style: minor adjustment

README.md

@@ -110,6 +110,8 @@ https://github.com/user-attachments/assets/e5933d8e-3c8a-400e-870b-4e452f5321d9
 **D‑FINE‑X** | Objects365+COCO | **59.3** | 62M | 12.89ms | 202 | [yml](./configs/dfine/objects365/dfine_hgnetv2_x_obj2coco.yml) | [59.3](https://github.com/Peterande/storage/releases/download/dfinev1.0/dfine_x_obj2coco.pth) | [url](https://raw.githubusercontent.com/Peterande/storage/refs/heads/master/logs/obj2coco/dfine_x_obj2coco_log.txt)
 
 **We highly recommend that you use the Objects365 pre-trained model for fine-tuning:**
+
+⚠️ **Important**: Please note that this is generally beneficial for complex scene understanding. If your categories are very simple, it might lead to overfitting and suboptimal performance.
 <details>
 <summary><strong> 🔥 Pretrained Models on Objects365 (Best generalization) </strong></summary>
 

README_cn.md

@@ -99,6 +99,8 @@ https://github.com/user-attachments/assets/e5933d8e-3c8a-400e-870b-4e452f5321d9
 
 **我们强烈推荐您使用 Objects365 预训练模型进行微调：**
 
+⚠️ 重要提醒：通常这种预训练模型对复杂场景的理解非常有用。如果您的类别非常简单，请注意，这可能会导致过拟合和次优性能。
+
 <details> <summary><strong> 🔥 Objects365 预训练模型（泛化性最好）</strong></summary>
 
 | 模型 | 数据集 | AP<sup>5000</sup> | 参数量 | 时延 (ms) | GFLOPs | 配置 | 权重 | 日志 |

README_ja.md

@@ -114,6 +114,8 @@ https://github.com/user-attachments/assets/e5933d8e-3c8a-400e-870b-4e452f5321d9
 
 **微調整のために Objects365 の事前学習モデルを使用することを強くお勧めします：**
 
+⚠️ 重要なお知らせ：このプリトレインモデルは複雑なシーンの理解に有益ですが、カテゴリが非常に単純な場合、過学習や最適ではない性能につながる可能性がありますので、ご注意ください。
+
 <details> <summary><strong> 🔥 Objects365で事前トレーニングされたモデル（最良の汎化性能）</strong></summary>
 
 | モデル | データセット | AP<sup>5000</sup> | パラメータ数 | レイテンシ | GFLOPs | config | checkpoint | logs |

configs/dfine/custom/dfine_hgnetv2_n_custom.yml

@@ -0,0 +1,82 @@
+__include__: [
+  '../../dataset/custom_detection.yml',
+  '../../runtime.yml',
+  '../include/dataloader.yml',
+  '../include/optimizer.yml',
+  '../include/dfine_hgnetv2.yml',
+]
+
+output_dir: ./output/dfine_hgnetv2_n_custom
+
+
+DFINE:
+  backbone: HGNetv2
+
+HGNetv2:
+  name: 'B0'
+  return_idx: [2, 3]
+  freeze_at: -1
+  freeze_norm: False
+  use_lab: True
+
+
+HybridEncoder:
+  in_channels: [512, 1024]
+  feat_strides: [16, 32]
+
+  # intra
+  hidden_dim: 128
+  use_encoder_idx: [1]
+  dim_feedforward: 512
+
+  # cross
+  expansion: 0.34
+  depth_mult: 0.5
+
+
+DFINETransformer:
+  feat_channels: [128, 128]
+  feat_strides: [16, 32]
+  hidden_dim: 128
+  dim_feedforward: 512
+  num_levels: 2
+
+  num_layers: 3
+  eval_idx: -1
+
+  num_points: [6, 6]
+
+optimizer:
+  type: AdamW
+  params:
+    -
+      params: '^(?=.*backbone)(?!.*norm|bn).*$'
+      lr: 0.0004
+    -
+      params: '^(?=.*backbone)(?=.*norm|bn).*$'
+      lr: 0.0004
+      weight_decay: 0.
+    -
+      params: '^(?=.*(?:encoder|decoder))(?=.*(?:norm|bn|bias)).*$'
+      weight_decay: 0.
+
+  lr: 0.0008
+  betas: [0.9, 0.999]
+  weight_decay: 0.0001
+
+
+# Increase to search for the optimal ema
+epoches: 300
+train_dataloader:
+  total_batch_size: 128
+  dataset:
+    transforms:
+      policy:
+        epoch: 280
+  collate_fn:
+    stop_epoch: 280
+    ema_restart_decay: 0.9999
+    base_size_repeat: ~
+
+val_dataloader:
+  total_batch_size: 256

configs/dfine/custom/objects365/dfine_hgnetv2_l_obj2custom.yml

@@ -1,9 +1,9 @@
 __include__: [
-  '../../dataset/custom_detection.yml',
-  '../../runtime.yml',
-  '../include/dataloader.yml',
-  '../include/optimizer.yml',
-  '../include/dfine_hgnetv2.yml',
+  '../../../dataset/custom_detection.yml',
+  '../../../runtime.yml',
+  '../../include/dataloader.yml',
+  '../../include/optimizer.yml',
+  '../../include/dfine_hgnetv2.yml',
 ]
 
 output_dir: ./output/dfine_hgnetv2_l_obj2custom
@@ -18,6 +18,7 @@ HGNetv2:
   freeze_stem_only: True
   freeze_at: 0
   freeze_norm: True
+  pretrained: False
 
 optimizer:
   type: AdamW

configs/dfine/custom/objects365/dfine_hgnetv2_m_obj2custom.yml

@@ -1,9 +1,9 @@
 __include__: [
-  '../../dataset/custom_detection.yml',
-  '../../runtime.yml',
-  '../include/dataloader.yml',
-  '../include/optimizer.yml',
-  '../include/dfine_hgnetv2.yml',
+  '../../../dataset/custom_detection.yml',
+  '../../../runtime.yml',
+  '../../include/dataloader.yml',
+  '../../include/optimizer.yml',
+  '../../include/dfine_hgnetv2.yml',
 ]
 
 output_dir: ./output/dfine_hgnetv2_m_obj2custom
@@ -18,6 +18,7 @@ HGNetv2:
   freeze_at: -1
   freeze_norm: False
   use_lab: True
+  pretrained: False
 
 DFINETransformer:
   num_layers: 4  # 5 6

configs/dfine/custom/objects365/dfine_hgnetv2_s_obj2custom.yml

@@ -1,9 +1,9 @@
 __include__: [
-  '../../dataset/custom_detection.yml',
-  '../../runtime.yml',
-  '../include/dataloader.yml',
-  '../include/optimizer.yml',
-  '../include/dfine_hgnetv2.yml',
+  '../../../dataset/custom_detection.yml',
+  '../../../runtime.yml',
+  '../../include/dataloader.yml',
+  '../../include/optimizer.yml',
+  '../../include/dfine_hgnetv2.yml',
 ]
 
 output_dir: ./output/dfine_hgnetv2_s_obj2custom
@@ -18,6 +18,7 @@ HGNetv2:
   freeze_at: -1
   freeze_norm: False
   use_lab: True
+  pretrained: False
 
 DFINETransformer:
   num_layers: 3  # 4 5 6

configs/dfine/custom/objects365/dfine_hgnetv2_x_obj2custom.yml

@@ -1,9 +1,9 @@
 __include__: [
-  '../../dataset/custom_detection.yml',
-  '../../runtime.yml',
-  '../include/dataloader.yml',
-  '../include/optimizer.yml',
-  '../include/dfine_hgnetv2.yml',
+  '../../../dataset/custom_detection.yml',
+  '../../../runtime.yml',
+  '../../include/dataloader.yml',
+  '../../include/optimizer.yml',
+  '../../include/dfine_hgnetv2.yml',
 ]
 
 output_dir: ./output/dfine_hgnetv2_x_obj2custom
@@ -18,6 +18,7 @@ HGNetv2:
   freeze_stem_only: True
   freeze_at: 0
   freeze_norm: True
+  pretrained: False
 
 HybridEncoder:
   # intra

configs/dfine/include/dataloader.yml

@@ -24,7 +24,7 @@ train_dataloader:
     stop_epoch: 72 # epoch in [72, ~) stop `multiscales`
 
   shuffle: True
-  total_batch_size: 32 # total batch size equals to 32 s(4 * 8)
+  total_batch_size: 32 # total batch size equals to 32 (4 * 8)
   num_workers: 4
 
 

src/core/_config.py

@@ -151,7 +151,8 @@ def val_dataloader(self) -> DataLoader:
                                 num_workers=self.num_workers,
                                 drop_last=False,
                                 collate_fn=self.collate_fn,
-                                shuffle=self.val_shuffle)
+                                shuffle=self.val_shuffle,
+                                persistent_workers=True)
             loader.shuffle = self.val_shuffle
             self._val_dataloader = loader
 

src/data/dataset/coco_dataset.py

@@ -19,6 +19,7 @@
 
 torchvision.disable_beta_transforms_warning()
 faster_coco_eval.init_as_pycocotools()
+Image.MAX_IMAGE_PIXELS = None
 
 __all__ = ['CocoDetection']
 
@@ -50,7 +51,6 @@ def load_item(self, idx):
 
         if self.remap_mscoco_category:
             image, target = self.prepare(image, target, category2label=mscoco_category2label)
-            # image, target = self.prepare(image, target, category2label=self.category2label)
         else:
             image, target = self.prepare(image, target)
 

src/solver/det_solver.py

@@ -44,7 +44,8 @@ def fit(self, ):
                 best_stat[k] = test_stats[k][0]
                 top1 = test_stats[k][0]
                 print(f'best_stat: {best_stat}')
-
+
+        best_stat_print = best_stat.copy()
         start_time = time.time()
         start_epoch = self.last_epoch + 1
         for epoch in range(start_epoch, args.epoches):
@@ -109,22 +110,34 @@ def fit(self, ):
                 else:
                     best_stat['epoch'] = epoch
                     best_stat[k] = test_stats[k][0]
-
+
+                if best_stat[k] > top1:
+                    best_stat_print['epoch'] = epoch
+                    top1 = best_stat[k]
+                    if self.output_dir:
+                        if epoch >= self.train_dataloader.collate_fn.stop_epoch:
+                            dist_utils.save_on_master(self.state_dict(), self.output_dir / 'best_stg2.pth')
+                        else:
+                            dist_utils.save_on_master(self.state_dict(), self.output_dir / 'best_stg1.pth')
+
+                best_stat_print[k] = max(best_stat[k], top1) 
+                print(f'best_stat: {best_stat_print}')  # global best
+
                 if best_stat['epoch'] == epoch and self.output_dir:
                     if epoch >= self.train_dataloader.collate_fn.stop_epoch:
                         if test_stats[k][0] > top1:
                             top1 = test_stats[k][0]
                             dist_utils.save_on_master(self.state_dict(), self.output_dir / 'best_stg2.pth')
                     else:
+                        top1 = max(test_stats[k][0], top1)
                         dist_utils.save_on_master(self.state_dict(), self.output_dir / 'best_stg1.pth')
 
                 elif epoch >= self.train_dataloader.collate_fn.stop_epoch:
                     best_stat = {'epoch': -1, }
                     self.ema.decay -= 0.0001
                     self.load_resume_state(str(self.output_dir / 'best_stg1.pth'))
                     print(f'Refresh EMA at epoch {epoch} with decay {self.ema.decay}')
-
-            print(f'best_stat: {best_stat}')
+
 
             log_stats = {
                 **{f'train_{k}': v for k, v in train_stats.items()},

tools/benchmark/dataset.py

@@ -13,6 +13,7 @@
 import torchvision.transforms as T
 import torchvision.transforms.functional as F
 
+Image.MAX_IMAGE_PIXELS = None
 
 class ToTensor(T.ToTensor):
     def __init__(self) -> None: