add RTMDet-v2

config/__init__.py

@@ -32,7 +32,7 @@ from .data_config.transform_config import (
     yolox_huge_trans_config,
     # SSD-Style
     ssd_trans_config,
-    # YOLOvx-Style
+    # RTMDet-v1-Style
     rtmdet_v1_pico_trans_config,
     rtmdet_v1_nano_trans_config,
     rtmdet_v1_small_trans_config,
@@ -77,7 +77,7 @@ def build_trans_config(trans_config='ssd'):
     elif trans_config == 'yolox_huge':
         cfg = yolox_huge_trans_config
 
-    # YOLOvx-style transform 
+    # RTMDetv1-style transform 
     elif trans_config == 'rtmdet_v1_pico':
         cfg = rtmdet_v1_pico_trans_config
     elif trans_config == 'rtmdet_v1_nano':
@@ -105,7 +105,9 @@ from .model_config.yolov4_config import yolov4_cfg
 from .model_config.yolov5_config import yolov5_cfg
 from .model_config.yolov7_config import yolov7_cfg
 from .model_config.yolox_config import yolox_cfg
+## My RTMDet series
 from .model_config.rtmdet_v1_config import rtmdet_v1_cfg
+from .model_config.rtmdet_v2_config import rtmdet_v2_cfg
 
 
 def build_model_config(args):
@@ -132,9 +134,12 @@ def build_model_config(args):
     # YOLOX
     elif args.model in ['yolox_n', 'yolox_s', 'yolox_m', 'yolox_l', 'yolox_x']:
         cfg = yolox_cfg[args.model]
-    # YOLOvx
+    # My RTMDet-v1
     elif args.model in ['rtmdet_v1_n', 'rtmdet_v1_t', 'rtmdet_v1_s', 'rtmdet_v1_m', 'rtmdet_v1_l', 'rtmdet_v1_x']:
         cfg = rtmdet_v1_cfg[args.model]
+    # My RTMDet-v2
+    elif args.model in ['rtmdet_v2_n', 'rtmdet_v2_t', 'rtmdet_v2_s', 'rtmdet_v2_m', 'rtmdet_v2_l', 'rtmdet_v2_x']:
+        cfg = rtmdet_v2_cfg[args.model]
 
     return cfg
 

config/model_config/rtmdet_v2_config.py

@@ -0,0 +1,123 @@
+# YOLOvx Config
+
+
+rtmdet_v2_cfg = {
+    'rtmdet_v2_n':{
+        # ---------------- Model config ----------------
+        ## Backbone
+        'backbone': 'mcnet',
+        'pretrained': True,
+        'bk_act': 'silu',
+        'bk_norm': 'BN',
+        'bk_depthwise': False,
+        'bk_num_heads': 4,
+        'width': 0.25,
+        'depth': 0.34,
+        'stride': [8, 16, 32],  # P3, P4, P5
+        'max_stride': 32,
+        ## Neck: SPP
+        'neck': 'sppf',
+        'neck_expand_ratio': 0.5,
+        'pooling_size': 5,
+        'neck_act': 'silu',
+        'neck_norm': 'BN',
+        'neck_depthwise': False,
+        ## Neck: PaFPN
+        'fpn': 'rtmdet_pafpn',
+        'fpn_reduce_layer': 'conv',
+        'fpn_downsample_layer': 'conv',
+        'fpn_core_block': 'mcblock',
+        'fpn_num_heads': 4,
+        'fpn_act': 'silu',
+        'fpn_norm': 'BN',
+        'fpn_depthwise': False,
+        ## Head
+        'head': 'decoupled_head',
+        'head_act': 'silu',
+        'head_norm': 'BN',
+        'num_cls_head': 2,
+        'num_reg_head': 2,
+        'head_depthwise': False,
+        'reg_max': 16,
+        # ---------------- Train config ----------------
+        ## Input
+        'multi_scale': [0.5, 1.5],   # 320 -> 960
+        'trans_type': 'rtmdet_v1_nano',
+        # ---------------- Assignment config ----------------
+        ## Matcher
+        'matcher': {'tal': {'topk': 10,
+                            'alpha': 0.5,
+                            'beta': 6.0},
+                    'ota': {'center_sampling_radius': 2.5,
+                             'topk_candidate': 10},
+                    },
+        # ---------------- Loss config ----------------
+        ## Loss weight
+        'ema_update': False,
+        'loss_cls_weight': 1.0,
+        'loss_box_weight': 5.0,
+        'loss_dfl_weight': 1.0,
+        # ---------------- Train config ----------------
+        'trainer_type': 'rtmdet',
+    },
+
+    'rtmdet_v2_l':{
+        # ---------------- Model config ----------------
+        ## Backbone
+        'backbone': 'mcnet',
+        'pretrained': True,
+        'bk_act': 'silu',
+        'bk_norm': 'BN',
+        'bk_depthwise': False,
+        'bk_num_heads': 4,
+        'width': 1.0,
+        'depth': 1.0,
+        'stride': [8, 16, 32],  # P3, P4, P5
+        'max_stride': 32,
+        ## Neck: SPP
+        'neck': 'sppf',
+        'neck_expand_ratio': 0.5,
+        'pooling_size': 5,
+        'neck_act': 'silu',
+        'neck_norm': 'BN',
+        'neck_depthwise': False,
+        ## Neck: PaFPN
+        'fpn': 'rtmdet_pafpn',
+        'fpn_reduce_layer': 'conv',
+        'fpn_downsample_layer': 'conv',
+        'fpn_core_block': 'mcblock',
+        'fpn_num_heads': 4,
+        'fpn_act': 'silu',
+        'fpn_norm': 'BN',
+        'fpn_depthwise': False,
+        ## Head
+        'head': 'decoupled_head',
+        'head_act': 'silu',
+        'head_norm': 'BN',
+        'num_cls_head': 2,
+        'num_reg_head': 2,
+        'head_depthwise': False,
+        'reg_max': 16,
+        # ---------------- Train config ----------------
+        ## Input
+        'multi_scale': [0.5, 1.25],   # 320 -> 800
+        'trans_type': 'rtmdet_v1_large',
+        # ---------------- Assignment config ----------------
+        ## Matcher
+        'matcher': {'tal': {'topk': 10,
+                            'alpha': 0.5,
+                            'beta': 6.0},
+                    'ota': {'center_sampling_radius': 2.5,
+                             'topk_candidate': 10},
+                    },
+        # ---------------- Loss config ----------------
+        ## Loss weight
+        'ema_update': False,
+        'loss_cls_weight': 1.0,
+        'loss_box_weight': 5.0,
+        'loss_dfl_weight': 1.0,
+        # ---------------- Train config ----------------
+        'trainer_type': 'rtmdet',
+    },
+
+}

models/detectors/__init__.py

@@ -10,6 +10,7 @@ from .yolov4.build import build_yolov4
 from .yolov5.build import build_yolov5
 from .yolov7.build import build_yolov7
 from .rtmdet_v1.build import build_rtmdet_v1
+from .rtmdet_v2.build import build_rtmdet_v2
 # My custom YOLO
 from .yolox.build import build_yolox
 
@@ -49,10 +50,14 @@ def build_model(args,
     elif args.model in ['yolox_n', 'yolox_s', 'yolox_m', 'yolox_l', 'yolox_x']:
         model, criterion = build_yolox(
             args, model_cfg, device, num_classes, trainable, deploy)
-    # My RTMDet
+    # My RTMDet-v1
     elif args.model in ['rtmdet_v1_n', 'rtmdet_v1_t', 'rtmdet_v1_s', 'rtmdet_v1_m', 'rtmdet_v1_l', 'rtmdet_v1_x']:
         model, criterion = build_rtmdet_v1(
             args, model_cfg, device, num_classes, trainable, deploy)
+    # My RTMDet-v2
+    elif args.model in ['rtmdet_v2_n', 'rtmdet_v2_t', 'rtmdet_v2_s', 'rtmdet_v2_m', 'rtmdet_v2_l', 'rtmdet_v2_x']:
+        model, criterion = build_rtmdet_v2(
+            args, model_cfg, device, num_classes, trainable, deploy)
 
     if trainable:
         # Load pretrained weight

models/detectors/rtmdet_v2/README.md

@@ -0,0 +1,16 @@
+# RTMDet-v2: My Second Empirical Study of Real-Time General Object Detectors.
+
+|   Model    | Scale | Batch | AP<sup>test<br>0.5:0.95 | AP<sup>test<br>0.5 | AP<sup>val<br>0.5:0.95 | AP<sup>val<br>0.5 | FLOPs<br><sup>(G) | Params<br><sup>(M) | Weight |
+|------------|-------|-------|-------------------------|--------------------|------------------------|-------------------|-------------------|--------------------|--------|
+| RTMDetv2-N |  640  | 8xb16 |                         |                    |                        |                   |                   |                    |  |
+| RTMDetv2-T |  640  | 8xb16 |                         |                    |                        |                   |                   |                    |  |
+| RTMDetv2-S |  640  | 8xb16 |                         |                    |                        |                   |                   |                    |  |
+| RTMDetv2-M |  640  | 8xb16 |                         |                    |                        |                   |                   |                    |  |
+| RTMDetv2-L |  640  | 8xb16 |                         |                    |                        |                   |                   |                    |  |
+| RTMDetv2-X |  640  |       |                         |                    |                        |                   |                   |                    |  |
+
+- For training, we train my RTMDetv2 series series with 300 epochs on COCO.
+- For data augmentation, we use the large scale jitter (LSJ), Mosaic augmentation and Mixup augmentation, following the setting of [YOLOX](https://github.com/ultralytics/yolov5), but we remove the rotation transformation which is used in YOLOX's strong augmentation.
+- For optimizer, we use AdamW with weight decay 0.05 and base per image lr 0.001 / 64.
+- For learning rate scheduler, we use linear decay scheduler.
+- Due to my limited computing resources, I can not train `RTMDetv2-X` with the setting of `batch size=128`.

models/detectors/rtmdet_v2/build.py

@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+# -*- coding:utf-8 -*-
+
+import torch
+import torch.nn as nn
+
+from .loss import build_criterion
+from .rtmdet_v2 import RTMDet
+
+
+# build object detector
+def build_rtmdet_v2(args, cfg, device, num_classes=80, trainable=False, deploy=False):
+    print('==============================')
+    print('Build {} ...'.format(args.model.upper()))
+        
+    # -------------- Build RTMDet --------------
+    model = RTMDet(
+        cfg=cfg,
+        device=device, 
+        num_classes=num_classes,
+        trainable=trainable,
+        conf_thresh=args.conf_thresh,
+        nms_thresh=args.nms_thresh,
+        topk=args.topk,
+        deploy=deploy
+        )
+
+    # -------------- Initialize RTMDet --------------
+    for m in model.modules():
+        if isinstance(m, nn.BatchNorm2d):
+            m.eps = 1e-3
+            m.momentum = 0.03    
+            
+    # -------------- Build criterion --------------
+    criterion = None
+    if trainable:
+        # build criterion for training
+        criterion = build_criterion(args, cfg, device, num_classes)
+    return model, criterion

models/detectors/rtmdet_v2/loss.py

@@ -0,0 +1,337 @@
+import torch
+import torch.nn.functional as F
+
+from utils.box_ops import bbox2dist, get_ious
+from utils.distributed_utils import get_world_size, is_dist_avail_and_initialized
+
+from .matcher import TaskAlignedAssigner, AlignedSimOTA
+
+
+class Criterion(object):
+    def __init__(self, args, cfg, device, num_classes=80):
+        self.cfg = cfg
+        self.args = args
+        self.device = device
+        self.num_classes = num_classes
+        self.use_ema_update = cfg['ema_update']
+        # loss weight
+        self.loss_cls_weight = cfg['loss_cls_weight']
+        self.loss_box_weight = cfg['loss_box_weight']
+        self.loss_dfl_weight = cfg['loss_dfl_weight']
+        # matcher
+        matcher_config = cfg['matcher']
+        self.tal_matcher = TaskAlignedAssigner(
+            topk=matcher_config['tal']['topk'],
+            alpha=matcher_config['tal']['alpha'],
+            beta=matcher_config['tal']['beta'],
+            num_classes=num_classes
+            )
+        self.ota_matcher = AlignedSimOTA(
+            center_sampling_radius=matcher_config['ota']['center_sampling_radius'],
+            topk_candidate=matcher_config['ota']['topk_candidate'],
+            num_classes=num_classes
+        )
+
+    def __call__(self, outputs, targets, epoch=0):
+        if epoch < self.args.wp_epoch:
+            return self.ota_loss(outputs, targets)
+        else:
+            return self.tal_loss(outputs, targets)
+
+    def ema_update(self, name: str, value, initial_value, momentum=0.9):
+        if hasattr(self, name):
+            old = getattr(self, name)
+        else:
+            old = initial_value
+        new = old * momentum + value * (1 - momentum)
+        setattr(self, name, new)
+        return new
+
+    # ----------------- Loss functions -----------------
+    def loss_classes(self, pred_cls, gt_score, gt_label=None, vfl=False):
+        if vfl:
+            assert gt_label is not None
+            # compute varifocal loss
+            alpha, gamma = 0.75, 2.0
+            focal_weight = alpha * pred_cls.sigmoid().pow(gamma) * (1 - gt_label) + gt_score * gt_label
+            bce_loss = F.binary_cross_entropy_with_logits(pred_cls, gt_score, reduction='none')
+            loss_cls = bce_loss * focal_weight
+        else:
+            # compute bce loss
+            loss_cls = F.binary_cross_entropy_with_logits(pred_cls, gt_score, reduction='none')
+
+        return loss_cls
+
+    def loss_bboxes(self, pred_box, gt_box, bbox_weight=None):
+        # regression loss
+        ious = get_ious(pred_box, gt_box, 'xyxy', 'giou')
+        loss_box = 1.0 - ious
+
+        if bbox_weight is not None:
+            loss_box *= bbox_weight
+
+        return loss_box
+
+    def loss_dfl(self, pred_reg, gt_box, anchor, stride, bbox_weight=None):
+        # rescale coords by stride
+        gt_box_s = gt_box / stride
+        anchor_s = anchor / stride
+
+        # compute deltas
+        gt_ltrb_s = bbox2dist(anchor_s, gt_box_s, self.cfg['reg_max'] - 1)
+
+        gt_left = gt_ltrb_s.to(torch.long)
+        gt_right = gt_left + 1
+
+        weight_left = gt_right.to(torch.float) - gt_ltrb_s
+        weight_right = 1 - weight_left
+
+        # loss left
+        loss_left = F.cross_entropy(
+            pred_reg.view(-1, self.cfg['reg_max']),
+            gt_left.view(-1),
+            reduction='none').view(gt_left.shape) * weight_left
+        # loss right
+        loss_right = F.cross_entropy(
+            pred_reg.view(-1, self.cfg['reg_max']),
+            gt_right.view(-1),
+            reduction='none').view(gt_left.shape) * weight_right
+
+        loss_dfl = (loss_left + loss_right).mean(-1)
+        
+        if bbox_weight is not None:
+            loss_dfl *= bbox_weight
+
+        return loss_dfl
+    
+    # ----------------- Loss with TAL assigner -----------------
+    def tal_loss(self, outputs, targets):
+        """ Compute loss with TAL assigner """
+        bs = outputs['pred_cls'][0].shape[0]
+        device = outputs['pred_cls'][0].device
+        anchors = torch.cat(outputs['anchors'], dim=0)
+        num_anchors = anchors.shape[0]
+        # preds: [B, M, C]
+        cls_preds = torch.cat(outputs['pred_cls'], dim=1)
+        reg_preds = torch.cat(outputs['pred_reg'], dim=1)
+        box_preds = torch.cat(outputs['pred_box'], dim=1)
+
+        # --------------- label assignment ---------------
+        gt_label_targets = []
+        gt_score_targets = []
+        gt_bbox_targets = []
+        fg_masks = []
+        for batch_idx in range(bs):
+            tgt_labels = targets[batch_idx]["labels"].to(device)
+            tgt_bboxes = targets[batch_idx]["boxes"].to(device)
+
+            # check target
+            if len(tgt_labels) == 0 or tgt_bboxes.max().item() == 0.:
+                # There is no valid gt
+                fg_mask = cls_preds.new_zeros(1, num_anchors).bool()               #[1, M,]
+                gt_label = cls_preds.new_zeros((1, num_anchors,))                  #[1, M,]
+                gt_score = cls_preds.new_zeros((1, num_anchors, self.num_classes)) #[1, M, C]
+                gt_box = cls_preds.new_zeros((1, num_anchors, 4))                  #[1, M, 4]
+            else:
+                tgt_labels = tgt_labels[None, :, None]      # [1, Mp, 1]
+                tgt_bboxes = tgt_bboxes[None]                   # [1, Mp, 4]
+                (
+                    gt_label,   #[1, M]
+                    gt_box,     #[1, M, 4]
+                    gt_score,   #[1, M, C]
+                    fg_mask,    #[1, M,]
+                    _
+                ) = self.tal_matcher(
+                    pd_scores = cls_preds[batch_idx:batch_idx+1].detach().sigmoid(), 
+                    pd_bboxes = box_preds[batch_idx:batch_idx+1].detach(),
+                    anc_points = anchors,
+                    gt_labels = tgt_labels,
+                    gt_bboxes = tgt_bboxes
+                    )
+            gt_label_targets.append(gt_label)
+            gt_score_targets.append(gt_score)
+            gt_bbox_targets.append(gt_box)
+            fg_masks.append(fg_mask)
+
+        # List[B, 1, M, C] -> Tensor[B, M, C] -> Tensor[BM, C]
+        fg_masks = torch.cat(fg_masks, 0).view(-1)                                    # [BM,]
+        gt_score_targets = torch.cat(gt_score_targets, 0).view(-1, self.num_classes)  # [BM, C]
+        gt_bbox_targets = torch.cat(gt_bbox_targets, 0).view(-1, 4)                   # [BM, 4]
+        gt_label_targets = torch.cat(gt_label_targets, 0).view(-1)                    # [BM,]
+        gt_label_targets = torch.where(fg_masks > 0, gt_label_targets, torch.full_like(gt_label_targets, self.num_classes))
+        gt_labels_one_hot = F.one_hot(gt_label_targets.long(), self.num_classes + 1)[..., :-1]
+        bbox_weight = gt_score_targets[fg_masks].sum(-1)
+        num_fgs = max(gt_score_targets.sum(), 1)
+
+        # average loss normalizer across all the GPUs
+        if is_dist_avail_and_initialized():
+            torch.distributed.all_reduce(num_fgs)
+        num_fgs = max(num_fgs / get_world_size(), 1.0)
+
+        # update loss normalizer with EMA
+        if self.use_ema_update:
+            normalizer = self.ema_update("loss_normalizer", max(num_fgs, 1), 100)
+        else:
+            normalizer = num_fgs
+
+        # ------------------ Classification loss ------------------
+        cls_preds = cls_preds.view(-1, self.num_classes)
+        loss_cls = self.loss_classes(cls_preds, gt_score_targets, gt_labels_one_hot, vfl=False)
+        loss_cls = loss_cls.sum() / normalizer
+
+        # ------------------ Regression loss ------------------
+        box_preds_pos = box_preds.view(-1, 4)[fg_masks]
+        box_targets_pos = gt_bbox_targets[fg_masks]
+        loss_box = self.loss_bboxes(box_preds_pos, box_targets_pos, bbox_weight)
+        loss_box = loss_box.sum() / normalizer
+
+        # ------------------ Distribution focal loss  ------------------
+        ## process anchors
+        anchors = anchors[None].repeat(bs, 1, 1).view(-1, 2)
+        ## process stride tensors
+        strides = torch.cat(outputs['stride_tensor'], dim=0)
+        strides = strides.unsqueeze(0).repeat(bs, 1, 1).view(-1, 1)
+        ## fg preds
+        reg_preds_pos = reg_preds.view(-1, 4*self.cfg['reg_max'])[fg_masks]
+        anchors_pos = anchors[fg_masks]
+        strides_pos = strides[fg_masks]
+        ## compute dfl
+        loss_dfl = self.loss_dfl(reg_preds_pos, box_targets_pos, anchors_pos, strides_pos, bbox_weight)
+        loss_dfl = loss_dfl.sum() / normalizer
+
+        # total loss
+        losses = self.loss_cls_weight * loss_cls + \
+                 self.loss_box_weight * loss_box + \
+                 self.loss_dfl_weight * loss_dfl
+
+        loss_dict = dict(
+                loss_cls = loss_cls,
+                loss_box = loss_box,
+                loss_dfl = loss_dfl,
+                losses = losses
+        )
+
+        return loss_dict
+    
+    # ----------------- Loss with SimOTA assigner -----------------
+    def ota_loss(self, outputs, targets):
+        """ Compute loss with SimOTA assigner """
+        bs = outputs['pred_cls'][0].shape[0]
+        device = outputs['pred_cls'][0].device
+        fpn_strides = outputs['strides']
+        anchors = outputs['anchors']
+        num_anchors = sum([ab.shape[0] for ab in anchors])
+        # preds: [B, M, C]
+        cls_preds = torch.cat(outputs['pred_cls'], dim=1)
+        reg_preds = torch.cat(outputs['pred_reg'], dim=1)
+        box_preds = torch.cat(outputs['pred_box'], dim=1)
+
+        # --------------- label assignment ---------------
+        cls_targets = []
+        box_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.:
+                # There is no valid gt
+                cls_target = cls_preds.new_zeros((num_anchors, self.num_classes))
+                box_target = cls_preds.new_zeros((0, 4))
+                fg_mask = cls_preds.new_zeros(num_anchors).bool()
+            else:
+                (
+                    fg_mask,
+                    assigned_labels,
+                    assigned_ious,
+                    assigned_indexs
+                ) = self.ota_matcher(
+                    fpn_strides = fpn_strides,
+                    anchors = anchors,
+                    pred_cls = cls_preds[batch_idx], 
+                    pred_box = box_preds[batch_idx],
+                    tgt_labels = tgt_labels,
+                    tgt_bboxes = tgt_bboxes
+                    )
+                # prepare cls targets
+                assigned_labels = F.one_hot(assigned_labels.long(), self.num_classes)
+                assigned_labels = assigned_labels * assigned_ious.unsqueeze(-1)
+                cls_target = assigned_labels.new_zeros((num_anchors, self.num_classes))
+                cls_target[fg_mask] = assigned_labels
+                # prepare box targets
+                box_target = tgt_bboxes[assigned_indexs]
+
+            cls_targets.append(cls_target)
+            box_targets.append(box_target)
+            fg_masks.append(fg_mask)
+
+        cls_targets = torch.cat(cls_targets, 0)
+        box_targets = torch.cat(box_targets, 0)
+        fg_masks = torch.cat(fg_masks, 0)
+        num_fgs = fg_masks.sum()
+
+        # average loss normalizer across all the GPUs
+        if is_dist_avail_and_initialized():
+            torch.distributed.all_reduce(num_fgs)
+        num_fgs = (num_fgs / get_world_size()).clamp(1.0)
+
+        # update loss normalizer with EMA
+        if self.use_ema_update:
+            normalizer = self.ema_update("loss_normalizer", max(num_fgs, 1), 100)
+        else:
+            normalizer = num_fgs
+        
+        # ------------------ Classification loss ------------------
+        cls_preds = cls_preds.view(-1, self.num_classes)
+        loss_cls = self.loss_classes(cls_preds, cls_targets)
+        loss_cls = loss_cls.sum() / normalizer
+
+        # ------------------ 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() / normalizer
+
+        # ------------------ Distribution focal loss  ------------------
+        ## process anchors
+        anchors = torch.cat(anchors, dim=0)
+        anchors = anchors[None].repeat(bs, 1, 1).view(-1, 2)
+        ## process stride tensors
+        strides = torch.cat(outputs['stride_tensor'], dim=0)
+        strides = strides.unsqueeze(0).repeat(bs, 1, 1).view(-1, 1)
+        ## fg preds
+        reg_preds_pos = reg_preds.view(-1, 4*self.cfg['reg_max'])[fg_masks]
+        anchors_pos = anchors[fg_masks]
+        strides_pos = strides[fg_masks]
+        ## compute dfl
+        loss_dfl = self.loss_dfl(reg_preds_pos, box_targets, anchors_pos, strides_pos)
+        loss_dfl = loss_dfl.sum() / normalizer
+
+        # total loss
+        losses = self.loss_cls_weight * loss_cls + \
+                 self.loss_box_weight * loss_box + \
+                 self.loss_dfl_weight * loss_dfl
+
+        loss_dict = dict(
+                loss_cls = loss_cls,
+                loss_box = loss_box,
+                loss_dfl = loss_dfl,
+                losses = losses
+        )
+
+        return loss_dict
+
+
+def build_criterion(args, cfg, device, num_classes):
+    criterion = Criterion(
+        args=args,
+        cfg=cfg,
+        device=device,
+        num_classes=num_classes
+        )
+
+    return criterion
+
+
+if __name__ == "__main__":
+    pass

models/detectors/rtmdet_v2/matcher.py

@@ -0,0 +1,372 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from utils.box_ops import box_iou, bbox_iou
+
+
+# -------------------------- Basic Functions --------------------------
+def select_candidates_in_gts(xy_centers, gt_bboxes, eps=1e-9):
+    """select the positive anchors's center in gt
+    Args:
+        xy_centers (Tensor): shape(bs*n_max_boxes, num_total_anchors, 4)
+        gt_bboxes (Tensor): shape(bs, n_max_boxes, 4)
+    Return:
+        (Tensor): shape(bs, n_max_boxes, num_total_anchors)
+    """
+    n_anchors = xy_centers.size(0)
+    bs, n_max_boxes, _ = gt_bboxes.size()
+    _gt_bboxes = gt_bboxes.reshape([-1, 4])
+    xy_centers = xy_centers.unsqueeze(0).repeat(bs * n_max_boxes, 1, 1)
+    gt_bboxes_lt = _gt_bboxes[:, 0:2].unsqueeze(1).repeat(1, n_anchors, 1)
+    gt_bboxes_rb = _gt_bboxes[:, 2:4].unsqueeze(1).repeat(1, n_anchors, 1)
+    b_lt = xy_centers - gt_bboxes_lt
+    b_rb = gt_bboxes_rb - xy_centers
+    bbox_deltas = torch.cat([b_lt, b_rb], dim=-1)
+    bbox_deltas = bbox_deltas.reshape([bs, n_max_boxes, n_anchors, -1])
+    return (bbox_deltas.min(axis=-1)[0] > eps).to(gt_bboxes.dtype)
+
+def select_highest_overlaps(mask_pos, overlaps, n_max_boxes):
+    """if an anchor box is assigned to multiple gts,
+        the one with the highest iou will be selected.
+    Args:
+        mask_pos (Tensor): shape(bs, n_max_boxes, num_total_anchors)
+        overlaps (Tensor): shape(bs, n_max_boxes, num_total_anchors)
+    Return:
+        target_gt_idx (Tensor): shape(bs, num_total_anchors)
+        fg_mask (Tensor): shape(bs, num_total_anchors)
+        mask_pos (Tensor): shape(bs, n_max_boxes, num_total_anchors)
+    """
+    fg_mask = mask_pos.sum(axis=-2)
+    if fg_mask.max() > 1:
+        mask_multi_gts = (fg_mask.unsqueeze(1) > 1).repeat([1, n_max_boxes, 1])
+        max_overlaps_idx = overlaps.argmax(axis=1)
+        is_max_overlaps = F.one_hot(max_overlaps_idx, n_max_boxes)
+        is_max_overlaps = is_max_overlaps.permute(0, 2, 1).to(overlaps.dtype)
+        mask_pos = torch.where(mask_multi_gts, is_max_overlaps, mask_pos)
+        fg_mask = mask_pos.sum(axis=-2)
+    target_gt_idx = mask_pos.argmax(axis=-2)
+    return target_gt_idx, fg_mask , mask_pos
+
+def iou_calculator(box1, box2, eps=1e-9):
+    """Calculate iou for batch
+    Args:
+        box1 (Tensor): shape(bs, n_max_boxes, 1, 4)
+        box2 (Tensor): shape(bs, 1, num_total_anchors, 4)
+    Return:
+        (Tensor): shape(bs, n_max_boxes, num_total_anchors)
+    """
+    box1 = box1.unsqueeze(2)  # [N, M1, 4] -> [N, M1, 1, 4]
+    box2 = box2.unsqueeze(1)  # [N, M2, 4] -> [N, 1, M2, 4]
+    px1y1, px2y2 = box1[:, :, :, 0:2], box1[:, :, :, 2:4]
+    gx1y1, gx2y2 = box2[:, :, :, 0:2], box2[:, :, :, 2:4]
+    x1y1 = torch.maximum(px1y1, gx1y1)
+    x2y2 = torch.minimum(px2y2, gx2y2)
+    overlap = (x2y2 - x1y1).clip(0).prod(-1)
+    area1 = (px2y2 - px1y1).clip(0).prod(-1)
+    area2 = (gx2y2 - gx1y1).clip(0).prod(-1)
+    union = area1 + area2 - overlap + eps
+
+    return overlap / union
+
+
+# -------------------------- Task Aligned Assigner --------------------------
+class TaskAlignedAssigner(nn.Module):
+    def __init__(self, topk=10, alpha=0.5, beta=6.0, eps=1e-9, num_classes=80):
+        super(TaskAlignedAssigner, self).__init__()
+        self.topk = topk
+        self.num_classes = num_classes
+        self.bg_idx = num_classes
+        self.alpha = alpha
+        self.beta = beta
+        self.eps = eps
+
+    @torch.no_grad()
+    def forward(self,
+                pd_scores,
+                pd_bboxes,
+                anc_points,
+                gt_labels,
+                gt_bboxes):
+        """This code referenced to
+           https://github.com/Nioolek/PPYOLOE_pytorch/blob/master/ppyoloe/assigner/tal_assigner.py
+        Args:
+            pd_scores (Tensor): shape(bs, num_total_anchors, num_classes)
+            pd_bboxes (Tensor): shape(bs, num_total_anchors, 4)
+            anc_points (Tensor): shape(num_total_anchors, 2)
+            gt_labels (Tensor): shape(bs, n_max_boxes, 1)
+            gt_bboxes (Tensor): shape(bs, n_max_boxes, 4)
+        Returns:
+            target_labels (Tensor): shape(bs, num_total_anchors)
+            target_bboxes (Tensor): shape(bs, num_total_anchors, 4)
+            target_scores (Tensor): shape(bs, num_total_anchors, num_classes)
+            fg_mask (Tensor): shape(bs, num_total_anchors)
+        """
+        self.bs = pd_scores.size(0)
+        self.n_max_boxes = gt_bboxes.size(1)
+
+        mask_pos, align_metric, overlaps = self.get_pos_mask(
+            pd_scores, pd_bboxes, gt_labels, gt_bboxes, anc_points)
+
+        target_gt_idx, fg_mask, mask_pos = select_highest_overlaps(
+            mask_pos, overlaps, self.n_max_boxes)
+
+        # assigned target
+        target_labels, target_bboxes, target_scores = self.get_targets(
+            gt_labels, gt_bboxes, target_gt_idx, fg_mask)
+
+        # normalize
+        align_metric *= mask_pos
+        pos_align_metrics = align_metric.amax(axis=-1, keepdim=True)  # b, max_num_obj
+        pos_overlaps = (overlaps * mask_pos).amax(axis=-1, keepdim=True)  # b, max_num_obj
+        norm_align_metric = (align_metric * pos_overlaps / (pos_align_metrics + self.eps)).amax(-2).unsqueeze(-1)
+        target_scores = target_scores * norm_align_metric
+
+        return target_labels, target_bboxes, target_scores, fg_mask.bool(), target_gt_idx
+
+
+    def get_pos_mask(self, pd_scores, pd_bboxes, gt_labels, gt_bboxes, anc_points):
+        # get anchor_align metric, (b, max_num_obj, h*w)
+        align_metric, overlaps = self.get_box_metrics(pd_scores, pd_bboxes, gt_labels, gt_bboxes)
+        # get in_gts mask, (b, max_num_obj, h*w)
+        mask_in_gts = select_candidates_in_gts(anc_points, gt_bboxes)
+        # get topk_metric mask, (b, max_num_obj, h*w)
+        mask_topk = self.select_topk_candidates(align_metric * mask_in_gts)
+        # merge all mask to a final mask, (b, max_num_obj, h*w)
+        mask_pos = mask_topk * mask_in_gts
+
+        return mask_pos, align_metric, overlaps
+
+
+    def get_box_metrics(self, pd_scores, pd_bboxes, gt_labels, gt_bboxes):
+        ind = torch.zeros([2, self.bs, self.n_max_boxes], dtype=torch.long)  # 2, b, max_num_obj
+        ind[0] = torch.arange(end=self.bs).view(-1, 1).repeat(1, self.n_max_boxes)  # b, max_num_obj
+        ind[1] = gt_labels.long().squeeze(-1)  # b, max_num_obj
+        # get the scores of each grid for each gt cls
+        bbox_scores = pd_scores[ind[0], :, ind[1]]  # b, max_num_obj, h*w
+
+        overlaps = bbox_iou(gt_bboxes.unsqueeze(2), pd_bboxes.unsqueeze(1), xywh=False).squeeze(3).clamp(0)
+        align_metric = bbox_scores.pow(self.alpha) * overlaps.pow(self.beta)
+
+        return align_metric, overlaps
+
+
+    def select_topk_candidates(self, metrics, largest=True):
+        """
+        Args:
+            metrics: (b, max_num_obj, h*w).
+            topk_mask: (b, max_num_obj, topk) or None
+        """
+
+        num_anchors = metrics.shape[-1]  # h*w
+        # (b, max_num_obj, topk)
+        topk_metrics, topk_idxs = torch.topk(metrics, self.topk, dim=-1, largest=largest)
+        topk_mask = (topk_metrics.max(-1, keepdim=True)[0] > self.eps).tile([1, 1, self.topk])
+        # (b, max_num_obj, topk)
+        topk_idxs[~topk_mask] = 0
+        # (b, max_num_obj, topk, h*w) -> (b, max_num_obj, h*w)
+        is_in_topk = F.one_hot(topk_idxs, num_anchors).sum(-2)
+        # filter invalid bboxes
+        is_in_topk = torch.where(is_in_topk > 1, 0, is_in_topk)
+        return is_in_topk.to(metrics.dtype)
+
+
+    def get_targets(self, gt_labels, gt_bboxes, target_gt_idx, fg_mask):
+        """
+        Args:
+            gt_labels: (b, max_num_obj, 1)
+            gt_bboxes: (b, max_num_obj, 4)
+            target_gt_idx: (b, h*w)
+            fg_mask: (b, h*w)
+        """
+
+        # assigned target labels, (b, 1)
+        batch_ind = torch.arange(end=self.bs, dtype=torch.int64, device=gt_labels.device)[..., None]
+        target_gt_idx = target_gt_idx + batch_ind * self.n_max_boxes  # (b, h*w)
+        target_labels = gt_labels.long().flatten()[target_gt_idx]  # (b, h*w)
+
+        # assigned target boxes, (b, max_num_obj, 4) -> (b, h*w)
+        target_bboxes = gt_bboxes.view(-1, 4)[target_gt_idx]
+
+        # assigned target scores
+        target_labels.clamp(0)
+        target_scores = F.one_hot(target_labels, self.num_classes)  # (b, h*w, 80)
+        fg_scores_mask = fg_mask[:, :, None].repeat(1, 1, self.num_classes)  # (b, h*w, 80)
+        target_scores = torch.where(fg_scores_mask > 0, target_scores, 0)
+
+        return target_labels, target_bboxes, target_scores
+    
+
+# -------------------------- Aligned SimOTA Assigner --------------------------
+class AlignedSimOTA(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_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)
+        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] -> [N, Mp, C]
+            score_preds = cls_preds.sigmoid_().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
+    

models/detectors/rtmdet_v2/rtmdet_v2.py

@@ -0,0 +1,176 @@
+# --------------- Torch components ---------------
+import torch
+import torch.nn as nn
+
+# --------------- Model components ---------------
+from .rtmdet_v2_backbone import build_backbone
+from .rtmdet_v2_neck import build_neck
+from .rtmdet_v2_pafpn import build_fpn
+from .rtmdet_v2_head import build_det_head
+from .rtmdet_v2_pred import build_pred_layer
+
+# --------------- External components ---------------
+from utils.misc import multiclass_nms
+
+
+# My RTMDet
+class RTMDet(nn.Module):
+    def __init__(self, 
+                 cfg,
+                 device, 
+                 num_classes = 20, 
+                 conf_thresh = 0.05,
+                 nms_thresh = 0.6,
+                 trainable = False, 
+                 topk = 1000,
+                 deploy = False):
+        super(RTMDet, self).__init__()
+        # ---------------------- Basic Parameters ----------------------
+        self.cfg = cfg
+        self.device = device
+        self.stride = cfg['stride']
+        self.reg_max = cfg['reg_max']
+        self.num_classes = num_classes
+        self.trainable = trainable
+        self.conf_thresh = conf_thresh
+        self.nms_thresh = nms_thresh
+        self.topk = topk
+        self.deploy = deploy
+        self.head_dim = round(256*cfg['width'])
+        
+        # ---------------------- Network Parameters ----------------------
+        ## ----------- Backbone -----------
+        self.backbone, feats_dim = build_backbone(cfg, trainable&cfg['pretrained'])
+
+        ## ----------- Neck: SPP -----------
+        self.neck = build_neck(cfg, feats_dim[-1], feats_dim[-1])
+        feats_dim[-1] = self.neck.out_dim
+        
+        ## ----------- Neck: FPN -----------
+        self.fpn = build_fpn(cfg, feats_dim, round(256*cfg['width']))
+        self.fpn_dims = self.fpn.out_dim
+
+        ## ----------- Heads -----------
+        self.det_heads = build_det_head(
+            cfg, self.fpn_dims, self.head_dim, num_classes, num_levels=len(self.stride))
+
+        ## ----------- Preds -----------
+        self.pred_layers = build_pred_layer(
+            self.det_heads.cls_head_dim, self.det_heads.reg_head_dim,
+            self.stride, num_classes, num_coords=4, num_levels=len(self.stride))
+
+
+    ## post-process
+    def post_process(self, cls_preds, box_preds):
+        """
+        Input:
+            cls_preds: List(Tensor) [[H x W, C], ...]
+            box_preds: List(Tensor) [[H x W, 4], ...]
+        """
+        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]
+            
+            # (H x W x C,)
+            scores_i = cls_pred_i.sigmoid().flatten()
+
+            # Keep top k top scoring indices only.
+            num_topk = min(self.topk, 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)
+        labels = torch.cat(all_labels)
+        bboxes = torch.cat(all_bboxes)
+
+        # 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, False)
+
+        return bboxes, scores, labels
+
+
+    # ---------------------- Main Process for Inference ----------------------
+    @torch.no_grad()
+    def inference_single_image(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.det_heads(pyramid_feats)
+
+        # ---------------- Preds ----------------
+        outputs = self.pred_layers(cls_feats, reg_feats)
+
+        all_cls_preds = outputs['pred_cls']
+        all_box_preds = outputs['pred_box']
+
+        if self.deploy:
+            cls_preds = torch.cat(all_cls_preds, dim=1)[0]
+            box_preds = torch.cat(all_box_preds, dim=1)[0]
+            scores = cls_preds.sigmoid()
+            bboxes = box_preds
+            # [n_anchors_all, 4 + C]
+            outputs = torch.cat([bboxes, scores], dim=-1)
+
+            return outputs
+        else:
+            # post process
+            bboxes, scores, labels = self.post_process(all_cls_preds, all_box_preds)
+        
+            return bboxes, scores, labels
+
+
+    def forward(self, x):
+        if not self.trainable:
+            return self.inference_single_image(x)
+        else:
+            # ---------------- 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.det_heads(pyramid_feats)
+
+            # ---------------- Preds ----------------
+            outputs = self.pred_layers(cls_feats, reg_feats)
+            
+            return outputs 
+        

models/detectors/rtmdet_v2/rtmdet_v2_backbone.py

@@ -0,0 +1,161 @@
+import torch
+import torch.nn as nn
+try:
+    from .rtmdet_v2_basic import Conv, MCBlock, DSBlock
+except:
+    from rtmdet_v2_basic import Conv, MCBlock, DSBlock
+
+
+
+model_urls = {
+    'mcnet_p': None,
+    'mcnet_n': None,
+    'mcnet_t': None,
+    'mcnet_s': None,
+    'mcnet_m': None,
+    'mcnet_l': None,
+    'mcnet_x': None,
+}
+
+
+# ---------------------------- Backbones ----------------------------
+class MixedConvNet(nn.Module):
+    def __init__(self, width=1.0, depth=1.0, num_heads=4, act_type='silu', norm_type='BN', depthwise=False):
+        super(MixedConvNet, self).__init__()
+        # ------------------ Basic parameters ------------------
+        self.feat_dims_base = [64, 128, 256, 512, 1024]
+        self.nblocks_base = [3, 6, 9, 3]
+        self.feat_dims = [round(dim * width) for dim in self.feat_dims_base]
+        self.nblocks = [round(nblock * depth) for nblock in self.nblocks_base]
+        self.num_heads = num_heads
+        self.act_type = act_type
+        self.norm_type = norm_type
+        self.depthwise = depthwise
+        
+        # ------------------ Network parameters ------------------
+        ## P1/2
+        self.layer_1 = nn.Sequential(
+            Conv(3, self.feat_dims[0], k=3, p=1, s=2, act_type=self.act_type, norm_type=self.norm_type),
+            Conv(self.feat_dims[0], self.feat_dims[0], k=3, p=1, act_type=self.act_type, norm_type=self.norm_type, depthwise=self.depthwise),
+        )
+        ## P2/4
+        self.layer_2 = nn.Sequential(   
+            Conv(self.feat_dims[0], self.feat_dims[1], k=3, p=1, s=2, act_type=self.act_type, norm_type=self.norm_type),
+            MCBlock(self.feat_dims[1], self.feat_dims[1], self.nblocks[0], self.num_heads, True, self.act_type, self.norm_type, self.depthwise)
+        )
+        ## P3/8
+        self.layer_3 = nn.Sequential(
+            DSBlock(self.feat_dims[1], self.feat_dims[2], self.num_heads, self.act_type, self.norm_type, self.depthwise),             
+            MCBlock(self.feat_dims[2], self.feat_dims[2], self.nblocks[1], self.num_heads, True, self.act_type, self.norm_type, self.depthwise)
+        )
+        ## P4/16
+        self.layer_4 = nn.Sequential(
+            DSBlock(self.feat_dims[2], self.feat_dims[3], self.num_heads, self.act_type, self.norm_type, self.depthwise),             
+            MCBlock(self.feat_dims[3], self.feat_dims[3], self.nblocks[2], self.num_heads, True, self.act_type, self.norm_type, self.depthwise)
+        )
+        ## P5/32
+        self.layer_5 = nn.Sequential(
+            DSBlock(self.feat_dims[3], self.feat_dims[4], self.num_heads, self.act_type, self.norm_type, self.depthwise),             
+            MCBlock(self.feat_dims[4], self.feat_dims[4], self.nblocks[3], self.num_heads, True, self.act_type, self.norm_type, self.depthwise)
+        )
+
+
+    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
+
+
+# ---------------------------- Functions ----------------------------
+## load pretrained weight
+def load_weight(model, model_name):
+    # load weight
+    print('Loading pretrained weight ...')
+    url = model_urls[model_name]
+    if url is not None:
+        checkpoint = torch.hub.load_state_dict_from_url(
+            url=url, map_location="cpu", check_hash=True)
+        # checkpoint state dict
+        checkpoint_state_dict = checkpoint.pop("model")
+        # model state dict
+        model_state_dict = model.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(k)
+
+        model.load_state_dict(checkpoint_state_dict)
+    else:
+        print('No pretrained for {}'.format(model_name))
+
+    return model
+
+
+## build MCNet
+def build_backbone(cfg, pretrained=False):
+    # model
+    backbone = MixedConvNet(cfg['width'], cfg['depth'], cfg['bk_num_heads'], cfg['bk_act'], cfg['bk_norm'], cfg['bk_depthwise'])
+
+    # check whether to load imagenet pretrained weight
+    if pretrained:
+        if cfg['width'] == 0.25 and cfg['depth'] == 0.34 and cfg['bk_depthwise']:
+            backbone = load_weight(backbone, model_name='mcnet_p')
+        elif cfg['width'] == 0.25 and cfg['depth'] == 0.34:
+            backbone = load_weight(backbone, model_name='mcnet_n')
+        elif cfg['width'] == 0.375 and cfg['depth'] == 0.34:
+            backbone = load_weight(backbone, model_name='mcnet_t')
+        elif cfg['width'] == 0.5 and cfg['depth'] == 0.34:
+            backbone = load_weight(backbone, model_name='mcnet_s')
+        elif cfg['width'] == 0.75 and cfg['depth'] == 0.67:
+            backbone = load_weight(backbone, model_name='mcnet_m')
+        elif cfg['width'] == 1.0 and cfg['depth'] == 1.0:
+            backbone = load_weight(backbone, model_name='mcnet_l')
+        elif cfg['width'] == 1.25 and cfg['depth'] == 1.34:
+            backbone = load_weight(backbone, model_name='mcnet_x')
+    feat_dims = backbone.feat_dims[-3:]
+
+    return backbone, feat_dims
+
+
+if __name__ == '__main__':
+    import time
+    from thop import profile
+    cfg = {
+        ## Backbone
+        'backbone': 'mcnet',
+        'pretrained': True,
+        'bk_act': 'silu',
+        'bk_norm': 'BN',
+        'bk_depthwise': False,
+        'bk_num_heads': 4,
+        'width': 0.25,
+        'depth': 0.34,
+        'stride': [8, 16, 32],  # P3, P4, P5
+        'max_stride': 32,
+    }
+    model, feats = build_backbone(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)
+
+    print('==============================')
+    flops, params = profile(model, inputs=(x, ), verbose=False)
+    print('==============================')
+    print('GFLOPs : {:.2f}'.format(flops / 1e9 * 2))
+    print('Params : {:.2f} M'.format(params / 1e6))

models/detectors/rtmdet_v2/rtmdet_v2_basic.py

@@ -0,0 +1,210 @@
+import numpy as np
+import torch
+import torch.nn as nn
+
+
+# ---------------------------- Base Conv Module ----------------------------
+class SiLU(nn.Module):
+    """export-friendly version of nn.SiLU()"""
+
+    @staticmethod
+    def forward(x):
+        return x * torch.sigmoid(x)
+
+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()
+
+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)
+
+class Conv(nn.Module):
+    def __init__(self, 
+                 c1,                   # in channels
+                 c2,                   # out channels 
+                 k=1,                  # kernel size 
+                 p=0,                  # padding
+                 s=1,                  # padding
+                 d=1,                  # dilation
+                 act_type='lrelu',     # activation
+                 norm_type='BN',       # normalization
+                 depthwise=False):
+        super(Conv, self).__init__()
+        convs = []
+        add_bias = False if norm_type else True
+        p = p if d == 1 else d
+        if depthwise:
+            convs.append(get_conv2d(c1, c1, k=k, p=p, s=s, d=d, g=c1, bias=add_bias))
+            # depthwise conv
+            if norm_type:
+                convs.append(get_norm(norm_type, c1))
+            if act_type:
+                convs.append(get_activation(act_type))
+            # pointwise conv
+            convs.append(get_conv2d(c1, c2, k=1, p=0, s=1, d=d, g=1, bias=add_bias))
+            if norm_type:
+                convs.append(get_norm(norm_type, c2))
+            if act_type:
+                convs.append(get_activation(act_type))
+
+        else:
+            convs.append(get_conv2d(c1, c2, k=k, p=p, s=s, d=d, g=1, bias=add_bias))
+            if norm_type:
+                convs.append(get_norm(norm_type, c2))
+            if act_type:
+                convs.append(get_activation(act_type))
+            
+        self.convs = nn.Sequential(*convs)
+
+
+    def forward(self, x):
+        return self.convs(x)
+
+
+# ---------------------------- Base Modules ----------------------------
+## Multi-head Mixed Conv (MHMC)
+class MultiHeadMixedConv(nn.Module):
+    def __init__(self, in_dim, out_dim, num_heads=4, shortcut=False, act_type='silu', norm_type='BN', depthwise=False):
+        super().__init__()
+        # -------------- Basic parameters --------------
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+        self.num_heads = num_heads
+        self.shortcut = shortcut
+        self.head_dim = in_dim // num_heads
+        # -------------- Network parameters --------------
+        ## Scale Modulation
+        self.mixed_convs = nn.ModuleList()
+        for i in range(num_heads):
+            self.mixed_convs.append(
+                Conv(self.head_dim, self.head_dim, k=2*i+1, p=i, act_type=act_type, norm_type=norm_type, depthwise=depthwise)
+            )
+        ## Out-proj
+        self.out_proj = Conv(in_dim, out_dim, k=1, act_type=act_type, norm_type=norm_type)
+
+
+    def forward(self, x):
+        xs = torch.chunk(x, self.num_heads, dim=1)
+        ys = [mixed_conv(x_h) for x_h, mixed_conv in zip(xs, self.mixed_convs)]
+        out = self.out_proj(torch.cat(ys, dim=1))
+
+        return out + x if self.shortcut else out
+
+# ---------------------------- Base Blocks ----------------------------
+## Mixed Convolution Block
+class MCBlock(nn.Module):
+    def __init__(self, in_dim, out_dim, nblocks=1, num_heads=4, shortcut=False, act_type='silu', norm_type='BN', depthwise=False):
+        super().__init__()
+        # -------------- Basic parameters --------------
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+        self.nblocks = nblocks
+        self.num_heads = num_heads
+        self.shortcut = shortcut
+        self.inter_dim = in_dim // 2
+        # -------------- Network parameters --------------
+        ## branch-1
+        self.cv1 = Conv(self.in_dim, self.inter_dim, k=1, act_type=act_type, norm_type=norm_type)
+        self.cv2 = Conv(self.in_dim, self.inter_dim, k=1, act_type=act_type, norm_type=norm_type)
+        ## branch-2
+        self.smblocks = nn.Sequential(*[
+            MultiHeadMixedConv(self.inter_dim, self.inter_dim, self.num_heads, self.shortcut, act_type, norm_type, depthwise)
+            for _ in range(nblocks)])
+        ## out proj
+        self.out_proj = Conv(self.inter_dim*2, out_dim, k=1, act_type=act_type, norm_type=norm_type)
+
+
+    def forward(self, x):
+        # branch-1
+        x1 = self.cv1(x)
+        # branch-2
+        x2 = self.smblocks(self.cv2(x))
+        # output
+        out = torch.cat([x1, x2], dim=1)
+        out = self.out_proj(out)
+
+        return out
+
+## DownSample Block
+class DSBlock(nn.Module):
+    def __init__(self, in_dim, out_dim, num_heads=4, act_type='silu', norm_type='BN', depthwise=False):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+        self.inter_dim = out_dim // 2
+        self.num_heads = num_heads
+        # branch-1
+        self.maxpool = nn.Sequential(
+            Conv(in_dim, self.inter_dim, k=1, act_type=act_type, norm_type=norm_type),
+            nn.MaxPool2d((2, 2), 2)
+        )
+        # branch-2
+        self.ds_conv = nn.Sequential(
+            Conv(in_dim, self.inter_dim, k=1, act_type=act_type, norm_type=norm_type),
+            Conv(self.inter_dim, self.inter_dim, k=3, p=1, s=2, act_type=act_type, norm_type=norm_type, depthwise=depthwise)
+        ) 
+
+
+    def forward(self, x):
+        # branch-1
+        x1 = self.maxpool(x)
+        # branch-2
+        x2 = self.ds_conv(x)
+        # out-proj
+        out = torch.cat([x1, x2], dim=1)
+
+        return out
+
+
+# ---------------------------- FPN Modules ----------------------------
+## build fpn's core block
+def build_fpn_block(cfg, in_dim, out_dim):
+    if cfg['fpn_core_block'] == 'mcblock':
+        layer = MCBlock(in_dim=in_dim,
+                        out_dim=out_dim,
+                        nblocks=round(cfg['depth'] * 3),
+                        num_heads=cfg['fpn_num_heads'],
+                        shortcut=False,
+                        act_type=cfg['fpn_act'],
+                        norm_type=cfg['fpn_norm'],
+                        depthwise=cfg['fpn_depthwise']
+                        )
+        
+    return layer
+
+## build fpn's reduce layer
+def build_reduce_layer(cfg, in_dim, out_dim):
+    if cfg['fpn_reduce_layer'] == 'conv':
+        layer = Conv(in_dim, out_dim, k=1, act_type=cfg['fpn_act'], norm_type=cfg['fpn_norm'])
+        
+    return layer
+
+## build fpn's downsample layer
+def build_downsample_layer(cfg, in_dim, out_dim):
+    if cfg['fpn_downsample_layer'] == 'conv':
+        layer = Conv(in_dim, out_dim, k=3, s=2, p=1,
+                     act_type=cfg['fpn_act'], norm_type=cfg['fpn_norm'], depthwise=cfg['fpn_depthwise'])
+    elif cfg['fpn_downsample_layer'] == 'maxpool':
+        assert in_dim == out_dim
+        layer = nn.MaxPool2d((2, 2), stride=2)
+    elif cfg['fpn_downsample_layer'] == 'dsblock':
+        layer = DSBlock(in_dim, out_dim, num_heads=cfg['fpn_num_heads'],
+                        act_type=cfg['fpn_act'], norm_type=cfg['fpn_norm'], depthwise=cfg['fpn_depthwise'])
+        
+    return layer

models/detectors/rtmdet_v2/rtmdet_v2_head.py

@@ -0,0 +1,117 @@
+import torch
+import torch.nn as nn
+
+from .rtmdet_v2_basic import Conv
+
+
+# Single-level Head
+class SingleLevelHead(nn.Module):
+    def __init__(self, in_dim, out_dim, num_classes, num_cls_head, num_reg_head, act_type, norm_type, depthwise):
+        super().__init__()
+        # --------- Basic Parameters ----------
+        self.in_dim = in_dim
+        self.num_classes = num_classes
+        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 = max(out_dim, num_classes)
+        for i in range(num_cls_head):
+            if i == 0:
+                cls_feats.append(
+                    Conv(in_dim, self.cls_head_dim, k=3, p=1, s=1, 
+                         act_type=act_type,
+                         norm_type=norm_type,
+                         depthwise=depthwise)
+                        )
+            else:
+                cls_feats.append(
+                    Conv(self.cls_head_dim, self.cls_head_dim, k=3, p=1, s=1, 
+                        act_type=act_type,
+                        norm_type=norm_type,
+                        depthwise=depthwise)
+                        )      
+        ## reg head
+        reg_feats = []
+        self.reg_head_dim = out_dim
+        for i in range(num_reg_head):
+            if i == 0:
+                reg_feats.append(
+                    Conv(in_dim, self.reg_head_dim, k=3, p=1, s=1, 
+                         act_type=act_type,
+                         norm_type=norm_type,
+                         depthwise=depthwise)
+                        )
+            else:
+                reg_feats.append(
+                    Conv(self.reg_head_dim, self.reg_head_dim, k=3, p=1, s=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)
+
+
+    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 Head
+class MultiLevelHead(nn.Module):
+    def __init__(self, cfg, in_dims, out_dim, num_classes=80, num_levels=3):
+        super().__init__()
+        ## ----------- Network Parameters -----------
+        self.multi_level_heads = nn.ModuleList(
+            [SingleLevelHead(
+                in_dims[level],
+                out_dim,
+                num_classes,
+                cfg['num_cls_head'],
+                cfg['num_reg_head'],
+                cfg['head_act'],
+                cfg['head_norm'],
+                cfg['head_depthwise'])
+                for level in range(num_levels)
+            ])
+        # --------- Basic Parameters ----------
+        self.in_dims = in_dims
+        self.num_classes = num_classes
+
+        self.cls_head_dim = self.multi_level_heads[0].cls_head_dim
+        self.reg_head_dim = self.multi_level_heads[0].reg_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
+    
+
+# build detection head
+def build_det_head(cfg, in_dim, out_dim, num_classes=80, num_levels=3):
+    if cfg['head'] == 'decoupled_head':
+        head = MultiLevelHead(cfg, in_dim, out_dim, num_classes, num_levels) 
+
+    return head

models/detectors/rtmdet_v2/rtmdet_v2_neck.py

@@ -0,0 +1,72 @@
+import torch
+import torch.nn as nn
+
+from .rtmdet_v2_basic import Conv
+
+
+# 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, expand_ratio=0.5):
+        super().__init__()
+        inter_dim = int(in_dim * expand_ratio)
+        self.out_dim = out_dim
+        self.cv1 = Conv(in_dim, inter_dim, k=1, act_type=cfg['neck_act'], norm_type=cfg['neck_norm'])
+        self.cv2 = Conv(inter_dim * 4, out_dim, k=1, act_type=cfg['neck_act'], norm_type=cfg['neck_norm'])
+        self.m = nn.MaxPool2d(kernel_size=cfg['pooling_size'], stride=1, padding=cfg['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, expand_ratio):
+        super(SPPFBlockCSP, self).__init__()
+        inter_dim = int(in_dim * expand_ratio)
+        self.out_dim = out_dim
+        self.cv1 = Conv(in_dim, inter_dim, k=1, act_type=cfg['neck_act'], norm_type=cfg['neck_norm'])
+        self.cv2 = Conv(in_dim, inter_dim, k=1, act_type=cfg['neck_act'], norm_type=cfg['neck_norm'])
+        self.m = nn.Sequential(
+            Conv(inter_dim, inter_dim, k=3, p=1, 
+                 act_type=cfg['neck_act'], norm_type=cfg['neck_norm'], 
+                 depthwise=cfg['neck_depthwise']),
+            SPPF(cfg, inter_dim, inter_dim, expand_ratio=1.0),
+            Conv(inter_dim, inter_dim, k=3, p=1, 
+                 act_type=cfg['neck_act'], norm_type=cfg['neck_norm'], 
+                 depthwise=cfg['neck_depthwise'])
+        )
+        self.cv3 = Conv(inter_dim * 2, self.out_dim, k=1, act_type=cfg['neck_act'], norm_type=cfg['neck_norm'])
+
+        
+    def forward(self, x):
+        x1 = self.cv1(x)
+        x2 = self.cv2(x)
+        x3 = self.m(x2)
+        y = self.cv3(torch.cat([x1, x3], dim=1))
+
+        return y
+
+
+# build neck
+def build_neck(cfg, in_dim, out_dim):
+    model = cfg['neck']
+    print('==============================')
+    print('Neck: {}'.format(model))
+    # build neck
+    if model == 'sppf':
+        neck = SPPF(cfg, in_dim, out_dim, cfg['neck_expand_ratio'])
+    elif model == 'csp_sppf':
+        neck = SPPFBlockCSP(cfg, in_dim, out_dim, cfg['neck_expand_ratio'])
+
+    return neck
+        

models/detectors/rtmdet_v2/rtmdet_v2_pafpn.py

@@ -0,0 +1,90 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .rtmdet_v2_basic import (Conv, build_reduce_layer, build_downsample_layer, build_fpn_block)
+
+
+# RTMDet-Style PaFPN
+class RTMDetPaFPN(nn.Module):
+    def __init__(self, cfg, in_dims=[256, 512, 1024], out_dim=None):
+        super(RTMDetPaFPN, self).__init__()
+        # --------------------------- Basic Parameters ---------------------------
+        self.in_dims = in_dims
+        self.fpn_dims = in_dims
+        
+        # --------------------------- Top-down FPN---------------------------
+        ## P5 -> P4
+        self.reduce_layer_1 = build_reduce_layer(cfg, self.fpn_dims[2], self.fpn_dims[2]//2)
+        self.top_down_layer_1 = build_fpn_block(cfg, self.fpn_dims[1] + self.fpn_dims[2]//2, self.fpn_dims[1])
+
+        ## P4 -> P3
+        self.reduce_layer_2 = build_reduce_layer(cfg, self.fpn_dims[1], self.fpn_dims[1]//2)
+        self.top_down_layer_2 = build_fpn_block(cfg, self.fpn_dims[0] + self.fpn_dims[1]//2, self.fpn_dims[0])
+
+        # --------------------------- Bottom-up FPN ---------------------------
+        ## P3 -> P4
+        self.downsample_layer_1 = build_downsample_layer(cfg, self.fpn_dims[0], self.fpn_dims[0])
+        self.bottom_up_layer_1 = build_fpn_block(cfg, self.fpn_dims[0] + self.fpn_dims[1]//2, self.fpn_dims[1])
+
+        ## P4 -> P5
+        self.downsample_layer_2 = build_downsample_layer(cfg, self.fpn_dims[1], self.fpn_dims[1])
+        self.bottom_up_layer_2 = build_fpn_block(cfg, self.fpn_dims[1] + self.fpn_dims[2]//2, self.fpn_dims[2])
+                
+        # --------------------------- Output proj ---------------------------
+        if out_dim is not None:
+            self.out_layers = nn.ModuleList([
+                Conv(in_dim, out_dim, k=1,
+                     act_type=cfg['fpn_act'], norm_type=cfg['fpn_norm'])
+                     for in_dim in self.fpn_dims
+                     ])
+            self.out_dim = [out_dim] * 3
+        else:
+            self.out_layers = None
+            self.out_dim = self.fpn_dims
+
+
+    def forward(self, fpn_feats):
+        c3, c4, c5 = fpn_feats
+
+        # Top down
+        ## P5 -> P4
+        c6 = self.reduce_layer_1(c5)
+        c7 = F.interpolate(c6, scale_factor=2.0)
+        c8 = torch.cat([c7, c4], dim=1)
+        c9 = self.top_down_layer_1(c8)
+        ## P4 -> P3
+        c10 = self.reduce_layer_2(c9)
+        c11 = F.interpolate(c10, scale_factor=2.0)
+        c12 = torch.cat([c11, c3], dim=1)
+        c13 = self.top_down_layer_2(c12)
+
+        # Bottom up
+        ## p3 -> P4
+        c14 = self.downsample_layer_1(c13)
+        c15 = torch.cat([c14, c10], dim=1)
+        c16 = self.bottom_up_layer_1(c15)
+        ## P4 -> P5
+        c17 = self.downsample_layer_2(c16)
+        c18 = torch.cat([c17, c6], dim=1)
+        c19 = self.bottom_up_layer_2(c18)
+
+        out_feats = [c13, c16, c19] # [P3, P4, P5]
+        
+        # output proj layers
+        if self.out_layers is not None:
+            out_feats_proj = []
+            for feat, layer in zip(out_feats, self.out_layers):
+                out_feats_proj.append(layer(feat))
+            return out_feats_proj
+
+        return out_feats
+
+
+def build_fpn(cfg, in_dims, out_dim=None):
+    model = cfg['fpn']
+    # build pafpn
+    if model == 'rtmdet_pafpn':
+        fpn_net = RTMDetPaFPN(cfg, in_dims, out_dim)
+
+    return fpn_net

models/detectors/rtmdet_v2/rtmdet_v2_pred.py

@@ -0,0 +1,153 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+# Single-level pred layer
+class SingleLevelPredLayer(nn.Module):
+    def __init__(self, cls_dim, reg_dim, num_classes, num_coords=4):
+        super().__init__()
+        # --------- Basic Parameters ----------
+        self.cls_dim = cls_dim
+        self.reg_dim = reg_dim
+        self.num_classes = num_classes
+        self.num_coords = num_coords
+
+        # --------- Network Parameters ----------
+        self.cls_pred = nn.Conv2d(cls_dim, num_classes, kernel_size=1)
+        self.reg_pred = nn.Conv2d(reg_dim, num_coords, kernel_size=1)                
+
+        self.init_bias()
+        
+
+    def init_bias(self):
+        # Init bias
+        init_prob = 0.01
+        bias_value = -torch.log(torch.tensor((1. - init_prob) / init_prob))
+        # cls pred
+        b = self.cls_pred.bias.view(1, -1)
+        b.data.fill_(bias_value.item())
+        self.cls_pred.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+        # reg pred
+        b = self.reg_pred.bias.view(-1, )
+        b.data.fill_(1.0)
+        self.reg_pred.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+        w = self.reg_pred.weight
+        w.data.fill_(0.)
+        self.reg_pred.weight = torch.nn.Parameter(w, requires_grad=True)
+
+
+    def forward(self, cls_feat, reg_feat):
+        """
+            in_feats: (Tensor) [B, C, H, W]
+        """
+        cls_pred = self.cls_pred(cls_feat)
+        reg_pred = self.reg_pred(reg_feat)
+
+        return cls_pred, reg_pred
+    
+
+# Multi-level pred layer
+class MultiLevelPredLayer(nn.Module):
+    def __init__(self, cls_dim, reg_dim, strides, num_classes, num_coords=4, num_levels=3, reg_max=16):
+        super().__init__()
+        # --------- Basic Parameters ----------
+        self.cls_dim = cls_dim
+        self.reg_dim = reg_dim
+        self.strides = strides
+        self.num_classes = num_classes
+        self.num_coords = num_coords
+        self.num_levels = num_levels
+        self.reg_max = reg_max
+
+        # ----------- Network Parameters -----------
+        ## pred layers
+        self.multi_level_preds = nn.ModuleList(
+            [SingleLevelPredLayer(
+                cls_dim,
+                reg_dim,
+                num_classes,
+                num_coords * self.reg_max)
+                for _ in range(num_levels)
+            ])
+        ## proj conv
+        self.proj = nn.Parameter(torch.linspace(0, reg_max, reg_max), requires_grad=False)
+        self.proj_conv = nn.Conv2d(self.reg_max, 1, kernel_size=1, bias=False)
+        self.proj_conv.weight = nn.Parameter(self.proj.view([1, reg_max, 1, 1]).clone().detach(), requires_grad=False)
+
+
+    def generate_anchors(self, level, fmp_size):
+        """
+            fmp_size: (List) [H, W]
+        """
+        # generate grid cells
+        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 += 0.5  # add center offset
+        anchors *= self.strides[level]
+
+        return anchors
+        
+
+    def forward(self, cls_feats, reg_feats):
+        all_anchors = []
+        all_strides = []
+        all_cls_preds = []
+        all_reg_preds = []
+        all_box_preds = []
+        for level in range(self.num_levels):
+            # pred
+            cls_pred, reg_pred = self.multi_level_preds[level](cls_feats[level], reg_feats[level])
+
+            # generate anchor boxes: [M, 4]
+            B, _, H, W = cls_pred.size()
+            fmp_size = [H, W]
+            anchors = self.generate_anchors(level, fmp_size)
+            anchors = anchors.to(cls_pred.device)
+            # stride tensor: [M, 1]
+            stride_tensor = torch.ones_like(anchors[..., :1]) * self.strides[level]
+            
+            # [B, C, H, W] -> [B, H, W, C] -> [B, M, 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*self.reg_max)
+
+            # ----------------------- Decode bbox -----------------------
+            B, M = reg_pred.shape[:2]
+            # [B, M, 4*(reg_max)] -> [B, M, 4, reg_max] -> [B, 4, M, reg_max]
+            reg_pred_ = reg_pred.reshape([B, M, 4, self.reg_max])
+            # [B, M, 4, reg_max] -> [B, reg_max, 4, M]
+            reg_pred_ = reg_pred_.permute(0, 3, 2, 1).contiguous()
+            # [B, reg_max, 4, M] -> [B, 1, 4, M]
+            reg_pred_ = self.proj_conv(F.softmax(reg_pred_, dim=1))
+            # [B, 1, 4, M] -> [B, 4, M] -> [B, M, 4]
+            reg_pred_ = reg_pred_.view(B, 4, M).permute(0, 2, 1).contiguous()
+            ## tlbr -> xyxy
+            x1y1_pred = anchors[None] - reg_pred_[..., :2] * self.strides[level]
+            x2y2_pred = anchors[None] + reg_pred_[..., 2:] * self.strides[level]
+            box_pred = torch.cat([x1y1_pred, x2y2_pred], dim=-1)
+
+            all_cls_preds.append(cls_pred)
+            all_reg_preds.append(reg_pred)
+            all_box_preds.append(box_pred)
+            all_anchors.append(anchors)
+            all_strides.append(stride_tensor)
+        
+        # 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]
+                   "anchors": all_anchors,           # List(Tensor) [M, 2]
+                   "strides": self.strides,          # List(Int) = [8, 16, 32]
+                   "stride_tensor": all_strides      # List(Tensor) [M, 1]
+                   }
+
+        return outputs
+    
+
+# build detection head
+def build_pred_layer(cls_dim, reg_dim, strides, num_classes, num_coords=4, num_levels=3):
+    pred_layers = MultiLevelPredLayer(cls_dim, reg_dim, strides, num_classes, num_coords, num_levels) 
+
+    return pred_layers