本文提出了一个简单的 MLP-like 的架构 CycleMLP,它是视觉识别和密集预测的通用主干,不同于现代 MLP 架构,例如 MLP-Mixer、ResMLP 和 gMLP,其架构与图像大小相关,因此是在目标检测和分割中不可行。
☞☞☞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☜☜☜

这是一个PaddlePaddle实现的CycleMLP。
本文提出了一个简单的 MLP-like 的架构 CycleMLP,它是视觉识别和密集预测的通用主干,不同于现代 MLP 架构,例如 MLP-Mixer、ResMLP 和 gMLP,其架构与图像大小相关,因此是在目标检测和分割中不可行。
论文: CycleMLP: A MLP-like Architecture for Dense Prediction
参考repo: CycleMLP
在此非常感谢ShoufaChen贡献的CycleMLP,提高了本repo复现论文的效率。
数据集为ImageNet,训练集包含1281167张图像,验证集包含50000张图像。
│imagenet/ ├──train/ │ ├── n01440764 │ │ ├── n01440764_10026.JPEG │ │ ├── n01440764_10027.JPEG │ │ ├── ......│ ├── ......├──val/ │ ├── n01440764 │ │ ├── ILSVRC2012_val_00000293.JPEG │ │ ├── ILSVRC2012_val_00002138.JPEG │ │ ├── ......│ ├── ......
您可以从ImageNet 官网申请下载数据。
| 模型 | top1 acc (参考精度) | top1 acc (复现精度) | 权重 | 训练日志 |
|---|---|---|---|
| CycleMLP-B1 | 0.789 | 0.790 | checkpoint-best.pd | train.log |
权重及训练日志下载地址:百度网盘
硬件和框架版本等环境的要求如下:
# 需要安装2.2及以上版本的Paddle,如果# 安装GPU版本的Paddlepip install paddlepaddle-gpu==2.2.0# 安装CPU版本的Paddlepip install paddlepaddle==2.2.0
更多安装方法可以参考:Paddle安装指南。
%cd /home/aistudio !git clone https://github.com/flytocc/CycleMLP-paddle.git
%cd /home/aistudio/CycleMLP-paddle !pip install -r requirements.txt
如果您已经ImageNet1k数据集,那么该步骤可以跳过,如果您没有,则可以从ImageNet官网申请下载。
CycleFC模块
与现代方法相比,CycleMLP 有两个优势。
(1) 可以应对各种图像尺寸。
(2) 利用局部窗口实现对图像大小的线性计算复杂度。

class CycleFC(nn.Layer):
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size, # re-defined kernel_size, represent the spatial area of staircase FC
stride: int = 1,
padding: int = 0,
dilation: int = 1,
groups: int = 1,
bias: bool = True, ):
super(CycleFC, self).__init__() if in_channels % groups != 0: raise ValueError('in_channels must be divisible by groups') if out_channels % groups != 0: raise ValueError('out_channels must be divisible by groups') if stride != 1: raise ValueError('stride must be 1') if padding != 0: raise ValueError('padding must be 0')
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = to_2tuple(stride)
self.padding = to_2tuple(padding)
self.dilation = to_2tuple(dilation)
self.groups = groups
self.deformable_groups = self.in_channels
self.weight = self.create_parameter(
shape=[out_channels, in_channels // groups, 1, 1]) # kernel size == 1
if bias:
self.bias = self.create_parameter(shape=[out_channels]) else:
self.bias = None
self.register_buffer('offset', self.gen_offset()) def gen_offset(self):
"""
offset (Tensor[batch_size, 2 * offset_groups * kernel_height * kernel_width,
out_height, out_width]): offsets to be applied for each position in the
convolution kernel.
"""
offset = paddle.empty([1, self.in_channels * 2, 1, 1])
start_idx = (self.kernel_size[0] * self.kernel_size[1]) // 2
assert self.kernel_size[0] == 1 or self.kernel_size[1] == 1, self.kernel_size for i in range(self.in_channels): if self.kernel_size[0] == 1:
offset[0, 2 * i + 0, 0, 0] = 0
offset[0, 2 * i + 1, 0, 0] = (i + start_idx) % self.kernel_size[1] - (self.kernel_size[1] // 2) else:
offset[0, 2 * i + 0, 0, 0] = (i + start_idx) % self.kernel_size[0] - (self.kernel_size[0] // 2)
offset[0, 2 * i + 1, 0, 0] = 0
return offset def forward(self, input):
"""
Args:
input (Tensor[batch_size, in_channels, in_height, in_width]): input tensor
"""
B, C, H, W = input.shape return deform_conv2d(input, self.offset.expand([B, -1, H, W]), self.weight, bias=self.bias, stride=self.stride,
padding=self.padding, dilation=self.dilation, deformable_groups=self.deformable_groups)构建CycleMLP模块
class CycleMLP(nn.Layer):
def __init__(self,
dim,
qkv_bias=False,
qk_scale=None,
attn_drop=0.,
proj_drop=0.):
super().__init__()
self.mlp_c = nn.Linear(dim, dim, bias_attr=qkv_bias)
self.sfc_h = CycleFC(dim, dim, (1, 3), 1, 0)
self.sfc_w = CycleFC(dim, dim, (3, 1), 1, 0)
self.reweight = Mlp(dim, dim // 4, dim * 3)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop) def forward(self, x):
B, H, W, C = x.shape
h = self.sfc_h(x.transpose([0, 3, 1, 2])).transpose([0, 2, 3, 1])
w = self.sfc_w(x.transpose([0, 3, 1, 2])).transpose([0, 2, 3, 1])
c = self.mlp_c(x)
a = (h + w + c).transpose([0, 3, 1, 2]).flatten(2).mean(2)
a = self.reweight(a).reshape([B, C, 3]).transpose([2, 0, 1])
a = F.softmax(a, axis=0).unsqueeze(2).unsqueeze(2)
x = h * a[0] + w * a[1] + c * a[2]
x = self.proj(x)
x = self.proj_drop(x) return x参考论文及official code,主要超参如下:
| total batxh size | learning rate | epochs |
|---|---|---|
| 1024 | 1e-3 | 300 |

%cd /home/aistudio/CycleMLP-paddle
%run infer.py \
--model=CycleMLP_B1 \
--infer_imgs=/home/aistudio/CycleMLP-paddle/demo/ILSVRC2012_val_00020010.JPEG \
--resume=/home/aistudio/CycleMLP_B1.pdparams最终输出结果为
[{'class_ids': [178, 211, 209, 210, 246], 'scores': [0.9213957190513611, 0.006610415875911713, 0.0018257270567119122, 0.0013606979046016932, 0.001132593140937388], 'file_name': '/home/aistudio/CycleMLP-paddle/demo/ILSVRC2012_val_00020010.JPEG', 'label_names': ['Weimaraner', 'vizsla, Hungarian pointer', 'Chesapeake Bay retriever', 'German short-haired pointer', 'Great Dane']}]表示预测的类别为Weimaraner(魏玛猎狗),ID是178,置信度为0.9213957190513611。
export CUDA_VISIBLE_DEVICES=0,1,2,3
python -m paddle.distributed.launch --gpus="0,1,2,3" \
main.py \
--model=CycleMLP_B1 \
--batch_size=256 \
--data_path=/path/to/imagenet/ \
--output_dir=./output/ \
--dist_eval部分训练日志如下所示。

[16:56:29.233819] Epoch: [261] [ 920/1251] eta: 0:05:50 lr: 0.000052 loss: 3.4592 (3.3812) time: 1.0303 data: 0.0012[16:56:49.578909] Epoch: [261] [ 940/1251] eta: 0:05:29 lr: 0.000052 loss: 3.7399 (3.3853) time: 1.0171 data: 0.0015
export CUDA_VISIBLE_DEVICES=0,1,2,3
python -m paddle.distributed.launch --gpus="0,1,2,3" \
eval.py \
--model=CycleMLP_B1 \
--batch_size=256 \
--data_path=/path/to/imagenet/ \
--dist_eval \
--resume=$TRAINED_MODELpython export_model.py \
--model=CycleMLP_B1 \
--output_dir=./output/ \
--resume=$TRAINED_MODEL├── cycle_mlp.py├── demo ├── engine.py├── eval.py├── export_model.py├── infer.py├── main.py├── README.md├── requirements.txt├── test_tipc └── util
详细日志在test_tipc/output
TIPC: test_tipc/README.md
首先安装auto_log,需要进行安装,安装方式如下: auto_log的详细介绍参考https://github.com/LDOUBLEV/AutoLog。
git clone https://github.com/LDOUBLEV/AutoLog cd AutoLog/ pip3 install -r requirements.txt python3 setup.py bdist_wheel pip3 install ./dist/auto_log-1.2.0-py3-none-any.whl
进行TIPC:
bash test_tipc/prepare.sh test_tipc/config/CycleMLP/CycleMLP_B1.txt 'lite_train_lite_infer'bash test_tipc/test_train_inference_python.sh test_tipc/config/CycleMLP/CycleMLP_B1.txt 'lite_train_lite_infer'
TIPC结果:
如果运行成功,在终端中会显示下面的内容,具体的日志也会输出到test_tipc/output/文件夹中的文件中。
Run successfully with command - python3.7 eval.py --model=CycleMLP_B1 --data_path=./dataset/ILSVRC2012/ --cls_label_path=./dataset/ILSVRC2012/val_list.txt --resume=./test_tipc/output/norm_train_gpus_0_autocast_null/CycleMLP_B1/checkpoint-latest.pd ! Run successfully with command - python3.7 export_model.py --model=CycleMLP_B1 --resume=./test_tipc/output/norm_train_gpus_0_autocast_null/CycleMLP_B1/checkpoint-latest.pd --output=./test_tipc/output/norm_train_gpus_0_autocast_null ! Run successfully with command - python3.7 inference.py --use_gpu=True --use_tensorrt=False --precision=fp32 --model_file=./test_tipc/output/norm_train_gpus_0_autocast_null/model.pdmodel --batch_size=2 --input_file=./dataset/ILSVRC2012/val --params_file=./test_tipc/output/norm_train_gpus_0_autocast_null/model.pdiparams > ./test_tipc/output/python_infer_gpu_usetrt_False_precision_fp32_batchsize_2.log 2>&1 ! ...
以上就是基于PaddlePaddle复现的CycleMLP的详细内容,更多请关注php中文网其它相关文章!
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。
Copyright 2014-2025 https://www.php.cn/ All Rights Reserved | php.cn | 湘ICP备2023035733号