首页 > 科技周边 > 人工智能 > 正文

DETR:基于transformer实现端到端目标检测

P粉084495128
发布： 2025-07-24 11:39:47
原创
200人浏览过
本文复现DETR目标检测方案，基于Transformer实现端到端检测，无需NMS或anchor生成。模型含CNN（Res50）+Transformer+FFN结构，用二分图匹配与匈牙利算法处理预测框。适配Paddle 2.0，修复BUG并对齐精度，DETR-DC5变体在COCO2017验证集MAP达0.431。

☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 免费无限量使用 DeepSeek R1 模型☜☜☜

detr:基于transformer实现端到端目标检测 - php中文网

模型介绍

论文复现第三期DETR方案， 精确度MAP为0.431

End-to-End Object Detection with Transformers

参考代码: https://aistudio.baidu.com/aistudio/projectdetail/1327221

  • 在大佬的基础上进行修改，包括Paddle 2.0适配， BUG修复，以及精度对齐等工作

github pytorch代码: https://github.com/facebookresearch/detr

论文地址: https://arxiv.org/pdf/1706.03762.pdf

  • 基于transformer实现端到端目标检测, 两大主要组成，一者是预测的匹配损失函数，再者预测目标和他们的关系。
  • 不需要使用NMS或者anchor生成， 本文能将该任务的先验知识进行编码。DETR结合目标关联关系和上下文直接并行输出最终的预测集。
  • Bipartite Matching二分图最大匹配， 将节点集V分割为互不相割的子集。训练时将预测框和GT boxes进行匹配。若没有找到匹配的预测框作为"no object"。 最终计算匹配损失。
  • transformer的自注意力机制，专门对序列中成对交互的元素进行建模，所以也很好解决去除重复框
  • 推理出的box坐标是直接基于输入图片而不是基于anchor
  • 先预设固定数量的预测量，通常比图像的目标数量多。南调是对预测目标进行评分，使用最低成对匹配成本，用到匈牙利算法。线性组合IOU损失和L1损失。
  • 网络结构由cnn(res50)+transformer+FFN. 常用的backbone都可以用。cnn到transformer通过1x1卷积进行降维，空间维度压缩到一维， FFN由3层MLP+relu组成。
  • backbone和transformer分开训练，学习率设置不一样。具体是怎样的方式？
  • 关于transformer，
    • 其实就是全连接(或一维卷积)加上attention结合体，对于局部特征捕获能力稍欠缺。
    • 编解码结构较大的局限性是固定长度的语义向量， 编码器压缩成一个固长向量无法表示整个序列信息, 前期信息容易被后序信息覆盖，解码时输入每个单词权重不一致。
    • attention解决编解码结构的弊端，编码器时将正序和逆序的隐藏层状态结合起来， 解码时步骤，1.计算编码器隐藏层状态和解码器隐藏层状态间的相关程度并用softmax归一化得到权重值， 2. 计算语义编码向量的加权和，3. 根据加权和计算隐藏状态和解码器输出。不同编解码结构的是解码时选择性的从向量序列中挑选。
    • soft attention缺点是每次decode都会计算所有编码器隐藏层状态向量，计算复杂度较高。hard attention每次近选择一个source进行计算，缺点是不可微，无法反向传播。
    • global attetion也属于soft attention， 只是计算权重公式上有差别。实验表明general方式好一些。
    • local attention, 每次选择一部分source进行计算， 既减少计算量又能可微。思路是为decoder预测一个source位置，然后基于位置选择一个窗口用于计算编码向量。关键是怎么确定pt位置， monotoic或predictive。
    • self attention, 传统attention是基于target和source间的相似度，而self是发生在target内部或者source内部的相似关联。self更容易捕获长距依赖特征。query和key点积相当于给每个单词打分，决定了编码单词时重视句子其他部分的程度。softmax决定了其他单词对当前单词编码的贡献，self作为最大，其他单词也很有参考作用。
    • transformer和attention机制一样，只是更复杂，多个encoder和docoder堆叠一起， encoder包含self-attention(不仅仅当前词还有上下文)和神经网络层(可并行)，decoder多了一层attention(当前需关注的重点内容)。encoder前要对输入数据进行embedding操作。
    • 位置编码， transformer缺少解释序列中单词顺序的方法，positional encoding在维度上和embedding一样，偶数位置正弦编码，奇数位置余弦编码，决定当前词的位置或不同词间距离，并且模型可学习到。
    • 多头注意力，主要扩展了模型专注不同位置的能力，给出了注意力层的多个表示子空间。8个头就有8套qkv矩阵，每个头的qkv矩阵权重独立，将这些矩阵拼接在一起并用附加权重矩阵相乘。
    • transformer还使用到了残差和layer norm。LN是在每一个样本上计算均值和方差，而不是BN那种在批方向计算均值和方差。还用到了mask，掩盖值，使其参数更新时不产生效果， padding mask在较短序列填充0，Sequence mask只依赖t之前的信息，将t之后信息掩盖起来。

关于数据集COCO2017

COCO的全称是Common Objects in Context，是微软团队提供的一个可以用来进行图像识别的数据集。MS COCO数据集中的图像分为训练、验证和测试集。其行业地位就不再多少了，本文主要梳理一下该数据集包含的内容。下图是官网给出的可下载的数据集（更新时间2020年01月09日），从这里可看出其数据集主要包括有标注的和无标注的数据。

DETR:基于transformer实现端到端目标检测 - php中文网        

In [ ]
#解压数据集%cd /home/aistudio/data/data7122/ 
!unzip train2017.zip!unzip val2017.zip !unzip annotations_trainval2017.zip
登录后复制
   
In [ ]
#加载数据集%cd ~/my_detr
!python coco_dataset.py
登录后复制
   

模型结构搭建

  1. Backbone在resnet50基础上修改，支持DC5变体
  2. 搭建transformer结构
  3. 搭建后处理包括匈牙利匹配算法
  4. 损失函数
  5. 后处理

核心代码主要有:

  • model.py
  • resnet.py
  • backbone.py
  • transformer.py

DETR:基于transformer实现端到端目标检测 - php中文网        

In [ ]
<br/>
登录后复制
   

精度对齐

因为loss函数以及后处理较为复杂，所以这里特定还用了loss精度对齐

神卷标书
神卷标书

神卷标书，专注于AI智能标书制作、管理与咨询服务，提供高效、专业的招投标解决方案。支持一站式标书生成、模板下载，助力企业轻松投标，提升中标率。

神卷标书 39
查看详情 神卷标书

输出结果结果精度对齐

模拟tensor输入: image = [paddle.ones(shape=[3, 800, 1199])] samples.tensors = torch.ones(1, 3, 800, 1199) 因为resize对图像输入有一定差距，所以使用ones的张量, 比如使用样例数据, 0.348(torch) vs 0.31(paddle)

pytorch结果:

DETR:基于transformer实现端到端目标检测 - php中文网        

paddle结果:

DETR:基于transformer实现端到端目标检测 - php中文网        

LOSS 精度对齐

相差小数点2位

Pytorch的结果:

DETR:基于transformer实现端到端目标检测 - php中文网        

PaddlePaddle的结果:

DETR:基于transformer实现端到端目标检测 - php中文网        

训练DETR

In [ ]
#开始训练%cd ~/my_detr
!python train_val.py train
登录后复制
       
/home/aistudio/my_detr
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/utils.py:26: DeprecationWarning: `np.int` is a deprecated alias for the builtin `int`. To silence this warning, use `int` by itself. Doing this will not modify any behavior and is safe. When replacing `np.int`, you may wish to use e.g. `np.int64` or `np.int32` to specify the precision. If you wish to review your current use, check the release note link for additional information.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  def convert_to_list(value, n, name, dtype=np.int):
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/__init__.py:107: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import MutableMapping
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/rcsetup.py:20: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Iterable, Mapping
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/colors.py:53: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Sized
---->
{'hidden_dim': 256, 'lr_backbone': -1, 'masks': False, 'dilation': False, 'backbone': 'resnet50', 'num_classes': 91, 'dropout': 0.1, 'nheads': 8, 'dim_feedforward': 2048, 'enc_layers': 6, 'dec_layers': 6, 'pre_norm': False, 'num_queries': 100, 'aux_loss': True, 'set_cost_class': 1, 'set_cost_bbox': 5, 'set_cost_giou': 2, 'bbox_loss_coef': 5, 'giou_loss_coef': 2, 'eos_coef': 0.1, 'coco_path': '/home/aistudio/data/data7122', 'lr': 1e-06, 'clip_max_norm': 0.1, 'batch_size': 8, 'epochs': 2}
{'num_classes': 91, 'norm_layer': <class 'backbone.FrozenBatchNorm2d'>}
block <class 'resnet.BottleneckBlock'>
[False, False, False]
W0514 19:16:37.620630 26619 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.0, Runtime API Version: 10.1
W0514 19:16:37.627422 26619 device_context.cc:372] device: 0, cuDNN Version: 7.6.
debug: 1 1
debug: 2 1
debug: 2 1
debug: 2 1
Epoch 0: StepDecay set learning rate to 1e-06.
loading annotations into memory...
Done (t=0.56s)
creating index...
index created!
loading annotations into memory...
Done (t=0.69s)
creating index...
index created!
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/framework.py:687: DeprecationWarning: `np.bool` is a deprecated alias for the builtin `bool`. To silence this warning, use `bool` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.bool_` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  elif dtype == np.bool:
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/tensor/creation.py:143: DeprecationWarning: `np.object` is a deprecated alias for the builtin `object`. To silence this warning, use `object` by itself. Doing this will not modify any behavior and is safe. 
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  if data.dtype == np.object:
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py:238: UserWarning: The dtype of left and right variables are not the same, left dtype is VarType.FP32, but right dtype is VarType.INT64, the right dtype will convert to VarType.FP32
  format(lhs_dtype, rhs_dtype, lhs_dtype))
epoch: 0, batch_id: 0, loss: 10.05978012084961
epoch: 0, batch_id: 100, loss: 6.360683917999268
epoch: 0, batch_id: 200, loss: 6.134881973266602
epoch: 0, batch_id: 300, loss: 7.114040851593018
epoch: 0, batch_id: 400, loss: 6.500746250152588
epoch: 0, batch_id: 500, loss: 6.7673797607421875
epoch: 0, batch_id: 600, loss: 5.745387554168701
epoch: 1, batch_id: 0, loss: 7.610352993011475
epoch: 1, batch_id: 100, loss: 5.726753234863281
epoch: 1, batch_id: 200, loss: 5.837918758392334
epoch: 1, batch_id: 300, loss: 6.997137069702148
epoch: 1, batch_id: 400, loss: 6.292409420013428
epoch: 1, batch_id: 500, loss: 6.677578926086426
epoch: 1, batch_id: 600, loss: 5.692938327789307
登录后复制
       

验证DETR

In [14]
%cd ~/my_detr
!python train_val.py eval
登录后复制
       
/home/aistudio/my_detr
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/utils.py:26: DeprecationWarning: `np.int` is a deprecated alias for the builtin `int`. To silence this warning, use `int` by itself. Doing this will not modify any behavior and is safe. When replacing `np.int`, you may wish to use e.g. `np.int64` or `np.int32` to specify the precision. If you wish to review your current use, check the release note link for additional information.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  def convert_to_list(value, n, name, dtype=np.int):
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/__init__.py:107: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import MutableMapping
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/rcsetup.py:20: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Iterable, Mapping
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/colors.py:53: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Sized
---->
{'hidden_dim': 256, 'lr_backbone': -1, 'masks': False, 'dilation': False, 'backbone': 'resnet50', 'num_classes': 91, 'dropout': 0.1, 'nheads': 8, 'dim_feedforward': 2048, 'enc_layers': 6, 'dec_layers': 6, 'pre_norm': False, 'num_queries': 100, 'aux_loss': True, 'set_cost_class': 1, 'set_cost_bbox': 5, 'set_cost_giou': 2, 'bbox_loss_coef': 5, 'giou_loss_coef': 2, 'eos_coef': 0.1, 'coco_path': '/home/aistudio/data/data7122', 'lr': 1e-06, 'clip_max_norm': 0.1, 'batch_size': 8, 'epochs': 2}
{'num_classes': 91, 'norm_layer': <class 'backbone.FrozenBatchNorm2d'>}
block <class 'resnet.BottleneckBlock'>
[False, False, False]
W0514 21:08:15.278102  7932 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.0, Runtime API Version: 10.1
W0514 21:08:15.283421  7932 device_context.cc:372] device: 0, cuDNN Version: 7.6.
debug: 1 1
debug: 2 1
debug: 2 1
debug: 2 1
Epoch 0: StepDecay set learning rate to 1e-06.
loading annotations into memory...
Done (t=0.55s)
creating index...
index created!
loading annotations into memory...
Done (t=0.68s)
creating index...
index created!
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/framework.py:687: DeprecationWarning: `np.bool` is a deprecated alias for the builtin `bool`. To silence this warning, use `bool` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.bool_` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  elif dtype == np.bool:
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py:238: UserWarning: The dtype of left and right variables are not the same, left dtype is VarType.FP32, but right dtype is VarType.INT64, the right dtype will convert to VarType.FP32
  format(lhs_dtype, rhs_dtype, lhs_dtype))
Accumulating evaluation results...
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/pycocotools/cocoeval.py:378: DeprecationWarning: `np.float` is a deprecated alias for the builtin `float`. To silence this warning, use `float` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.float64` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  tp_sum = np.cumsum(tps, axis=1).astype(dtype=np.float)
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/pycocotools/cocoeval.py:379: DeprecationWarning: `np.float` is a deprecated alias for the builtin `float`. To silence this warning, use `float` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.float64` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  fp_sum = np.cumsum(fps, axis=1).astype(dtype=np.float)
DONE (t=11.61s).
IoU metric: bbox
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.420
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.624
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.442
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.213
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.460
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.611
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.334
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.533
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.575
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.325
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.631
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.809
登录后复制
       

训练DETR-DC5

这个模型非常耗显存, 很难训练

In [ ]
#开始训练DC5%cd ~/my_detr
!python train_val_dc5.py train
登录后复制
       
/home/aistudio/my_detr
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/utils.py:26: DeprecationWarning: `np.int` is a deprecated alias for the builtin `int`. To silence this warning, use `int` by itself. Doing this will not modify any behavior and is safe. When replacing `np.int`, you may wish to use e.g. `np.int64` or `np.int32` to specify the precision. If you wish to review your current use, check the release note link for additional information.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  def convert_to_list(value, n, name, dtype=np.int):
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/__init__.py:107: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import MutableMapping
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/rcsetup.py:20: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Iterable, Mapping
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/colors.py:53: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Sized
---->
{'hidden_dim': 256, 'lr_backbone': -1, 'masks': False, 'dilation': True, 'backbone': 'resnet50', 'num_classes': 91, 'dropout': 0.1, 'nheads': 8, 'dim_feedforward': 2048, 'enc_layers': 6, 'dec_layers': 6, 'pre_norm': False, 'num_queries': 100, 'aux_loss': True, 'set_cost_class': 1, 'set_cost_bbox': 5, 'set_cost_giou': 2, 'bbox_loss_coef': 5, 'giou_loss_coef': 2, 'eos_coef': 0.1, 'coco_path': '/home/aistudio/data/data7122', 'lr': 1e-06, 'clip_max_norm': 0.1, 'batch_size': 1, 'epochs': 1}
{'num_classes': 91, 'replace_stride_with_dilation': [False, False, True], 'norm_layer': <class 'backbone.FrozenBatchNorm2d'>}
block <class 'resnet.BottleneckBlock'>
[False, False, True]
W0514 17:11:00.507395 24269 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.0, Runtime API Version: 10.1
W0514 17:11:00.512784 24269 device_context.cc:372] device: 0, cuDNN Version: 7.6.
debug: 1 1
debug: 2 1
debug: 2 1
debug: 1 2
530
Epoch 0: StepDecay set learning rate to 1e-06.
loading annotations into memory...
Done (t=0.57s)
creating index...
index created!
loading annotations into memory...
Done (t=0.70s)
creating index...
index created!
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/framework.py:687: DeprecationWarning: `np.bool` is a deprecated alias for the builtin `bool`. To silence this warning, use `bool` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.bool_` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  elif dtype == np.bool:
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/tensor/creation.py:143: DeprecationWarning: `np.object` is a deprecated alias for the builtin `object`. To silence this warning, use `object` by itself. Doing this will not modify any behavior and is safe. 
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  if data.dtype == np.object:
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py:238: UserWarning: The dtype of left and right variables are not the same, left dtype is VarType.FP32, but right dtype is VarType.INT64, the right dtype will convert to VarType.FP32
  format(lhs_dtype, rhs_dtype, lhs_dtype))
epoch: 0, batch_id: 0, loss: 5.015964031219482
epoch: 0, batch_id: 100, loss: 14.247769355773926
epoch: 0, batch_id: 200, loss: 7.298154830932617
epoch: 0, batch_id: 300, loss: 1.0471426248550415
epoch: 0, batch_id: 400, loss: 7.584997653961182
epoch: 0, batch_id: 500, loss: 3.577380895614624
epoch: 0, batch_id: 600, loss: 4.61794900894165
epoch: 0, batch_id: 700, loss: 5.049403667449951
epoch: 0, batch_id: 800, loss: 5.122508525848389
epoch: 0, batch_id: 900, loss: 3.216052770614624
epoch: 0, batch_id: 1000, loss: 3.3042514324188232
epoch: 0, batch_id: 1100, loss: 4.413068771362305
epoch: 0, batch_id: 1200, loss: 7.288424015045166
epoch: 0, batch_id: 1300, loss: 6.1409735679626465
epoch: 0, batch_id: 1400, loss: 10.504143714904785
epoch: 0, batch_id: 1500, loss: 3.685210704803467
epoch: 0, batch_id: 1600, loss: 2.6168665885925293
epoch: 0, batch_id: 1700, loss: 22.14373016357422
epoch: 0, batch_id: 1800, loss: 8.267280578613281
epoch: 0, batch_id: 1900, loss: 1.4486600160598755
epoch: 0, batch_id: 2000, loss: 4.107017993927002
epoch: 0, batch_id: 2100, loss: 9.582965850830078
epoch: 0, batch_id: 2200, loss: 6.967478275299072
epoch: 0, batch_id: 2300, loss: 15.338693618774414
epoch: 0, batch_id: 2400, loss: 13.399685859680176
epoch: 0, batch_id: 2500, loss: 1.4326478242874146
epoch: 0, batch_id: 2600, loss: 6.990074157714844
epoch: 0, batch_id: 2700, loss: 8.32422161102295
epoch: 0, batch_id: 2800, loss: 5.453993797302246
epoch: 0, batch_id: 2900, loss: 7.272365093231201
epoch: 0, batch_id: 3000, loss: 8.217702865600586
epoch: 0, batch_id: 3100, loss: 2.2091081142425537
epoch: 0, batch_id: 3200, loss: 0.9168111085891724
epoch: 0, batch_id: 3300, loss: 1.9928405284881592
epoch: 0, batch_id: 3400, loss: 6.245678424835205
epoch: 0, batch_id: 3500, loss: 8.91486930847168
epoch: 0, batch_id: 3600, loss: 6.916267395019531
epoch: 0, batch_id: 3700, loss: 4.461080551147461
epoch: 0, batch_id: 3800, loss: 8.265044212341309
epoch: 0, batch_id: 3900, loss: 8.603659629821777
epoch: 0, batch_id: 4000, loss: 4.085428714752197
epoch: 0, batch_id: 4100, loss: 10.250466346740723
epoch: 0, batch_id: 4200, loss: 2.327882766723633
epoch: 0, batch_id: 4300, loss: 3.273315906524658
epoch: 0, batch_id: 4400, loss: 6.719542026519775
epoch: 0, batch_id: 4500, loss: 3.232994794845581
epoch: 0, batch_id: 4600, loss: 4.106349945068359
epoch: 0, batch_id: 4700, loss: 1.6361501216888428
epoch: 0, batch_id: 4800, loss: 4.615266799926758
epoch: 0, batch_id: 4900, loss: 8.394232749938965
start evaluating....
Accumulating evaluation results...
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/pycocotools/cocoeval.py:378: DeprecationWarning: `np.float` is a deprecated alias for the builtin `float`. To silence this warning, use `float` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.float64` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  tp_sum = np.cumsum(tps, axis=1).astype(dtype=np.float)
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/pycocotools/cocoeval.py:379: DeprecationWarning: `np.float` is a deprecated alias for the builtin `float`. To silence this warning, use `float` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.float64` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  fp_sum = np.cumsum(fps, axis=1).astype(dtype=np.float)
DONE (t=9.37s).
IoU metric: bbox
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.432
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.634
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.458
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.223
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.471
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.612
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.340
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.551
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.591
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.334
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.642
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.815
save weights with map:  0.43226377363430357
登录后复制
       

验证DETR-DC5

In [15]
#验证DETR-DC5%cd ~/my_detr
!python train_val_dc5.py eval
登录后复制
       
/home/aistudio/my_detr
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/utils.py:26: DeprecationWarning: `np.int` is a deprecated alias for the builtin `int`. To silence this warning, use `int` by itself. Doing this will not modify any behavior and is safe. When replacing `np.int`, you may wish to use e.g. `np.int64` or `np.int32` to specify the precision. If you wish to review your current use, check the release note link for additional information.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  def convert_to_list(value, n, name, dtype=np.int):
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/__init__.py:107: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import MutableMapping
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/rcsetup.py:20: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Iterable, Mapping
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/matplotlib/colors.py:53: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working
  from collections import Sized
---->
{'hidden_dim': 256, 'lr_backbone': -1, 'masks': False, 'dilation': True, 'backbone': 'resnet50', 'num_classes': 91, 'dropout': 0.1, 'nheads': 8, 'dim_feedforward': 2048, 'enc_layers': 6, 'dec_layers': 6, 'pre_norm': False, 'num_queries': 100, 'aux_loss': True, 'set_cost_class': 1, 'set_cost_bbox': 5, 'set_cost_giou': 2, 'bbox_loss_coef': 5, 'giou_loss_coef': 2, 'eos_coef': 0.1, 'coco_path': '/home/aistudio/data/data7122', 'lr': 1e-06, 'clip_max_norm': 0.1, 'batch_size': 1, 'epochs': 1}
{'num_classes': 91, 'replace_stride_with_dilation': [False, False, True], 'norm_layer': <class 'backbone.FrozenBatchNorm2d'>}
block <class 'resnet.BottleneckBlock'>
[False, False, True]
W0514 21:21:04.015319  9341 device_context.cc:362] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.0, Runtime API Version: 10.1
W0514 21:21:04.020784  9341 device_context.cc:372] device: 0, cuDNN Version: 7.6.
debug: 1 1
debug: 2 1
debug: 2 1
debug: 1 2
Epoch 0: StepDecay set learning rate to 1e-06.
loading annotations into memory...
Done (t=0.56s)
creating index...
index created!
loading annotations into memory...
Done (t=0.69s)
creating index...
index created!
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/framework.py:687: DeprecationWarning: `np.bool` is a deprecated alias for the builtin `bool`. To silence this warning, use `bool` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.bool_` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  elif dtype == np.bool:
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/dygraph/math_op_patch.py:238: UserWarning: The dtype of left and right variables are not the same, left dtype is VarType.FP32, but right dtype is VarType.INT64, the right dtype will convert to VarType.FP32
  format(lhs_dtype, rhs_dtype, lhs_dtype))
Accumulating evaluation results...
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/pycocotools/cocoeval.py:378: DeprecationWarning: `np.float` is a deprecated alias for the builtin `float`. To silence this warning, use `float` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.float64` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  tp_sum = np.cumsum(tps, axis=1).astype(dtype=np.float)
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/pycocotools/cocoeval.py:379: DeprecationWarning: `np.float` is a deprecated alias for the builtin `float`. To silence this warning, use `float` by itself. Doing this will not modify any behavior and is safe. If you specifically wanted the numpy scalar type, use `np.float64` here.
Deprecated in NumPy 1.20; for more details and guidance: https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations
  fp_sum = np.cumsum(fps, axis=1).astype(dtype=np.float)
DONE (t=9.90s).
IoU metric: bbox
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.431
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.630
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.458
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.223
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.470
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.609
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.340
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.549
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.593
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.339
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.645
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.812
登录后复制
       

以上就是DETR:基于transformer实现端到端目标检测的详细内容，更多请关注php中文网其它相关文章！

相关标签：
python git facebook ai cos red Object 堆 map padding github 算法 cnn pytorch paddlepaddle transformer https bug embedding

大家都在看：

最佳 Windows 性能的顶级免费优化软件
最佳 Windows 性能的顶级免费优化软件

每个人都需要一台速度更快、更稳定的 PC。随着时间的推移，垃圾文件、旧注册表数据和不必要的后台进程会占用资源并降低性能。幸运的是，许多工具可以让 Windows 保持平稳运行。

下载
来源：php中文网
收藏 点赞
上一篇：【特训营第三期 】基于PaddleClas的多场景杂草分类 下一篇：论文复现：WS-DAN细粒度分类问题经典之作
本文内容由网友自发贡献，版权归原作者所有，本站不承担相应法律责任。如您发现有涉嫌抄袭侵权的内容，请联系admin@php.cn
作者最新文章
最新问题
讯飞火星AI怎样生成高质量文本_讯飞火星AI文本生成技巧与优化方法 优化提示结构、调整生成参数、利用上下文记忆和后处理校对可提升讯飞火星AI文本质量。通过设定角色、明确任务与格式，控制温度、长度及重复惩罚，保持话题连贯，并结合人工润色，确保输出精准逻辑清晰。
2025-11-05 11:15:12
951
即梦可以用来写论文或者代码吗_即梦写论文与代码使用说明 即梦可辅助撰写论文或代码，提供思路、格式建议与初稿生成，但用户需自行验证内容准确性与原创性。
2025-11-05 10:58:02
472
夸克A眼镜如何恢复备份_夸克A眼镜数据备份与恢复操作详细教程 可通过夸克云服务、本地设备传输或恢复密钥三种方式恢复夸克A眼镜数据：首先确保设备联网并登录账号，选择云备份与恢复功能下载历史备份；若无云端备份，可将存储在手机或电脑的.qab文件通过USB传输至眼镜，并在系统工具中导入；对于加密备份，需在初始化时输入12位恢复密钥验证后选择内容恢复，全程保持供电稳定。
2025-11-05 10:52:02
944
星火大模型体验页_科大讯飞AI网页版服务平台 星火大模型体验页科大讯飞AI网页版服务平台为https://www.xfyun.cn/，提供多语种语音识别、深度推理能力及130多种语言实时切换功能；集成35000多个智能助手，支持个性化智能体搭建与API接入；内置PPT生成、虚拟人视频制作和写作辅助工具；开放API矩阵，提供免费试用与多版本模型选择，支持开源模型接入与定制微调。
2025-11-05 10:45:21
686
即梦生成的图片可以商用吗_即梦图片商用政策说明 是的，JIEMUAI生成的图像通常可用于商业用途。平台允许在遵守使用指南的前提下将内容用于电商视觉、宣传材料等商业场景，但禁止未经授权的转售或滥用行为。
2025-11-05 10:28:02
323
夸克AI平台智能助手官网 夸克AI官方主页直达访问官方地址 夸克AI平台智能助手官网是https://www.quark.cn/，该网站集成AI搜索、文档解析、网页提取和多轮对话功能，支持跨设备数据同步与网盘服务，提供极简高效的操作体验。
2025-11-05 09:37:03
264
DeepSeekOCR本地部署如何进行模型微调_DeepSeekOCR模型微调与自定义训练方法介绍 DeepSeekOCR可通过逆向分析模型结构并构建兼容数据流程实现本地微调。首先确认模型格式与框架支持，获取推理代码或反向工程网络结构；接着准备训练环境与标注数据集，选用PyTorch+MMOCR等框架进行数据预处理；然后加载预训练权重，采用冻结骨干网络、分阶段微调策略，结合CTC或DBLoss损失函数与小学习率优化；随后编写自定义训练循环，实现参数更新与学习率调度；最后在验证集上评估性能，保存模型为Checkpoint或ONNX/TorchScript格式，并结合TensorRT等工具部署。关
2025-11-05 09:06:02
709
瑞达写作平台登录入口 瑞达写作App官方正版应用商店 瑞达写作平台登录入口位于官网https://www.ruidawriting.com，用户可在此下载官方正版App并使用智能内容生成、文献整合、高效大纲构建、长文管理、查重优化与内容润色等功能。
2025-11-05 08:45:02
722
华为AI眼镜如何使用手势控制_华为AI眼镜手势识别与控制功能教程 华为AI眼镜支持手势控制音频播放与接听电话。一、双击镜腿可播放/暂停音乐，注意敲击速度与力度。二、通过App启用“轻滑”功能后，前后滑动镜腿可切换上下曲目。三、滑动镜腿还可调节音量，前滑增大、后滑减小，系统实时提示音量等级。四、双击左镜腿可唤醒语音助手，需在App中开启权限并保持蓝牙连接，实现语音拨号与信息查询等功能。
2025-11-05 08:19:02
676
讯飞火星AI怎样进行婚礼策划文案生成_讯飞火星AI婚礼流程与文案智能创作方法 讯飞星火AI可通过星火对话生成婚礼文案初稿，使用内容写作功能定制宣传文案，结合讯飞智文一键生成图文并茂的婚礼策划方案，提升创作效率与质量。
2025-11-05 07:19:10
734
相关专题
更多>
热门推荐
开源免费商场系统广告
热门教程
更多>
相关推荐
热门推荐
最新课程
最新下载
更多>
网站特效
网站源码
网站素材
前端模板
关于我们 免责申明 意见反馈 讲师合作 广告合作 最新更新 English
php中文网：公益在线php培训，帮助PHP学习者快速成长！
关注服务号 技术交流群
PHP中文网订阅号
每天精选资源文章推送
PHP中文网APP
随时随地碎片化学习

Copyright 2014-2025 https://www.php.cn/ All Rights Reserved | php.cn | 湘ICP备2023035733号

  • PHP学习

  • 技术支持

  • 返回顶部