train YOLOv8-S

config/__init__.py

@@ -93,8 +93,7 @@ from .model_config.yolov4_config import yolov4_cfg
 from .model_config.yolov5_config import yolov5_cfg, yolov5_adamw_cfg
 from .model_config.yolov7_config import yolov7_cfg
 from .model_config.yolov8_config import yolov8_cfg
-from .model_config.yolox_config import yolox_cfg, yolox_adamw_cfg
-from .model_config.rtcdet_config import rtcdet_cfg
+from .model_config.yolox_config  import yolox_cfg
 ## Real-time DETR series
 from .model_config.rtdetr_config import rtdetr_cfg
 from .model_config.rtpdetr_config import rtpdetr_cfg
@@ -129,12 +128,6 @@ 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]
-    # YOLOX-AdamW
-    elif args.model in ['yolox_n_adamw', 'yolox_s_adamw', 'yolox_m_adamw', 'yolox_l_adamw', 'yolox_x_adamw']:
-        cfg = yolox_adamw_cfg[args.model]
-    # RTCDet
-    elif args.model in ['rtcdet_p', 'rtcdet_n', 'rtcdet_t', 'rtcdet_s', 'rtcdet_m', 'rtcdet_l', 'rtcdet_x']:
-        cfg = rtcdet_cfg[args.model]
     # RT-DETR
     elif args.model in ['rtdetr_r18', 'rtdetr_r34', 'rtdetr_r50', 'rtdetr_r101']:
         cfg = rtdetr_cfg[args.model]

config/model_config/rtcdet_config.py

@@ -1,101 +0,0 @@
-# yolov8-v2 Config
-
-
-rtcdet_cfg = {
-    'rtcdet_n':{
-        # ---------------- Model config ----------------
-        ## Backbone
-        'bk_act': 'silu',
-        'bk_norm': 'BN',
-        'bk_depthwise': False,
-        'width': 0.25,
-        'depth': 0.34,
-        'stride': [8, 16, 32],  # P3, P4, P5
-        'max_stride': 32,
-        'reg_max': 16,
-        ## Neck: SPP
-        'neck': 'sppf',
-        'neck_expand_ratio': 0.5,
-        'pooling_size': 5,
-        'neck_act': 'silu',
-        'neck_norm': 'BN',
-        'neck_depthwise': False,
-        ## Neck: PaFPN
-        'fpn': 'rtc_pafpn',
-        '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,
-        # ---------------- Train config ----------------
-        ## Input
-        'multi_scale': [0.5, 1.5], # 320 -> 960
-        'trans_type': 'yolov5_s',
-        # ---------------- Assignment config ----------------
-        ## Matcher
-        'matcher': "simota",
-        'matcher_hpy': {'center_sampling_radius': 2.5,
-                        'topk_candidate': 10,
-                        },
-        # ---------------- Loss config ----------------
-        'loss_cls_weight': 0.5,
-        'loss_box_weight': 7.5,
-        'loss_dfl_weight': 1.5,
-        # ---------------- Train config ----------------
-        'trainer_type': 'rtcdet',
-    },
-
-    'rtcdet_s':{
-        # ---------------- Model config ----------------
-        ## Backbone
-        'bk_act': 'silu',
-        'bk_norm': 'BN',
-        'bk_depthwise': False,
-        'width': 0.50,
-        'depth': 0.34,
-        'stride': [8, 16, 32],  # P3, P4, P5
-        'max_stride': 32,
-        'reg_max': 16,
-        ## Neck: SPP
-        'neck': 'sppf',
-        'neck_expand_ratio': 0.5,
-        'pooling_size': 5,
-        'neck_act': 'silu',
-        'neck_norm': 'BN',
-        'neck_depthwise': False,
-        ## Neck: PaFPN
-        'fpn': 'rtc_pafpn',
-        '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,
-        # ---------------- Train config ----------------
-        ## Input
-        'multi_scale': [0.5, 1.5], # 320 -> 960
-        'trans_type': 'yolov5_s',
-        # ---------------- Assignment config ----------------
-        ## Matcher
-        'matcher': "simota",
-        'matcher_hpy': {'center_sampling_radius': 2.5,
-                        'topk_candidate': 10,
-                        },
-        # ---------------- Loss config ----------------
-        'loss_cls_weight': 0.5,
-        'loss_box_weight': 7.5,
-        'loss_dfl_weight': 1.5,
-        # ---------------- Train config ----------------
-        'trainer_type': 'rtcdet',
-    },
-
-}

config/model_config/yolox_config.py

@@ -217,221 +217,3 @@ yolox_cfg = {
         'trainer_type': 'yolox',
     },
 }
-
-
-yolox_adamw_cfg = {
-    'yolox_n_adamw':{
-        # ---------------- Model config ----------------
-        ## Backbone
-        'backbone': 'cspdarknet',
-        'bk_act': 'silu',
-        'bk_norm': 'BN',
-        'bk_dpw': False,
-        'width': 0.25,
-        'depth': 0.34,
-        'stride': [8, 16, 32],  # P3, P4, P5
-        'max_stride': 32,
-        ## FPN
-        'fpn': 'yolox_pafpn',
-        'fpn_reduce_layer': 'conv',
-        'fpn_downsample_layer': 'conv',
-        'fpn_core_block': 'cspblock',
-        '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,
-        # ---------------- Train config ----------------
-        ## input
-        'multi_scale': [0.5, 1.25],   # 320 -> 800
-        'trans_type': 'yolov5_n',
-        # ---------------- Assignment config ----------------
-        ## matcher
-        'matcher': {'center_sampling_radius': 2.5,
-                    'topk_candicate': 10},
-        # ---------------- Loss config ----------------
-        ## loss weight
-        'loss_obj_weight': 1.0,
-        'loss_cls_weight': 1.0,
-        'loss_box_weight': 5.0,
-        # ---------------- Train config ----------------
-        'trainer_type': 'rtcdet',
-    },
-
-    'yolox_s_adamw':{
-        # ---------------- Model config ----------------
-        ## Backbone
-        'backbone': 'cspdarknet',
-        'bk_act': 'silu',
-        'bk_norm': 'BN',
-        'bk_dpw': False,
-        'width': 0.50,
-        'depth': 0.34,
-        'stride': [8, 16, 32],  # P3, P4, P5
-        'max_stride': 32,
-        ## FPN
-        'fpn': 'yolox_pafpn',
-        'fpn_reduce_layer': 'conv',
-        'fpn_downsample_layer': 'conv',
-        'fpn_core_block': 'cspblock',
-        '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,
-        # ---------------- Train config ----------------
-        ## input
-        'multi_scale': [0.5, 1.25],   # 320 -> 800
-        'trans_type': 'yolov5_s',
-        # ---------------- Assignment config ----------------
-        ## matcher
-        'matcher': {'center_sampling_radius': 2.5,
-                    'topk_candicate': 10},
-        # ---------------- Loss config ----------------
-        ## loss weight
-        'loss_obj_weight': 1.0,
-        'loss_cls_weight': 1.0,
-        'loss_box_weight': 5.0,
-        # ---------------- Train config ----------------
-        'trainer_type': 'rtcdet',
-    },
-
-    'yolox_m_adamw':{
-        # ---------------- Model config ----------------
-        ## Backbone
-        'backbone': 'cspdarknet',
-        'bk_act': 'silu',
-        'bk_norm': 'BN',
-        'bk_dpw': False,
-        'width': 0.75,
-        'depth': 0.67,
-        'stride': [8, 16, 32],  # P3, P4, P5
-        'max_stride': 32,
-        ## FPN
-        'fpn': 'yolox_pafpn',
-        'fpn_reduce_layer': 'conv',
-        'fpn_downsample_layer': 'conv',
-        'fpn_core_block': 'cspblock',
-        '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,
-        # ---------------- Train config ----------------
-        ## input
-        'multi_scale': [0.5, 1.25],   # 320 -> 800
-        'trans_type': 'yolov5_m',
-        # ---------------- Assignment config ----------------
-        ## matcher
-        'matcher': {'center_sampling_radius': 2.5,
-                    'topk_candicate': 10},
-        # ---------------- Loss config ----------------
-        ## loss weight
-        'loss_obj_weight': 1.0,
-        'loss_cls_weight': 1.0,
-        'loss_box_weight': 5.0,
-        # ---------------- Train config ----------------
-        'trainer_type': 'rtcdet',
-    },
-
-    'yolox_l_adamw':{
-        # ---------------- Model config ----------------
-        ## Backbone
-        'backbone': 'cspdarknet',
-        'bk_act': 'silu',
-        'bk_norm': 'BN',
-        'bk_dpw': False,
-        'width': 1.0,
-        'depth': 1.0,
-        'stride': [8, 16, 32],  # P3, P4, P5
-        'max_stride': 32,
-        ## FPN
-        'fpn': 'yolox_pafpn',
-        'fpn_reduce_layer': 'conv',
-        'fpn_downsample_layer': 'conv',
-        'fpn_core_block': 'cspblock',
-        '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,
-        # ---------------- Train config ----------------
-        ## input
-        'multi_scale': [0.5, 1.25],   # 320 -> 800
-        'trans_type': 'yolov5_l',
-        # ---------------- Assignment config ----------------
-        ## matcher
-        'matcher': {'center_sampling_radius': 2.5,
-                    'topk_candicate': 10},
-        # ---------------- Loss config ----------------
-        ## loss weight
-        'loss_obj_weight': 1.0,
-        'loss_cls_weight': 1.0,
-        'loss_box_weight': 5.0,
-        # ---------------- Train config ----------------
-        'trainer_type': 'rtcdet',
-    },
-
-    'yolox_x_adamw':{
-        # ---------------- Model config ----------------
-        ## Backbone
-        'backbone': 'cspdarknet',
-        'bk_act': 'silu',
-        'bk_norm': 'BN',
-        'bk_dpw': False,
-        'width': 1.25,
-        'depth': 1.34,
-        'stride': [8, 16, 32],  # P3, P4, P5
-        'max_stride': 32,
-        ## FPN
-        'fpn': 'yolox_pafpn',
-        'fpn_reduce_layer': 'conv',
-        'fpn_downsample_layer': 'conv',
-        'fpn_core_block': 'cspblock',
-        '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,
-        # ---------------- Train config ----------------
-        ## input
-        'multi_scale': [0.5, 1.25],   # 320 -> 800
-        'trans_type': 'yolov5_x',
-        # ---------------- Assignment config ----------------
-        ## matcher
-        'matcher': {'center_sampling_radius': 2.5,
-                    'topk_candicate': 10},
-        # ---------------- Loss config ----------------
-        ## loss weight
-        'loss_obj_weight': 1.0,
-        'loss_cls_weight': 1.0,
-        'loss_box_weight': 5.0,
-        # ---------------- Train config ----------------
-        'trainer_type': 'rtcdet',
-    },
-}

models/detectors/__init__.py

@@ -11,7 +11,6 @@ from .yolov5.build import build_yolov5
 from .yolov7.build import build_yolov7
 from .yolov8.build import build_yolov8
 from .yolox.build import build_yolox
-from .rtcdet.build import build_rtcdet
 # Real-time DETR series
 from .rtdetr.build import build_rtdetr
 from .rtpdetr.build import build_rtpdetr
@@ -64,10 +63,6 @@ def build_model(args,
     elif args.model in ['yolox_n_adamw', 'yolox_s_adamw', 'yolox_m_adamw', 'yolox_l_adamw', 'yolox_x_adamw']:
         model, criterion = build_yolox(
             args, model_cfg, device, num_classes, trainable, deploy)
-    # RTCDet
-    elif args.model in ['rtcdet_p', 'rtcdet_n', 'rtcdet_t', 'rtcdet_s', 'rtcdet_m', 'rtcdet_l', 'rtcdet_x']:
-        model, criterion = build_rtcdet(
-            args, model_cfg, device, num_classes, trainable, deploy)
     # RT-DETR
     elif args.model in ['rtdetr_r18', 'rtdetr_r34', 'rtdetr_r50', 'rtdetr_r101']:
         model, criterion = build_rtdetr(

models/detectors/rtcdet/README.md

@@ -1,55 +0,0 @@
-# RTCDet:
-
-|   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 |
-|-----------|--------|-------|------------------------|-------------------|-------------------|--------------------|--------|
-| RTCDet-N  | 8xb16  |  640  |                        |                   |                   |                    |  |
-| RTCDet-S  | 8xb16  |  640  |                        |                   |                   |                    |  |
-| RTCDet-M  | 8xb16  |  640  |                        |                   |                   |                    |  |
-| RTCDet-L  | 8xb16  |  640  |                        |                   |                   |                    |  |
-
-- For training, we train RTCDet series with 500 epochs on COCO.
-- For data augmentation, we use the large scale jitter (LSJ), Mosaic augmentation and Mixup augmentation, following the setting of [RTCDet](https://github.com/ultralytics/RTCDet).
-- For optimizer, we use AdamW with weight decay 0.05 and base per image lr 0.001 / 64, which is different from the official RTCDet. We have tried SGD, but it has weakened performance. For example, when using SGD, RTCDet-N's AP was only 35.8%, lower than the current result (36.8 %), perhaps because some hyperparameters were not set properly.
-- For learning rate scheduler, we use linear decay scheduler.
-
-
-## Train RTCDet
-### Single GPU
-Taking training RTCDet-S on COCO as the example,
-```Shell
-python train.py --cuda -d coco --root path/to/coco -m rtcdet_s -bs 16 -size 640 --wp_epoch 3 --max_epoch 500 --eval_epoch 10 --no_aug_epoch 20 --ema --fp16 --multi_scale 
-```
-
-### Multi GPU
-Taking training RTCDet on COCO as the example,
-```Shell
-python -m torch.distributed.run --nproc_per_node=8 train.py --cuda -dist -d coco --root /data/datasets/ -m rtcdet_s -bs 128 -size 640 --wp_epoch 3 --max_epoch 500  --eval_epoch 10 --no_aug_epoch 20 --ema --fp16 --sybn --multi_scale --save_folder weights/ 
-```
-
-## Test RTCDet
-Taking testing RTCDet on COCO-val as the example,
-```Shell
-python test.py --cuda -d coco --root path/to/coco -m rtcdet_s --weight path/to/RTCDet.pth -size 640 -vt 0.4 --show 
-```
-
-## Evaluate RTCDet
-Taking evaluating RTCDet on COCO-val as the example,
-```Shell
-python eval.py --cuda -d coco-val --root path/to/coco -m rtcdet_s --weight path/to/RTCDet.pth 
-```
-
-## Demo
-### Detect with Image
-```Shell
-python demo.py --mode image --path_to_img path/to/image_dirs/ --cuda -m rtcdet_s --weight path/to/weight -size 640 -vt 0.4 --show
-```
-
-### Detect with Video
-```Shell
-python demo.py --mode video --path_to_vid path/to/video --cuda -m rtcdet_s --weight path/to/weight -size 640 -vt 0.4 --show --gif
-```
-
-### Detect with Camera
-```Shell
-python demo.py --mode camera --cuda -m rtcdet_s --weight path/to/weight -size 640 -vt 0.4 --show --gif
-```

models/detectors/rtcdet/build.py

@@ -1,44 +0,0 @@
-#!/usr/bin/env python3
-# -*- coding:utf-8 -*-
-
-import torch
-import torch.nn as nn
-
-from .loss import build_criterion
-from .rtcdet import RTCDet
-
-
-# build object detector
-def build_rtcdet(args, cfg, device, num_classes=80, trainable=False, deploy=False):
-    print('==============================')
-    print('Build {} ...'.format(args.model.upper()))
-    
-    print('==============================')
-    print('Model Configuration: \n', cfg)
-    
-    # -------------- Build YOLO --------------
-    model = RTCDet(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,
-                   no_multi_labels    = args.no_multi_labels,
-                   nms_class_agnostic = args.nms_class_agnostic
-                   )
-
-    # -------------- Initialize RTCDet --------------
-    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(cfg, device, num_classes)
-        
-    return model, criterion

models/detectors/rtcdet/loss.py

@@ -1,195 +0,0 @@
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-
-from utils.box_ops import bbox2dist, bbox_iou
-from utils.distributed_utils import get_world_size, is_dist_avail_and_initialized
-
-from .matcher import SimOTA
-
-
-class Criterion(object):
-    def __init__(self, cfg, device, num_classes=80):
-        # --------------- Basic parameters ---------------
-        self.cfg = cfg
-        self.device = device
-        self.num_classes = num_classes
-        self.reg_max = cfg['reg_max']
-        # --------------- Loss config ---------------
-        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 config ---------------
-        self.matcher_hpy = cfg['matcher_hpy']
-        self.matcher = SimOTA(center_sampling_radius = self.matcher_hpy['center_sampling_radius'],
-                              topk_candidate         = self.matcher_hpy['topk_candidate'],
-                              num_classes            = num_classes)
-
-    def loss_classes(self, pred_cls, gt_score):
-        # 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):
-        # regression loss
-        ious = bbox_iou(pred_box, gt_box, xywh=False, CIoU=True)
-        loss_box = (1.0 - ious.squeeze(-1)) * 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
-
-    def __call__(self, outputs, targets, epoch=0):        
-        """
-            outputs['pred_cls']: List(Tensor) [B, M, C]
-            outputs['pred_reg']: List(Tensor) [B, M, 4*(reg_max+1)]
-            outputs['pred_box']: List(Tensor) [B, M, 4]
-            outputs['anchors']: List(Tensor) [M, 2]
-            outputs['strides']: List(Int) [8, 16, 32] output stride
-            outputs['stride_tensor']: List(Tensor) [M, 1]
-            targets: (List) [dict{'boxes': [...], 
-                                 'labels': [...], 
-                                 'orig_size': ...}, ...]
-        """
-        bs = outputs['pred_cls'][0].shape[0]
-        device = outputs['pred_cls'][0].device
-        anchors = outputs['anchors']
-        fpn_strides = outputs['strides']
-
-        # 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)
-        num_anchors = box_preds.shape[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.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 = cls_targets.sum()
-        
-        # Average loss normalizer across all the GPUs
-        if is_dist_avail_and_initialized():
-            torch.distributed.all_reduce(num_fgs)
-        num_fgs = (num_fgs / get_world_size()).clamp(1.0)
-
-        # ------------------ Classification loss ------------------
-        cls_preds = cls_preds.view(-1, self.num_classes)
-        loss_cls = self.loss_classes(cls_preds, cls_targets)
-        loss_cls = loss_cls.sum() / num_fgs
-
-        # ------------------ Regression loss ------------------
-        box_preds_pos = box_preds.view(-1, 4)[fg_masks]
-        bbox_weight = cls_targets[fg_masks].sum(-1)
-        loss_box = self.loss_bboxes(box_preds_pos, box_targets, bbox_weight)
-        loss_box = loss_box.sum() / num_fgs
-
-        # ------------------ 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, bbox_weight)
-        loss_dfl = loss_dfl.sum() / num_fgs
-
-        # total loss
-        losses = loss_cls * self.loss_cls_weight + loss_box * self.loss_box_weight + loss_dfl * self.loss_dfl_weight
-        loss_dict = dict(
-                loss_cls = loss_cls,
-                loss_box = loss_box,
-                loss_dfl = loss_dfl,
-                losses = losses
-        )
-
-        return loss_dict
-    
-
-def build_criterion(cfg, device, num_classes):
-    criterion = Criterion(
-        cfg=cfg,
-        device=device,
-        num_classes=num_classes
-        )
-
-    return criterion
-
-
-if __name__ == "__main__":
-    pass

models/detectors/rtcdet/matcher.py

@@ -1,179 +0,0 @@
-import torch
-import torch.nn.functional as F
-from utils.box_ops import box_iou
-
-
-# -------------------------- YOLOX's SimOTA Assigner --------------------------
-## Simple OTA
-class SimOTA(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]
-            cls_preds_expand = cls_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, cls_preds_expand.size(1), 1)
-            cls_targets *= pair_wise_ious.unsqueeze(-1)  # iou-aware
-            # [N, Mp]
-            cls_cost = F.binary_cross_entropy_with_logits(cls_preds_expand, cls_targets, reduction="none").sum(-1)
-        del cls_preds_expand
-
-        #----------------------- 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/rtcdet/rtcdet.py

@@ -1,183 +0,0 @@
-# --------------- Torch components ---------------
-import torch
-import torch.nn as nn
-
-# --------------- Model components ---------------
-from .rtcdet_backbone import build_backbone
-from .rtcdet_neck     import build_neck
-from .rtcdet_pafpn    import build_fpn
-from .rtcdet_head     import build_head
-from .rtcdet_pred     import build_predictor
-
-# --------------- External components ---------------
-from utils.misc import multiclass_nms
-
-
-# YOLOv8
-class RTCDet(nn.Module):
-    def __init__(self,
-                 cfg,
-                 device,
-                 num_classes = 80,
-                 conf_thresh = 0.01,
-                 nms_thresh  = 0.5,
-                 topk        = 1000,
-                 trainable   = False,
-                 deploy      = False,
-                 no_multi_labels    = False,
-                 nms_class_agnostic = False):
-        super(RTCDet, self).__init__()
-        # ---------------------- Basic Parameters ----------------------
-        self.cfg = cfg
-        self.device = device
-        self.strides = 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.num_levels = len(self.strides)
-        self.num_classes = num_classes
-        self.topk_candidates = topk
-        self.deploy = deploy
-        self.no_multi_labels = no_multi_labels
-        self.nms_class_agnostic = nms_class_agnostic
-        
-        # ---------------------- Network Parameters ----------------------
-        ## ----------- Backbone -----------
-        self.backbone, feat_dims = build_backbone(cfg)
-
-        ## ----------- Neck: SPP -----------
-        self.neck = build_neck(cfg, feat_dims[-1], feat_dims[-1])
-        feat_dims[-1] = self.neck.out_dim
-        
-        ## ----------- Neck: FPN -----------
-        self.fpn = build_fpn(cfg, feat_dims, round(256 * cfg['width']))
-        self.fpn_dims = self.fpn.out_dims
-
-        ## ----------- Heads -----------
-        self.det_heads = build_head(cfg, self.fpn_dims, self.num_levels, num_classes, self.reg_max)
-        cls_head_dim = self.det_heads.cls_head_dim
-        reg_head_dim = self.det_heads.reg_head_dim
-
-        ## ----------- Preds -----------
-        self.pred_layers = build_predictor(cls_head_dim, reg_head_dim, self.strides, num_classes, 4, self.num_levels, self.reg_max)
-
-    ## post-process
-    def post_process(self, cls_preds, box_preds):
-        """
-        Input:
-            cls_preds: List[np.array] -> [[M, C], ...]
-            box_preds: List[np.array] -> [[M, 4], ...]
-        Output:
-            bboxes: np.array -> [N, 4]
-            scores: np.array -> [N,]
-            labels: np.array -> [N,]
-        """
-        assert len(cls_preds) == self.num_levels
-        all_scores = []
-        all_labels = []
-        all_bboxes = []
-        
-        for cls_pred_i, box_pred_i in zip(cls_preds, box_preds):
-            cls_pred_i = cls_pred_i[0]
-            box_pred_i = box_pred_i[0]
-            if self.no_multi_labels:
-                # [M,]
-                scores, labels = torch.max(cls_pred_i.sigmoid(), dim=1)
-
-                # Keep top k top scoring indices only.
-                num_topk = min(self.topk_candidates, box_pred_i.size(0))
-
-                # topk candidates
-                predicted_prob, topk_idxs = scores.sort(descending=True)
-                topk_scores = predicted_prob[:num_topk]
-                topk_idxs = topk_idxs[:num_topk]
-
-                # filter out the proposals with low confidence score
-                keep_idxs = topk_scores > self.conf_thresh
-                scores = topk_scores[keep_idxs]
-                topk_idxs = topk_idxs[keep_idxs]
-
-                labels = labels[topk_idxs]
-                bboxes = box_pred_i[topk_idxs]
-            else:
-                # [M, C] -> [MC,]
-                scores_i = cls_pred_i.sigmoid().flatten()
-
-                # Keep top k top scoring indices only.
-                num_topk = min(self.topk_candidates, box_pred_i.size(0))
-
-                # torch.sort is actually faster than .topk (at least on GPUs)
-                predicted_prob, topk_idxs = scores_i.sort(descending=True)
-                topk_scores = predicted_prob[:num_topk]
-                topk_idxs = topk_idxs[:num_topk]
-
-                # filter out the proposals with low confidence score
-                keep_idxs = topk_scores > self.conf_thresh
-                scores = topk_scores[keep_idxs]
-                topk_idxs = topk_idxs[keep_idxs]
-
-                anchor_idxs = torch.div(topk_idxs, self.num_classes, rounding_mode='floor')
-                labels = topk_idxs % self.num_classes
-
-                bboxes = box_pred_i[anchor_idxs]
-
-            all_scores.append(scores)
-            all_labels.append(labels)
-            all_bboxes.append(bboxes)
-
-        scores = torch.cat(all_scores, dim=0)
-        labels = torch.cat(all_labels, dim=0)
-        bboxes = torch.cat(all_bboxes, dim=0)
-
-        # to cpu & numpy
-        scores = scores.cpu().numpy()
-        labels = labels.cpu().numpy()
-        bboxes = bboxes.cpu().numpy()
-
-        # nms
-        scores, labels, bboxes = multiclass_nms(
-            scores, labels, bboxes, self.nms_thresh, self.num_classes, self.nms_class_agnostic)
-
-        return bboxes, scores, labels
-    
-    # ---------------------- Main Process for Inference ----------------------
-    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.det_heads(pyramid_feats)
-
-        # ---------------- Preds ----------------
-        outputs = self.pred_layers(cls_feats, reg_feats)
-
-        if not self.training:
-            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)
-
-            else:
-                # post process
-                bboxes, scores, labels = self.post_process(all_cls_preds, all_box_preds)
-                outputs = {
-                    "scores": scores,
-                    "labels": labels,
-                    "bboxes": bboxes
-                }
-        
-        return outputs 

models/detectors/rtcdet/rtcdet_backbone.py

@@ -1,133 +0,0 @@
-import torch
-import torch.nn as nn
-
-try:
-    from .rtcdet_basic import BasicConv, RTCBlock
-except:
-    from  rtcdet_basic import BasicConv, RTCBlock
-
-
-# ---------------------------- Basic functions ----------------------------
-## YOLOv8's backbone
-class RTCBackbone(nn.Module):
-    def __init__(self, width=1.0, depth=1.0, act_type='silu', norm_type='BN', depthwise=False):
-        super(RTCBackbone, self).__init__()
-        self.feat_dims = [round(64 * width), round(128 * width), round(256 * width), round(512 * width), round(1024 * width)]
-        # P1/2
-        self.layer_1 = BasicConv(3, self.feat_dims[0], kernel_size=6, padding=2, stride=2, act_type=act_type, norm_type=norm_type)
-        # P2/4
-        self.layer_2 = nn.Sequential(
-            BasicConv(self.feat_dims[0], self.feat_dims[1],
-                      kernel_size=3, padding=1, stride=2,
-                      act_type=act_type, norm_type=norm_type, depthwise=depthwise),
-            RTCBlock(in_dim     = self.feat_dims[1],
-                     out_dim    = self.feat_dims[1],
-                     num_blocks = round(3*depth),
-                     shortcut   = True,
-                     act_type   = act_type,
-                     norm_type  = norm_type,
-                     depthwise  = depthwise)
-        )
-        # P3/8
-        self.layer_3 = nn.Sequential(
-            BasicConv(self.feat_dims[1], self.feat_dims[2],
-                      kernel_size=3, padding=1, stride=2,
-                      act_type=act_type, norm_type=norm_type, depthwise=depthwise),
-            RTCBlock(in_dim     = self.feat_dims[2],
-                     out_dim    = self.feat_dims[2],
-                     num_blocks = round(9*depth),
-                     shortcut   = True,
-                     act_type   = act_type,
-                     norm_type  = norm_type,
-                     depthwise  = depthwise)
-        )
-        # P4/16
-        self.layer_4 = nn.Sequential(
-            BasicConv(self.feat_dims[2], self.feat_dims[3],
-                      kernel_size=3, padding=1, stride=2,
-                      act_type=act_type, norm_type=norm_type, depthwise=depthwise),
-            RTCBlock(in_dim     = self.feat_dims[3],
-                     out_dim    = self.feat_dims[3],
-                     num_blocks = round(9*depth),
-                     shortcut   = True,
-                     act_type   = act_type,
-                     norm_type  = norm_type,
-                     depthwise  = depthwise)
-        )
-        # P5/32
-        self.layer_5 = nn.Sequential(
-            BasicConv(self.feat_dims[3], self.feat_dims[4],
-                      kernel_size=3, padding=1, stride=2,
-                      act_type=act_type, norm_type=norm_type, depthwise=depthwise),
-            RTCBlock(in_dim     = self.feat_dims[4],
-                     out_dim    = self.feat_dims[4],
-                     num_blocks = round(3*depth),
-                     shortcut   = True,
-                     act_type   = act_type,
-                     norm_type  = norm_type,
-                     depthwise  = depthwise)
-        )
-
-        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):
-        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 ----------------------------
-## build Yolov8's Backbone
-def build_backbone(cfg): 
-    # model
-    backbone = RTCBackbone(width=cfg['width'],
-                           depth=cfg['depth'],
-                           act_type=cfg['bk_act'],
-                           norm_type=cfg['bk_norm'],
-                           depthwise=cfg['bk_depthwise']
-                           )
-    feat_dims = backbone.feat_dims[-3:]
-        
-    return backbone, feat_dims
-
-
-if __name__ == '__main__':
-    import time
-    from thop import profile
-    cfg = {
-        'bk_act': 'silu',
-        'bk_norm': 'BN',
-        'bk_depthwise': False,
-        'width': 1.0,
-        'depth': 1.0,
-        'ratio': 1.0,
-    }
-    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)
-
-    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))

models/detectors/rtcdet/rtcdet_basic.py

@@ -1,128 +0,0 @@
-import torch
-import torch.nn as nn
-
-
-# --------------------- Basic modules ---------------------
-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, g, bias=False):
-    conv = nn.Conv2d(c1, c2, k, stride=s, padding=p, 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 :int = 1,       # kernel size 
-                 padding     :int = 0,       # padding
-                 stride      :int = 1,       # padding
-                 act_type    :str = 'silu',  # activation
-                 norm_type   :str = 'BN',    # normalization
-                 depthwise   :bool = False,
-                ):
-        super(BasicConv, self).__init__()
-        self.depthwise = depthwise
-        add_bias = False if norm_type else True
-        if not depthwise:
-            self.conv = get_conv2d(in_dim, out_dim, k=kernel_size, p=padding, s=stride, g=1, bias=add_bias)
-            self.norm = get_norm(norm_type, out_dim)
-            self.act  = get_activation(act_type)
-        else:
-            self.conv1 = get_conv2d(in_dim, in_dim, k=kernel_size, p=padding, s=stride, g=in_dim, bias=add_bias)
-            self.norm1 = get_norm(norm_type, in_dim)
-            self.conv2 = get_conv2d(in_dim, out_dim, k=1, d=0, s=1, g=1, bias=add_bias)
-            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:
-            return self.act(self.norm2(self.conv2(self.norm1(self.conv1(x)))))
-
-
-# --------------------- Yolov8 modules ---------------------
-## Yolov8 BottleNeck
-class Bottleneck(nn.Module):
-    def __init__(self,
-                 in_dim,
-                 out_dim,
-                 expand_ratio = 0.5,
-                 kernel_sizes = [3, 3],
-                 shortcut     = True,
-                 act_type     = 'silu',
-                 norm_type    = 'BN',
-                 depthwise    = False,):
-        super(Bottleneck, self).__init__()
-        inter_dim = int(out_dim * expand_ratio)  # hidden channels 
-        padding_sizes = [k // 2 for k in kernel_sizes]           
-        self.cv1 = BasicConv(in_dim, inter_dim, kernel_size=kernel_sizes[0], padding=padding_sizes[0], act_type=act_type, norm_type=norm_type, depthwise=depthwise)
-        self.cv2 = BasicConv(inter_dim, out_dim, kernel_size=kernel_sizes[1], padding=padding_sizes[1], act_type=act_type, norm_type=norm_type, depthwise=depthwise)
-        self.shortcut = shortcut and in_dim == out_dim
-
-    def forward(self, x):
-        h = self.cv2(self.cv1(x))
-
-        return x + h if self.shortcut else h
-
-# Yolov8 StageBlock
-class RTCBlock(nn.Module):
-    def __init__(self,
-                 in_dim,
-                 out_dim,
-                 num_blocks = 1,
-                 shortcut   = False,
-                 act_type   = 'silu',
-                 norm_type  = 'BN',
-                 depthwise  = False,):
-        super(RTCBlock, self).__init__()
-        self.inter_dim = out_dim // 2
-        self.input_proj = BasicConv(in_dim, out_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
-        self.m = nn.ModuleList([
-            Bottleneck(self.inter_dim, self.inter_dim, 1.0, [1, 3], shortcut, act_type, norm_type, depthwise)
-            for _ in range(num_blocks)])
-        self.output_proj = BasicConv((2 + num_blocks) * self.inter_dim, out_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
-
-    def forward(self, x):
-        # Input proj
-        x1, x2 = torch.chunk(self.input_proj(x), 2, dim=1)
-        out = list([x1, x2])
-
-        # Bottleneck
-        out.extend(m(out[-1]) for m in self.m)
-
-        # Output proj
-        out = self.output_proj(torch.cat(out, dim=1))
-
-        return out
-    

models/detectors/rtcdet/rtcdet_head.py

@@ -1,157 +0,0 @@
-import torch
-import torch.nn as nn
-
-try:
-    from .rtcdet_basic import BasicConv
-except:
-    from  rtcdet_basic import BasicConv
-
-
-# Single-level Head
-class SingleLevelHead(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 Head
-class MultiLevelHead(nn.Module):
-    def __init__(self, cfg, in_dims, num_levels=3, num_classes=80, reg_max=16):
-        super().__init__()
-        ## ----------- Network Parameters -----------
-        self.multi_level_heads = nn.ModuleList(
-            [SingleLevelHead(in_dim       = in_dims[level],
-                             cls_head_dim = max(in_dims[0], min(num_classes, 100)),
-                             reg_head_dim = max(in_dims[0]//4, 16, 4*reg_max),
-                             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(num_levels)
-                             ])
-        # --------- Basic Parameters ----------
-        self.in_dims = in_dims
-        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_head(cfg, in_dims, num_levels=3, num_classes=80, reg_max=16):
-    if cfg['head'] == 'decoupled_head':
-        head = MultiLevelHead(cfg, in_dims, num_levels, num_classes, reg_max)
-
-    return head
-
-
-if __name__ == '__main__':
-    import time
-    from thop import profile
-    cfg = {
-        'head': 'decoupled_head',
-        'num_cls_head': 2,
-        'num_reg_head': 2,
-        'head_act': 'silu',
-        'head_norm': 'BN',
-        'head_depthwise': False,
-        'reg_max': 16,
-    }
-    fpn_dims = [256, 256, 256]
-    cls_out_dim = 256
-    reg_out_dim = 256
-    # Head-1
-    model = build_head(cfg, fpn_dims, num_levels=3, num_classes=80, reg_max=16)
-    print(model)
-    fpn_feats = [torch.randn(1, fpn_dims[0], 80, 80), torch.randn(1, fpn_dims[1], 40, 40), torch.randn(1, fpn_dims[2], 20, 20)]
-    t0 = time.time()
-    outputs = model(fpn_feats)
-    t1 = time.time()
-    print('Time: ', t1 - t0)
-    # for out in outputs:
-    #     print(out.shape)
-
-    print('==============================')
-    flops, params = profile(model, inputs=(fpn_feats, ), verbose=False)
-    print('==============================')
-    print('Head-1: GFLOPs : {:.2f}'.format(flops / 1e9 * 2))
-    print('Head-1: Params : {:.2f} M'.format(params / 1e6))

models/detectors/rtcdet/rtcdet_neck.py

@@ -1,103 +0,0 @@
-import torch
-import torch.nn as nn
-
-try:
-    from .rtcdet_basic import BasicConv
-except:
-    from  rtcdet_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, expand_ratio=0.5):
-        super().__init__()
-        inter_dim = int(in_dim * 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(inter_dim * 4, out_dim, kernel_size=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 = 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.m = 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, expand_ratio=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.cv2(x)
-        x3 = self.m(x2)
-        y = self.cv3(torch.cat([x1, x3], dim=1))
-
-        return y
-
-
-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
-
-
-if __name__ == '__main__':
-    import time
-    from thop import profile
-    cfg = {
-        ## Neck: SPP
-        'neck': 'sppf',
-        'neck_expand_ratio': 0.5,
-        'pooling_size': 5,
-        'neck_act': 'silu',
-        'neck_norm': 'BN',
-        'neck_depthwise': False,
-    }
-    in_dim = 512
-    out_dim = 512
-    # Head-1
-    model = build_neck(cfg, in_dim, out_dim)
-    feat = torch.randn(1, in_dim, 20, 20)
-    t0 = time.time()
-    outputs = model(feat)
-    t1 = time.time()
-    print('Time: ', t1 - t0)
-    # for out in outputs:
-    #     print(out.shape)
-
-    print('==============================')
-    flops, params = profile(model, inputs=(feat, ), verbose=False)
-    print('==============================')
-    print('FPN: GFLOPs : {:.2f}'.format(flops / 1e9 * 2))
-    print('FPN: Params : {:.2f} M'.format(params / 1e6))

models/detectors/rtcdet/rtcdet_pafpn.py

@@ -1,178 +0,0 @@
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-
-try:
-    from .rtcdet_basic import BasicConv, RTCBlock
-except:
-    from  rtcdet_basic import BasicConv, RTCBlock
-
-
-# PaFPN-ELAN
-class RTCPaFPN(nn.Module):
-    def __init__(self, 
-                 in_dims   = [256, 512, 1024],
-                 out_dim   = 256,
-                 width     = 1.0,
-                 depth     = 1.0,
-                 act_type  = 'silu',
-                 norm_type = 'BN',
-                 depthwise = False):
-        super(RTCPaFPN, self).__init__()
-        print('==============================')
-        print('FPN: {}'.format("RTC PaFPN"))
-        # ---------------- Basic parameters ----------------
-        self.in_dims = in_dims
-        self.width = width
-        self.depth = depth
-        c3, c4, c5 = in_dims
-
-        # ---------------- Top-dwon FPN----------------
-        ## P5 -> P4
-        self.reduce_layer_1   = BasicConv(c5, round(512*width),
-                                          kernel_size=1, padding=0, stride=1,
-                                          act_type=act_type, norm_type=norm_type)
-        self.top_down_layer_1 = RTCBlock(in_dim      = round(512*width) + c4,
-                                         out_dim     = round(512*width),
-                                         num_blocks  = round(3*depth),
-                                         shortcut    = False,
-                                         act_type    = act_type,
-                                         norm_type   = norm_type,
-                                         depthwise   = depthwise,
-                                         )
-
-        ## P4 -> P3
-        self.reduce_layer_2   = BasicConv(round(512*width), round(256*width),
-                                          kernel_size=1, padding=0, stride=1,
-                                          act_type=act_type, norm_type=norm_type)
-        self.top_down_layer_2 = RTCBlock(in_dim      = round(256*width) + c3,
-                                         out_dim     = round(256*width),
-                                         num_blocks  = round(3*depth),
-                                         shortcut    = False,
-                                         act_type    = act_type,
-                                         norm_type   = norm_type,
-                                         depthwise   = depthwise,
-                                         )
-
-        # ---------------- Bottom-up PAN ----------------
-        ## P3 -> P4
-        self.dowmsample_layer_1 = BasicConv(round(256*width), round(256*width),
-                                            kernel_size=3, padding=1, stride=2,
-                                            act_type=act_type, norm_type=norm_type, depthwise=depthwise)
-        self.bottom_up_layer_1  = RTCBlock(in_dim      = round(256*width) + round(256*width),
-                                           out_dim     = round(512*width),
-                                           num_blocks  = round(3*depth),
-                                           shortcut    = False,
-                                           act_type    = act_type,
-                                           norm_type   = norm_type,
-                                           depthwise   = depthwise,
-                                           )
-
-        ## P4 -> P5
-        self.dowmsample_layer_2 = BasicConv(round(512*width), round(512*width),
-                                            kernel_size=3, padding=1, stride=2,
-                                            act_type=act_type, norm_type=norm_type, depthwise=depthwise)
-        self.bottom_up_layer_2  = RTCBlock(in_dim      = round(512*width) + round(512*width),
-                                           out_dim     = round(1024*width),
-                                           num_blocks  = round(3*depth),
-                                           shortcut    = False,
-                                           act_type    = act_type,
-                                           norm_type   = norm_type,
-                                           depthwise   = depthwise,
-                                           )
-
-        # ---------------- Output projection ----------------
-        ## Output projs
-        self.out_layers = nn.ModuleList([
-            BasicConv(in_dim, out_dim, kernel_size=1, act_type=act_type, norm_type=norm_type)
-            for in_dim in [round(256*width), round(512*width), round(1024*width)]
-            ])
-        self.out_dims = [out_dim] * 3
-
-        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, features):
-        c3, c4, c5 = features
-
-        # 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.dowmsample_layer_1(c13)
-        c15 = torch.cat([c14, c10], dim=1)
-        c16 = self.bottom_up_layer_1(c15)
-        ## P4 -> P5
-        c17 = self.dowmsample_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
-        out_feats_proj = []
-        for feat, layer in zip(out_feats, self.out_layers):
-            out_feats_proj.append(layer(feat))
-
-        return out_feats_proj
-
-
-def build_fpn(cfg, in_dims, out_dim):
-    # build neck
-    if cfg['fpn'] == 'rtc_pafpn':
-        fpn_net = RTCPaFPN(in_dims   = in_dims,
-                           out_dim   = out_dim,
-                           width     = cfg['width'],
-                           depth     = cfg['depth'],
-                           act_type  = cfg['fpn_act'],
-                           norm_type = cfg['fpn_norm'],
-                           depthwise = cfg['fpn_depthwise']
-                           )
-    else:
-        raise NotImplementedError("Unknown fpn: {}".format(cfg['fpn']))
-    return fpn_net
-
-
-if __name__ == '__main__':
-    import time
-    from thop import profile
-    cfg = {
-        'fpn': 'rtc_pafpn',
-        'fpn_act': 'silu',
-        'fpn_norm': 'BN',
-        'fpn_depthwise': False,
-        'width': 1.0,
-        'depth': 1.0,
-        'ratio': 1.0,
-    }
-    model = build_fpn(cfg, in_dims=[256, 512, 1024], out_dim=256)
-    pyramid_feats = [torch.randn(1, 256, 80, 80), torch.randn(1, 512, 40, 40), torch.randn(1, 1024, 20, 20)]
-    t0 = time.time()
-    outputs = model(pyramid_feats)
-    t1 = time.time()
-    print('Time: ', t1 - t0)
-    for out in outputs:
-        print(out.shape)
-
-    print('==============================')
-    flops, params = profile(model, inputs=(pyramid_feats, ), verbose=False)
-    print('==============================')
-    print('GFLOPs : {:.2f}'.format(flops / 1e9 * 2))
-    print('Params : {:.2f} M'.format(params / 1e6))

models/detectors/rtcdet/rtcdet_pred.py

@@ -1,169 +0,0 @@
-import math
-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     :int = 256,
-                 reg_dim     :int = 256,
-                 stride      :int = 32,
-                 reg_max     :int = 16,
-                 num_classes :int = 80,
-                 num_coords  :int = 4):
-        super().__init__()
-        # --------- Basic Parameters ----------
-        self.stride = stride
-        self.cls_dim = cls_dim
-        self.reg_dim = reg_dim
-        self.reg_max = reg_max
-        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):
-        # cls pred bias
-        b = self.cls_pred.bias.view(1, -1)
-        b.data.fill_(math.log(5 / self.num_classes / (640. / self.stride) ** 2))
-        self.cls_pred.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
-        # reg pred bias
-        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]
-        """
-        # 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.stride
-
-        return anchors
-        
-    def forward(self, cls_feat, reg_feat):
-        # pred
-        cls_pred = self.cls_pred(cls_feat)
-        reg_pred = self.reg_pred(reg_feat)
-
-        # generate anchor boxes: [M, 4]
-        B, _, H, W = cls_pred.size()
-        fmp_size = [H, W]
-        anchors = self.generate_anchors(fmp_size)
-        anchors = anchors.to(cls_pred.device)
-        # stride tensor: [M, 1]
-        stride_tensor = torch.ones_like(anchors[..., :1]) * self.stride
-        
-        # [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)
-        
-        # output dict
-        outputs = {"pred_cls": cls_pred,            # List(Tensor) [B, M, C]
-                   "pred_reg": reg_pred,            # List(Tensor) [B, M, 4*(reg_max)]
-                   "anchors": anchors,              # List(Tensor) [M, 2]
-                   "strides": self.stride,          # List(Int) = [8, 16, 32]
-                   "stride_tensor": stride_tensor   # List(Tensor) [M, 1]
-                   }
-
-        return outputs
-
-# Multi-level pred layer
-class MultiLevelPredLayer(nn.Module):
-    def __init__(self,
-                 cls_dim,
-                 reg_dim,
-                 strides,
-                 num_classes :int = 80,
-                 num_coords  :int = 4,
-                 num_levels  :int = 3,
-                 reg_max     :int = 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     = cls_dim,
-                                  reg_dim     = reg_dim,
-                                  stride      = strides[level],
-                                  reg_max     = reg_max,
-                                  num_classes = num_classes,
-                                  num_coords  = num_coords * reg_max)
-                                  for level in range(num_levels)
-                                  ])
-        ## proj conv
-        proj_init = torch.arange(reg_max, dtype=torch.float)
-        self.proj_conv = nn.Conv2d(self.reg_max, 1, kernel_size=1, bias=False).requires_grad_(False)
-        self.proj_conv.weight.data[:] = nn.Parameter(proj_init.view([1, reg_max, 1, 1]))
-
-    def forward(self, cls_feats, reg_feats):
-        all_anchors = []
-        all_strides = []
-        all_cls_preds = []
-        all_reg_preds = []
-        all_box_preds = []
-        all_delta_preds = []
-        for level in range(self.num_levels):
-            # -------------- Single-level prediction --------------
-            outputs = self.multi_level_preds[level](cls_feats[level], reg_feats[level])
-
-            # -------------- Decode bbox --------------
-            B, M = outputs["pred_reg"].shape[:2]
-            # [B, M, 4*(reg_max)] -> [B, M, 4, reg_max]
-            delta_pred = outputs["pred_reg"].reshape([B, M, 4, self.reg_max])
-            # [B, M, 4, reg_max] -> [B, reg_max, 4, M]
-            delta_pred = delta_pred.permute(0, 3, 2, 1).contiguous()
-            # [B, reg_max, 4, M] -> [B, 1, 4, M]
-            delta_pred = self.proj_conv(F.softmax(delta_pred, dim=1))
-            # [B, 1, 4, M] -> [B, 4, M] -> [B, M, 4]
-            delta_pred = delta_pred.view(B, 4, M).permute(0, 2, 1).contiguous()
-            ## tlbr -> xyxy
-            x1y1_pred = outputs["anchors"][None] - delta_pred[..., :2] * self.strides[level]
-            x2y2_pred = outputs["anchors"][None] + delta_pred[..., 2:] * self.strides[level]
-            box_pred = torch.cat([x1y1_pred, x2y2_pred], dim=-1)
-
-            # collect results
-            all_cls_preds.append(outputs["pred_cls"])
-            all_reg_preds.append(outputs["pred_reg"])
-            all_box_preds.append(box_pred)
-            all_delta_preds.append(delta_pred)
-            all_anchors.append(outputs["anchors"])
-            all_strides.append(outputs["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]
-                   "pred_delta": all_delta_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_predictor(cls_dim, reg_dim, strides, num_classes, num_coords=4, num_levels=3, reg_max=16):
-    pred_layers = MultiLevelPredLayer(cls_dim, reg_dim, strides, num_classes, num_coords, num_levels, reg_max) 
-
-    return pred_layers

models/detectors/yolov8/README.md

@@ -8,7 +8,7 @@
 | YOLOv8-L  | 8xb16  |  640  |          50.7          |        68.3       |       165.7       |         43.7       | [ckpt](https://github.com/yjh0410/RT-ODLab/releases/download/yolo_tutorial_ckpt/yolov8_l_coco.pth) |
 
 - For training, we train YOLOv8 series with 500 epochs on COCO.
-- For data augmentation, we use the large scale jitter (LSJ), Mosaic augmentation and Mixup augmentation, following the setting of [YOLOv8](https://github.com/ultralytics/yolov8).
+- For data augmentation, we use the random affine, hsv augmentation, mosaic augmentation and mixup augmentation, following the setting of [YOLOv8](https://github.com/ultralytics/yolov8).
 - For optimizer, we use AdamW with weight decay 0.05 and base per image lr 0.001 / 64, which is different from the official YOLOv8. We have tried SGD, but it has weakened performance. For example, when using SGD, YOLOv8-N's AP was only 35.8%, lower than the current result (36.8 %), perhaps because some hyperparameters were not set properly.
 - For learning rate scheduler, we use linear decay scheduler.
 

models/detectors/yolov8/yolov8.py

@@ -4,10 +4,10 @@ import torch.nn as nn
 
 # --------------- Model components ---------------
 from .yolov8_backbone import build_backbone
-from .yolov8_neck import build_neck
-from .yolov8_pafpn import build_fpn
-from .yolov8_head import build_det_head
-from .yolov8_pred import build_pred_layer
+from .yolov8_neck     import build_neck
+from .yolov8_pafpn    import build_fpn
+from .yolov8_head     import build_det_head
+from .yolov8_pred     import build_pred_layer
 
 # --------------- External components ---------------
 from utils.misc import multiclass_nms
@@ -67,7 +67,6 @@ class YOLOv8(nn.Module):
                                             num_levels  = self.num_levels,
                                             reg_max     = self.reg_max)
 
-    ## post-process
     def post_process(self, cls_preds, box_preds):
         """
         Input:
@@ -146,9 +145,7 @@ class YOLOv8(nn.Module):
 
         return bboxes, scores, labels
     
-    # ---------------------- Main Process for Inference ----------------------
-    @torch.no_grad()
-    def inference_single_image(self, x):
+    def forward(self, x):
         # ---------------- Backbone ----------------
         pyramid_feats = self.backbone(x)
 
@@ -164,45 +161,25 @@ class YOLOv8(nn.Module):
         # ---------------- 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)
-
-        else:
-            # post process
-            bboxes, scores, labels = self.post_process(all_cls_preds, all_box_preds)
-            outputs = {
-                "scores": scores,
-                "labels": labels,
-                "bboxes": bboxes
-            }
-        
-        return outputs
-
-    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)
+        if not self.training:
+            all_cls_preds = outputs['pred_cls']
+            all_box_preds = outputs['pred_box']
 
-            # ---------------- Heads ----------------
-            cls_feats, reg_feats = self.det_heads(pyramid_feats)
+            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)
 
-            # ---------------- Preds ----------------
-            outputs = self.pred_layers(cls_feats, reg_feats)
-            
-            return outputs 
+            else:
+                # post process
+                bboxes, scores, labels = self.post_process(all_cls_preds, all_box_preds)
+                outputs = {
+                    "scores": scores,
+                    "labels": labels,
+                    "bboxes": bboxes
+                }
+        
+        return outputs 

train.sh

@@ -17,12 +17,6 @@ if [[ $MODEL == *"rtdetr"* ]]; then
     EVAL_EPOCH=1
     NO_AUG_EPOCH=-1
     FIND_UNUSED_PARAMS=True
-elif [[ $MODEL == *"rtcdet"* ]]; then
-    # Epoch setting
-    MAX_EPOCH=400
-    WP_EPOCH=3
-    EVAL_EPOCH=10
-    NO_AUG_EPOCH=20
 elif [[ $MODEL == *"yolov8"* ]]; then
     # Epoch setting
     MAX_EPOCH=500