add yolov7_af & remove yolov6 as it is ugly

yolo/config/__init__.py

@@ -5,6 +5,7 @@ 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
 from .rtdetr_config    import build_rtdetr_config
@@ -25,6 +26,8 @@ def build_config(args):
         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:
         cfg = build_yolov8_config(args)
     elif 'gelan' in args.model:

yolo/config/yolov7_af_config.py

@@ -0,0 +1,150 @@
+# yolo Config
+
+
+def build_yolov7af_config(args):
+    if   args.model == 'yolov7_af_t':
+        return Yolov7AFTConfig()
+    elif args.model == 'yolov7_af_l':
+        return Yolov7AFLConfig()
+    else:
+        raise NotImplementedError("No config for model: {}".format(args.model))
+    
+# YOLOv7AF-Base config
+class Yolov7AFBaseConfig(object):
+    def __init__(self) -> None:
+        # ---------------- Model config ----------------
+        self.width    = 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 = False
+        ## 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
+        self.fpn_expansions = [0.5, 0.5]
+        self.fpn_block_bw = 4
+        self.fpn_block_dw = 1
+        ## 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_center_sampling_radius = 2.5
+        self.ota_topk_candidate = 10
+        ## Loss weight
+        self.loss_obj = 1.0
+        self.loss_cls = 1.0
+        self.loss_box = 5.0
+
+        # ---------------- 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))
+
+# YOLOv7-S
+class Yolov7AFTConfig(Yolov7AFBaseConfig):
+    def __init__(self) -> None:
+        super().__init__()
+        # ---------------- Model config ----------------
+        self.width = 0.50
+        self.scale = "t"
+        self.fpn_expansions = [0.5, 1.0]
+        self.fpn_block_bw = 2
+        self.fpn_block_dw = 1
+
+        # ---------------- Data process config ----------------
+        self.mosaic_prob = 1.0
+        self.mixup_prob  = 0.0
+        self.copy_paste  = 0.5
+
+# YOLOv7-L
+class Yolov7AFLConfig(Yolov7AFBaseConfig):
+    def __init__(self) -> None:
+        super().__init__()
+        # ---------------- Model config ----------------
+        self.width = 1.0
+        self.scale = "l"
+        self.fpn_expansions = [0.5, 0.5]
+        self.fpn_block_bw = 4
+        self.fpn_block_dw = 1
+
+        # ---------------- Data process config ----------------
+        self.mosaic_prob = 1.0
+        self.mixup_prob  = 0.1
+        self.copy_paste  = 0.5

yolo/models/__init__.py

@@ -8,6 +8,7 @@ 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
 from .rtdetr.build    import build_rtdetr
@@ -33,6 +34,9 @@ def build_model(args, cfg, is_val=False):
     ## 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)
     ## YOLOv8
     elif 'yolov8' in args.model:
         model, criterion = build_yolov8(cfg, is_val)

yolo/models/yolov7_af/README.md

@@ -0,0 +1,61 @@
+# Anchor-free YOLOv7:
+
+- VOC
+
+|     Model   | Batch | Scale | AP<sup>val<br>0.5 | Weight |  Logs  |
+|-------------|-------|-------|-------------------|--------|--------|
+| YOLOv7-AF-S | 1xb16 |  640  |       82.7        | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v7/releases/download/yolo_tutorial_ckpt/yolov7_af_s_voc.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v7/releases/download/yolo_tutorial_ckpt/YOLOv7-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  |
+|-------------|-------|-------|------------------------|-------------------|-------------------|--------------------|--------|--------|
+| YOLOv7-AF-S | 1xb16 |  640  |                    |               |   26.9            |   8.9             |  |  |
+
+- For training, we train redesigned YOLOv7-AF with 500 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 YOLOv7-AF
+### Single GPU
+Taking training YOLOv7-AF-S on COCO as the example,
+```Shell
+python train.py --cuda -d coco --root path/to/coco -m yolov7_af_s -bs 16 --fp16 
+```
+
+### Multi GPU
+Taking training YOLOv7-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 yolov7_af_s -bs 16 --fp16 
+```
+
+## Test YOLOv7-AF
+Taking testing YOLOv7-AF-S on COCO-val as the example,
+```Shell
+python test.py --cuda -d coco --root path/to/coco -m yolov7_af_s --weight path/to/yolov7.pth --show 
+```
+
+## Evaluate YOLOv7-AF
+Taking evaluating YOLOv7-AF-S on COCO-val as the example,
+```Shell
+python eval.py --cuda -d coco --root path/to/coco -m yolov7_af_s --weight path/to/yolov7.pth 
+```
+
+## Demo
+### Detect with Image
+```Shell
+python demo.py --mode image --path_to_img path/to/image_dirs/ --cuda -m yolov7_af_s --weight path/to/weight --show
+```
+
+### Detect with Video
+```Shell
+python demo.py --mode video --path_to_vid path/to/video --cuda -m yolov7_af_s --weight path/to/weight --show --gif
+```
+
+### Detect with Camera
+```Shell
+python demo.py --mode camera --cuda -m yolov7_af_s --weight path/to/weight --show --gif
+```

yolo/models/yolov7_af/build.py

@@ -0,0 +1,24 @@
+import torch.nn as nn
+
+from .loss import SetCriterion
+from .yolov7_af import Yolov7AF
+
+
+# build object detector
+def build_yolov7af(cfg, is_val=False):
+    # -------------- Build YOLO --------------
+    model = Yolov7AF(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/yolov7_af/loss.py

@@ -0,0 +1,141 @@
+import torch
+import torch.nn.functional as F
+from utils.box_ops import get_ious
+from utils.distributed_utils import get_world_size, is_dist_avail_and_initialized
+
+from .matcher import YoloxMatcher
+
+
+class SetCriterion(object):
+    def __init__(self, cfg):
+        self.cfg = cfg
+        self.num_classes = cfg.num_classes
+        self.loss_obj_weight = cfg.loss_obj
+        self.loss_cls_weight = cfg.loss_cls
+        self.loss_box_weight = cfg.loss_box
+        # matcher
+        self.matcher = YoloxMatcher(cfg.num_classes, cfg.ota_center_sampling_radius, cfg.ota_topk_candidate)
+
+    def loss_objectness(self, pred_obj, gt_obj):
+        loss_obj = F.binary_cross_entropy_with_logits(pred_obj, gt_obj, reduction='none')
+
+        return loss_obj
+    
+    def loss_classes(self, pred_cls, gt_label):
+        loss_cls = F.binary_cross_entropy_with_logits(pred_cls, gt_label, reduction='none')
+
+        return loss_cls
+
+    def loss_bboxes(self, pred_box, gt_box):
+        # regression loss
+        ious = get_ious(pred_box, gt_box, "xyxy", 'giou')
+        loss_box = 1.0 - ious
+
+        return loss_box
+
+    def __call__(self, outputs, targets):        
+        """
+            outputs['pred_obj']: List(Tensor) [B, M, 1]
+            outputs['pred_cls']: List(Tensor) [B, M, C]
+            outputs['pred_reg']: 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]
+        obj_preds = torch.cat(outputs['pred_obj'], dim=1)
+        cls_preds = torch.cat(outputs['pred_cls'], dim=1)
+        box_preds = torch.cat(outputs['pred_box'], dim=1)
+
+        # label assignment
+        cls_targets = []
+        box_targets = []
+        obj_targets = []
+        fg_masks = []
+
+        for batch_idx in range(bs):
+            tgt_labels = targets[batch_idx]["labels"].to(device)
+            tgt_bboxes = targets[batch_idx]["boxes"].to(device)
+
+            # check target
+            if len(tgt_labels) == 0 or tgt_bboxes.max().item() == 0.:
+                num_anchors = sum([ab.shape[0] for ab in anchors])
+                # There is no valid gt
+                cls_target = obj_preds.new_zeros((0, self.num_classes))
+                box_target = obj_preds.new_zeros((0, 4))
+                obj_target = obj_preds.new_zeros((num_anchors, 1))
+                fg_mask = obj_preds.new_zeros(num_anchors).bool()
+            else:
+                (
+                    fg_mask,
+                    assigned_labels,
+                    assigned_ious,
+                    assigned_indexs
+                ) = self.matcher(
+                    fpn_strides = fpn_strides,
+                    anchors = anchors,
+                    pred_obj = obj_preds[batch_idx],
+                    pred_cls = cls_preds[batch_idx], 
+                    pred_box = box_preds[batch_idx],
+                    tgt_labels = tgt_labels,
+                    tgt_bboxes = tgt_bboxes
+                    )
+
+                obj_target = fg_mask.unsqueeze(-1)
+                cls_target = F.one_hot(assigned_labels.long(), self.num_classes)
+                cls_target = cls_target * assigned_ious.unsqueeze(-1)
+                box_target = tgt_bboxes[assigned_indexs]
+
+            cls_targets.append(cls_target)
+            box_targets.append(box_target)
+            obj_targets.append(obj_target)
+            fg_masks.append(fg_mask)
+
+        cls_targets = torch.cat(cls_targets, 0)
+        box_targets = torch.cat(box_targets, 0)
+        obj_targets = torch.cat(obj_targets, 0)
+        fg_masks = torch.cat(fg_masks, 0)
+        num_fgs = fg_masks.sum()
+
+        if is_dist_avail_and_initialized():
+            torch.distributed.all_reduce(num_fgs)
+        num_fgs = (num_fgs / get_world_size()).clamp(1.0)
+
+        # ------------------ Objecntness loss ------------------
+        loss_obj = self.loss_objectness(obj_preds.view(-1, 1), obj_targets.float())
+        loss_obj = loss_obj.sum() / num_fgs
+        
+        # ------------------ Classification loss ------------------
+        cls_preds_pos = cls_preds.view(-1, self.num_classes)[fg_masks]
+        loss_cls = self.loss_classes(cls_preds_pos, cls_targets)
+        loss_cls = loss_cls.sum() / num_fgs
+
+        # ------------------ Regression loss ------------------
+        box_preds_pos = box_preds.view(-1, 4)[fg_masks]
+        loss_box = self.loss_bboxes(box_preds_pos, box_targets)
+        loss_box = loss_box.sum() / num_fgs
+
+        # total loss
+        losses = self.loss_obj_weight * loss_obj + \
+                 self.loss_cls_weight * loss_cls + \
+                 self.loss_box_weight * loss_box
+
+        # Loss dict
+        loss_dict = dict(
+                loss_obj = loss_obj,
+                loss_cls = loss_cls,
+                loss_box = loss_box,
+                losses = losses
+        )
+
+        return loss_dict
+
+
+if __name__ == "__main__":
+    pass

yolo/models/yolov7_af/matcher.py

@@ -0,0 +1,185 @@
+# ---------------------------------------------------------------------
+# Copyright (c) Megvii Inc. All rights reserved.
+# ---------------------------------------------------------------------
+
+
+import torch
+import torch.nn.functional as F
+from utils.box_ops import *
+
+
+class YoloxMatcher(object):
+    """
+        This code referenced to https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/models/yolo_head.py
+    """
+    def __init__(self, num_classes, center_sampling_radius, topk_candidate ):
+        self.num_classes = num_classes
+        self.center_sampling_radius = center_sampling_radius
+        self.topk_candidate = topk_candidate
+
+
+    @torch.no_grad()
+    def __call__(self, 
+                 fpn_strides, 
+                 anchors, 
+                 pred_obj, 
+                 pred_cls, 
+                 pred_box, 
+                 tgt_labels,
+                 tgt_bboxes):
+        # [M,]
+        strides_tensor = 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]
+        anchors = torch.cat(anchors, dim=0)
+        num_anchor = anchors.shape[0]        
+        num_gt = len(tgt_labels)
+
+        # ----------------------- Find inside points -----------------------
+        fg_mask, is_in_boxes_and_center = self.get_in_boxes_info(
+            tgt_bboxes, anchors, strides_tensor, num_anchor, num_gt)
+        obj_preds = pred_obj[fg_mask].float()   # [Mp, 1]
+        cls_preds = pred_cls[fg_mask].float()   # [Mp, C]
+        box_preds = pred_box[fg_mask].float()   # [Mp, 4]
+
+        # ----------------------- Reg cost -----------------------
+        pair_wise_ious, _ = box_iou(tgt_bboxes, box_preds)      # [N, Mp]
+        reg_cost = -torch.log(pair_wise_ious + 1e-8)            # [N, Mp]
+
+        # ----------------------- Cls cost -----------------------
+        with torch.cuda.amp.autocast(enabled=False):
+            # [Mp, C]
+            score_preds = torch.sqrt(obj_preds.sigmoid_()* cls_preds.sigmoid_())
+            # [N, Mp, C]
+            score_preds = score_preds.unsqueeze(0).repeat(num_gt, 1, 1)
+            # prepare cls_target
+            cls_targets = F.one_hot(tgt_labels.long(), self.num_classes).float()
+            cls_targets = cls_targets.unsqueeze(1).repeat(1, score_preds.size(1), 1)
+            # [N, Mp]
+            cls_cost = F.binary_cross_entropy(score_preds, cls_targets, reduction="none").sum(-1)
+        del score_preds
+
+        #----------------------- Dynamic K-Matching -----------------------
+        cost_matrix = (
+            cls_cost
+            + 3.0 * reg_cost
+            + 100000.0 * (~is_in_boxes_and_center)
+        ) # [N, Mp]
+
+        (
+            assigned_labels,         # [num_fg,]
+            assigned_ious,           # [num_fg,]
+            assigned_indexs,         # [num_fg,]
+        ) = self.dynamic_k_matching(
+            cost_matrix,
+            pair_wise_ious,
+            tgt_labels,
+            num_gt,
+            fg_mask
+            )
+        del cls_cost, cost_matrix, pair_wise_ious, reg_cost
+
+        return fg_mask, assigned_labels, assigned_ious, assigned_indexs
+
+    def get_in_boxes_info(
+        self,
+        gt_bboxes,   # [N, 4]
+        anchors,     # [M, 2]
+        strides,     # [M,]
+        num_anchors, # M
+        num_gt,      # N
+        ):
+        # anchor center
+        x_centers = anchors[:, 0]
+        y_centers = anchors[:, 1]
+
+        # [M,] -> [1, M] -> [N, M]
+        x_centers = x_centers.unsqueeze(0).repeat(num_gt, 1)
+        y_centers = y_centers.unsqueeze(0).repeat(num_gt, 1)
+
+        # [N,] -> [N, 1] -> [N, M]
+        gt_bboxes_l = gt_bboxes[:, 0].unsqueeze(1).repeat(1, num_anchors) # x1
+        gt_bboxes_t = gt_bboxes[:, 1].unsqueeze(1).repeat(1, num_anchors) # y1
+        gt_bboxes_r = gt_bboxes[:, 2].unsqueeze(1).repeat(1, num_anchors) # x2
+        gt_bboxes_b = gt_bboxes[:, 3].unsqueeze(1).repeat(1, num_anchors) # y2
+
+        b_l = x_centers - gt_bboxes_l
+        b_r = gt_bboxes_r - x_centers
+        b_t = y_centers - gt_bboxes_t
+        b_b = gt_bboxes_b - y_centers
+        bbox_deltas = torch.stack([b_l, b_t, b_r, b_b], 2)
+
+        is_in_boxes = bbox_deltas.min(dim=-1).values > 0.0
+        is_in_boxes_all = is_in_boxes.sum(dim=0) > 0
+        # in fixed center
+        center_radius = self.center_sampling_radius
+
+        # [N, 2]
+        gt_centers = (gt_bboxes[:, :2] + gt_bboxes[:, 2:]) * 0.5
+        
+        # [1, M]
+        center_radius_ = center_radius * strides.unsqueeze(0)
+
+        gt_bboxes_l = gt_centers[:, 0].unsqueeze(1).repeat(1, num_anchors) - center_radius_ # x1
+        gt_bboxes_t = gt_centers[:, 1].unsqueeze(1).repeat(1, num_anchors) - center_radius_ # y1
+        gt_bboxes_r = gt_centers[:, 0].unsqueeze(1).repeat(1, num_anchors) + center_radius_ # x2
+        gt_bboxes_b = gt_centers[:, 1].unsqueeze(1).repeat(1, num_anchors) + center_radius_ # y2
+
+        c_l = x_centers - gt_bboxes_l
+        c_r = gt_bboxes_r - x_centers
+        c_t = y_centers - gt_bboxes_t
+        c_b = gt_bboxes_b - y_centers
+        center_deltas = torch.stack([c_l, c_t, c_r, c_b], 2)
+        is_in_centers = center_deltas.min(dim=-1).values > 0.0
+        is_in_centers_all = is_in_centers.sum(dim=0) > 0
+
+        # in boxes and in centers
+        is_in_boxes_anchor = is_in_boxes_all | is_in_centers_all
+
+        is_in_boxes_and_center = (
+            is_in_boxes[:, is_in_boxes_anchor] & is_in_centers[:, is_in_boxes_anchor]
+        )
+        return is_in_boxes_anchor, is_in_boxes_and_center
+
+    def dynamic_k_matching(
+        self, 
+        cost, 
+        pair_wise_ious, 
+        gt_classes, 
+        num_gt, 
+        fg_mask
+        ):
+        # Dynamic K
+        # ---------------------------------------------------------------
+        matching_matrix = torch.zeros_like(cost, dtype=torch.uint8)
+
+        ious_in_boxes_matrix = pair_wise_ious
+        n_candidate_k = min(self.topk_candidate, ious_in_boxes_matrix.size(1))
+        topk_ious, _ = torch.topk(ious_in_boxes_matrix, n_candidate_k, dim=1)
+        dynamic_ks = torch.clamp(topk_ious.sum(1).int(), min=1)
+        dynamic_ks = dynamic_ks.tolist()
+        for gt_idx in range(num_gt):
+            _, pos_idx = torch.topk(
+                cost[gt_idx], k=dynamic_ks[gt_idx], largest=False
+            )
+            matching_matrix[gt_idx][pos_idx] = 1
+
+        del topk_ious, dynamic_ks, pos_idx
+
+        anchor_matching_gt = matching_matrix.sum(0)
+        if (anchor_matching_gt > 1).sum() > 0:
+            _, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0)
+            matching_matrix[:, anchor_matching_gt > 1] *= 0
+            matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1
+        fg_mask_inboxes = matching_matrix.sum(0) > 0
+
+        fg_mask[fg_mask.clone()] = fg_mask_inboxes
+
+        assigned_indexs = matching_matrix[:, fg_mask_inboxes].argmax(0)
+        assigned_labels = gt_classes[assigned_indexs]
+
+        assigned_ious = (matching_matrix * pair_wise_ious).sum(0)[
+            fg_mask_inboxes
+        ]
+        return assigned_labels, assigned_ious, assigned_indexs
+    

yolo/models/yolov7_af/yolov7_af.py

@@ -0,0 +1,157 @@
+# --------------- Torch components ---------------
+import torch
+import torch.nn as nn
+
+# --------------- Model components ---------------
+from .yolov7_af_backbone import Yolov7Backbone
+from .yolov7_af_neck     import SPPF
+from .yolov7_af_pafpn    import Yolov7PaFPN
+from .yolov7_af_head     import Yolov7DetHead
+from .yolov7_af_pred     import Yolov7AFDetPredLayer
+
+# --------------- External components ---------------
+from utils.misc import multiclass_nms
+
+
+# Yolov7AF
+class Yolov7AF(nn.Module):
+    def __init__(self,
+                 cfg,
+                 is_val = False,
+                 ) -> None:
+        super(Yolov7AF, 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 = Yolov7Backbone(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]//2)
+        self.pyramid_feat_dims[-1] = self.neck.out_dim
+        ## Neck: FPN
+        self.fpn      = Yolov7PaFPN(cfg, self.pyramid_feat_dims)
+        ## Head
+        self.head     = Yolov7DetHead(cfg, self.fpn.out_dims)
+        ## Pred
+        self.pred     = Yolov7AFDetPredLayer(cfg)
+
+    def post_process(self, obj_preds, cls_preds, box_preds):
+        """
+        We process predictions at each scale hierarchically
+        Input:
+            obj_preds: List[torch.Tensor] -> [[B, M, 1], ...], B=1
+            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 obj_pred_i, cls_pred_i, box_pred_i in zip(obj_preds, cls_preds, box_preds):
+            obj_pred_i = obj_pred_i[0]
+            cls_pred_i = cls_pred_i[0]
+            box_pred_i = box_pred_i[0]
+            if self.no_multi_labels:
+                # [M,]
+                scores, labels = torch.max(
+                    torch.sqrt(obj_pred_i.sigmoid() * 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 = torch.sqrt(obj_pred_i.sigmoid() * 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_obj_preds = outputs['pred_obj']
+            all_cls_preds = outputs['pred_cls']
+            all_box_preds = outputs['pred_box']
+
+            # post process
+            bboxes, scores, labels = self.post_process(all_obj_preds, all_cls_preds, all_box_preds)
+            outputs = {
+                "scores": scores,
+                "labels": labels,
+                "bboxes": bboxes
+            }
+        
+        return outputs 

yolo/models/yolov7_af/yolov7_af_backbone.py

@@ -0,0 +1,126 @@
+import torch
+import torch.nn as nn
+
+try:
+    from .yolov7_af_basic import BasicConv, MDown, ELANLayer
+except:
+    from  yolov7_af_basic import BasicConv, MDown, ELANLayer
+
+
+# ELANNet
+class Yolov7Backbone(nn.Module):
+    def __init__(self, cfg):
+        super(Yolov7Backbone, self).__init__()
+        # ---------------- Basic parameters ----------------
+        self.model_scale = cfg.scale
+        self.bk_act = cfg.bk_act
+        self.bk_norm = cfg.bk_norm
+        self.bk_depthwise = cfg.bk_depthwise
+        if self.model_scale in ["l", "x"]:
+            self.elan_depth = 2
+            self.feat_dims = [round(64   * cfg.width), round(128  * cfg.width), round(256  * cfg.width),
+                              round(512  * cfg.width), round(1024 * cfg.width), round(1024 * cfg.width)]
+            self.last_stage_eratio = 0.25
+        if self.model_scale in ["t"]:
+            self.elan_depth = 1
+            self.feat_dims = [round(64  * cfg.width), round(128  * cfg.width),
+                              round(256  * cfg.width), round(512 * cfg.width), round(1024 * cfg.width)]
+            self.last_stage_eratio = 0.5
+
+        # ---------------- Model parameters ----------------
+        self.layer_1 = self.make_stem(3, self.feat_dims[0])
+        self.layer_2 = self.make_block(self.feat_dims[0], self.feat_dims[1], expansion=0.5)
+        self.layer_3 = self.make_block(self.feat_dims[1], self.feat_dims[2], expansion=0.5)
+        self.layer_4 = self.make_block(self.feat_dims[2], self.feat_dims[3], expansion=0.5)
+        self.layer_5 = self.make_block(self.feat_dims[3], self.feat_dims[4], expansion=self.last_stage_eratio)
+
+        # Initialize all layers
+        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 make_stem(self, in_dim, out_dim):
+        if self.model_scale in ["l", "x"]:
+            stem = nn.Sequential(
+                BasicConv(in_dim, out_dim//2, kernel_size=3, padding=1, stride=1,
+                          act_type=self.bk_act, norm_type=self.bk_norm, depthwise=self.bk_depthwise),
+                BasicConv(out_dim//2, out_dim, kernel_size=3, padding=1, stride=2,
+                          act_type=self.bk_act, norm_type=self.bk_norm, depthwise=self.bk_depthwise),
+                BasicConv(out_dim, out_dim, kernel_size=3, padding=1, stride=1,
+                          act_type=self.bk_act, norm_type=self.bk_norm, depthwise=self.bk_depthwise)
+
+            )
+        elif self.model_scale in ["t"]:
+            stem = BasicConv(in_dim, out_dim, kernel_size=6, padding=2, stride=2,
+                              act_type=self.bk_act, norm_type=self.bk_norm, depthwise=self.bk_depthwise)
+        else:
+            raise NotImplementedError("Unknown model scale: {}".format(self.model_scale))
+        
+        return stem
+
+    def make_block(self, in_dim, out_dim, expansion=0.5):
+        if self.model_scale in ["l", "x"]:
+            block = nn.Sequential(
+                MDown(in_dim, out_dim,
+                    act_type=self.bk_act, norm_type=self.bk_norm, depthwise=self.bk_depthwise),             
+                ELANLayer(out_dim, out_dim,
+                        expansion=expansion, num_blocks=self.elan_depth,
+                        act_type=self.bk_act, norm_type=self.bk_norm, depthwise=self.bk_depthwise),
+            )
+        elif self.model_scale in ["t"]:
+            block = nn.Sequential(
+                nn.MaxPool2d((2, 2), stride=2),             
+                ELANLayer(in_dim, out_dim,
+                        expansion=expansion, num_blocks=self.elan_depth,
+                        act_type=self.bk_act, norm_type=self.bk_norm, depthwise=self.bk_depthwise),
+            )
+        else:
+            raise NotImplementedError("Unknown model scale: {}".format(self.model_scale))
+        
+        return block
+    
+    def forward(self, x):
+        c1 = self.layer_1(x)
+        c2 = self.layer_2(c1)
+        c3 = self.layer_3(c2)
+        c4 = self.layer_4(c3)
+        c5 = self.layer_5(c4)
+        outputs = [c3, c4, c5]
+
+        return outputs
+
+
+if __name__ == '__main__':
+    import time
+    from thop import profile
+    class BaseConfig(object):
+        def __init__(self) -> None:
+            self.bk_act = 'silu'
+            self.bk_norm = 'BN'
+            self.bk_depthwise = False
+            self.width = 0.5
+            self.depth = 0.34
+            self.scale = "t"
+
+    cfg = BaseConfig()
+    model = Yolov7Backbone(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/yolov7_af/yolov7_af_basic.py

@@ -0,0 +1,190 @@
+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 MDown(nn.Module):
+    def __init__(self,
+                 in_dim    :int,
+                 out_dim   :int,
+                 act_type  :str   = 'silu',
+                 norm_type :str   = 'BN',
+                 depthwise :bool  = False,
+                 ) -> None:
+        super().__init__()
+        inter_dim = out_dim // 2
+        self.downsample_1 = nn.Sequential(
+            nn.MaxPool2d((2, 2), stride=2),
+            BasicConv(in_dim, inter_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
+        )
+        self.downsample_2 = nn.Sequential(
+            BasicConv(in_dim, inter_dim, kernel_size=1, act_type=act_type, norm_type=norm_type),
+            BasicConv(inter_dim, inter_dim,
+                      kernel_size=3, padding=1, stride=2,
+                      act_type=act_type, norm_type=norm_type, depthwise=depthwise)
+        )
+        if in_dim == out_dim:
+            self.output_proj = nn.Identity()
+        else:
+            self.output_proj = BasicConv(inter_dim * 2, out_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
+
+    def forward(self, x):
+        x1 = self.downsample_1(x)
+        x2 = self.downsample_2(x)
+
+        out = self.output_proj(torch.cat([x1, x2], dim=1))
+
+        return out
+
+class ELANLayer(nn.Module):
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 expansion  :float = 0.5,
+                 num_blocks :int   = 1,
+                 act_type   :str   = 'silu',
+                 norm_type  :str   = 'BN',
+                 depthwise  :bool  = False,
+                 ) -> None:
+        super(ELANLayer, self).__init__()
+        self.inter_dim = round(in_dim * expansion)
+        self.conv_layer_1 = BasicConv(in_dim, self.inter_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
+        self.conv_layer_2 = BasicConv(in_dim, self.inter_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
+        self.conv_layer_3 = BasicConv(self.inter_dim * 4, out_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
+        self.elan_layer_1 = nn.Sequential(*[BasicConv(self.inter_dim, self.inter_dim,
+                                                      kernel_size=3, padding=1,
+                                                      act_type=act_type, norm_type=norm_type, depthwise=depthwise)
+                                                      for _ in range(num_blocks)])
+        self.elan_layer_2 = nn.Sequential(*[BasicConv(self.inter_dim, self.inter_dim,
+                                                      kernel_size=3, padding=1,
+                                                      act_type=act_type, norm_type=norm_type, depthwise=depthwise)
+                                                      for _ in range(num_blocks)])
+
+    def forward(self, x):
+        # Input proj
+        x1 = self.conv_layer_1(x)
+        x2 = self.conv_layer_2(x)
+        x3 = self.elan_layer_1(x2)
+        x4 = self.elan_layer_2(x3)
+    
+        out = self.conv_layer_3(torch.cat([x1, x2, x3, x4], dim=1))
+
+        return out
+
+## PaFPN's ELAN-Block proposed by YOLOv7
+class ELANLayerFPN(nn.Module):
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 expansions   :List = [0.5, 0.5],
+                 branch_width :int  = 4,
+                 branch_depth :int  = 1,
+                 act_type     :str  = 'silu',
+                 norm_type    :str  = 'BN',
+                 depthwise=False):
+        super(ELANLayerFPN, self).__init__()
+        # Basic parameters
+        inter_dim  = round(in_dim * expansions[0])
+        inter_dim2 = round(inter_dim * expansions[1]) 
+        # Network structure
+        self.cv1 = BasicConv(in_dim, inter_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
+        self.cv2 = BasicConv(in_dim, inter_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
+        self.cv3 = nn.ModuleList()
+        for idx in range(round(branch_width)):
+            if idx == 0:
+                cvs = [BasicConv(inter_dim, inter_dim2,
+                                 kernel_size=3, padding=1,
+                                 act_type=act_type, norm_type=norm_type, depthwise=depthwise)]
+            else:
+                cvs = [BasicConv(inter_dim2, inter_dim2,
+                                 kernel_size=3, padding=1,
+                                 act_type=act_type, norm_type=norm_type, depthwise=depthwise)]
+            # deeper
+            if round(branch_depth) > 1:
+                for _ in range(1, round(branch_depth)):
+                    cvs.append(BasicConv(inter_dim2, inter_dim2, kernel_size=3, padding=1, act_type=act_type, norm_type=norm_type, depthwise=depthwise))
+                self.cv3.append(nn.Sequential(*cvs))
+            else:
+                self.cv3.append(cvs[0])
+
+        self.output_proj = BasicConv(inter_dim*2+inter_dim2*len(self.cv3), out_dim,
+                                     kernel_size=1, act_type=act_type, norm_type=norm_type)
+
+
+    def forward(self, x):
+        x1 = self.cv1(x)
+        x2 = self.cv2(x)
+        inter_outs = [x1, x2]
+        for m in self.cv3:
+            y1 = inter_outs[-1]
+            y2 = m(y1)
+            inter_outs.append(y2)
+        out = self.output_proj(torch.cat(inter_outs, dim=1))
+
+        return out

yolo/models/yolov7_af/yolov7_af_head.py

@@ -0,0 +1,171 @@
+import torch
+import torch.nn as nn
+
+try:
+    from .yolov7_af_basic import BasicConv
+except:
+    from  yolov7_af_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 Yolov7DetHead(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 = Yolov7DetHead(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/yolov7_af/yolov7_af_neck.py

@@ -0,0 +1,60 @@
+import torch
+import torch.nn as nn
+from .yolov7_af_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, expansion=0.5):
+        super().__init__()
+        ## ----------- Basic Parameters -----------
+        inter_dim = round(in_dim * expansion)
+        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))
+
+# SPPF block with CSP module
+class SPPFBlockCSP(nn.Module):
+    """
+        CSP Spatial Pyramid Pooling Block
+    """
+    def __init__(self, cfg, in_dim, out_dim):
+        super(SPPFBlockCSP, self).__init__()
+        inter_dim = int(in_dim * cfg.neck_expand_ratio)
+        self.out_dim = out_dim
+        self.cv1 = BasicConv(in_dim, inter_dim, kernel_size=1, act_type=cfg.neck_act, norm_type=cfg.neck_norm)
+        self.cv2 = BasicConv(in_dim, inter_dim, kernel_size=1, act_type=cfg.neck_act, norm_type=cfg.neck_norm)
+        self.module = nn.Sequential(
+            BasicConv(inter_dim, inter_dim, kernel_size=3, padding=1, 
+                      act_type=cfg.neck_act, norm_type=cfg.neck_norm, depthwise=cfg.neck_depthwise),
+            SPPF(cfg, inter_dim, inter_dim, expansion=1.0),
+            BasicConv(inter_dim, inter_dim, kernel_size=3, padding=1, 
+                      act_type=cfg.neck_act, norm_type=cfg.neck_norm, depthwise=cfg.neck_depthwise),
+                      )
+        self.cv3 = BasicConv(inter_dim * 2, self.out_dim, kernel_size=1, act_type=cfg.neck_act, norm_type=cfg.neck_norm)
+
+        
+    def forward(self, x):
+        x1 = self.cv1(x)
+        x2 = self.module(self.cv2(x))
+        y = self.cv3(torch.cat([x1, x2], dim=1))
+
+        return y

yolo/models/yolov7_af/yolov7_af_pafpn.py

@@ -0,0 +1,114 @@
+from typing import List
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .yolov7_af_basic import BasicConv, ELANLayerFPN, MDown
+
+
+# PaFPN-ELAN (YOLOv7's)
+class Yolov7PaFPN(nn.Module):
+    def __init__(self, cfg, in_dims: List = [512, 1024, 512]):
+        super(Yolov7PaFPN, self).__init__()
+        # ----------------------------- Basic parameters -----------------------------
+        self.in_dims = in_dims
+        self.out_dims = [round(256*cfg.width), round(512*cfg.width), round(1024*cfg.width)]
+        c3, c4, c5 = in_dims
+
+        # ----------------------------- Yolov7's Top-down FPN -----------------------------
+        ## P5 -> P4
+        self.reduce_layer_1 = BasicConv(c5, round(256*cfg.width),
+                                        kernel_size=1, act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+        self.reduce_layer_2 = BasicConv(c4, round(256*cfg.width),
+                                        kernel_size=1, act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+        self.top_down_layer_1 = ELANLayerFPN(in_dim     = round(256*cfg.width) + round(256*cfg.width),
+                                             out_dim    = round(256*cfg.width),
+                                             expansions   = cfg.fpn_expansions,
+                                             branch_width = cfg.fpn_block_bw,
+                                             branch_depth = cfg.fpn_block_dw,
+                                             act_type   = cfg.fpn_act,
+                                             norm_type  = cfg.fpn_norm,
+                                             depthwise  = cfg.fpn_depthwise,
+                                             )
+        ## P4 -> P3
+        self.reduce_layer_3 = BasicConv(round(256*cfg.width), round(128*cfg.width),
+                                        kernel_size=1, act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+        self.reduce_layer_4 = BasicConv(c3, round(128*cfg.width),
+                                        kernel_size=1, act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+        self.top_down_layer_2 = ELANLayerFPN(in_dim     = round(128*cfg.width) + round(128*cfg.width),
+                                             out_dim    = round(128*cfg.width),
+                                             expansions   = cfg.fpn_expansions,
+                                             branch_width = cfg.fpn_block_bw,
+                                             branch_depth = cfg.fpn_block_dw,
+                                             act_type   = cfg.fpn_act,
+                                             norm_type  = cfg.fpn_norm,
+                                             depthwise  = cfg.fpn_depthwise,
+                                             )
+        # ----------------------------- Yolov7's Bottom-up PAN -----------------------------
+        ## P3 -> P4
+        self.downsample_layer_1 = MDown(round(128*cfg.width), round(256*cfg.width),
+                                        act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+        self.bottom_up_layer_1 = ELANLayerFPN(in_dim     = round(256*cfg.width) + round(256*cfg.width),
+                                              out_dim    = round(256*cfg.width),
+                                              expansions   = cfg.fpn_expansions,
+                                              branch_width = cfg.fpn_block_bw,
+                                              branch_depth = cfg.fpn_block_dw,
+                                              act_type     = cfg.fpn_act,
+                                              norm_type    = cfg.fpn_norm,
+                                              depthwise    = cfg.fpn_depthwise,
+                                              )
+        ## P4 -> P5
+        self.downsample_layer_2 = MDown(round(256*cfg.width), round(512*cfg.width),
+                                        act_type=cfg.fpn_act, norm_type=cfg.fpn_norm)
+        self.bottom_up_layer_2 = ELANLayerFPN(in_dim     = round(512*cfg.width) + c5,
+                                              out_dim    = round(512*cfg.width),
+                                              expansions   = cfg.fpn_expansions,
+                                              branch_width = cfg.fpn_block_bw,
+                                              branch_depth = cfg.fpn_block_dw,
+                                              act_type   = cfg.fpn_act,
+                                              norm_type  = cfg.fpn_norm,
+                                              depthwise  = cfg.fpn_depthwise,
+                                              )
+
+        # ----------------------------- Head conv layers -----------------------------
+        ## Head convs
+        self.head_conv_1 = BasicConv(round(128*cfg.width), round(256*cfg.width),
+                                     kernel_size=3, padding=1, stride=1,
+                                     act_type=cfg.fpn_act, norm_type=cfg.fpn_norm, depthwise=cfg.fpn_depthwise)
+        self.head_conv_2 = BasicConv(round(256*cfg.width), round(512*cfg.width),
+                                     kernel_size=3, padding=1, stride=1,
+                                     act_type=cfg.fpn_act, norm_type=cfg.fpn_norm, depthwise=cfg.fpn_depthwise)
+        self.head_conv_3 = BasicConv(round(512*cfg.width), round(1024*cfg.width),
+                                     kernel_size=3, padding=1, stride=1,
+                                     act_type=cfg.fpn_act, norm_type=cfg.fpn_norm, depthwise=cfg.fpn_depthwise)
+
+    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.reduce_layer_2(c4)
+        p4 = self.top_down_layer_1(torch.cat([p5_up, p4], dim=1))
+
+        ## P4 -> P3
+        p4_in = self.reduce_layer_3(p4)
+        p4_up = F.interpolate(p4_in, scale_factor=2.0)
+        p3 = self.reduce_layer_4(c3)
+        p3 = self.top_down_layer_2(torch.cat([p4_up, p3], dim=1))
+
+        # ------------------ Bottom up PAN ------------------
+        ## P3 -> P4
+        p3_ds = self.downsample_layer_1(p3)
+        p4 = torch.cat([p3_ds, p4], dim=1)
+        p4 = self.bottom_up_layer_1(p4)
+
+        ## P4 -> P5
+        p4_ds = self.downsample_layer_2(p4)
+        p5 = torch.cat([p4_ds, c5], dim=1)
+        p5 = self.bottom_up_layer_2(p5)
+
+        out_feats = [self.head_conv_1(p3), self.head_conv_2(p4), self.head_conv_3(p5)]
+            
+        return out_feats

yolo/models/yolov7_af/yolov7_af_pred.py

@@ -0,0 +1,142 @@
+import torch
+import torch.nn as nn
+from typing import List
+
+# -------------------- 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.obj_pred = nn.Conv2d(self.cls_dim, 1, kernel_size=1)
+        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))
+        # obj pred
+        b = self.obj_pred.bias.view(1, -1)
+        b.data.fill_(bias_value.item())
+        self.obj_pred.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+        # 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
+        obj_pred = self.obj_pred(reg_feat)
+        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]
+        obj_pred = obj_pred.permute(0, 2, 3, 1).contiguous().view(B, -1, 1)
+        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_obj": obj_pred,       # (torch.Tensor) [B, M, 1]
+                   "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 Yolov7AFDetPredLayer(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_obj_preds = []
+        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_obj_preds.append(outputs["pred_obj"])
+            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_obj":  all_obj_preds,         # List(Tensor) [B, M, 1]
+                   "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