目标检测：使用自己的数据集微调DEIMv2进行物体检测

FriendshipT2025-10-21 23:53

目标检测：使用自己的数据集微调DEIMv2进行物体检测


前言

环境要求

bash 复制代码 
Package                  Version
------------------------ ------------
absl-py                  2.3.1
accelerate               1.10.1
calflops                 0.3.2
certifi                  2025.10.5
charset-normalizer       3.4.4
colorama                 0.4.6
coloredlogs              15.0.1
faster-coco-eval         1.6.8
filelock                 3.20.0
flatbuffers              25.9.23
fsspec                   2025.9.0
grpcio                   1.75.1
huggingface-hub          0.35.3
humanfriendly            10.0
idna                     3.11
Jinja2                   3.1.6
Markdown                 3.9
markdown-it-py           4.0.0
MarkupSafe               3.0.3
mdurl                    0.1.2
ml_dtypes                0.5.3
mpmath                   1.3.0
networkx                 3.5
numpy                    1.26.4
nvidia-cuda-runtime-cu13 0.0.0a0
onnx                     1.19.1
onnxruntime              1.23.1
onnxruntime-gpu          1.15.0
onnxsim                  0.4.36
opencv-python            4.6.0.66
packaging                25.0
pillow                   12.0.0
pip                      24.0
protobuf                 6.33.0
psutil                   7.1.0
Pygments                 2.19.2
pyreadline3              3.5.4
PyYAML                   6.0.3
regex                    2025.9.18
requests                 2.32.5
rich                     14.2.0
safetensors              0.6.2
scipy                    1.16.2
setuptools               65.5.0
sympy                    1.13.1
tensorboard              2.20.0
tensorboard-data-server  0.7.2
tensorrt                 8.6.1
tensorrt-dispatch        8.6.1
tensorrt-lean            8.6.1
tokenizers               0.22.1
torch                    2.5.1+cu118
torchaudio               2.5.1+cu118
torchvision              0.20.1+cu118
tqdm                     4.67.1
transformers             4.57.1
typing_extensions        4.15.0
urllib3                  2.5.0
Werkzeug                 3.1.3

相关介绍

  • Python是一种跨平台的计算机程序设计语言。是一个高层次的结合了解释性、编译性、互动性和面向对象的脚本语言。最初被设计用于编写自动化脚本(shell)，随着版本的不断更新和语言新功能的添加，越多被用于独立的、大型项目的开发。
  • PyTorch 是一个深度学习框架，封装好了很多网络和深度学习相关的工具方便我们调用，而不用我们一个个去单独写了。它分为 CPU 和 GPU 版本，其他框架还有 TensorFlow、Caffe 等。PyTorch 是由 Facebook 人工智能研究院（FAIR）基于 Torch 推出的，它是一个基于 Python 的可续计算包，提供两个高级功能：1、具有强大的 GPU 加速的张量计算（如 NumPy）；2、构建深度神经网络时的自动微分机制。
  • DEIMv2是一种基于DINOv3特性的实时对象检测器，其核心特点是：
    • 多种模型规模：DEIMv2设计了多种模型规模，从满足GPU到移动设备部署需求的大型模型到超轻量级模型。
    • Spatial Tuning Adapter（STA）：引入STA高效融合单尺度输出为多尺度特征。
    • 简化解码器和增强的密集物体对齐（Dense O2O）策略：进一步提升性能。
  • DEIMv2的优点是能够在不同的场景下实现高效的实时对象检测，尤其是在大模型和紧凑型模型方面表现出色。同时，DEIMv2还能够提高实时检测性能和效率。
  • DEIMv2也有一些局限性，例如需要大量的计算资源和训练时间。此外，DEIMv2在处理小目标时可能存在一定的误检率。
  • 未来，可以考虑针对DEIMv2的局限性进行改进，如优化训练算法、增加数据集样本等，以进一步提高DEIMv2的性能和适用范围。
  • 官方源代码： https://github.com/Intellindust-AI-Lab/DEIMv2.git
  • Huang, Shihua and Hou, Yongjie and Liu, Longfei and Yu, Xuanlong and Shen, Xi. Real-Time Object Detection Meets DINOv3. 2025

微调DEIMv2模型

下载DEIMv2项目

Windows

Linux

bash 复制代码 
git clone https://github.com/Intellindust-AI-Lab/DEIMv2.git

准备数据集

准备一个COCO格式的目标检测数据集。

bash 复制代码 
dataset/
├── images/
│   ├── train/
│   │   ├── image1.jpg
│   │   ├── image2.jpg
│   │   └── ...
│   ├── val/
│   │   ├── image1.jpg
│   │   ├── image2.jpg
│   │   └── ...
└── annotations/
    ├── instances_train.json
    ├── instances_val.json
    └── ...

下载预训练模型

对于 DINOv3 S 和 S+模型，请按照 https://github.com/facebookresearch/dinov3 中的指南下载。

对于蒸馏的 ViT-Tiny 和 ViT-Tiny+模型，您可以从 ViT-Tiny 和 ViT-Tiny+ 下载。

然后将它们放入 ./ckpts 中：

shell 复制代码 
ckpts/
├── dinov3_vits16.pth
├── vitt_distill.pt
├── vittplus_distill.pt
└── ...

配置文件

deimv2_dinov3_s_coco.yml

yaml 复制代码 
__include__: [
  '../dataset/coco_detection.yml',
  '../runtime.yml',
  '../base/dataloader.yml',
  '../base/optimizer.yml',
  '../base/deimv2.yml',
]

output_dir: ./outputs/deimv2_dinov3_s_coco

DEIM:
  backbone: DINOv3STAs

DINOv3STAs:
  name: vit_tiny
  embed_dim: 192
  weights_path: ./ckpts/vitt_distill.pt
  interaction_indexes: [3, 7, 11]   # only need the [1/8, 1/16, 1/32]
  num_heads: 3

HybridEncoder:
  in_channels: [192, 192, 192]
  depth_mult: 0.67
  expansion: 0.34
  hidden_dim: 192
  dim_feedforward: 512

DEIMTransformer:
  feat_channels: [192, 192, 192]
  hidden_dim: 192
  dim_feedforward: 512
  num_layers: 4  # 4 5 6
  eval_idx: -1  # -2 -3 -4


## Optimizer
optimizer:
  type: AdamW

  params: 
    -
      # except norm/bn/bias in self.dinov3
      params: '^(?=.*.dinov3)(?!.*(?:norm|bn|bias)).*$'  
      lr: 0.000025
    -
      # including all norm/bn/bias in self.dinov3
      params: '^(?=.*.dinov3)(?=.*(?:norm|bn|bias)).*$'    
      lr: 0.000025
      weight_decay: 0.
    - 
      # including all norm/bn/bias except for the self.dinov3
      params: '^(?=.*(?:sta|encoder|decoder))(?=.*(?:norm|bn|bias)).*$'
      weight_decay: 0.

  lr: 0.0005
  betas: [0.9, 0.999]
  weight_decay: 0.0001

# Increase to search for the optimal ema
epoches: 132 # 120 + 4n

## Our LR-Scheduler
flat_epoch: 64    # 4 + epoch // 2, e.g., 40 = 4 + 72 / 2
no_aug_epoch: 12

## Our DataAug
train_dataloader:
  dataset: 
    transforms:
      ops:
        - {type: Mosaic, output_size: 320, rotation_range: 10, translation_range: [0.1, 0.1], scaling_range: [0.5, 1.5],
           probability: 1.0, fill_value: 0, use_cache: True, max_cached_images: 50, random_pop: True}
        - {type: RandomPhotometricDistort, p: 0.5}
        - {type: RandomZoomOut, fill: 0}
        - {type: RandomIoUCrop, p: 0.8}
        - {type: SanitizeBoundingBoxes, min_size: 1}
        - {type: RandomHorizontalFlip}
        - {type: Resize, size: [640, 640], }
        - {type: SanitizeBoundingBoxes, min_size: 1}
        - {type: ConvertPILImage, dtype: 'float32', scale: True}
        - {type: Normalize, mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225]}
        - {type: ConvertBoxes, fmt: 'cxcywh', normalize: True}
      policy:
        epoch: [4, 64, 120]   # list 

  collate_fn:
    base_size: 640
    mixup_prob: 0.5
    ema_restart_decay: 0.9999
    base_size_repeat: 20
    mixup_epochs: [4, 64]
    stop_epoch: 120
    copyblend_epochs: [4, 120]

val_dataloader:
  dataset:
    transforms:
      ops:
        - {type: Resize, size: [640, 640], }
        - {type: ConvertPILImage, dtype: 'float32', scale: True}
        - {type: Normalize, mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225]}

DEIMCriterion:
  matcher:
    # change matcher
    change_matcher: True
    iou_order_alpha: 4.0
    matcher_change_epoch: 100

coco_detection.yml

yaml 复制代码 
task: detection

evaluator:
  type: CocoEvaluator
  iou_types: ['bbox', ]

num_classes: 80
remap_mscoco_category: True

train_dataloader: 
  type: DataLoader
  dataset: 
    type: CocoDetection
    img_folder: 你自己的数据路径/images/train/
    ann_file: 你自己的数据路径/annotations/instances_train.json
    return_masks: False
    transforms:
      type: Compose
      ops: ~
  shuffle: True
  num_workers: 4
  drop_last: True 
  collate_fn:
    type: BatchImageCollateFunction


val_dataloader:
  type: DataLoader
  dataset: 
    type: CocoDetection
    img_folder: 你自己的数据路径/images/val/
    ann_file: 你自己的数据路径/annotations/instances_val.json
    return_masks: False
    transforms:
      type: Compose
      ops: ~ 
  shuffle: False
  num_workers: 4
  drop_last: False
  collate_fn:
    type: BatchImageCollateFunction

dataloader.yml

yaml 复制代码 
train_dataloader:
  dataset:
    transforms:
      ops:
        - {type: RandomPhotometricDistort, p: 0.5}
        - {type: RandomZoomOut, fill: 0}
        - {type: RandomIoUCrop, p: 0.8}
        - {type: SanitizeBoundingBoxes, min_size: 1}
        - {type: RandomHorizontalFlip}
        - {type: Resize, size: [640, 640], }
        - {type: SanitizeBoundingBoxes, min_size: 1}
        - {type: ConvertPILImage, dtype: 'float32', scale: True}
        - {type: ConvertBoxes, fmt: 'cxcywh', normalize: True}
      policy:
        name: stop_epoch
        epoch: 72 # epoch in [71, ~) stop `ops`
        ops: ['Mosaic', 'RandomPhotometricDistort', 'RandomZoomOut', 'RandomIoUCrop']

  collate_fn:
    type: BatchImageCollateFunction
    base_size: 640
    base_size_repeat: 3
    stop_epoch: 72 # epoch in [72, ~) stop `multiscales`

  shuffle: True
  # total_batch_size: 32 # total batch size equals to 32 (4 * 8)
  total_batch_size: 2 # 单卡，8G显存
  num_workers: 4


val_dataloader:
  dataset:
    transforms:
      ops:
        - {type: Resize, size: [640, 640], }
        - {type: ConvertPILImage, dtype: 'float32', scale: True}
  shuffle: False
  # total_batch_size: 64 # 8G显存不够
  total_batch_size: 2 # 单卡，8G显存
  num_workers: 4

进行训练

bash 复制代码 
python train.py -c configs/deimv2/deimv2_dinov3_s_coco.yml --use-amp --seed=0 -t deimv2_dinov3_s_coco.pth
bash 复制代码 
$python train.py -c configs/deimv2/deimv2_dinov3_s_coco.yml --use-amp --seed=0 -t deimv2_dinov3_s_coco.pth
Not init distributed mode.
cfg:  {'task': 'detection', '_model': None, '_postprocessor': None, '_criterion': None, '_optimizer': None, '_lr_scheduler': None, '_lr_warmup_scheduler': None, '_train_dataloader': None, '_val_dataloader': None, '_ema': None, '_scaler': None, '_train_dataset': None, '_val_dataset': None, '_collate_fn': None, '_evaluator': None, '_writer': None, 'num_workers': 0, 'batch_size': None, '_train_batch_size': None, '_val_batch_size': None, '_train_shuffle': None, '_val_shuffle': None, 'resume': None, 'tuning': 'deimv2_dinov3_s_coco.pth', 'epoches': 132, 'last_epoch': -1, 'lrsheduler': 'flatcosine', 'lr_gamma': 0.5, 'no_aug_epoch': 12, 'warmup_iter': 2000, 'flat_epoch': 64, 'use_amp': True, 'use_ema': True, 'ema_decay': 0.9999, 'ema_warmups': 2000, 'sync_bn': True, 'clip_max_norm': 0.1, 'find_unused_parameters': True, 'seed': 0, 'print_freq': 500, 'checkpoint_freq': 5, 'output_dir': './outputs/deimv2_dinov3_s_coco', 'summary_dir': None, 'device': '', 'yaml_cfg': {'task': 'detection', 'evaluator': {'type': 'CocoEvaluator', 'iou_types': ['bbox']}, 'num_classes': 80, 'remap_mscoco_category': True, 'train_dataloader': {'type': 'DataLoader', 'dataset': {'type': 'CocoDetection', 'img_folder': 'F:/user/mytest/DEIMv2_test/coco_mini_dataset/images/train/', 'ann_file': 'F:/user/mytest/DEIMv2_test/coco_mini_dataset/annotations/instances_train.json', 'return_masks': False, 'transforms': {'type': 'Compose', 'ops': [{'type': 'Mosaic', 'output_size': 320, 'rotation_range': 10, 'translation_range': [0.1, 0.1], 'scaling_range': [0.5, 1.5], 'probability': 1.0, 'fill_value': 0, 'use_cache': True, 'max_cached_images': 50, 'random_pop': True}, {'type': 'RandomPhotometricDistort', 'p': 0.5}, {'type': 'RandomZoomOut', 'fill': 0}, {'type': 'RandomIoUCrop', 'p': 0.8}, {'type': 'SanitizeBoundingBoxes', 'min_size': 1}, {'type': 'RandomHorizontalFlip'}, {'type': 'Resize', 'size': [640, 640]}, {'type': 'SanitizeBoundingBoxes', 'min_size': 1}, {'type': 'ConvertPILImage', 'dtype': 'float32', 'scale': True}, {'type': 'Normalize', 'mean': [0.485, 0.456, 0.406], 'std': [0.229, 0.224, 0.225]}, {'type': 'ConvertBoxes', 'fmt': 'cxcywh', 'normalize': True}], 'policy': {'name': 'stop_epoch', 'epoch': [4, 64, 120], 'ops': ['Mosaic', 'RandomPhotometricDistort', 'RandomZoomOut', 'RandomIoUCrop']}, 'mosaic_prob': 0.5}}, 'shuffle': True, 'num_workers': 4, 'drop_last': True, 'collate_fn': {'type': 'BatchImageCollateFunction', 'base_size': 640, 'base_size_repeat': 20, 'stop_epoch': 120, 'mixup_prob': 0.5, 'mixup_epochs': [4, 64], 'copyblend_prob': 0.5, 'copyblend_epochs': [4, 120], 'area_threshold': 100, 'num_objects': 3, 'with_expand': True, 'expand_ratios': [0.1, 0.25], 'ema_restart_decay': 0.9999}, 'total_batch_size': 2}, 'val_dataloader': {'type': 'DataLoader', 'dataset': {'type': 'CocoDetection', 'img_folder': 'F:/user/mytest/DEIMv2_test/coco_mini_dataset/images/val/', 'ann_file': 'F:/user/mytest/DEIMv2_test/coco_mini_dataset/annotations/instances_val.json', 'return_masks': False, 'transforms': {'type': 'Compose', 'ops': [{'type': 'Resize', 'size': [640, 640]}, {'type': 'ConvertPILImage', 'dtype': 'float32', 'scale': True}, {'type': 'Normalize', 'mean': [0.485, 0.456, 0.406], 'std': [0.229, 0.224, 0.225]}]}}, 'shuffle': False, 'num_workers': 4, 'drop_last': False, 'collate_fn': {'type': 'BatchImageCollateFunction'}, 'total_batch_size': 2}, 'print_freq': 500, 'output_dir': './outputs/deimv2_dinov3_s_coco', 'checkpoint_freq': 5, 'sync_bn': True, 'find_unused_parameters': True, 'use_amp': True, 'scaler': {'type': 'GradScaler', 'enabled': True}, 'use_ema': True, 'ema': {'type': 'ModelEMA', 'decay': 0.9999, 'warmups': 1000, 'start': 0}, 'epoches': 132, 'clip_max_norm': 0.1, 'optimizer': {'type': 'AdamW', 'params': [{'params': '^(?=.*.dinov3)(?!.*(?:norm|bn|bias)).*$', 'lr': 2.5e-05}, {'params': '^(?=.*.dinov3)(?=.*(?:norm|bn|bias)).*$', 'lr': 2.5e-05, 'weight_decay': 0.0}, {'params': '^(?=.*(?:sta|encoder|decoder))(?=.*(?:norm|bn|bias)).*$', 'weight_decay': 0.0}], 'lr': 0.0005, 'betas': [0.9, 0.999], 'weight_decay': 0.0001}, 'lr_scheduler': {'type': 'MultiStepLR', 'milestones': [500], 'gamma': 0.1}, 'lr_warmup_scheduler': {'type': 'LinearWarmup', 'warmup_duration': 500}, 'model': 'DEIM', 'criterion': 'DEIMCriterion', 'postprocessor': 'PostProcessor', 'use_focal_loss': True, 'eval_spatial_size': [640, 640], 'DEIM': {'backbone': 'DINOv3STAs', 'encoder': 'HybridEncoder', 'decoder': 'DEIMTransformer'}, 'HGNetv2': {'name': 'B4', 'return_idx': [1, 2, 3], 'freeze_at': -1, 'freeze_stem_only': True, 'freeze_norm': False, 'pretrained': False, 'local_model_dir': './weight/hgnetv2/'}, 'HybridEncoder': {'in_channels': [192, 192, 192], 'feat_strides': [8, 16, 32], 'hidden_dim': 192, 'use_encoder_idx': [2], 'num_encoder_layers': 1, 'nhead': 8, 'dim_feedforward': 512, 'dropout': 0.0, 'enc_act': 'gelu', 'expansion': 0.34, 'depth_mult': 0.67, 'act': 'silu', 'version': 'deim', 'csp_type': 'csp2', 'fuse_op': 'sum'}, 'DEIMTransformer': {'feat_channels': [192, 192, 192], 'feat_strides': [8, 16, 32], 'hidden_dim': 192, 'num_levels': 3, 'num_layers': 4, 'eval_idx': -1, 'num_queries': 300, 'num_denoising': 100, 'label_noise_ratio': 0.5, 'box_noise_scale': 1.0, 'reg_max': 32, 'reg_scale': 4, 'layer_scale': 1, 'num_points': [3, 6, 3], 'cross_attn_method': 'default', 'query_select_method': 'default', 'activation': 'silu', 'mlp_act': 'silu', 'dim_feedforward': 512}, 'PostProcessor': {'num_top_queries': 300}, 'lrsheduler': 'flatcosine', 'lr_gamma': 0.5, 'warmup_iter': 2000, 'flat_epoch': 64, 'no_aug_epoch': 12, 'DEIMCriterion': {'weight_dict': {'loss_mal': 1, 'loss_bbox': 5, 'loss_giou': 2, 'loss_fgl': 0.15, 'loss_ddf': 1.5}, 'losses': ['mal', 'boxes', 'local'], 'gamma': 1.5, 'alpha': 0.75, 'reg_max': 32, 'matcher': {'type': 'HungarianMatcher', 'weight_dict': {'cost_class': 2, 'cost_bbox': 5, 'cost_giou': 2}, 'alpha': 0.25, 'gamma': 2.0, 'change_matcher': True, 'iou_order_alpha': 4.0, 'matcher_change_epoch': 100}}, '__include__': ['../dataset/coco_detection.yml', '../runtime.yml', '../base/dataloader.yml', '../base/optimizer.yml', '../base/deimv2.yml'], 'DINOv3STAs': {'name': 'vit_tiny', 'embed_dim': 192, 'weights_path': './ckpts/vitt_distill.pt', 'interaction_indexes': [3, 7, 11], 'num_heads': 3}, 'config': 'configs/deimv2/deimv2_dinov3_s_coco.yml', 'tuning': 'deimv2_dinov3_s_coco.pth', 'seed': 0, 'test_only': False, 'print_method': 'builtin', 'print_rank': 0}}
Training ViT-Tiny from scratch...
Using Lite Spatial Prior Module with inplanes=16
     --- Use Gateway@True ---
     --- Use Share Bbox Head@False ---
     --- Use Share Score Head@False ---
     --- Wide Layer@1 ---
Tuning checkpoint from deimv2_dinov3_s_coco.pth
F:\user\mytest\DEIMv2_test\DEIMv2-main\engine\solver\_solver.py:169: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  state = torch.load(path, map_location='cpu')
Load model.state_dict, {'missed': [], 'unmatched': []}
Using the new matching cost with iou_order_alpha = 4.0 at epoch 100
F:\user\mytest\DEIMv2_test\DEIMv2-main\engine\core\workspace.py:177: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
  return module(**module_kwargs)
Initial lr: [2.5e-05, 2.5e-05, 0.0005, 0.0005]
building train_dataloader with batch_size=2...
     ### Transform @Mosaic ###
     ### Transform @RandomPhotometricDistort ###
     ### Transform @RandomZoomOut ###
     ### Transform @RandomIoUCrop ###
     ### Transform @SanitizeBoundingBoxes ###
     ### Transform @RandomHorizontalFlip ###
     ### Transform @Resize ###
     ### Transform @SanitizeBoundingBoxes ###
     ### Transform @ConvertPILImage ###
     ### Transform @Normalize ###
     ### Transform @ConvertBoxes ###
     ### Mosaic with Prob.@0.5 and ZoomOut/IoUCrop existed ###
     ### ImgTransforms Epochs: [4, 64, 120] ###
     ### Policy_ops@['Mosaic', 'RandomPhotometricDistort', 'RandomZoomOut', 'RandomIoUCrop'] ###
     ### Using MixUp with Prob@0.5 in [4, 64] epochs ###
     ### Using CopyBlend-blend with Prob@0.5 in [4, 120] epochs ###
     ### CopyBlend -- area threshold@100 and num of object@3 ###
     ### CopyBlend -- expand@[0.1, 0.25] ###
     ### Multi-scale Training until 120 epochs ###
     ### Multi-scales@ [480, 512, 544, 576, 608, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 640, 800, 768, 736, 704, 672] ###
building val_dataloader with batch_size=2...
     ### Transform @Resize ###
     ### Transform @ConvertPILImage ###
     ### Transform @Normalize ###

------------------------------------- Calculate Flops Results -------------------------------------
Notations:
number of parameters (Params), number of multiply-accumulate operations(MACs),
number of floating-point operations (FLOPs), floating-point operations per second (FLOPS),
fwd FLOPs (model forward propagation FLOPs), bwd FLOPs (model backward propagation FLOPs),
default model backpropagation takes 2.00 times as much computation as forward propagation.

Total Training Params:                                                  9.71 M
fwd MACs:                                                               12.7678 GMACs
fwd FLOPs:                                                              25.6242 GFLOPS
fwd+bwd MACs:                                                           38.3033 GMACs
fwd+bwd FLOPs:                                                          76.8725 GFLOPS
---------------------------------------------------------------------------------------------------
{'Model FLOPs:25.6242 GFLOPS   MACs:12.7678 GMACs   Params:9707403'}
------------------------------------------Start training-------------------------------------------
Index 194: decoder.decoder.layers.1.cross_attn.attention_weights.weight - requires_grad: True
Index 195: decoder.decoder.layers.1.cross_attn.attention_weights.bias - requires_grad: True
     ## Using Self-defined Scheduler-flatcosine ##
[2.5e-05, 2.5e-05, 0.0005, 0.0005] [1.25e-05, 1.25e-05, 0.00025, 0.00025] 13200 2000 6400 1200
number of trainable parameters: 9779836
number of non-trainable parameters: 2
Epoch: [0]  [  0/100]  eta: 1:10:51  lr: 0.000000  loss: 24.4887 (24.4887)  loss_mal: 1.3350 (1.3350)  loss_bbox: 0.0748 (0.0748)  loss_giou: 0.4123 (0.4123)  loss_fgl: 1.1017 (1.1017)  loss_mal_aux_0: 1.2637 (1.2637)  loss_bbox_aux_0: 0.0727 (0.0727)  loss_giou_aux_0: 0.4149 (0.4149)  loss_fgl_aux_0: 1.1103 (1.1103)  loss_ddf_aux_0: 0.0201 (0.0201)  loss_mal_aux_1: 1.2578 (1.2578)  loss_bbox_aux_1: 0.0753 (0.0753)  loss_giou_aux_1: 0.4027 (0.4027)  loss_fgl_aux_1: 1.1033 (1.1033)  loss_ddf_aux_1: 0.0044 (0.0044)  loss_mal_aux_2: 1.3594 (1.3594)  loss_bbox_aux_2: 0.0758 (0.0758)  loss_giou_aux_2: 0.4157 (0.4157)  loss_fgl_aux_2: 1.1035 (1.1035)  loss_ddf_aux_2: 0.0013 (0.0013)  loss_mal_pre: 1.2695 (1.2695)  loss_bbox_pre: 0.0724 (0.0724)  loss_giou_pre: 0.4064 (0.4064)  loss_mal_enc_0: 1.3691 (1.3691)  loss_bbox_enc_0: 0.0837 (0.0837)  loss_giou_enc_0: 0.4584 (0.4584)  loss_mal_dn_0: 0.6572 (0.6572)  loss_bbox_dn_0: 0.0996 (0.0996)  loss_giou_dn_0: 0.4012 (0.4012)  loss_fgl_dn_0: 1.1022 (1.1022)  loss_ddf_dn_0: 0.1698 (0.1698)  loss_mal_dn_1: 0.5098 (0.5098)  loss_bbox_dn_1: 0.0670 (0.0670)  loss_giou_dn_1: 0.2808 (0.2808)  loss_fgl_dn_1: 1.0306 (1.0306)  loss_ddf_dn_1: 0.0210 (0.0210)  loss_mal_dn_2: 0.4963 (0.4963)  loss_bbox_dn_2: 0.0623 (0.0623)  loss_giou_dn_2: 0.2722 (0.2722)  loss_fgl_dn_2: 1.0261 (1.0261)  loss_ddf_dn_2: 0.0015 (0.0015)  loss_mal_dn_3: 0.4963 (0.4963)  loss_bbox_dn_3: 0.0616 (0.0616)  loss_giou_dn_3: 0.2690 (0.2690)  loss_fgl_dn_3: 1.0251 (1.0251)  loss_mal_dn_pre: 0.6597 (0.6597)  loss_bbox_dn_pre: 0.1088 (0.1088)  loss_giou_dn_pre: 0.4065 (0.4065)  time: 42.5199  data: 40.4094  max mem: 1979
......


输出结果

进行验证

bash 复制代码 
python train.py -c configs/deimv2/deimv2_dinov3_s_coco.yml --test-only -r ./outputs/deimv2_dinov3_s_coco/best_stg1.pth

输出结果



进行预测

bash 复制代码 
python tools/inference/torch_inf.py -c configs/deimv2/deimv2_dinov3_s_coco.yml -r ./outputs/deimv2_dinov3_s_coco/best_stg1.pth --input ./test_imgs/1.png --device cuda:0

输出结果

导出ONNX

bash 复制代码 
python tools/deployment/export_onnx.py --check -c configs/deimv2/deimv2_dinov3_s_coco.yml -r ./outputs/deimv2_dinov3_s_coco/best_stg1.pth

输出结果

推理ONNX

bash 复制代码 
python tools/inference/onnx_inf.py --onnx ./outputs/deimv2_dinov3_s_coco/best_stg1.onnx --input ./test_imgs/1.png

输出结果


参考

1 Huang, Shihua and Hou, Yongjie and Liu, Longfei and Yu, Xuanlong and Shen, Xi. Real-Time Object Detection Meets DINOv3. 2025

2 https://github.com/xinntao/Real-ESRGAN.git

bash 复制代码 
@article{huang2025deimv2,
  title={Real-Time Object Detection Meets DINOv3},
  author={Huang, Shihua and Hou, Yongjie and Liu, Longfei and Yu, Xuanlong and Shen, Xi},
  journal={arXiv},
  year={2025}
}
相关推荐
小鹿研究点东西7 分钟前
AI直播工具实操：从直播录制、AI剪辑去重到直播伴侣开播完整流程
人工智能·自动化·音视频·语音识别
碳基硅坊10 分钟前
Spring AI：把大模型接进 Spring 应用
java·人工智能·spring ai
才兄说10 分钟前
机器人二次开发机器狗巡检？全环境稳定感知
人工智能·机器人
一一哥Sun22 分钟前
第06课：Transformer与注意力机制——大模型背后的秘密武器
人工智能·深度学习·transformer
landyjzlai23 分钟前
蓝迪哥玩转Ai(10)---Harness工程说透1。
人工智能·harness
onething36524 分钟前
Spring Boot + Spring AI 从入门到实战：7天转型计划 Day 3 —— 消息表设计 + 级联删除 + 事务管理
人工智能·后端
王某某人30 分钟前
LangChain4j 入门：Java 程序员的第一个 AI 对话程序
人工智能·后端
海兰34 分钟前
【实用程序】电商销售分析仪表盘 — 从零搭建一个AI参与的全栈数据洞察系统
人工智能·学习·算法
枫糖浆AI34 分钟前
openclaw页面无法访问解决方法
人工智能
浩子coding43 分钟前
通过 Spring AI Alibaba 源码，看如何玩转 ReAct 智能体范式
人工智能·后端
热门推荐
01HTTP 与 HTTPS 的区别：从原理到实战详解022026 AI 编程工具终极实战指南：Cursor vs Claude Code vs Copilot，开发者该怎么选？03【AI】2026 年具身智能模型和世界模型总结042026年6月AI行业全景：从百模大战到Agent元年，这30天发生了什么？05AI科技热点日报 | 2026年6月1日06《置身钉内》原文-可播放阅读07GitHub 镜像站点08Claude Code、Codex、Cursor三分天下：2026年AI编程Agent生态全景剖析092026 年 AI 编程工具终极横评：Cursor vs Claude Code vs Copilot vs Windsurf10AI一周事件 · 2026-06-03 至 2026-06-09