def find_inside_points(self, gt_bboxes, anchors):
+        """
+            gt_bboxes: Tensor -> [N, 2]
+            anchors:   Tensor -> [M, 2]
+        """
+        num_anchors = anchors.shape[0]
+        num_gt = gt_bboxes.shape[0]
+
+        anchors_expand = anchors.unsqueeze(0).repeat(num_gt, 1, 1)           # [N, M, 2]
+        gt_bboxes_expand = gt_bboxes.unsqueeze(1).repeat(1, num_anchors, 1)  # [N, M, 4]
+
+        # offset
+        lt = anchors_expand - gt_bboxes_expand[..., :2]
+        rb = gt_bboxes_expand[..., 2:] - anchors_expand
+        bbox_deltas = torch.cat([lt, rb], dim=-1)
+
+        is_in_gts = bbox_deltas.min(dim=-1).values > 0
+
+        return is_in_gts
+    
+    def dynamic_k_matching(self, cost_matrix, pairwise_ious, num_gt):
+        """Use IoU and matching cost to calculate the dynamic top-k positive
+        targets.
+
+        Args:
+            cost_matrix (Tensor): Cost matrix.
+            pairwise_ious (Tensor): Pairwise iou matrix.
+            num_gt (int): Number of gt.
+            valid_mask (Tensor): Mask for valid bboxes.
+        Returns:
+            tuple: matched ious and gt indexes.
+        """
+        matching_matrix = torch.zeros_like(cost_matrix, dtype=torch.uint8)
+        # select candidate topk ious for dynamic-k calculation
+        candidate_topk = min(self.topk_candidates, pairwise_ious.size(1))
+        topk_ious, _ = torch.topk(pairwise_ious, candidate_topk, dim=1)
+        # calculate dynamic k for each gt
+        dynamic_ks = torch.clamp(topk_ious.sum(1).int(), min=1)
+
+        # sorting the batch cost matirx is faster than topk
+        _, sorted_indices = torch.sort(cost_matrix, dim=1)
+        for gt_idx in range(num_gt):
+            topk_ids = sorted_indices[gt_idx, :dynamic_ks[gt_idx]]
+            matching_matrix[gt_idx, :][topk_ids] = 1
+
+        del topk_ious, dynamic_ks, topk_ids
+
+        prior_match_gt_mask = matching_matrix.sum(0) > 1
+        if prior_match_gt_mask.sum() > 0:
+            cost_min, cost_argmin = torch.min(
+                cost_matrix[:, prior_match_gt_mask], dim=0)
+            matching_matrix[:, prior_match_gt_mask] *= 0
+            matching_matrix[cost_argmin, prior_match_gt_mask] = 1
+
+        # get foreground mask inside box and center prior
+        fg_mask_inboxes = matching_matrix.sum(0) > 0
+        matched_pred_ious = (matching_matrix *
+                             pairwise_ious).sum(0)[fg_mask_inboxes]
+        matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)
+
+        return matched_pred_ious, matched_gt_inds, fg_mask_inboxes
+    

yolo/models/yolox2/yolox2.py

@@ -0,0 +1,152 @@
+# --------------- Torch components ---------------
+import torch
+import torch.nn as nn
+
+# --------------- Model components ---------------
+from .yolox2_backbone import Yolov5Backbone
+from .yolox2_neck     import SPPF
+from .yolox2_pafpn    import Yolov5PaFPN
+from .yolox2_head     import Yolov5DetHead
+from .yolox2_pred     import Yolov5AFDetPredLayer
+
+# --------------- External components ---------------
+from utils.misc import multiclass_nms
+
+
+# Yolov5AF
+class Yolox2(nn.Module):
+    def __init__(self,
+                 cfg,
+                 is_val = False,
+                 ) -> None:
+        super(Yolox2, 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 = Yolov5Backbone(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      = Yolov5PaFPN(cfg, self.pyramid_feat_dims)
+        ## Head
+        self.head     = Yolov5DetHead(cfg, self.fpn.out_dims)
+        ## Pred
+        self.pred     = Yolov5AFDetPredLayer(cfg)
+
+    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/yolox2/yolox2_backbone.py

@@ -0,0 +1,172 @@
+import torch
+import torch.nn as nn
+
+try:
+    from .yolox2_basic import BasicConv, CSPBlock
+except:
+    from  yolox2_basic import BasicConv, CSPBlock
+
+# IN1K pretrained weight
+pretrained_urls = {
+    'n': None,
+    's': None,
+    'm': None,
+    'l': None,
+    'x': None,
+}
+
+# --------------------- Yolov3's Backbone -----------------------
+## Modified DarkNet
+class Yolov5Backbone(nn.Module):
+    def __init__(self, cfg):
+        super(Yolov5Backbone, self).__init__()
+        # ------------------ Basic setting ------------------
+        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 = BasicConv(3, self.feat_dims[0],
+                                 kernel_size=6, padding=2, stride=2,
+                                 act_type=cfg.bk_act, norm_type=cfg.bk_norm, depthwise=cfg.bk_depthwise)
+        # P2/4
+        self.layer_2 = nn.Sequential(
+            BasicConv(self.feat_dims[0], self.feat_dims[1],
+                      kernel_size=3, padding=1, stride=2,
+                      act_type=cfg.bk_act, norm_type=cfg.bk_norm, depthwise=cfg.bk_depthwise),
+            CSPBlock(in_dim     = self.feat_dims[1],
+                     out_dim    = self.feat_dims[1],
+                     num_blocks = round(3*cfg.depth),
+                     expansion  = 0.5,
+                     shortcut   = True,
+                     act_type   = cfg.bk_act,
+                     norm_type  = cfg.bk_norm,
+                     depthwise  = cfg.bk_depthwise)
+        )
+        # P3/8
+        self.layer_3 = nn.Sequential(
+            BasicConv(self.feat_dims[1], self.feat_dims[2],
+                      kernel_size=3, padding=1, stride=2,
+                      act_type=cfg.bk_act, norm_type=cfg.bk_norm, depthwise=cfg.bk_depthwise),
+            CSPBlock(in_dim     = self.feat_dims[2],
+                     out_dim    = self.feat_dims[2],
+                     num_blocks = round(9*cfg.depth),
+                     expansion  = 0.5,
+                     shortcut   = True,
+                     act_type   = cfg.bk_act,
+                     norm_type  = cfg.bk_norm,
+                     depthwise  = cfg.bk_depthwise)
+        )
+        # P4/16
+        self.layer_4 = nn.Sequential(
+            BasicConv(self.feat_dims[2], self.feat_dims[3],
+                      kernel_size=3, padding=1, stride=2,
+                      act_type=cfg.bk_act, norm_type=cfg.bk_norm, depthwise=cfg.bk_depthwise),
+            CSPBlock(in_dim     = self.feat_dims[3],
+                     out_dim    = self.feat_dims[3],
+                     num_blocks = round(9*cfg.depth),
+                     expansion  = 0.5,
+                     shortcut   = True,
+                     act_type   = cfg.bk_act,
+                     norm_type  = cfg.bk_norm,
+                     depthwise  = cfg.bk_depthwise)
+        )
+        # P5/32
+        self.layer_5 = nn.Sequential(
+            BasicConv(self.feat_dims[3], self.feat_dims[4],
+                      kernel_size=3, padding=1, stride=2,
+                      act_type=cfg.bk_act, norm_type=cfg.bk_norm, depthwise=cfg.bk_depthwise),
+            CSPBlock(in_dim     = self.feat_dims[4],
+                     out_dim    = self.feat_dims[4],
+                     num_blocks = round(3*cfg.depth),
+                     expansion  = 0.5,
+                     shortcut   = True,
+                     act_type   = cfg.bk_act,
+                     norm_type  = cfg.bk_norm,
+                     depthwise  = cfg.bk_depthwise)
+        )
+
+        # 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 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_act = 'silu'
+            self.bk_norm = 'BN'
+            self.bk_depthwise = False
+            self.width = 0.5
+            self.depth = 0.34
+            self.scale = "s"
+            self.use_pretrained = True
+
+    cfg = BaseConfig()
+    model = Yolov5Backbone(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)
+
+    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/yolox2/yolox2_basic.py

@@ -0,0 +1,137 @@
+import torch
+import torch.nn as nn
+from typing import List
+
+
+# --------------------- 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)
+            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)
+            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)
+            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
+
+
+# ---------------------------- Basic Modules ----------------------------
+class YoloBottleneck(nn.Module):
+    def __init__(self,
+                 in_dim       :int,
+                 out_dim      :int,
+                 kernel_size  :List  = [1, 3],
+                 expansion    :float = 0.5,
+                 shortcut     :bool  = False,
+                 act_type     :str   = 'silu',
+                 norm_type    :str   = 'BN',
+                 depthwise    :bool  = False,
+                 ) -> None:
+        super(YoloBottleneck, self).__init__()
+        inter_dim = int(out_dim * expansion)
+        # ----------------- Network setting -----------------
+        self.conv_layer1 = BasicConv(in_dim, inter_dim,
+                                     kernel_size=kernel_size[0], padding=kernel_size[0]//2, stride=1,
+                                     act_type=act_type, norm_type=norm_type, depthwise=depthwise)
+        self.conv_layer2 = BasicConv(inter_dim, out_dim,
+                                     kernel_size=kernel_size[1], padding=kernel_size[1]//2, stride=1,
+                                     act_type=act_type, norm_type=norm_type, depthwise=depthwise)
+        self.shortcut = shortcut and in_dim == out_dim
+
+    def forward(self, x):
+        h = self.conv_layer2(self.conv_layer1(x))
+
+        return x + h if self.shortcut else h
+
+class CSPBlock(nn.Module):
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 num_blocks   :int   = 1,
+                 expansion    :float = 0.5,
+                 shortcut     :bool  = False,
+                 act_type     :str   = 'silu',
+                 norm_type    :str   = 'BN',
+                 depthwise    :bool  = False,
+                 ):
+        super(CSPBlock, self).__init__()
+        # ---------- Basic parameters ----------
+        self.num_blocks = num_blocks
+        self.expansion = expansion
+        self.shortcut = shortcut
+        inter_dim = round(out_dim * expansion)
+        # ---------- Model parameters ----------
+        self.conv_layer_1 = BasicConv(in_dim, inter_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
+        self.conv_layer_2 = BasicConv(in_dim, inter_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
+        self.conv_layer_3 = BasicConv(inter_dim * 2, out_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
+        self.module       = nn.Sequential(*[YoloBottleneck(inter_dim,
+                                                           inter_dim,
+                                                           kernel_size  = [1, 3],
+                                                           expansion    = 1.0,
+                                                           shortcut     = shortcut,
+                                                           act_type     = act_type,
+                                                           norm_type    = norm_type,
+                                                           depthwise    = depthwise)
+                                                           for _ in range(num_blocks)
+                                                           ])
+
+    def forward(self, x):
+        x1 = self.conv_layer_1(x)
+        x2 = self.module(self.conv_layer_2(x))
+        out = self.conv_layer_3(torch.cat([x1, x2], dim=1))
+
+        return out
+    

yolo/models/yolox2/yolox2_head.py

@@ -0,0 +1,171 @@
+import torch
+import torch.nn as nn
+
+try:
+    from .yolox2_basic import BasicConv
+except:
+    from  yolox2.yolox2_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 Yolov5DetHead(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 = round(cfg.head_dim * cfg.width),
+                     reg_head_dim = round(cfg.head_dim * cfg.width),
+                     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
+        self.cls_head_dim = cfg.head_dim
+        self.reg_head_dim = cfg.head_dim
+
+
+    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 YoloxBaseConfig(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 = YoloxBaseConfig()
+    # 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 = Yolov5DetHead(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/yolox2/yolox2_neck.py

@@ -0,0 +1,33 @@
+import torch
+import torch.nn as nn
+
+from .yolox2_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/yolox2/yolox2_pafpn.py

@@ -0,0 +1,108 @@
+from typing import List
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .yolox2_basic import BasicConv, CSPBlock
+
+
+# Yolov5FPN
+class Yolov5PaFPN(nn.Module):
+    def __init__(self, cfg, in_dims: List = [256, 512, 1024],
+                 ):
+        super(Yolov5PaFPN, self).__init__()
+        self.in_dims = in_dims
+        c3, c4, c5 = in_dims
+
+        # ---------------------- Yolox's Top down FPN ----------------------
+        ## P5 -> P4
+        self.reduce_layer_1   = BasicConv(c5, round(512*cfg.width),
+                                          kernel_size=1, padding=0, stride=1,
+                                          act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+        self.top_down_layer_1 = CSPBlock(in_dim     = c4 + round(512*cfg.width),
+                                         out_dim    = round(512*cfg.width),
+                                         num_blocks = round(3*cfg.depth),
+                                         expansion  = 0.5,
+                                         shortcut   = False,
+                                         act_type   = cfg.fpn_act,
+                                         norm_type  = cfg.fpn_norm,
+                                         depthwise  = cfg.fpn_depthwise)
+
+        ## P4 -> P3
+        self.reduce_layer_2   = BasicConv(round(512*cfg.width), 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_2 = CSPBlock(in_dim     = c3 + round(256*cfg.width),
+                                         out_dim    = round(256*cfg.width),
+                                         num_blocks = round(3*cfg.depth),
+                                         expansion  = 0.5,
+                                         shortcut   = False,
+                                         act_type   = cfg.fpn_act,
+                                         norm_type  = cfg.fpn_norm,
+                                         depthwise  = cfg.fpn_depthwise)
+        
+        # ---------------------- Yolox's Bottom up PAN ----------------------
+        ## P3 -> P4
+        self.downsample_layer_1 = 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_1  = CSPBlock(in_dim     = round(256*cfg.width) + round(256*cfg.width),
+                                           out_dim    = round(512*cfg.width),
+                                           num_blocks = round(3*cfg.depth),
+                                           expansion  = 0.5,
+                                           shortcut   = False,
+                                           act_type   = cfg.fpn_act,
+                                           norm_type  = cfg.fpn_norm,
+                                           depthwise  = cfg.fpn_depthwise)
+        ## P4 -> P5
+        self.downsample_layer_2 = BasicConv(round(512*cfg.width), round(512*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  = CSPBlock(in_dim     = round(512*cfg.width) + round(512*cfg.width),
+                                           out_dim    = round(1024*cfg.width),
+                                           num_blocks = round(3*cfg.depth),
+                                           expansion  = 0.5,
+                                           shortcut   = False,
+                                           act_type   = cfg.fpn_act,
+                                           norm_type  = cfg.fpn_norm,
+                                           depthwise  = cfg.fpn_depthwise)
+
+        # ---------------------- Yolox's output projection ----------------------
+        self.out_layers = nn.ModuleList([
+            BasicConv(in_dim, round(cfg.head_dim*cfg.width), kernel_size=1,
+                      act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+                      for in_dim in [round(256*cfg.width), round(512*cfg.width), round(1024*cfg.width)]
+                      ])
+        self.out_dims = [round(cfg.head_dim*cfg.width)] * 3
+
+    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

yolo/models/yolox2/yolox2_pred.py

@@ -0,0 +1,130 @@
+import torch
+import torch.nn as nn
+
+# -------------------- Detection Pred Layer --------------------
+## Single-level pred layer
+class AFDetPredLayer(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 Yolov5AFDetPredLayer(nn.Module):
+    def __init__(self, cfg):
+        super().__init__()
+        # --------- Basic Parameters ----------
+        self.cfg = cfg
+
+        # ----------- Network Parameters -----------
+        ## pred layers
+        self.multi_level_preds = nn.ModuleList(
+            [AFDetPredLayer(cls_dim      = round(cfg.head_dim * cfg.width),
+                            reg_dim      = round(cfg.head_dim * cfg.width),
+                            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

yolo/train.py

@@ -75,7 +75,7 @@ def parse_args():
                         help='data root')
     parser.add_argument('-d', '--dataset', default='coco',
                         help='coco, voc')
-    parser.add_argument('--num_workers', default=16, type=int, 
+    parser.add_argument('--num_workers', default=4, type=int, 
                         help='Number of workers used in dataloading')
     
     # DDP train