update

PaddlePaddle · Jul 29, 2024 · 423aee8 · 423aee8
1 parent b6db227
commit 423aee8
Show file tree

Hide file tree

Showing 12 changed files with 101 additions and 67 deletions.
diff --git a/configs/rtdetrv2/README.md b/configs/rtdetrv2/README.md
@@ -22,10 +22,10 @@ FPS。而且，混合精度训练策略的使用使得训练速度提高了 15%
 
 |        Model         | Epoch | Backbone  | Input shape | $AP^{val}$ | $AP^{val}_{50}$ | Params(M) | FLOPs(G) | T4 TensorRT FP16(FPS) |                                      Pretrained Model                                       |                    config                    |
 |:--------------------:|:-----:|:---------:|:-----------:|:----------:|:---------------:|:---------:|:--------:|:---------------------:|:-------------------------------------------------------------------------------------------:|:--------------------------------------------:|
-|  *RT-DETRv2-R18-dsp  |  120  | ResNet-18 |     640     |    47.4    |      64.8       |    20     |    60    |          217          |  [download](https://bj.bcebos.com/v1/paddledet/models/rtdetrv2_r18vd_dsp_3x_coco.pdparams)  |  [config](./rtdetrv2_r18vd_dsp_3x_coco.yml)  |
-|  *RT-DETRv2-R34-dsp  |  120  | ResNet-34 |     640     |    49.2    |      67.2       |    31     |    92    |          161          |  [download](https://bj.bcebos.com/v1/paddledet/models/rtdetrv2_r34vd_dsp_1x_coco.pdparams)  |  [config](./rtdetrv2_r34vd_dsp_1x_coco.yml)  |
-| *RT-DETRv2-R50-m-dsp |  84   | ResNet-50 |     640     |    51.3    |      69.7       |    36     |   100    |          145          | [download](https://bj.bcebos.com/v1/paddledet/models/rtdetrv2_r50vd_m_dsp_3x_coco.pdparams) | [config](./rtdetrv2_r50vd_m_dsp_3x_coco.yml) |
-|  *RT-DETRv2-R50-dsp  |  72   | ResNet-50 |     640     |    52.8    |      71.3       |    42     |   136    |          108          |  [download](https://bj.bcebos.com/v1/paddledet/models/rtdetrv2_r50vd_dsp_1x_coco.pdparams)  |  [config](./rtdetrv2_r50vd_dsp_1x_coco.yml)  |
+|  *RT-DETRv2-R18-dsp  |  36   | ResNet-18 |     640     |    47.4    |      64.8       |    20     |    60    |          217          |  [download](https://bj.bcebos.com/v1/paddledet/models/rtdetrv2_r18vd_dsp_3x_coco.pdparams)  |  [config](./rtdetrv2_r18vd_dsp_3x_coco.yml)  |
+|  *RT-DETRv2-R34-dsp  |  12   | ResNet-34 |     640     |    49.2    |      67.2       |    31     |    92    |          161          |  [download](https://bj.bcebos.com/v1/paddledet/models/rtdetrv2_r34vd_dsp_1x_coco.pdparams)  |  [config](./rtdetrv2_r34vd_dsp_1x_coco.yml)  |
+| *RT-DETRv2-R50-m-dsp |  36   | ResNet-50 |     640     |    51.3    |      69.7       |    36     |   100    |          145          | [download](https://bj.bcebos.com/v1/paddledet/models/rtdetrv2_r50vd_m_dsp_3x_coco.pdparams) | [config](./rtdetrv2_r50vd_m_dsp_3x_coco.yml) |
+|  *RT-DETRv2-R50-dsp  |  12   | ResNet-50 |     640     |    52.8    |      71.3       |    42     |   136    |          108          |  [download](https://bj.bcebos.com/v1/paddledet/models/rtdetrv2_r50vd_dsp_1x_coco.pdparams)  |  [config](./rtdetrv2_r50vd_dsp_1x_coco.yml)  |
 
 **注意事项:**
 
@@ -129,7 +129,7 @@ paddle2onnx --model_dir=./output_inference/rtdetrv2_r50vd_6x_coco/ \
 <details>
 <summary>3. 转换成TensorRT（可选） </summary>
 
-- 确保TensorRT的版本>=8.5.1
+- 基础模型请确保TensorRT的版本>=8.5.1，离散采样模型支持TensorRT的版本==8.4甚至一些更早的版本
 - TRT推理可以参考[RT-DETR](https://github.com/lyuwenyu/RT-DETR)的部分代码或者其他网络资源
 
 ```shell

diff --git a/configs/rtdetrv2/_base_/rtdetrv2_r50vd.yml b/configs/rtdetrv2/_base_/rtdetrv2_r50vd.yml
@@ -8,6 +8,7 @@ ema_filter_no_grad: True
 hidden_dim: 256
 use_focal_loss: True
 eval_size: [640, 640]
+reset_norm_param_attr: True
 
 
 DETR:
@@ -22,7 +23,6 @@ ResNet:
   depth: 50
   variant: d
   norm_type: bn
-  norm_decay: 0.
   freeze_at: 0
   return_idx: [1, 2, 3]
   num_stages: 4

diff --git a/configs/rtdetrv2/_base_/rtdetrv2_reader.yml b/configs/rtdetrv2/_base_/rtdetrv2_reader.yml
@@ -2,8 +2,8 @@ worker_num: 4
 TrainReader:
   sample_transforms:
     - Decode: {}
-    - RandomDistort: {prob: 0.5}
-    - RandomExpand: {fill_value: [0., 0, 0.]}
+    - RandomDistort: {prob: 0.8}
+    - RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
     - RandomCrop: {prob: 0.8}
     - RandomFlip: {}
     - Resize: {target_size: [640, 640], keep_ratio: False, interp: 1}
@@ -18,11 +18,13 @@ TrainReader:
   drop_last: true
   collate_batch: false
   use_shared_memory: true
-  transform_schedulers: # [start_epoch, stop_epoch)
-    - RandomDistort: {start_epoch: 0, stop_epoch: 71}
-    - RandomExpand: {start_epoch: 0, stop_epoch: 71}
-    - RandomCrop: {start_epoch: 0, stop_epoch: 71}
-    - BatchRandomResize: {start_epoch: 0, stop_epoch: 71}
+  transform_schedulers:
+    # [start_epoch, stop_epoch) | [0, stop_epoch) | [start_epoch, inf)
+    - RandomDistort: {stop_epoch: 71}
+    - RandomExpand: {stop_epoch: 71}
+    - RandomCrop: {stop_epoch: 71}
+    - Resize: {start_epoch: 71}
+    - BatchRandomResize: {stop_epoch: 71}
 
 
 EvalReader:

diff --git a/configs/rtdetrv2/rtdetrv2_r18vd_120e_coco.yml b/configs/rtdetrv2/rtdetrv2_r18vd_120e_coco.yml
@@ -14,7 +14,6 @@ pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet18_vd_
 
 ResNet:
   depth: 18
-  norm_decay: 0.
   freeze_at: -1
   freeze_norm: false
 
@@ -44,6 +43,6 @@ TrainReader:
     - BboxXYXY2XYWH: {}
     - Permute: {}
   transform_schedulers:
-    - RandomDistort: {start_epoch: 0, stop_epoch: 117}
-    - RandomExpand: {start_epoch: 0, stop_epoch: 117}
-    - RandomCrop: {start_epoch: 0, stop_epoch: 117}
+    - RandomDistort: {stop_epoch: 117}
+    - RandomExpand: {stop_epoch: 117}
+    - RandomCrop: {stop_epoch: 117}
diff --git a/configs/rtdetrv2/rtdetrv2_r18vd_dsp_3x_coco.yml b/configs/rtdetrv2/rtdetrv2_r18vd_dsp_3x_coco.yml
@@ -14,7 +14,6 @@ pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/rtdetrv2_r18vd_120e_
 
 ResNet:
   depth: 18
-  norm_decay: 0.
   freeze_at: -1
   freeze_norm: false
 
@@ -45,6 +44,6 @@ TrainReader:
     - BboxXYXY2XYWH: {}
     - Permute: {}
   transform_schedulers:
-    - RandomDistort: {start_epoch: 0, stop_epoch: 33}
-    - RandomExpand: {start_epoch: 0, stop_epoch: 33}
-    - RandomCrop: {start_epoch: 0, stop_epoch: 33}
+    - RandomDistort: {stop_epoch: 33}
+    - RandomExpand: {stop_epoch: 33}
+    - RandomCrop: {stop_epoch: 33}
diff --git a/configs/rtdetrv2/rtdetrv2_r34vd_120e_coco.yml b/configs/rtdetrv2/rtdetrv2_r34vd_120e_coco.yml
@@ -14,7 +14,6 @@ pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ResNet34_
 
 ResNet:
   depth: 34
-  norm_decay: 0.
   freeze_at: -1
   freeze_norm: false
 
@@ -44,7 +43,8 @@ OptimizerBuilder:
 
 TrainReader:
   transform_schedulers:
-    - RandomDistort: {start_epoch: 0, stop_epoch: 117}
-    - RandomExpand: {start_epoch: 0, stop_epoch: 117}
-    - RandomCrop: {start_epoch: 0, stop_epoch: 117}
-    - BatchRandomResize: {start_epoch: 0, stop_epoch: 117}
+    - RandomDistort: {stop_epoch: 117}
+    - RandomExpand: {stop_epoch: 117}
+    - RandomCrop: {stop_epoch: 117}
+    - Resize: {start_epoch: 117}
+    - BatchRandomResize: {stop_epoch: 117}
diff --git a/configs/rtdetrv2/rtdetrv2_r34vd_dsp_1x_coco.yml b/configs/rtdetrv2/rtdetrv2_r34vd_dsp_1x_coco.yml
@@ -14,7 +14,6 @@ pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/rtdetrv2_r34vd_120e_
 
 ResNet:
   depth: 34
-  norm_decay: 0.
   freeze_at: -1
   freeze_norm: false
 
@@ -45,7 +44,8 @@ OptimizerBuilder:
 
 TrainReader:
   transform_schedulers:
-    - RandomDistort: {start_epoch: 0, stop_epoch: 10}
-    - RandomExpand: {start_epoch: 0, stop_epoch: 10}
-    - RandomCrop: {start_epoch: 0, stop_epoch: 10}
-    - BatchRandomResize: {start_epoch: 0, stop_epoch: 10}
+    - RandomDistort: {stop_epoch: 10}
+    - RandomExpand: {stop_epoch: 10}
+    - RandomCrop: {stop_epoch: 10}
+    - Resize: {start_epoch: 10}
+    - BatchRandomResize: {stop_epoch: 10}
diff --git a/configs/rtdetrv2/rtdetrv2_r50vd_dsp_1x_coco.yml b/configs/rtdetrv2/rtdetrv2_r50vd_dsp_1x_coco.yml
@@ -31,7 +31,8 @@ OptimizerBuilder:
 
 TrainReader:
   transform_schedulers:
-    - RandomDistort: {start_epoch: 0, stop_epoch: 10}
-    - RandomExpand: {start_epoch: 0, stop_epoch: 10}
-    - RandomCrop: {start_epoch: 0, stop_epoch: 10}
-    - BatchRandomResize: {start_epoch: 0, stop_epoch: 10}
+    - RandomDistort: {stop_epoch: 10}
+    - RandomExpand: {stop_epoch: 10}
+    - RandomCrop: {stop_epoch: 10}
+    - Resize: {start_epoch: 10}
+    - BatchRandomResize: {stop_epoch: 10}
diff --git a/configs/rtdetrv2/rtdetrv2_r50vd_m_7x_coco.yml b/configs/rtdetrv2/rtdetrv2_r50vd_m_7x_coco.yml
@@ -33,7 +33,8 @@ OptimizerBuilder:
 
 TrainReader:
   transform_schedulers:
-    - RandomDistort: {start_epoch: 0, stop_epoch: 81}
-    - RandomExpand: {start_epoch: 0, stop_epoch: 81}
-    - RandomCrop: {start_epoch: 0, stop_epoch: 81}
-    - BatchRandomResize: {start_epoch: 0, stop_epoch: 81}
+    - RandomDistort: {stop_epoch: 81}
+    - RandomExpand: {stop_epoch: 81}
+    - RandomCrop: {stop_epoch: 81}
+    - Resize: {start_epoch: 81}
+    - BatchRandomResize: {stop_epoch: 81}
diff --git a/configs/rtdetrv2/rtdetrv2_r50vd_m_dsp_3x_coco.yml b/configs/rtdetrv2/rtdetrv2_r50vd_m_dsp_3x_coco.yml
@@ -36,7 +36,8 @@ OptimizerBuilder:
 
 TrainReader:
   transform_schedulers:
-    - RandomDistort: {start_epoch: 0, stop_epoch: 33}
-    - RandomExpand: {start_epoch: 0, stop_epoch: 33}
-    - RandomCrop: {start_epoch: 0, stop_epoch: 33}
-    - BatchRandomResize: {start_epoch: 0, stop_epoch: 33}
+    - RandomDistort: {stop_epoch: 33}
+    - RandomExpand: {stop_epoch: 33}
+    - RandomCrop: {stop_epoch: 33}
+    - Resize: {start_epoch: 33}
+    - BatchRandomResize: {stop_epoch: 33}
diff --git a/ppdet/data/reader.py b/ppdet/data/reader.py
@@ -44,7 +44,6 @@ class Compose(object):
     def __init__(self, transforms, num_classes=80):
         self.transforms = transforms
         self.transforms_cls = []
-        self.skip_transforms_cls = []
         for t in self.transforms:
             for k, v in t.items():
                 op_cls = getattr(transform, k)
@@ -54,27 +53,29 @@ def __init__(self, transforms, num_classes=80):
 
                 self.transforms_cls.append(f)
 
-    def _update_skip_transforms_cls(self, data):
+    def _update_transforms_cls(self, data):
         if 'transform_schedulers' in data:
-            self.skip_transforms_cls.clear()
-            transform_schedulers = data['transform_schedulers']
-            curr_epoch = data['curr_epoch']
-            for trans_op in self.transforms_cls:
-                trans_op_name = trans_op.__class__.__name__
-                for t in transform_schedulers:
+            def is_valid(op):
+                op_name = op.__class__.__name__
+                for t in data['transform_schedulers']:
                     for k, v in t.items():
-                        if trans_op_name == k:
+                        if op_name == k:
                             # [start_epoch, stop_epoch)
-                            if curr_epoch < v['start_epoch'] or curr_epoch >= v['stop_epoch']:
-                                self.skip_transforms_cls.append(trans_op)
+                            start_epoch = v.get('start_epoch', 0)
+                            if start_epoch > data['curr_epoch']:
+                                return False
+                            stop_epoch = v.get('stop_epoch', float('inf'))
+                            if stop_epoch <= data['curr_epoch']:
+                                return False
+                return True
+
+            return filter(is_valid, self.transforms_cls)
+        else:
+            return self.transforms_cls
 
     def __call__(self, data):
-        self._update_skip_transforms_cls(data)
-
-        for f in self.transforms_cls:
-            if f in self.skip_transforms_cls:
-                continue
-
+        transforms_cls = self._update_transforms_cls(data)
+        for f in transforms_cls:
             try:
                 data = f(data)
             except Exception as e:
@@ -93,12 +94,8 @@ def __init__(self, transforms, num_classes=80, collate_batch=True):
         self.collate_batch = collate_batch
 
     def __call__(self, data):
-        self._update_skip_transforms_cls(data[0])
-
-        for f in self.transforms_cls:
-            if f in self.skip_transforms_cls:
-                continue
-
+        transforms_cls = self._update_transforms_cls(data[0])
+        for f in transforms_cls:
             try:
                 data = f(data)
             except Exception as e:

diff --git a/ppdet/engine/trainer.py b/ppdet/engine/trainer.py
@@ -124,7 +124,12 @@ def __init__(self, cfg, mode='train'):
                     m._epsilon = 1e-3  # for amp(fp16)
                     m._momentum = 0.97  # 0.03 in pytorch
 
-        #normalize params for deploy
+        # reset norm param attr for setting them in optimizer
+        if cfg['reset_norm_param_attr']:
+            self.model = self.reset_norm_param_attr(
+                self.model, weight_attr=None, bias_attr=None)
+
+        # normalize params for deploy
         if 'slim' in cfg and cfg['slim_type'] == 'OFA':
             self.model.model.load_meanstd(cfg['TestReader'][
                 'sample_transforms'])
@@ -1450,3 +1455,32 @@ def setup_metrics_for_loader():
                     imshow_lanes(img, lanes, out_file=out_file)
 
         return results
+
+    def reset_norm_param_attr(self, layer, **kwargs):
+        if isinstance(layer, (nn.BatchNorm2D, nn.LayerNorm, nn.GroupNorm)):
+            src_state_dict = layer.state_dict()
+            if isinstance(layer, nn.BatchNorm2D):
+                layer = nn.BatchNorm2D(
+                    num_features=layer._num_features,
+                    momentum=layer._momentum,
+                    epsilon=layer._epsilon,
+                    **kwargs)
+            elif isinstance(layer, nn.LayerNorm):
+                layer = nn.LayerNorm(
+                    normalized_shape=layer._normalized_shape,
+                    epsilon=layer._epsilon,
+                    **kwargs)
+            else:
+                layer = nn.GroupNorm(
+                    num_groups=layer._num_groups,
+                    num_channels=layer._num_channels,
+                    epsilon=layer._epsilon,
+                    **kwargs)
+            layer.set_state_dict(src_state_dict)
+        else:
+            for name, sublayer in layer.named_children():
+                new_sublayer = self.reset_norm_param_attr(sublayer, **kwargs)
+                if new_sublayer is not sublayer:
+                    setattr(layer, name, new_sublayer)
+
+        return layer