Skip to content

P
PaddleX

PaddlePaddle / PaddleX

通知 138
Star 4
Fork 0
P
PaddleX
提交 4c631565 编写于 作者: J jack
浏览文件
操作

Merge branch 'develop' of github.com:joey12300/PaddleX into develop

上级 2bbff977 f5978528
显示空白变更内容
内联 并排
Showing with 1852 addition and 711 deletion
README.md
浏览文件 @ 4c631565
......@@ -6,6 +6,7 @@
[![Version](https://img.shields.io/github/release/PaddlePaddle/PaddleX.svg)](https://github.com/PaddlePaddle/PaddleX/releases)
![python version](https://img.shields.io/badge/python-3.6+-orange.svg)
![support os](https://img.shields.io/badge/os-linux%2C%20win%2C%20mac-yellow.svg)
![QQGroup](https://img.shields.io/badge/QQ_Group-1045148026-52B6EF?style=social&logo=tencent-qq&logoColor=000&logoWidth=20)
PaddleX是基于飞桨核心框架、开发套件和工具组件的深度学习全流程开发工具。具备**全流程打通**、**融合产业实践**、**易用易集成**三大特点。
......
docs/FAQ.md
浏览文件 @ 4c631565
......@@ -13,7 +13,7 @@
> 可以使用模型裁剪，参考文档[模型裁剪使用教程](slim/prune.md)，通过调整裁剪参数，可以控制模型裁剪后的大小，在实际实验中，如VOC检测数据，使用yolov3-mobilenet，原模型大小为XXM，裁剪后为XX M，精度基本保持不变
## 4. 如何配置训练时GPU的卡数
> 通过在终端export环境变量，或在Python代码中设置，可参考文档[CPU/多卡GPU训练](gpu_configure.md)
> 通过在终端export环境变量，或在Python代码中设置，可参考文档[CPU/多卡GPU训练](appendix/gpu_configure.md)
## 5. 想将之前训练的模型参数上继续训练
> 在训练调用`train`接口时，将`pretrain_weights`设为之前的模型保存路径即可
......@@ -52,7 +52,7 @@
> 1. 用户自行训练时，如不确定迭代的轮数，可以将轮数设高一些，同时注意设置`save_interval_epochs`，这样模型迭代每间隔相应轮数就会在验证集上进行评估和保存，可以根据不同轮数模型在验证集上的评估指标，判断模型是否已经收敛，若模型已收敛，可以自行结束训练进程
>
## 9. 只有CPU，没有GPU，如何提升训练速度
> 当没有GPU时，可以根据自己的CPU配置，选择是否使用多CPU进行训练，具体配置方式可以参考文档[多卡CPU/GPU训练](gpu_configure.md)
> 当没有GPU时，可以根据自己的CPU配置，选择是否使用多CPU进行训练，具体配置方式可以参考文档[多卡CPU/GPU训练](appendix/gpu_configure.md)
>
## 10. 电脑不能联网，训练时因为下载预训练模型失败，如何解决
> 可以预先通过其它方式准备好预训练模型，然后训练时自定义`pretrain_weights`即可，可参考文档[无联网模型训练](how_to_offline_run.md)
......@@ -61,7 +61,7 @@
> 1.可以按照9的方式来解决这个问题
> 2.每次训练前都设定`paddlex.pretrain_dir`路径，如设定`paddlex.pretrain_dir='/usrname/paddlex`，如此下载完的预训练模型会存放至`/usrname/paddlex`目录下，而已经下载在该目录的模型也不会再次重复下载
## 12. 程序启动时提示"Failed to execute script PaddleX"，如何解决？
## 12. PaddleX GUI启动时提示"Failed to execute script PaddleX"，如何解决？
> 1. 请检查目标机器上PaddleX程序所在路径是否包含中文。目前暂不支持中文路径，请尝试将程序移动到英文目录。
> 2. 如果您的系统是Windows 7或者Windows Server 2012时，原因是缺少MFPlat.DLL/MF.dll/MFReadWrite.dll等OpenCV依赖的DLL，请按如下方式安装桌面体验：通过“我的电脑”-->“属性”-->"管理"打开服务器管理器，点击右上角“管理”选择“添加角色和功能”。点击“服务器选择”-->“功能”，拖动滚动条到最下端，点开“用户界面和基础结构”，勾选“桌面体验”后点击“安装”，等安装完成尝试再次运行PaddleX。
> 3. 请检查目标机器上是否有其他的PaddleX程序或者进程在运行中，如有请退出或者重启机器看是否解决
......
docs/apis/models/classification.md
浏览文件 @ 4c631565
......@@ -80,7 +80,7 @@ predict(self, img_file, transforms=None, topk=5)
## 其它分类器类
PaddleX提供了共计22种分类器，所有分类器均提供同`ResNet50`相同的训练`train`，评估`evaluate`和预测`predict`接口，各模型效果可参考[模型库](../appendix/model_zoo.md)。
PaddleX提供了共计22种分类器，所有分类器均提供同`ResNet50`相同的训练`train`，评估`evaluate`和预测`predict`接口，各模型效果可参考[模型库](https://paddlex.readthedocs.io/zh_CN/latest/appendix/model_zoo.html)。
### ResNet18
```python
......
docs/apis/models/semantic_segmentation.md
浏览文件 @ 4c631565
......@@ -186,10 +186,10 @@ paddlex.seg.HRNet(num_classes=2, width=18, use_bce_loss=False, use_dice_loss=Fal
> **参数**
> > - **num_classes** (int): 类别数。
> > - **width** (int): 高分辨率分支中特征层的通道数量。默认值为18。可选择取值为[18, 30, 32, 40, 44, 48, 60, 64]。
> > - **width** (int|str): 高分辨率分支中特征层的通道数量。默认值为18。可选择取值为[18, 30, 32, 40, 44, 48, 60, 64, '18_small_v1']。'18_small_v1'是18的轻量级版本。
> > - **use_bce_loss** (bool): 是否使用bce loss作为网络的损失函数，只能用于两类分割。可与dice loss同时使用。默认False。
> > - **use_dice_loss** (bool): 是否使用dice loss作为网络的损失函数，只能用于两类分割，可与bce loss同时使用。当use_bce_loss和use_dice_loss都为False时，使用交叉熵损失函数。默认False。
> > - **class_weight** (list/str): 交叉熵损失函数各类损失的权重。当`class_weight`为list的时候，长度应为`num_classes`。当`class_weight`为str时， weight.lower()应为'dynamic'，这时会根据每一轮各类像素的比重自行计算相应的权重，每一类的权重为：每类的比例 * num_classes。class_weight取默认值None是，各类的权重1，即平时使用的交叉熵损失函数。
> > - **class_weight** (list|str): 交叉熵损失函数各类损失的权重。当`class_weight`为list的时候，长度应为`num_classes`。当`class_weight`为str时， weight.lower()应为'dynamic'，这时会根据每一轮各类像素的比重自行计算相应的权重，每一类的权重为：每类的比例 * num_classes。class_weight取默认值None是，各类的权重1，即平时使用的交叉熵损失函数。
> > - **ignore_index** (int): label上忽略的值，label为`ignore_index`的像素不参与损失函数的计算。默认255。
### train 训练接口
......
docs/apis/transforms/seg_transforms.md
浏览文件 @ 4c631565
......@@ -200,7 +200,7 @@ ComposedSegTransforms.add_augmenters(augmenters)
import paddlex as pdx
from paddlex.seg import transforms
train_transforms = transforms.ComposedSegTransforms(mode='train', train_crop_size=[512, 512])
eval_transforms = transforms.ComposedYOLOTransforms(mode='eval')
eval_transforms = transforms.ComposedSegTransforms(mode='eval')
# 添加数据增强
import imgaug.augmenters as iaa
......
docs/apis/visualize.md
浏览文件 @ 4c631565
......@@ -146,10 +146,11 @@ paddlex.interpret.normlime(img_file,
dataset=None,
num_samples=3000,
batch_size=50,
save_dir='./')
save_dir='./',
normlime_weights_file=None)
```
使用NormLIME算法将模型预测结果的可解释性可视化。
NormLIME是利用一定数量的样本来出一个全局的解释。NormLIME会提前计算一定数量的测试样本的LIME结果，然后对相同的特征进行权重的归一化，这样来得到一个全局的输入和输出的关系。
NormLIME是利用一定数量的样本来出一个全局的解释。由于NormLIME计算量较大，此处采用一种简化的方式：使用一定数量的测试样本（目前默认使用所有测试样本），对每个样本进行特征提取，映射到同一个特征空间；然后以此特征做为输入，以模型输出做为输出，使用线性回归对其进行拟合，得到一个全局的输入和输出的关系。之后，对一测试样本进行解释时，使用NormLIME全局的解释，来对LIME的结果进行滤波，使最终的可视化结果更加稳定。
**注意：** 可解释性结果可视化目前只支持分类模型。
......@@ -160,8 +161,9 @@ NormLIME是利用一定数量的样本来出一个全局的解释。NormLIME会
>* **num_samples** (int): LIME用于学习线性模型的采样数，默认为3000。
>* **batch_size** (int): 预测数据batch大小，默认为50。
>* **save_dir** (str): 可解释性可视化结果（保存为png格式文件）和中间文件存储路径。
>* **normlime_weights_file** (str): NormLIME初始化文件名，若不存在，则计算一次，保存于该路径；若存在，则直接载入。
**注意：** dataset`读取的是一个数据集，该数据集不宜过大，否则计算时间会较长，但应包含所有类别的数据。
**注意：** dataset`读取的是一个数据集，该数据集不宜过大，否则计算时间会较长，但应包含所有类别的数据。NormLIME可解释性结果可视化目前只支持分类模型。
### 使用示例
> 对预测可解释性结果可视化的过程可参见[代码](https://github.com/PaddlePaddle/PaddleX/blob/develop/tutorials/interpret/normlime.py)。
docs/appendix/index.rst
浏览文件 @ 4c631565
......@@ -7,6 +7,7 @@
:caption: 目录:
model_zoo.md
slim_model_zoo.md
metrics.md
interpret.md
parameters.md
......
docs/appendix/interpret.md
浏览文件 @ 4c631565
......@@ -20,9 +20,20 @@ LIME的使用方式可参见[代码示例](https://github.com/PaddlePaddle/Paddl
## NormLIME
NormLIME是在LIME上的改进，LIME的解释是局部性的，是针对当前样本给的特定解释，而NormLIME是利用一定数量的样本对当前样本的一个全局性的解释，有一定的降噪效果。其实现步骤如下所示：
1. 下载Kmeans模型参数和ResNet50_vc网络前三层参数。（ResNet50_vc的参数是在ImageNet上训练所得网络的参数；使用ImageNet图像作为数据集，每张图像从ResNet50_vc的第三层输出提取对应超象素位置上的平均特征和质心上的特征，训练将得到此处的Kmeans模型）
2. 计算测试集中每张图像的LIME结果。（如无测试集，可用验证集代替）
3. 使用Kmeans模型对所有图像中的所有像素进行聚类。
4. 对在同一个簇的超像素（相同的特征）进行权重的归一化，得到每个超像素的权重，以此来解释模型。
2. 使用测试集中的数据计算normlime的权重信息（如无测试集，可用验证集代替）:
对每张图像的处理：
(1) 获取图像的超像素。
(2) 使用ResNet50_vc获取第三层特征，针对每个超像素位置，组合质心特征和均值特征`F`。
(3) 把`F`作为Kmeans模型的输入，计算每个超像素位置的聚类中心。
(4) 使用训练好的分类模型，预测该张图像的`label`。
对所有图像的处理：
(1) 以每张图像的聚类中心信息组成的向量（若某聚类中心出现在盖章途中设置为1，反之为0）为输入，
预测的`label`为输出，构建逻辑回归函数`regression_func`。
(2) 由`regression_func`可获得每个聚类中心不同类别下的权重，并对权重进行归一化。
3. 使用Kmeans模型获取需要可视化图像的每个超像素的聚类中心。
4. 对需要可视化的图像的超像素进行随机遮掩构成新的图像。
5. 对每张构造的图像使用预测模型预测label。
6. 根据normlime的权重信息，每个超像素可获不同的权重，选取最高的权重为最终的权重，以此来解释模型。
NormLIME的使用方式可参见[代码示例](https://github.com/PaddlePaddle/PaddleX/blob/develop/tutorials/interpret/normlime.py)和[api介绍](../apis/visualize.html#normlime)。在使用时，参数中的`num_samples`设置尤为重要，其表示上述步骤2中的随机采样的个数，若设置过小会影响可解释性结果的稳定性，若设置过大则将在上述步骤3耗费较长时间；参数`batch_size`则表示在计算上述步骤3时，预测的batch size，若设置过小将在上述步骤3耗费较长时间，而上限则根据机器配置决定；而`dataset`则是由测试集或验证集构造的数据。
......
docs/appendix/model_zoo.md
浏览文件 @ 4c631565
......@@ -40,8 +40,8 @@
|[FasterRCNN-ResNet101](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_r101_1x.tar)| 212.5MB | 582.911 | 38.3 |
|[FasterRCNN-ResNet50-FPN](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_r50_fpn_1x.tar)| 167.7MB | 83.189 | 37.2 |
|[FasterRCNN-ResNet50_vd-FPN](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_r50_vd_fpn_2x.tar)|167.8MB | 128.277 | 38.9 |
|[FasterRCNN-ResNet101-FPN](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_r101_fpn_1x.tar)| 244.2MB | 156.097 | 38.7 |
|[FasterRCNN-ResNet101_vd-FPN](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_r101_vd_fpn_2x.tar) |244.3MB | 119.788 | 40.5 |
|[FasterRCNN-ResNet101-FPN](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_r101_fpn_1x.tar)| 244.2MB | 119.788 | 38.7 |
|[FasterRCNN-ResNet101_vd-FPN](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_r101_vd_fpn_2x.tar) |244.3MB | 156.097 | 40.5 |
|[FasterRCNN-HRNet_W18-FPN](https://paddlemodels.bj.bcebos.com/object_detection/faster_rcnn_hrnetv2p_w18_1x.tar) |115.5MB | 81.592 | 36 |
|[YOLOv3-DarkNet53](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_darknet.tar)|249.2MB | 42.672 | 38.9 |
|[YOLOv3-MobileNetV1](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) |99.2MB | 15.442 | 29.3 |
......
docs/appendix/slim_model_zoo.md 0 → 100644
浏览文件 @ 4c631565
# PaddleX压缩模型库
## 图像分类
数据集：ImageNet-1000
### 量化
| 模型 | 压缩策略 | Top-1准确率 | 存储体积 | TensorRT时延(V100, ms) |
|:--:|:---:|:--:|:--:|:--:|
|MobileNetV1| 无 |70.99%| 17MB | -|
|MobileNetV1| 量化 |70.18% (-0.81%)| 4.4MB | - |
| MobileNetV2 | 无 |72.15%| 15MB | - |
| MobileNetV2 | 量化 | 71.15% (-1%)| 4.0MB | - |
|ResNet50| 无 |76.50%| 99MB | 2.71 |
|ResNet50| 量化 |76.33% (-0.17%)| 25.1MB | 1.19 |
分类模型Lite时延(ms)
| 设备 | 模型类型 | 压缩策略 | armv7 Thread 1 | armv7 Thread 2 | armv7 Thread 4 | armv8 Thread 1 | armv8 Thread 2 | armv8 Thread 4 |
| ------- | ----------- | ------------- | -------------- | -------------- | -------------- | -------------- | -------------- | -------------- |
| 高通835 | MobileNetV1 | 无 | 96.1942 | 53.2058 | 32.4468 | 88.4955 | 47.95 | 27.5189 |
| 高通835 | MobileNetV1 | 量化 | 60.5615 | 32.4016 | 16.6596 | 56.5266 | 29.7178 | 15.1459 |
| 高通835 | MobileNetV2 | 无 | 65.715 | 38.1346 | 25.155 | 61.3593 | 36.2038 | 22.849 |
| 高通835 | MobileNetV2 | 量化 | 48.3495 | 30.3069 | 22.1506 | 45.8715 | 27.4105 | 18.2223 |
| 高通835 | ResNet50 | 无 | 526.811 | 319.6486 | 205.8345 | 506.1138 | 335.1584 | 214.8936 |
| 高通835 | ResNet50 | 量化 | 476.0507 | 256.5963 | 139.7266 | 461.9176 | 248.3795 | 149.353 |
| 高通855 | MobileNetV1 | 无 | 33.5086 | 19.5773 | 11.7534 | 31.3474 | 18.5382 | 10.0811 |
| 高通855 | MobileNetV1 | 量化 | 37.0498 | 21.7081 | 11.0779 | 14.0947 | 8.1926 | 4.2934 |
| 高通855 | MobileNetV2 | 无 | 25.0396 | 15.2862 | 9.6609 | 22.909 | 14.1797 | 8.8325 |
| 高通855 | MobileNetV2 | 量化 | 28.1631 | 18.3917 | 11.8333 | 16.9399 | 11.1772 | 7.4176 |
| 高通855 | ResNet50 | 无 | 185.3705 | 113.0825 | 87.0741 | 177.7367 | 110.0433 | 74.4114 |
| 高通855 | ResNet50 | 量化 | 328.2683 | 201.9937 | 106.744 | 242.6397 | 150.0338 | 79.8659 |
| 麒麟970 | MobileNetV1 | 无 | 101.2455 | 56.4053 | 35.6484 | 94.8985 | 51.7251 | 31.9511 |
| 麒麟970 | MobileNetV1 | 量化 | 62.4412 | 32.2585 | 16.6215 | 57.825 | 29.2573 | 15.1206 |
| 麒麟970 | MobileNetV2 | 无 | 70.4176 | 42.0795 | 25.1939 | 68.9597 | 39.2145 | 22.6617 |
| 麒麟970 | MobileNetV2 | 量化 | 53.0961 | 31.7987 | 21.8334 | 49.383 | 28.2358 | 18.3642 |
| 麒麟970 | ResNet50 | 无 | 586.8943 | 344.0858 | 228.2293 | 573.3344 | 351.4332 | 225.8006 |
| 麒麟970 | ResNet50 | 量化 | 489.6188 | 258.3279 | 142.6063 | 480.0064 | 249.5339 | 138.5284 |
### 剪裁
PaddleLite推理耗时说明：
环境：Qualcomm SnapDragon 845 + armv8
速度指标：Thread1/Thread2/Thread4耗时
| 模型 | 压缩策略 | Top-1 | 存储体积 |PaddleLite推理耗时|TensorRT推理速度(FPS)|
|:--:|:---:|:--:|:--:|:--:|:--:|
| MobileNetV1 | 无 | 70.99% | 17MB | 66.052\35.8014\19.5762|-|
| MobileNetV1 | 剪裁 -30% | 70.4% (-0.59%) | 12MB | 46.5958\25.3098\13.6982|-|
| MobileNetV1 | 剪裁 -50% | 69.8% (-1.19%) | 9MB | 37.9892\20.7882\11.3144|-|
## 目标检测
### 量化
数据集： COCO2017
| 模型 | 压缩策略 | 数据集 | Image/GPU | 输入608 Box AP | 存储体积 | TensorRT时延(V100, ms) |
| :----------------------------: | :---------: | :----: | :-------: | :------------: | :------------: | :----------: |
| MobileNet-V1-YOLOv3 | 无 | COCO | 8 | 29.3 | 95MB | - |
| MobileNet-V1-YOLOv3 | 量化 | COCO | 8 | 27.9 (-1.4)| 25MB | - |
| R34-YOLOv3 | 无 | COCO | 8 | 36.2 | 162MB | - |
| R34-YOLOv3 | 量化 | COCO | 8 | 35.7 (-0.5) | 42.7MB | - |
### 剪裁
数据集：Pasacl VOC & COCO2017
PaddleLite推理耗时说明：
环境：Qualcomm SnapDragon 845 + armv8
速度指标：Thread1/Thread2/Thread4耗时
| 模型 | 压缩策略 | 数据集 | Image/GPU | 输入608 Box mmAP | 存储体积 | PaddleLite推理耗时(ms)(608*608) | TensorRT推理速度(FPS)(608*608) |
| :----------------------------: | :---------------: | :--------: | :-------: | :------------: | :----------: | :--------------: | :--------------: |
| MobileNet-V1-YOLOv3 | 无 | Pascal VOC | 8 | 76.2 | 94MB | 1238\796.943\520.101|60.04|
| MobileNet-V1-YOLOv3 | 剪裁 -52.88% | Pascal VOC | 8 | 77.6 (+1.4) | 31MB | 602.497\353.759\222.427 |99.36|
| MobileNet-V1-YOLOv3 | 无 | COCO | 8 | 29.3 | 95MB |-|-|
| MobileNet-V1-YOLOv3 | 剪裁 -51.77% | COCO | 8 | 26.0 (-3.3) | 32MB |-|73.93|
## 语义分割
数据集：Cityscapes
### 量化
| 模型 | 压缩策略 | mIoU | 存储体积 |
| :--------------------: | :---------: | :-----------: | :------------: |
| DeepLabv3-MobileNetv2 | 无 | 69.81 | 7.4MB |
| DeepLabv3-MobileNetv2 | 量化 | 67.59 (-2.22) | 2.1MB |
图像分割模型Lite时延(ms), 输入尺寸769 x 769
| 设备 | 模型类型 | 压缩策略 | armv7 Thread 1 | armv7 Thread 2 | armv7 Thread 4 | armv8 Thread 1 | armv8 Thread 2 | armv8 Thread 4 |
| ------- | ---------------------- | ------------- | -------------- | -------------- | -------------- | -------------- | -------------- | -------------- |
| 高通835 | Deeplabv3-MobileNetV2 | 无 | 1282.8126 | 793.2064 | 653.6538 | 1193.9908 | 737.1827 | 593.4522 |
| 高通835 | Deeplabv3-MobileNetV2 | 量化 | 981.44 | 658.4969 | 538.6166 | 885.3273 | 586.1284 | 484.0018 |
| 高通855 | Deeplabv3-MobileNetV2 | 无 | 639.4425 | 390.1851 | 322.7014 | 477.7667 | 339.7411 | 262.2847 |
| 高通855 | Deeplabv3-MobileNetV2 | 量化 | 705.7589 | 474.4076 | 427.2951 | 394.8352 | 297.4035 | 264.6724 |
| 麒麟970 | Deeplabv3-MobileNetV2 | 无 | 1771.1301 | 1746.0569 | 1222.4805 | 1448.9739 | 1192.4491 | 760.606 |
| 麒麟970 | Deeplabv3-MobileNetV2 | 量化 | 1320.386 | 918.5328 | 672.2481 | 1020.753 | 820.094 | 591.4114 |
### 剪裁
PaddleLite推理耗时说明：
环境：Qualcomm SnapDragon 845 + armv8
速度指标：Thread1/Thread2/Thread4耗时
| 模型 | 压缩方法 | mIoU | 存储体积 | PaddleLite推理耗时 | TensorRT推理速度(FPS) |
| :-------: | :---------------: | :-----------: | :------: | :------------: | :----: |
| FastSCNN | 无 | 69.64 | 11MB | 1226.36\682.96\415.664 |39.53|
| FastSCNN | 剪裁 -47.60% | 66.68 (-2.96) | 5.7MB | 866.693\494.467\291.748 |51.48|
docs/cv_solutions.md
浏览文件 @ 4c631565
# PaddleX视觉方案介绍
PaddleX目前提供了4种视觉任务解决方案，分别为图像分类、目标检测、实例分割和语义分割。用户可以根据自己的任务类型按需选取。
PaddleX针对图像分类、目标检测、实例分割和语义分割4种视觉任务提供了包含模型选择、压缩策略选择、部署方案选择在内的解决方案。用户根据自己的需求选择合适的模型，选择合适的压缩策略来减小模型的计算量和存储体积、加速模型预测推理，最后选择合适的部署方案将模型部署在移动端或者服务器端。
## 图像分类
## 模型选择
### 图像分类
图像分类任务指的是输入一张图片，模型预测图片的类别，如识别为风景、动物、车等。
![](./images/image_classification.png)
对于图像分类任务，针对不同的应用场景，PaddleX提供了百度改进的模型，见下表所示
对于图像分类任务，针对不同的应用场景，PaddleX提供了百度改进的模型，见下表所示:
> 表中GPU预测速度是使用PaddlePaddle Python预测接口测试得到（测试GPU型号为Nvidia Tesla P40）。
> 表中CPU预测速度 (测试CPU型号为)。
> 表中骁龙855预测速度是使用处理器为骁龙855的手机测试得到。
> 测速时模型输入大小为224 x 224，Top1准确率为ImageNet-1000数据集上评估所得。
| 模型 | 模型大小 | GPU预测速度 | CPU预测速度 | ARM芯片预测速度 | 准确率 | 备注 |
| :--------- | :------ | :---------- | :-----------| :------------- | :----- | :--- |
| MobileNetV3_small_ssld | 12M | - | - | - | 71.3% |适用于移动端场景 |
| MobileNetV3_large_ssld | 21M | - | - | - | 79.0% | 适用于移动端/服务端场景 |
| ResNet50_vd_ssld | 102.8MB | - | - | - | 82.4% | 适用于服务端场景 |
| ResNet101_vd_ssld | 179.2MB | - | - | - |83.7% | 适用于服务端场景 |
| 模型 | 模型特点 | 存储体积 | GPU预测速度（毫秒） | CPU(x86)预测速度（毫秒） | 骁龙855(ARM)预测速度 （毫秒）| Top1准确率 |
| :--------- | :------ | :---------- | :-----------| :------------- | :------------- |:--- |
| MobileNetV3_small_ssld | 轻量高速，适用于追求高速的实时移动端场景 | 12.5MB | 7.08837 | - | 6.546 | 71.3.0% |
| ShuffleNetV2 | 轻量级模型，精度相对偏低，适用于要求更小存储体积的实时移动端场景 | 10.2MB | 15.40 | - | 10.941 | 68.8% |
| MobileNetV3_large_ssld | 轻量级模型，在存储方面优势不大，在速度和精度上表现适中，适合于移动端场景 | 22.8MB | 8.06651 | - | 19.803 | 79.0% |
| MobileNetV2 | 轻量级模型，适用于使用GPU预测的移动端场景 | 15.0MB | 5.92667 | - | 23.318| 72.2 % |
| ResNet50_vd_ssld | 高精度模型，预测时间较短，适用于大多数的服务器端场景 | 103.5MB | 7.79264 | - | - | 82.4% |
| ResNet101_vd_ssld | 超高精度模型，预测时间相对较长，适用于有大数据量时的服务器端场景 | 180.5MB | 13.34580 | - | -| 83.7% |
| Xception65 | 超高精度模型，预测时间更长，在处理较大数据量时有较高的精度，适用于服务器端场景 | 161.6MB | 13.87017 | - | - | 80.3% |
除上述模型外，PaddleX还支持近20种图像分类模型，模型列表可参考[PaddleX模型库](../appendix/model_zoo.md)
包括上述模型，PaddleX支持近20种图像分类模型，其余模型可参考[PaddleX模型库](../appendix/model_zoo.md)
## 目标检测
### 目标检测
目标检测任务指的是输入图像，模型识别出图像中物体的位置（用矩形框框出来，并给出框的位置），和物体的类别，如在手机等零件质检中，用于检测外观上的瑕疵等。
![](./images/object_detection.png)
对于目标检测，针对不同的应用场景，PaddleX提供了主流的YOLOv3模型和Faster-RCNN模型，见下表所示
| 模型 | 模型大小 | GPU预测速度 | CPU预测速度 |ARM芯片预测速度 | BoxMAP | 备注 |
| :------- | :------- | :--------- | :---------- | :------------- | :----- | :--- |
| YOLOv3-MobileNetV1 | 101.2M | - | - | - | 29.3 | |
| YOLOv3-MobileNetV3 | 94.6M | - | - | - | 31.6 | |
| YOLOv3-ResNet34 | 169.7M | - | - | - | 36.2 | |
| YOLOv3-DarkNet53 | 252.4 | - | - | - | 38.9 | |
除YOLOv3模型外，PaddleX同时也支持FasterRCNN模型，支持FPN结构和5种backbone网络，详情可参考[PaddleX模型库](../appendix/model_zoo.md)
## 实例分割
> 表中GPU预测速度是使用PaddlePaddle Python预测接口测试得到（测试GPU型号为Nvidia Tesla P40）。
> 表中CPU预测速度 (测试CPU型号为)。
> 表中骁龙855预测速度是使用处理器为骁龙855的手机测试得到。
> 测速时YOLOv3的输入大小为608 x 608，FasterRCNN的输入大小为800 x 1333，Box mmAP为COCO2017数据集上评估所得。
| 模型 | 模型特点 | 存储体积 | GPU预测速度 | CPU(x86)预测速度（毫秒） | 骁龙855(ARM)预测速度 （毫秒）| Box mmAP |
| :------- | :------- | :--------- | :---------- | :------------- | :------------- |:--- |
| YOLOv3-MobileNetV3_larget | 适用于追求高速预测的移动端场景 | 100.7MB | 143.322 | - | - | 31.6 |
| YOLOv3-MobileNetV1 | 精度相对偏低，适用于追求高速预测的服务器端场景 | 99.2MB| 15.422 | - | - | 29.3 |
| YOLOv3-DarkNet53 | 在预测速度和模型精度上都有较好的表现，适用于大多数的服务器端场景| 249.2MB | 42.672 | - | - | 38.9 |
| FasterRCNN-ResNet50-FPN | 经典的二阶段检测器，预测速度相对较慢，适用于重视模型精度的服务器端场景 | 167.MB | 83.189 | - | -| 37.2 |
| FasterRCNN-HRNet_W18-FPN | 适用于对图像分辨率较为敏感、对目标细节预测要求更高的服务器端场景 | 115.5MB | 81.592 | - | - | 36 |
| FasterRCNN-ResNet101_vd-FPN | 超高精度模型，预测时间更长，在处理较大数据量时有较高的精度，适用于服务器端场景 | 244.3MB | 156.097 | - | - | 40.5 |
除上述模型外，YOLOv3和Faster RCNN还支持其他backbone，详情可参考[PaddleX模型库](../appendix/model_zoo.md)
### 实例分割
在目标检测中，模型识别出图像中物体的位置和物体的类别。而实例分割则是在目标检测的基础上，做了像素级的分类，将框内的属于目标物体的像素识别出来。
![](./images/instance_segmentation.png)
PaddleX目前提供了实例分割MaskRCNN模型，支持5种不同的backbone网络，详情可参考[PaddleX模型库](../appendix/model_zoo.md)
| 模型 | 模型大小 | GPU预测速度 | CPU预测速度 | ARM芯片预测速度 | BoxMAP | SegMAP | 备注 |
| :---- | :------- | :---------- | :---------- | :------------- | :----- | :----- | :--- |
| MaskRCNN-ResNet50_vd-FPN | 185.5M | - | - | - | 39.8 | 35.4 | |
| MaskRCNN-ResNet101_vd-FPN | 268.6M | - | - | - | 41.4 | 36.8 | |
## 语义分割
> 表中GPU预测速度是使用PaddlePaddle Python预测接口测试得到（测试GPU型号为Nvidia Tesla P40）。
> 表中CPU预测速度 (测试CPU型号为)。
> 表中骁龙855预测速度是使用处理器为骁龙855的手机测试得到。
> 测速时MaskRCNN的输入大小为800 x 1333，Box mmAP和Seg mmAP为COCO2017数据集上评估所得。
| 模型 | 模型特点 | 存储体积 | GPU预测速度 | CPU(x86)预测速度（毫秒） | 骁龙855(ARM)预测速度 （毫秒）| Box mmAP | Seg mmAP |
| :---- | :------- | :---------- | :---------- | :----- | :----- | :--- |:--- |
| MaskRCNN-HRNet_W18-FPN | 适用于对图像分辨率较为敏感、对目标细节预测要求更高的服务器端场景 | - | - | - | - | 37.0 | 33.4 |
| MaskRCNN-ResNet50-FPN | 精度较高，适合大多数的服务器端场景| 185.5M | - | - | - | 37.9 | 34.2 |
| MaskRCNN-ResNet101_vd-FPN | 高精度但预测时间更长，在处理较大数据量时有较高的精度，适用于服务器端场景 | 268.6M | - | - | - | 41.4 | 36.8 |
### 语义分割
语义分割用于对图像做像素级的分类，应用在人像分类、遥感图像识别等场景。
![](./images/semantic_segmentation.png)
对于语义分割，PaddleX也针对不同的应用场景，提供了不同的模型选择，如下表所示
> 表中GPU预测速度是使用PaddlePaddle Python预测接口测试得到（测试GPU型号为Nvidia Tesla P40）。
> 表中CPU预测速度 (测试CPU型号为)。
> 表中骁龙855预测速度是使用处理器为骁龙855的手机测试得到。
> 测速时模型的输入大小为1024 x 2048，mIOU为Cityscapes数据集上评估所得。
| 模型 | 模型特点 | 存储体积 | GPU预测速度 | CPU(x86)预测速度（毫秒） | 骁龙855(ARM)预测速度 （毫秒）| mIOU |
| :---- | :------- | :---------- | :---------- | :----- | :----- |:--- |
| DeepLabv3p-MobileNetV2_x1.0 | 轻量级模型，适用于移动端场景| - | - | - | 69.8% |
| HRNet_W18_Small_v1 | 轻量高速，适用于移动端场景 | - | - | - | - |
| FastSCNN | 轻量高速，适用于追求高速预测的移动端或服务器端场景 | - | - | - | 69.64 |
| HRNet_W18 | 高精度模型，适用于对图像分辨率较为敏感、对目标细节预测要求更高的服务器端场景| - | - | - | 79.36 |
| DeepLabv3p-Xception65 | 高精度但预测时间更长，在处理较大数据量时有较高的精度，适用于服务器且背景复杂的场景| - | - | - | 79.3% |
## 压缩策略选择
PaddleX提供包含模型剪裁、定点量化的模型压缩策略来减小模型的计算量和存储体积，加快模型部署后的预测速度。使用不同压缩策略在图像分类、目标检测和语义分割模型上的模型精度和预测速度详见以下内容，用户可以选择根据自己的需求选择合适的压缩策略，进一步优化模型的性能。
| 压缩策略 | 策略特点 |
| :---- | :------- |
| 量化 | 较为显著地减少模型的存储体积，适用于移动端或服务期端TensorRT部署，在移动端对于MobileNet系列模型有明显的加速效果 |
| 剪裁 | 能够去除冗余的参数，达到显著减少参数计算量和模型体积的效果，提升模型的预测性能，适用于CPU部署或移动端部署（GPU上无明显加速效果） |
| 先剪裁后量化 | 可以进一步提升模型的预测性能，适用于移动端或服务器端TensorRT部署 |
### 性能对比
* 表中各指标的格式为XXX/YYY，XXX表示未采取压缩策略时的指标，YYY表示压缩后的指标
* 分类模型的准确率指的是ImageNet-1000数据集上的Top1准确率(模型输入大小为224x224)，检测模型的准确率指的是COCO2017数据集上的mmAP(模型输入大小为608x608)，分割模型的准确率指的是Cityscapes数据集上mIOU(模型输入大小为769x769)
* 量化策略中，PaddleLiter推理环境为Qualcomm SnapDragon 855 + armv8，速度指标为Thread4耗时
* 剪裁策略中，PaddleLiter推理环境为Qualcomm SnapDragon 845 + armv8，速度指标为Thread4耗时
| 模型 | 压缩策略 | 存储体积(MB) | 准确率(%) | PaddleLite推理耗时(ms) |
| :--: | :------: | :------: | :----: | :----------------: |
| MobileNetV1 | 量化 | 17/4.4 | 70.99/70.18 | 10.0811/4.2934 |
| MobileNetV1 | 剪裁 -30% | 17/12 | 70.99/70.4 | 19.5762/13.6982 |
| YOLOv3-MobileNetV1 | 量化 | 95/25 | 29.3/27.9 | - |
| YOLOv3-MobileNetV1 | 剪裁 -51.77% | 95/25 | 29.3/26 | - |
| Deeplabv3-MobileNetV2 | 量化 | 7.4/1.8 | 63.26/62.03 | 593.4522/484.0018 |
| FastSCNN | 剪裁 -47.60% | 11/5.7 | 69.64/66.68 | 415.664/291.748 |
更多模型在不同设备上压缩前后的指标对比详见[PaddleX压缩模型库](appendix/slim_model_zoo.md)
压缩策略的具体使用流程详见[模型压缩](tutorials/compress)
**注意：PaddleX中全部图像分类模型和语义分割模型都支持量化和剪裁操作，目标检测仅有YOLOv3支持量化和剪裁操作。**
## 模型部署
PaddleX提供服务器端python部署、服务器端c++部署、服务器端加密部署、OpenVINO部署、移动端部署共5种部署方案，用户可以根据自己的需求选择合适的部署方案，点击以下链接了解部署的具体流程。
| 模型 | 模型大小 | GPU预测速度 | CPU预测速度 | ARM芯片预测速度 | mIOU | 备注 |
| :---- | :------- | :---------- | :---------- | :------------- | :----- | :----- |
| DeepLabv3p-MobileNetV2_x0.25 | | - | - | - | - | - |
| DeepLabv3p-MobileNetV2_x1.0 | | - | - | - | - | - |
| DeepLabv3p-Xception65 | | - | - | - | - | - |
| UNet | | - | - | - | - | - |
| 部署方案 | 部署流程 |
| :------: | :------: |
| 服务器端python部署 | [部署流程](tutorials/deploy/deploy_server/deploy_python.html)|
| 服务器端c++部署 | [部署流程](tutorials/deploy/deploy_server/deploy_cpp/) |
| 服务器端加密部署 | [部署流程](tutorials/deploy/deploy_server/encryption.html) |
| OpenVINO部署 | [部署流程](tutorials/deploy/deploy_openvino.html) |
| 移动端部署 | [部署流程](tutorials/deploy/deploy_lite.html) |
docs/images/lime.png
查看替换文件 @ 2bbff977 浏览文件 @ 4c631565
docs/images/lime.png

802.8 KB | W: | H:

docs/images/lime.png

423.0 KB | W: | H:

docs/images/lime.png
docs/images/lime.png
docs/images/lime.png
docs/images/lime.png
  • 2-up
  • Swipe
  • Onion skin
docs/images/normlime.png
查看替换文件 @ 2bbff977 浏览文件 @ 4c631565
docs/images/normlime.png

1.5 MB | W: | H:

docs/images/normlime.png

546.2 KB | W: | H:

docs/images/normlime.png
docs/images/normlime.png
docs/images/normlime.png
docs/images/normlime.png
  • 2-up
  • Swipe
  • Onion skin
docs/tutorials/deploy/deploy_lite.md
浏览文件 @ 4c631565
......@@ -21,7 +21,7 @@ step 2: 将PaddleX模型导出为inference模型
step 3: 将inference模型转换成PaddleLite模型
```
python /path/to/PaddleX/deploy/lite/export_lite.py --model_dir /path/to/inference_model --save_file /path/to/onnx_model --place place/to/run
python /path/to/PaddleX/deploy/lite/export_lite.py --model_dir /path/to/inference_model --save_file /path/to/lite_model --place place/to/run
```
......
docs/tutorials/deploy/deploy_server/deploy_cpp/deploy_cpp_linux.md
浏览文件 @ 4c631565
......@@ -30,7 +30,7 @@ PaddlePaddle C++ 预测库针对不同的`CPU`，`CUDA`，以及是否支持Tens
| ubuntu14.04_cuda10.0_cudnn7_avx_mkl | [fluid_inference.tgz](https://paddle-inference-lib.bj.bcebos.com/1.8.2-gpu-cuda10-cudnn7-avx-mkl/fluid_inference.tgz ) |
| ubuntu14.04_cuda10.1_cudnn7.6_avx_mkl_trt6 | [fluid_inference.tgz](https://paddle-inference-lib.bj.bcebos.com/1.8.2-gpu-cuda10.1-cudnn7.6-avx-mkl-trt6%2Ffluid_inference.tgz) |
更多和更新的版本，请根据实际情况下载: [C++预测库下载列表](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/advanced_guide/inference_deployment/inference/windows_cpp_inference.html#id1)
更多和更新的版本，请根据实际情况下载: [C++预测库下载列表](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html)
下载并解压后`/root/projects/fluid_inference`目录包含内容为：
```
......@@ -42,7 +42,7 @@ fluid_inference
└── version.txt # 版本和编译信息
```
**注意:** 预编译版本除`nv-jetson-cuda10-cudnn7.5-trt5` 以外其它包都是基于`GCC 4.8.5`编译，使用高版本`GCC`可能存在 `ABI`兼容性问题，建议降级或[自行编译预测库](https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html#id12)。
**注意:** 预编译版本除`nv-jetson-cuda10-cudnn7.5-trt5` 以外其它包都是基于`GCC 4.8.5`编译，使用高版本`GCC`可能存在 `ABI`兼容性问题，建议降级或[自行编译预测库](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html#id12)。
### Step4: 编译
......
examples/human_segmentation/README.md 0 → 100644
浏览文件 @ 4c631565
# HumanSeg人像分割模型
本教程基于PaddleX核心分割网络，提供针对人像分割场景从预训练模型、Fine-tune、视频分割预测部署的全流程应用指南。
## 安装
**前置依赖**
* paddlepaddle >= 1.8.0
* python >= 3.5
```
pip install paddlex -i https://mirror.baidu.com/pypi/simple
```
安装的相关问题参考[PaddleX安装](https://paddlex.readthedocs.io/zh_CN/latest/install.html)
## 预训练模型
HumanSeg开放了在大规模人像数据上训练的两个预训练模型，满足多种使用场景的需求
| 模型类型 | Checkpoint Parameter | Inference Model | Quant Inference Model | 备注 |
| --- | --- | --- | ---| --- |
| HumanSeg-server | [humanseg_server_params](https://paddlex.bj.bcebos.com/humanseg/models/humanseg_server.pdparams) | [humanseg_server_inference](https://paddlex.bj.bcebos.com/humanseg/models/humanseg_server_inference.zip) | -- | 高精度模型，适用于服务端GPU且背景复杂的人像场景， 模型结构为Deeplabv3+/Xcetion65, 输入大小（512， 512） |
| HumanSeg-mobile | [humanseg_mobile_params](https://paddlex.bj.bcebos.com/humanseg/models/humanseg_mobile.pdparams) | [humanseg_mobile_inference](https://paddlex.bj.bcebos.com/humanseg/models/humanseg_mobile_inference.zip) | [humanseg_mobile_quant](https://paddlex.bj.bcebos.com/humanseg/models/humanseg_mobile_quant.zip) | 轻量级模型, 适用于移动端或服务端CPU的前置摄像头场景，模型结构为HRNet_w18_samll_v1，输入大小（192， 192） |
模型性能
| 模型 | 模型大小 | 计算耗时 |
| --- | --- | --- |
|humanseg_server_inference| 158M | - |
|humanseg_mobile_inference | 5.8 M | 42.35ms |
|humanseg_mobile_quant | 1.6M | 24.93ms |
计算耗时运行环境： 小米，cpu：骁龙855， 内存：6GB， 图片大小：192*192
**NOTE:**
其中Checkpoint Parameter为模型权重，用于Fine-tuning场景。
* Inference Model和Quant Inference Model为预测部署模型，包含`__model__`计算图结构、`__params__`模型参数和`model.yaml`基础的模型配置信息。
* 其中Inference Model适用于服务端的CPU和GPU预测部署，Qunat Inference Model为量化版本，适用于通过Paddle Lite进行移动端等端侧设备部署。
执行以下脚本进行HumanSeg预训练模型的下载
```bash
python pretrain_weights/download_pretrain_weights.py
```
## 下载测试数据
我们提供了[supervise.ly](https://supervise.ly/)发布人像分割数据集**Supervisely Persons**, 从中随机抽取一小部分并转化成PaddleX可直接加载数据格式。通过运行以下代码进行快速下载，其中包含手机前置摄像头的人像测试视频`video_test.mp4`.
```bash
python data/download_data.py
```
## 快速体验视频流人像分割
结合DIS（Dense Inverse Search-basedmethod）光流算法预测结果与分割结果，改善视频流人像分割
```bash
# 通过电脑摄像头进行实时分割处理
python video_infer.py --model_dir pretrain_weights/humanseg_mobile_inference
# 对人像视频进行分割处理
python video_infer.py --model_dir pretrain_weights/humanseg_mobile_inference --video_path data/video_test.mp4
```
视频分割结果如下：
<img src="https://paddleseg.bj.bcebos.com/humanseg/data/video_test.gif" width="20%" height="20%"><img src="https://paddleseg.bj.bcebos.com/humanseg/data/result.gif" width="20%" height="20%">
根据所选背景进行背景替换，背景可以是一张图片，也可以是一段视频。
```bash
# 通过电脑摄像头进行实时背景替换处理, 也可通过'--background_video_path'传入背景视频
python bg_replace.py --model_dir pretrain_weights/humanseg_mobile_inference --background_image_path data/background.jpg
# 对人像视频进行背景替换处理, 也可通过'--background_video_path'传入背景视频
python bg_replace.py --model_dir pretrain_weights/humanseg_mobile_inference --video_path data/video_test.mp4 --background_image_path data/background.jpg
# 对单张图像进行背景替换
python bg_replace.py --model_dir pretrain_weights/humanseg_mobile_inference --image_path data/human_image.jpg --background_image_path data/background.jpg
```
背景替换结果如下：
<img src="https://paddleseg.bj.bcebos.com/humanseg/data/video_test.gif" width="20%" height="20%"><img src="https://paddleseg.bj.bcebos.com/humanseg/data/bg_replace.gif" width="20%" height="20%">
**NOTE**:
视频分割处理时间需要几分钟，请耐心等待。
提供的模型适用于手机摄像头竖屏拍摄场景，宽屏效果会略差一些。
## 训练
使用下述命令基于与训练模型进行Fine-tuning，请确保选用的模型结构`model_type`与模型参数`pretrain_weights`匹配。
```bash
# 指定GPU卡号（以0号卡为例）
export CUDA_VISIBLE_DEVICES=0
# 若不使用GPU，则将CUDA_VISIBLE_DEVICES指定为空
# export CUDA_VISIBLE_DEVICES=
python train.py --model_type HumanSegMobile \
--save_dir output/ \
--data_dir data/mini_supervisely \
--train_list data/mini_supervisely/train.txt \
--val_list data/mini_supervisely/val.txt \
--pretrain_weights pretrain_weights/humanseg_mobile_params \
--batch_size 8 \
--learning_rate 0.001 \
--num_epochs 10 \
--image_shape 192 192
```
其中参数含义如下：
* `--model_type`: 模型类型，可选项为：HumanSegServer和HumanSegMobile
* `--save_dir`: 模型保存路径
* `--data_dir`: 数据集路径
* `--train_list`: 训练集列表路径
* `--val_list`: 验证集列表路径
* `--pretrain_weights`: 预训练模型路径
* `--batch_size`: 批大小
* `--learning_rate`: 初始学习率
* `--num_epochs`: 训练轮数
* `--image_shape`: 网络输入图像大小（w, h）
更多命令行帮助可运行下述命令进行查看：
```bash
python train.py --help
```
**NOTE**
可通过更换`--model_type`变量与对应的`--pretrain_weights`使用不同的模型快速尝试。
## 评估
使用下述命令进行评估
```bash
python eval.py --model_dir output/best_model \
--data_dir data/mini_supervisely \
--val_list data/mini_supervisely/val.txt \
--image_shape 192 192
```
其中参数含义如下：
* `--model_dir`: 模型路径
* `--data_dir`: 数据集路径
* `--val_list`: 验证集列表路径
* `--image_shape`: 网络输入图像大小（w, h）
## 预测
使用下述命令进行预测， 预测结果默认保存在`./output/result/`文件夹中。
```bash
python infer.py --model_dir output/best_model \
--data_dir data/mini_supervisely \
--test_list data/mini_supervisely/test.txt \
--save_dir output/result \
--image_shape 192 192
```
其中参数含义如下：
* `--model_dir`: 模型路径
* `--data_dir`: 数据集路径
* `--test_list`: 测试集列表路径
* `--image_shape`: 网络输入图像大小（w, h）
## 模型导出
```bash
paddlex --export_inference --model_dir output/best_model \
--save_dir output/export
```
其中参数含义如下：
* `--model_dir`: 模型路径
* `--save_dir`: 导出模型保存路径
## 离线量化
```bash
python quant_offline.py --model_dir output/best_model \
--data_dir data/mini_supervisely \
--quant_list data/mini_supervisely/val.txt \
--save_dir output/quant_offline \
--image_shape 192 192
```
其中参数含义如下：
* `--model_dir`: 待量化模型路径
* `--data_dir`: 数据集路径
* `--quant_list`: 量化数据集列表路径，一般直接选择训练集或验证集
* `--save_dir`: 量化模型保存路径
* `--image_shape`: 网络输入图像大小（w, h）
examples/human_segmentation/bg_replace.py 0 → 100644
浏览文件 @ 4c631565
# coding: utf8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import os
import os.path as osp
import cv2
import numpy as np
from postprocess import postprocess, threshold_mask
import paddlex as pdx
import paddlex.utils.logging as logging
from paddlex.seg import transforms
def parse_args():
parser = argparse.ArgumentParser(
description='HumanSeg inference for video')
parser.add_argument(
'--model_dir',
dest='model_dir',
help='Model path for inference',
type=str)
parser.add_argument(
'--image_path',
dest='image_path',
help='Image including human',
type=str,
default=None)
parser.add_argument(
'--background_image_path',
dest='background_image_path',
help='Background image for replacing',
type=str,
default=None)
parser.add_argument(
'--video_path',
dest='video_path',
help='Video path for inference',
type=str,
default=None)
parser.add_argument(
'--background_video_path',
dest='background_video_path',
help='Background video path for replacing',
type=str,
default=None)
parser.add_argument(
'--save_dir',
dest='save_dir',
help='The directory for saving the inference results',
type=str,
default='./output')
parser.add_argument(
"--image_shape",
dest="image_shape",
help="The image shape for net inputs.",
nargs=2,
default=[192, 192],
type=int)
return parser.parse_args()
def bg_replace(label_map, img, bg):
h, w, _ = img.shape
bg = cv2.resize(bg, (w, h))
label_map = np.repeat(label_map[:, :, np.newaxis], 3, axis=2)
comb = (label_map * img + (1 - label_map) * bg).astype(np.uint8)
return comb
def recover(img, im_info):
if im_info[0] == 'resize':
w, h = im_info[1][1], im_info[1][0]
img = cv2.resize(img, (w, h), cv2.INTER_LINEAR)
elif im_info[0] == 'padding':
w, h = im_info[1][0], im_info[1][0]
img = img[0:h, 0:w, :]
return img
def infer(args):
resize_h = args.image_shape[1]
resize_w = args.image_shape[0]
test_transforms = transforms.Compose([transforms.Normalize()])
model = pdx.load_model(args.model_dir)
if not osp.exists(args.save_dir):
os.makedirs(args.save_dir)
# 图像背景替换
if args.image_path is not None:
if not osp.exists(args.image_path):
raise Exception('The --image_path is not existed: {}'.format(
args.image_path))
if args.background_image_path is None:
raise Exception(
'The --background_image_path is not set. Please set it')
else:
if not osp.exists(args.background_image_path):
raise Exception(
'The --background_image_path is not existed: {}'.format(
args.background_image_path))
img = cv2.imread(args.image_path)
im_shape = img.shape
im_scale_x = float(resize_w) / float(im_shape[1])
im_scale_y = float(resize_h) / float(im_shape[0])
im = cv2.resize(
img,
None,
None,
fx=im_scale_x,
fy=im_scale_y,
interpolation=cv2.INTER_LINEAR)
image = im.astype('float32')
im_info = ('resize', im_shape[0:2])
pred = model.predict(image, test_transforms)
label_map = pred['label_map']
label_map = recover(label_map, im_info)
bg = cv2.imread(args.background_image_path)
save_name = osp.basename(args.image_path)
save_path = osp.join(args.save_dir, save_name)
result = bg_replace(label_map, img, bg)
cv2.imwrite(save_path, result)
# 视频背景替换，如果提供背景视频则以背景视频作为背景，否则采用提供的背景图片
else:
is_video_bg = False
if args.background_video_path is not None:
if not osp.exists(args.background_video_path):
raise Exception(
'The --background_video_path is not existed: {}'.format(
args.background_video_path))
is_video_bg = True
elif args.background_image_path is not None:
if not osp.exists(args.background_image_path):
raise Exception(
'The --background_image_path is not existed: {}'.format(
args.background_image_path))
else:
raise Exception(
'Please offer backgound image or video. You should set --backbground_iamge_paht or --background_video_path'
)
disflow = cv2.DISOpticalFlow_create(
cv2.DISOPTICAL_FLOW_PRESET_ULTRAFAST)
prev_gray = np.zeros((resize_h, resize_w), np.uint8)
prev_cfd = np.zeros((resize_h, resize_w), np.float32)
is_init = True
if args.video_path is not None:
logging.info('Please wait. It is computing......')
if not osp.exists(args.video_path):
raise Exception('The --video_path is not existed: {}'.format(
args.video_path))
cap_video = cv2.VideoCapture(args.video_path)
fps = cap_video.get(cv2.CAP_PROP_FPS)
width = int(cap_video.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap_video.get(cv2.CAP_PROP_FRAME_HEIGHT))
save_name = osp.basename(args.video_path)
save_name = save_name.split('.')[0]
save_path = osp.join(args.save_dir, save_name + '.avi')
cap_out = cv2.VideoWriter(
save_path,
cv2.VideoWriter_fourcc('M', 'J', 'P', 'G'), fps,
(width, height))
if is_video_bg:
cap_bg = cv2.VideoCapture(args.background_video_path)
frames_bg = cap_bg.get(cv2.CAP_PROP_FRAME_COUNT)
current_frame_bg = 1
else:
img_bg = cv2.imread(args.background_image_path)
while cap_video.isOpened():
ret, frame = cap_video.read()
if ret:
im_shape = frame.shape
im_scale_x = float(resize_w) / float(im_shape[1])
im_scale_y = float(resize_h) / float(im_shape[0])
im = cv2.resize(
frame,
None,
None,
fx=im_scale_x,
fy=im_scale_y,
interpolation=cv2.INTER_LINEAR)
image = im.astype('float32')
im_info = ('resize', im_shape[0:2])
pred = model.predict(image, test_transforms)
score_map = pred['score_map']
cur_gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
cur_gray = cv2.resize(cur_gray, (resize_w, resize_h))
score_map = 255 * score_map[:, :, 1]
optflow_map = postprocess(cur_gray, score_map, prev_gray, prev_cfd, \
disflow, is_init)
prev_gray = cur_gray.copy()
prev_cfd = optflow_map.copy()
is_init = False
optflow_map = cv2.GaussianBlur(optflow_map, (3, 3), 0)
optflow_map = threshold_mask(
optflow_map, thresh_bg=0.2, thresh_fg=0.8)
score_map = recover(optflow_map, im_info)
#循环读取背景帧
if is_video_bg:
ret_bg, frame_bg = cap_bg.read()
if ret_bg:
if current_frame_bg == frames_bg:
current_frame_bg = 1
cap_bg.set(cv2.CAP_PROP_POS_FRAMES, 0)
else:
break
current_frame_bg += 1
comb = bg_replace(score_map, frame, frame_bg)
else:
comb = bg_replace(score_map, frame, img_bg)
cap_out.write(comb)
else:
break
if is_video_bg:
cap_bg.release()
cap_video.release()
cap_out.release()
# 当没有输入预测图像和视频的时候，则打开摄像头
else:
cap_video = cv2.VideoCapture(0)
if not cap_video.isOpened():
raise IOError("Error opening video stream or file, "
"--video_path whether existing: {}"
" or camera whether working".format(
args.video_path))
return
if is_video_bg:
cap_bg = cv2.VideoCapture(args.background_video_path)
frames_bg = cap_bg.get(cv2.CAP_PROP_FRAME_COUNT)
current_frame_bg = 1
else:
img_bg = cv2.imread(args.background_image_path)
while cap_video.isOpened():
ret, frame = cap_video.read()
if ret:
im_shape = frame.shape
im_scale_x = float(resize_w) / float(im_shape[1])
im_scale_y = float(resize_h) / float(im_shape[0])
im = cv2.resize(
frame,
None,
None,
fx=im_scale_x,
fy=im_scale_y,
interpolation=cv2.INTER_LINEAR)
image = im.astype('float32')
im_info = ('resize', im_shape[0:2])
pred = model.predict(image, test_transforms)
score_map = pred['score_map']
cur_gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
cur_gray = cv2.resize(cur_gray, (resize_w, resize_h))
score_map = 255 * score_map[:, :, 1]
optflow_map = postprocess(cur_gray, score_map, prev_gray, prev_cfd, \
disflow, is_init)
prev_gray = cur_gray.copy()
prev_cfd = optflow_map.copy()
is_init = False
optflow_map = cv2.GaussianBlur(optflow_map, (3, 3), 0)
optflow_map = threshold_mask(
optflow_map, thresh_bg=0.2, thresh_fg=0.8)
score_map = recover(optflow_map, im_info)
#循环读取背景帧
if is_video_bg:
ret_bg, frame_bg = cap_bg.read()
if ret_bg:
if current_frame_bg == frames_bg:
current_frame_bg = 1
cap_bg.set(cv2.CAP_PROP_POS_FRAMES, 0)
else:
break
current_frame_bg += 1
comb = bg_replace(score_map, frame, frame_bg)
else:
comb = bg_replace(score_map, frame, img_bg)
cv2.imshow('HumanSegmentation', comb)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
else:
break
if is_video_bg:
cap_bg.release()
cap_video.release()
if __name__ == "__main__":
args = parse_args()
infer(args)
examples/human_segmentation/data/download_data.py 0 → 100644
浏览文件 @ 4c631565
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
import os
LOCAL_PATH = os.path.dirname(os.path.abspath(__file__))
import paddlex as pdx
def download_data(savepath):
url = "https://paddleseg.bj.bcebos.com/humanseg/data/mini_supervisely.zip"
pdx.utils.download_and_decompress(url=url, path=savepath)
url = "https://paddleseg.bj.bcebos.com/humanseg/data/video_test.zip"
pdx.utils.download_and_decompress(url=url, path=savepath)
if __name__ == "__main__":
download_data(LOCAL_PATH)
print("Data download finish!")
examples/human_segmentation/eval.py 0 → 100644
浏览文件 @ 4c631565
# coding: utf8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import paddlex as pdx
import paddlex.utils.logging as logging
from paddlex.seg import transforms
def parse_args():
parser = argparse.ArgumentParser(description='HumanSeg training')
parser.add_argument(
'--model_dir',
dest='model_dir',
help='Model path for evaluating',
type=str,
default='output/best_model')
parser.add_argument(
'--data_dir',
dest='data_dir',
help='The root directory of dataset',
type=str)
parser.add_argument(
'--val_list',
dest='val_list',
help='Val list file of dataset',
type=str,
default=None)
parser.add_argument(
'--batch_size',
dest='batch_size',
help='Mini batch size',
type=int,
default=128)
parser.add_argument(
"--image_shape",
dest="image_shape",
help="The image shape for net inputs.",
nargs=2,
default=[192, 192],
type=int)
return parser.parse_args()
def dict2str(dict_input):
out = ''
for k, v in dict_input.items():
try:
v = round(float(v), 6)
except:
pass
out = out + '{}={}, '.format(k, v)
return out.strip(', ')
def evaluate(args):
eval_transforms = transforms.Compose(
[transforms.Resize(args.image_shape), transforms.Normalize()])
eval_dataset = pdx.datasets.SegDataset(
data_dir=args.data_dir,
file_list=args.val_list,
transforms=eval_transforms)
model = pdx.load_model(args.model_dir)
metrics = model.evaluate(eval_dataset, args.batch_size)
logging.info('[EVAL] Finished, {} .'.format(dict2str(metrics)))
if __name__ == '__main__':
args = parse_args()
evaluate(args)
examples/human_segmentation/infer.py 0 → 100644
浏览文件 @ 4c631565
# coding: utf8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import os
import os.path as osp
import cv2
import numpy as np
import tqdm
import paddlex as pdx
from paddlex.seg import transforms
def parse_args():
parser = argparse.ArgumentParser(
description='HumanSeg prediction and visualization')
parser.add_argument(
'--model_dir',
dest='model_dir',
help='Model path for prediction',
type=str)
parser.add_argument(
'--data_dir',
dest='data_dir',
help='The root directory of dataset',
type=str)
parser.add_argument(
'--test_list',
dest='test_list',
help='Test list file of dataset',
type=str)
parser.add_argument(
'--save_dir',
dest='save_dir',
help='The directory for saving the inference results',
type=str,
default='./output/result')
parser.add_argument(
"--image_shape",
dest="image_shape",
help="The image shape for net inputs.",
nargs=2,
default=[192, 192],
type=int)
return parser.parse_args()
def infer(args):
def makedir(path):
sub_dir = osp.dirname(path)
if not osp.exists(sub_dir):
os.makedirs(sub_dir)
test_transforms = transforms.Compose(
[transforms.Resize(args.image_shape), transforms.Normalize()])
model = pdx.load_model(args.model_dir)
added_saved_path = osp.join(args.save_dir, 'added')
mat_saved_path = osp.join(args.save_dir, 'mat')
scoremap_saved_path = osp.join(args.save_dir, 'scoremap')
with open(args.test_list, 'r') as f:
files = f.readlines()
for file in tqdm.tqdm(files):
file = file.strip()
im_file = osp.join(args.data_dir, file)
im = cv2.imread(im_file)
result = model.predict(im_file, transforms=test_transforms)
# save added image
added_image = pdx.seg.visualize(
im_file, result, weight=0.6, save_dir=None)
added_image_file = osp.join(added_saved_path, file)
makedir(added_image_file)
cv2.imwrite(added_image_file, added_image)
# save score map
score_map = result['score_map'][:, :, 1]
score_map = (score_map * 255).astype(np.uint8)
score_map_file = osp.join(scoremap_saved_path, file)
makedir(score_map_file)
cv2.imwrite(score_map_file, score_map)
# save mat image
score_map = np.expand_dims(score_map, axis=-1)
mat_image = np.concatenate([im, score_map], axis=2)
mat_file = osp.join(mat_saved_path, file)
ext = osp.splitext(mat_file)[-1]
mat_file = mat_file.replace(ext, '.png')
makedir(mat_file)
cv2.imwrite(mat_file, mat_image)
if __name__ == '__main__':
args = parse_args()
infer(args)
examples/human_segmentation/postprocess.py 0 → 100644
浏览文件 @ 4c631565
# coding: utf8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
def cal_optical_flow_tracking(pre_gray, cur_gray, prev_cfd, dl_weights,
disflow):
"""计算光流跟踪匹配点和光流图
输入参数:
pre_gray: 上一帧灰度图
cur_gray: 当前帧灰度图
prev_cfd: 上一帧光流图
dl_weights: 融合权重图
disflow: 光流数据结构
返回值:
is_track: 光流点跟踪二值图，即是否具有光流点匹配
track_cfd: 光流跟踪图
"""
check_thres = 8
h, w = pre_gray.shape[:2]
track_cfd = np.zeros_like(prev_cfd)
is_track = np.zeros_like(pre_gray)
flow_fw = disflow.calc(pre_gray, cur_gray, None)
flow_bw = disflow.calc(cur_gray, pre_gray, None)
flow_fw = np.round(flow_fw).astype(np.int)
flow_bw = np.round(flow_bw).astype(np.int)
y_list = np.array(range(h))
x_list = np.array(range(w))
yv, xv = np.meshgrid(y_list, x_list)
yv, xv = yv.T, xv.T
cur_x = xv + flow_fw[:, :, 0]
cur_y = yv + flow_fw[:, :, 1]
# 超出边界不跟踪
not_track = (cur_x < 0) + (cur_x >= w) + (cur_y < 0) + (cur_y >= h)
flow_bw[~not_track] = flow_bw[cur_y[~not_track], cur_x[~not_track]]
not_track += (np.square(flow_fw[:, :, 0] + flow_bw[:, :, 0]) +
np.square(flow_fw[:, :, 1] + flow_bw[:, :, 1])
) >= check_thres
track_cfd[cur_y[~not_track], cur_x[~not_track]] = prev_cfd[~not_track]
is_track[cur_y[~not_track], cur_x[~not_track]] = 1
not_flow = np.all(np.abs(flow_fw) == 0,
axis=-1) * np.all(np.abs(flow_bw) == 0, axis=-1)
dl_weights[cur_y[not_flow], cur_x[not_flow]] = 0.05
return track_cfd, is_track, dl_weights
def fuse_optical_flow_tracking(track_cfd, dl_cfd, dl_weights, is_track):
"""光流追踪图和人像分割结构融合
输入参数:
track_cfd: 光流追踪图
dl_cfd: 当前帧分割结果
dl_weights: 融合权重图
is_track: 光流点匹配二值图
返回
cur_cfd: 光流跟踪图和人像分割结果融合图
"""
fusion_cfd = dl_cfd.copy()
is_track = is_track.astype(np.bool)
fusion_cfd[is_track] = dl_weights[is_track] * dl_cfd[is_track] + (
1 - dl_weights[is_track]) * track_cfd[is_track]
# 确定区域
index_certain = ((dl_cfd > 0.9) + (dl_cfd < 0.1)) * is_track
index_less01 = (dl_weights < 0.1) * index_certain
fusion_cfd[index_less01] = 0.3 * dl_cfd[index_less01] + 0.7 * track_cfd[
index_less01]
index_larger09 = (dl_weights >= 0.1) * index_certain
fusion_cfd[index_larger09] = 0.4 * dl_cfd[
index_larger09] + 0.6 * track_cfd[index_larger09]
return fusion_cfd
def threshold_mask(img, thresh_bg, thresh_fg):
dst = (img / 255.0 - thresh_bg) / (thresh_fg - thresh_bg)
dst[np.where(dst > 1)] = 1
dst[np.where(dst < 0)] = 0
return dst.astype(np.float32)
def postprocess(cur_gray, scoremap, prev_gray, pre_cfd, disflow, is_init):
"""光流优化
Args:
cur_gray : 当前帧灰度图
pre_gray : 前一帧灰度图
pre_cfd ：前一帧融合结果
scoremap : 当前帧分割结果
difflow : 光流
is_init : 是否第一帧
Returns:
fusion_cfd : 光流追踪图和预测结果融合图
"""
h, w = scoremap.shape
cur_cfd = scoremap.copy()
if is_init:
if h <= 64 or w <= 64:
disflow.setFinestScale(1)
elif h <= 160 or w <= 160:
disflow.setFinestScale(2)
else:
disflow.setFinestScale(3)
fusion_cfd = cur_cfd
else:
weights = np.ones((h, w), np.float32) * 0.3
track_cfd, is_track, weights = cal_optical_flow_tracking(
prev_gray, cur_gray, pre_cfd, weights, disflow)
fusion_cfd = fuse_optical_flow_tracking(track_cfd, cur_cfd, weights,
is_track)
return fusion_cfd
examples/human_segmentation/pretrain_weights/download_pretrain_weights.py 0 → 100644
浏览文件 @ 4c631565
# coding: utf8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
import os
LOCAL_PATH = os.path.dirname(os.path.abspath(__file__))
import paddlex as pdx
import paddlehub as hub
model_urls = {
"PaddleX_HumanSeg_Server_Params":
"https://bj.bcebos.com/paddlex/models/humanseg/humanseg_server_params.tar",
"PaddleX_HumanSeg_Server_Inference":
"https://bj.bcebos.com/paddlex/models/humanseg/humanseg_server_inference.tar",
"PaddleX_HumanSeg_Mobile_Params":
"https://bj.bcebos.com/paddlex/models/humanseg/humanseg_mobile_params.tar",
"PaddleX_HumanSeg_Mobile_Inference":
"https://bj.bcebos.com/paddlex/models/humanseg/humanseg_mobile_inference.tar",
"PaddleX_HumanSeg_Mobile_Quant":
"https://bj.bcebos.com/paddlex/models/humanseg/humanseg_mobile_quant.tar"
}
if __name__ == "__main__":
for model_name, url in model_urls.items():
pdx.utils.download_and_decompress(url=url, path=LOCAL_PATH)
print("Pretrained Model download success!")
examples/human_segmentation/quant_offline.py 0 → 100644
浏览文件 @ 4c631565
# coding: utf8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import paddlex as pdx
from paddlex.seg import transforms
def parse_args():
parser = argparse.ArgumentParser(description='HumanSeg training')
parser.add_argument(
'--model_dir',
dest='model_dir',
help='Model path for quant',
type=str,
default='output/best_model')
parser.add_argument(
'--batch_size',
dest='batch_size',
help='Mini batch size',
type=int,
default=1)
parser.add_argument(
'--batch_nums',
dest='batch_nums',
help='Batch number for quant',
type=int,
default=10)
parser.add_argument(
'--data_dir',
dest='data_dir',
help='the root directory of dataset',
type=str)
parser.add_argument(
'--quant_list',
dest='quant_list',
help='Image file list for model quantization, it can be vat.txt or train.txt',
type=str,
default=None)
parser.add_argument(
'--save_dir',
dest='save_dir',
help='The directory for saving the quant model',
type=str,
default='./output/quant_offline')
parser.add_argument(
"--image_shape",
dest="image_shape",
help="The image shape for net inputs.",
nargs=2,
default=[192, 192],
type=int)
return parser.parse_args()
def evaluate(args):
eval_transforms = transforms.Compose(
[transforms.Resize(args.image_shape), transforms.Normalize()])
eval_dataset = pdx.datasets.SegDataset(
data_dir=args.data_dir,
file_list=args.quant_list,
transforms=eval_transforms)
model = pdx.load_model(args.model_dir)
pdx.slim.export_quant_model(model, eval_dataset, args.batch_size,
args.batch_nums, args.save_dir)
if __name__ == '__main__':
args = parse_args()
evaluate(args)
examples/human_segmentation/train.py 0 → 100644
浏览文件 @ 4c631565
# coding: utf8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import paddlex as pdx
from paddlex.seg import transforms
MODEL_TYPE = ['HumanSegMobile', 'HumanSegServer']
def parse_args():
parser = argparse.ArgumentParser(description='HumanSeg training')
parser.add_argument(
'--model_type',
dest='model_type',
help="Model type for traing, which is one of ('HumanSegMobile', 'HumanSegServer')",
type=str,
default='HumanSegMobile')
parser.add_argument(
'--data_dir',
dest='data_dir',
help='The root directory of dataset',
type=str)
parser.add_argument(
'--train_list',
dest='train_list',
help='Train list file of dataset',
type=str)
parser.add_argument(
'--val_list',
dest='val_list',
help='Val list file of dataset',
type=str,
default=None)
parser.add_argument(
'--save_dir',
dest='save_dir',
help='The directory for saving the model snapshot',
type=str,
default='./output')
parser.add_argument(
'--num_classes',
dest='num_classes',
help='Number of classes',
type=int,
default=2)
parser.add_argument(
"--image_shape",
dest="image_shape",
help="The image shape for net inputs.",
nargs=2,
default=[192, 192],
type=int)
parser.add_argument(
'--num_epochs',
dest='num_epochs',
help='Number epochs for training',
type=int,
default=100)
parser.add_argument(
'--batch_size',
dest='batch_size',
help='Mini batch size',
type=int,
default=128)
parser.add_argument(
'--learning_rate',
dest='learning_rate',
help='Learning rate',
type=float,
default=0.01)
parser.add_argument(
'--pretrain_weights',
dest='pretrain_weights',
help='The path of pretrianed weight',
type=str,
default=None)
parser.add_argument(
'--resume_checkpoint',
dest='resume_checkpoint',
help='The path of resume checkpoint',
type=str,
default=None)
parser.add_argument(
'--use_vdl',
dest='use_vdl',
help='Whether to use visualdl',
action='store_true')
parser.add_argument(
'--save_interval_epochs',
dest='save_interval_epochs',
help='The interval epochs for save a model snapshot',
type=int,
default=5)
return parser.parse_args()
def train(args):
train_transforms = transforms.Compose([
transforms.Resize(args.image_shape), transforms.RandomHorizontalFlip(),
transforms.Normalize()
])
eval_transforms = transforms.Compose(
[transforms.Resize(args.image_shape), transforms.Normalize()])
train_dataset = pdx.datasets.SegDataset(
data_dir=args.data_dir,
file_list=args.train_list,
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.SegDataset(
data_dir=args.data_dir,
file_list=args.val_list,
transforms=eval_transforms)
if args.model_type == 'HumanSegMobile':
model = pdx.seg.HRNet(
num_classes=args.num_classes, width='18_small_v1')
elif args.model_type == 'HumanSegServer':
model = pdx.seg.DeepLabv3p(
num_classes=args.num_classes, backbone='Xception65')
else:
raise ValueError(
"--model_type: {} is set wrong, it shold be one of ('HumanSegMobile', "
"'HumanSegLite', 'HumanSegServer')".format(args.model_type))
model.train(
num_epochs=args.num_epochs,
train_dataset=train_dataset,
train_batch_size=args.batch_size,
eval_dataset=eval_dataset,
save_interval_epochs=args.save_interval_epochs,
learning_rate=args.learning_rate,
pretrain_weights=args.pretrain_weights,
resume_checkpoint=args.resume_checkpoint,
save_dir=args.save_dir,
use_vdl=args.use_vdl)
if __name__ == '__main__':
args = parse_args()
train(args)
examples/human_segmentation/video_infer.py 0 → 100644
浏览文件 @ 4c631565
# coding: utf8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import os
import os.path as osp
import cv2
import numpy as np
from postprocess import postprocess, threshold_mask
import paddlex as pdx
import paddlex.utils.logging as logging
from paddlex.seg import transforms
def parse_args():
parser = argparse.ArgumentParser(
description='HumanSeg inference for video')
parser.add_argument(
'--model_dir',
dest='model_dir',
help='Model path for inference',
type=str)
parser.add_argument(
'--video_path',
dest='video_path',
help='Video path for inference, camera will be used if the path not existing',
type=str,
default=None)
parser.add_argument(
'--save_dir',
dest='save_dir',
help='The directory for saving the inference results',
type=str,
default='./output')
parser.add_argument(
"--image_shape",
dest="image_shape",
help="The image shape for net inputs.",
nargs=2,
default=[192, 192],
type=int)
return parser.parse_args()
def recover(img, im_info):
if im_info[0] == 'resize':
w, h = im_info[1][1], im_info[1][0]
img = cv2.resize(img, (w, h), cv2.INTER_LINEAR)
elif im_info[0] == 'padding':
w, h = im_info[1][0], im_info[1][0]
img = img[0:h, 0:w, :]
return img
def video_infer(args):
resize_h = args.image_shape[1]
resize_w = args.image_shape[0]
model = pdx.load_model(args.model_dir)
test_transforms = transforms.Compose([transforms.Normalize()])
if not args.video_path:
cap = cv2.VideoCapture(0)
else:
cap = cv2.VideoCapture(args.video_path)
if not cap.isOpened():
raise IOError("Error opening video stream or file, "
"--video_path whether existing: {}"
" or camera whether working".format(args.video_path))
return
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
disflow = cv2.DISOpticalFlow_create(cv2.DISOPTICAL_FLOW_PRESET_ULTRAFAST)
prev_gray = np.zeros((resize_h, resize_w), np.uint8)
prev_cfd = np.zeros((resize_h, resize_w), np.float32)
is_init = True
fps = cap.get(cv2.CAP_PROP_FPS)
if args.video_path:
logging.info("Please wait. It is computing......")
# 用于保存预测结果视频
if not osp.exists(args.save_dir):
os.makedirs(args.save_dir)
out = cv2.VideoWriter(
osp.join(args.save_dir, 'result.avi'),
cv2.VideoWriter_fourcc('M', 'J', 'P', 'G'), fps, (width, height))
# 开始获取视频帧
while cap.isOpened():
ret, frame = cap.read()
if ret:
im_shape = frame.shape
im_scale_x = float(resize_w) / float(im_shape[1])
im_scale_y = float(resize_h) / float(im_shape[0])
im = cv2.resize(
frame,
None,
None,
fx=im_scale_x,
fy=im_scale_y,
interpolation=cv2.INTER_LINEAR)
image = im.astype('float32')
im_info = ('resize', im_shape[0:2])
pred = model.predict(image, test_transforms)
score_map = pred['score_map']
cur_gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
score_map = 255 * score_map[:, :, 1]
optflow_map = postprocess(cur_gray, score_map, prev_gray, prev_cfd, \
disflow, is_init)
prev_gray = cur_gray.copy()
prev_cfd = optflow_map.copy()
is_init = False
optflow_map = cv2.GaussianBlur(optflow_map, (3, 3), 0)
optflow_map = threshold_mask(
optflow_map, thresh_bg=0.2, thresh_fg=0.8)
img_matting = np.repeat(
optflow_map[:, :, np.newaxis], 3, axis=2)
img_matting = recover(img_matting, im_info)
bg_im = np.ones_like(img_matting) * 255
comb = (img_matting * frame +
(1 - img_matting) * bg_im).astype(np.uint8)
out.write(comb)
else:
break
cap.release()
out.release()
else:
while cap.isOpened():
ret, frame = cap.read()
if ret:
im_shape = frame.shape
im_scale_x = float(resize_w) / float(im_shape[1])
im_scale_y = float(resize_h) / float(im_shape[0])
im = cv2.resize(
frame,
None,
None,
fx=im_scale_x,
fy=im_scale_y,
interpolation=cv2.INTER_LINEAR)
image = im.astype('float32')
im_info = ('resize', im_shape[0:2])
pred = model.predict(image, test_transforms)
score_map = pred['score_map']
cur_gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
cur_gray = cv2.resize(cur_gray, (resize_w, resize_h))
score_map = 255 * score_map[:, :, 1]
optflow_map = postprocess(cur_gray, score_map, prev_gray, prev_cfd, \
disflow, is_init)
prev_gray = cur_gray.copy()
prev_cfd = optflow_map.copy()
is_init = False
optflow_map = cv2.GaussianBlur(optflow_map, (3, 3), 0)
optflow_map = threshold_mask(
optflow_map, thresh_bg=0.2, thresh_fg=0.8)
img_matting = np.repeat(
optflow_map[:, :, np.newaxis], 3, axis=2)
img_matting = recover(img_matting, im_info)
bg_im = np.ones_like(img_matting) * 255
comb = (img_matting * frame +
(1 - img_matting) * bg_im).astype(np.uint8)
cv2.imshow('HumanSegmentation', comb)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
else:
break
cap.release()
if __name__ == "__main__":
args = parse_args()
video_infer(args)
new_tutorials/train/README.md 已删除 100644 → 0
浏览文件 @ 2bbff977
# 使用教程——训练模型
本目录下整理了使用PaddleX训练模型的示例代码，代码中均提供了示例数据的自动下载，并均使用单张GPU卡进行训练。
|代码 | 模型任务 | 数据 |
|------|--------|---------|
|classification/mobilenetv2.py | 图像分类MobileNetV2 | 蔬菜分类 |
|classification/resnet50.py | 图像分类ResNet50 | 蔬菜分类 |
|detection/faster_rcnn_r50_fpn.py | 目标检测FasterRCNN | 昆虫检测 |
|detection/mask_rcnn_f50_fpn.py | 实例分割MaskRCNN | 垃圾分拣 |
|segmentation/deeplabv3p.py | 语义分割DeepLabV3| 视盘分割 |
|segmentation/unet.py | 语义分割UNet | 视盘分割 |
|segmentation/hrnet.py | 语义分割HRNet | 视盘分割 |
|segmentation/fast_scnn.py | 语义分割FastSCNN | 视盘分割 |
## 开始训练
在安装PaddleX后，使用如下命令开始训练
```
python classification/mobilenetv2.py
```
new_tutorials/train/classification/mobilenetv2.py 已删除 100644 → 0
浏览文件 @ 2bbff977
import os
# 选择使用0号卡
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
from paddlex.cls import transforms
import paddlex as pdx
# 下载和解压蔬菜分类数据集
veg_dataset = 'https://bj.bcebos.com/paddlex/datasets/vegetables_cls.tar.gz'
pdx.utils.download_and_decompress(veg_dataset, path='./')
# 定义训练和验证时的transforms
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/cls_transforms.html#composedclstransforms
train_transforms = transforms.ComposedClsTransforms(mode='train', crop_size=[224, 224])
eval_transforms = transforms.ComposedClsTransforms(mode='eval', crop_size=[224, 224])
# 定义训练和验证所用的数据集
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/datasets/classification.html#imagenet
train_dataset = pdx.datasets.ImageNet(
data_dir='vegetables_cls',
file_list='vegetables_cls/train_list.txt',
label_list='vegetables_cls/labels.txt',
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.ImageNet(
data_dir='vegetables_cls',
file_list='vegetables_cls/val_list.txt',
label_list='vegetables_cls/labels.txt',
transforms=eval_transforms)
# 初始化模型，并进行训练
# 可使用VisualDL查看训练指标
# VisualDL启动方式: visualdl --logdir output/mobilenetv2/vdl_log --port 8001
# 浏览器打开 https://0.0.0.0:8001即可
# 其中0.0.0.0为本机访问，如为远程服务, 改成相应机器IP
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/models/classification.html#resnet50
model = pdx.cls.MobileNetV2(num_classes=len(train_dataset.labels))
model.train(
num_epochs=10,
train_dataset=train_dataset,
train_batch_size=32,
eval_dataset=eval_dataset,
lr_decay_epochs=[4, 6, 8],
learning_rate=0.025,
save_dir='output/mobilenetv2',
use_vdl=True)
new_tutorials/train/classification/resnet50.py 已删除 100644 → 0
浏览文件 @ 2bbff977
import os
# 选择使用0号卡
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
import paddle.fluid as fluid
from paddlex.cls import transforms
import paddlex as pdx
# 下载和解压蔬菜分类数据集
veg_dataset = 'https://bj.bcebos.com/paddlex/datasets/vegetables_cls.tar.gz'
pdx.utils.download_and_decompress(veg_dataset, path='./')
# 定义训练和验证时的transforms
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/cls_transforms.html#composedclstransforms
train_transforms = transforms.ComposedClsTransforms(mode='train', crop_size=[224, 224])
eval_transforms = transforms.ComposedClsTransforms(mode='eval', crop_size=[224, 224])
# 定义训练和验证所用的数据集
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/datasets/classification.html#imagenet
train_dataset = pdx.datasets.ImageNet(
data_dir='vegetables_cls',
file_list='vegetables_cls/train_list.txt',
label_list='vegetables_cls/labels.txt',
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.ImageNet(
data_dir='vegetables_cls',
file_list='vegetables_cls/val_list.txt',
label_list='vegetables_cls/labels.txt',
transforms=eval_transforms)
# PaddleX支持自定义构建优化器
step_each_epoch = train_dataset.num_samples // 32
learning_rate = fluid.layers.cosine_decay(
learning_rate=0.025, step_each_epoch=step_each_epoch, epochs=10)
optimizer = fluid.optimizer.Momentum(
learning_rate=learning_rate,
momentum=0.9,
regularization=fluid.regularizer.L2Decay(4e-5))
# 初始化模型，并进行训练
# 可使用VisualDL查看训练指标
# VisualDL启动方式: visualdl --logdir output/resnet50/vdl_log --port 8001
# 浏览器打开 https://0.0.0.0:8001即可
# 其中0.0.0.0为本机访问，如为远程服务, 改成相应机器IP
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/models/classification.html#resnet50
model = pdx.cls.ResNet50(num_classes=len(train_dataset.labels))
model.train(
num_epochs=10,
train_dataset=train_dataset,
train_batch_size=32,
eval_dataset=eval_dataset,
optimizer=optimizer,
save_dir='output/resnet50',
use_vdl=True)
new_tutorials/train/detection/faster_rcnn_r50_fpn.py 已删除 100644 → 0
浏览文件 @ 2bbff977
import os
# 选择使用0号卡
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
from paddlex.det import transforms
import paddlex as pdx
# 下载和解压昆虫检测数据集
insect_dataset = 'https://bj.bcebos.com/paddlex/datasets/insect_det.tar.gz'
pdx.utils.download_and_decompress(insect_dataset, path='./')
# 定义训练和验证时的transforms
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/det_transforms.html#composedrcnntransforms
train_transforms = transforms.ComposedRCNNTransforms(mode='train', min_max_size=[800, 1333])
eval_transforms = transforms.ComposedRCNNTransforms(mode='eval', min_max_size=[800, 1333])
# 定义训练和验证所用的数据集
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/datasets/detection.html#vocdetection
train_dataset = pdx.datasets.VOCDetection(
data_dir='insect_det',
file_list='insect_det/train_list.txt',
label_list='insect_det/labels.txt',
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.VOCDetection(
data_dir='insect_det',
file_list='insect_det/val_list.txt',
label_list='insect_det/labels.txt',
transforms=eval_transforms)
# 初始化模型，并进行训练
# 可使用VisualDL查看训练指标
# VisualDL启动方式: visualdl --logdir output/faster_rcnn_r50_fpn/vdl_log --port 8001
# 浏览器打开 https://0.0.0.0:8001即可
# 其中0.0.0.0为本机访问，如为远程服务, 改成相应机器IP
# num_classes 需要设置为包含背景类的类别数，即: 目标类别数量 + 1
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/models/detection.html#fasterrcnn
num_classes = len(train_dataset.labels) + 1
model = pdx.det.FasterRCNN(num_classes=num_classes)
model.train(
num_epochs=12,
train_dataset=train_dataset,
train_batch_size=2,
eval_dataset=eval_dataset,
learning_rate=0.0025,
lr_decay_epochs=[8, 11],
save_dir='output/faster_rcnn_r50_fpn',
use_vdl=True)
new_tutorials/train/detection/mask_rcnn_r50_fpn.py 已删除 100644 → 0
浏览文件 @ 2bbff977
import os
# 选择使用0号卡
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
from paddlex.det import transforms
import paddlex as pdx
# 下载和解压小度熊分拣数据集
xiaoduxiong_dataset = 'https://bj.bcebos.com/paddlex/datasets/xiaoduxiong_ins_det.tar.gz'
pdx.utils.download_and_decompress(xiaoduxiong_dataset, path='./')
# 定义训练和验证时的transforms
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/det_transforms.html#composedrcnntransforms
train_transforms = transforms.ComposedRCNNTransforms(mode='train', min_max_size=[800, 1333])
eval_transforms = transforms.ComposedRCNNTransforms(mode='eval', min_max_size=[800, 1333])
# 定义训练和验证所用的数据集
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/datasets/detection.html#cocodetection
train_dataset = pdx.datasets.CocoDetection(
data_dir='xiaoduxiong_ins_det/JPEGImages',
ann_file='xiaoduxiong_ins_det/train.json',
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.CocoDetection(
data_dir='xiaoduxiong_ins_det/JPEGImages',
ann_file='xiaoduxiong_ins_det/val.json',
transforms=eval_transforms)
# 初始化模型，并进行训练
# 可使用VisualDL查看训练指标
# VisualDL启动方式: visualdl --logdir output/mask_rcnn_r50_fpn/vdl_log --port 8001
# 浏览器打开 https://0.0.0.0:8001即可
# 其中0.0.0.0为本机访问，如为远程服务, 改成相应机器IP
# num_classes 需要设置为包含背景类的类别数，即: 目标类别数量 + 1
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/models/instance_segmentation.html#maskrcnn
num_classes = len(train_dataset.labels) + 1
model = pdx.det.MaskRCNN(num_classes=num_classes)
model.train(
num_epochs=12,
train_dataset=train_dataset,
train_batch_size=1,
eval_dataset=eval_dataset,
learning_rate=0.00125,
warmup_steps=10,
lr_decay_epochs=[8, 11],
save_dir='output/mask_rcnn_r50_fpn',
use_vdl=True)
new_tutorials/train/detection/yolov3_darknet53.py 已删除 100644 → 0
浏览文件 @ 2bbff977
import os
# 选择使用0号卡
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
from paddlex.det import transforms
import paddlex as pdx
# 下载和解压昆虫检测数据集
insect_dataset = 'https://bj.bcebos.com/paddlex/datasets/insect_det.tar.gz'
pdx.utils.download_and_decompress(insect_dataset, path='./')
# 定义训练和验证时的transforms
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/det_transforms.html#composedyolotransforms
train_transforms = transforms.ComposedYOLOv3Transforms(mode='train', shape=[608, 608])
eval_transforms = transforms.ComposedYOLOv3Transforms(mode='eva', shape=[608, 608])
# 定义训练和验证所用的数据集
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/datasets/detection.html#vocdetection
train_dataset = pdx.datasets.VOCDetection(
data_dir='insect_det',
file_list='insect_det/train_list.txt',
label_list='insect_det/labels.txt',
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.VOCDetection(
data_dir='insect_det',
file_list='insect_det/val_list.txt',
label_list='insect_det/labels.txt',
transforms=eval_transforms)
# 初始化模型，并进行训练
# 可使用VisualDL查看训练指标
# VisualDL启动方式: visualdl --logdir output/yolov3_darknet/vdl_log --port 8001
# 浏览器打开 https://0.0.0.0:8001即可
# 其中0.0.0.0为本机访问，如为远程服务, 改成相应机器IP
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/models/detection.html#yolov3
num_classes = len(train_dataset.labels)
model = pdx.det.YOLOv3(num_classes=num_classes, backbone='DarkNet53')
model.train(
num_epochs=270,
train_dataset=train_dataset,
train_batch_size=8,
eval_dataset=eval_dataset,
learning_rate=0.000125,
lr_decay_epochs=[210, 240],
save_dir='output/yolov3_darknet53',
use_vdl=True)
new_tutorials/train/segmentation/deeplabv3p.py 已删除 100644 → 0
浏览文件 @ 2bbff977
import os
# 选择使用0号卡
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
import paddlex as pdx
from paddlex.seg import transforms
# 下载和解压视盘分割数据集
optic_dataset = 'https://bj.bcebos.com/paddlex/datasets/optic_disc_seg.tar.gz'
pdx.utils.download_and_decompress(optic_dataset, path='./')
# 定义训练和验证时的transforms
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/seg_transforms.html#composedsegtransforms
train_transforms = transforms.ComposedSegTransforms(mode='train', train_crop_size=[769, 769])
eval_transforms = transforms.ComposedSegTransforms(mode='eval')
train_transforms.add_augmenters([
transforms.RandomRotate()
])
# 定义训练和验证所用的数据集
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/datasets/semantic_segmentation.html#segdataset
train_dataset = pdx.datasets.SegDataset(
data_dir='optic_disc_seg',
file_list='optic_disc_seg/train_list.txt',
label_list='optic_disc_seg/labels.txt',
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.SegDataset(
data_dir='optic_disc_seg',
file_list='optic_disc_seg/val_list.txt',
label_list='optic_disc_seg/labels.txt',
transforms=eval_transforms)
# 初始化模型，并进行训练
# 可使用VisualDL查看训练指标
# VisualDL启动方式: visualdl --logdir output/deeplab/vdl_log --port 8001
# 浏览器打开 https://0.0.0.0:8001即可
# 其中0.0.0.0为本机访问，如为远程服务, 改成相应机器IP
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/models/semantic_segmentation.html#deeplabv3p
num_classes = len(train_dataset.labels)
model = pdx.seg.DeepLabv3p(num_classes=num_classes)
model.train(
num_epochs=40,
train_dataset=train_dataset,
train_batch_size=4,
eval_dataset=eval_dataset,
learning_rate=0.01,
save_dir='output/deeplab',
use_vdl=True)
new_tutorials/train/segmentation/hrnet.py 已删除 100644 → 0
浏览文件 @ 2bbff977
import os
# 选择使用0号卡
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
import paddlex as pdx
from paddlex.seg import transforms
# 下载和解压视盘分割数据集
optic_dataset = 'https://bj.bcebos.com/paddlex/datasets/optic_disc_seg.tar.gz'
pdx.utils.download_and_decompress(optic_dataset, path='./')
# 定义训练和验证时的transforms
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/seg_transforms.html#composedsegtransforms
train_transforms = transforms.ComposedSegTransforms(mode='train', train_crop_size=[769, 769])
eval_transforms = transforms.ComposedSegTransforms(mode='eval')
# 定义训练和验证所用的数据集
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/datasets/semantic_segmentation.html#segdataset
train_dataset = pdx.datasets.SegDataset(
data_dir='optic_disc_seg',
file_list='optic_disc_seg/train_list.txt',
label_list='optic_disc_seg/labels.txt',
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.SegDataset(
data_dir='optic_disc_seg',
file_list='optic_disc_seg/val_list.txt',
label_list='optic_disc_seg/labels.txt',
transforms=eval_transforms)
# 初始化模型，并进行训练
# 可使用VisualDL查看训练指标
# VisualDL启动方式: visualdl --logdir output/unet/vdl_log --port 8001
# 浏览器打开 https://0.0.0.0:8001即可
# 其中0.0.0.0为本机访问，如为远程服务, 改成相应机器IP
# https://paddlex.readthedocs.io/zh_CN/latest/apis/models/semantic_segmentation.html#hrnet
num_classes = len(train_dataset.labels)
model = pdx.seg.HRNet(num_classes=num_classes)
model.train(
num_epochs=20,
train_dataset=train_dataset,
train_batch_size=4,
eval_dataset=eval_dataset,
learning_rate=0.01,
save_dir='output/hrnet',
use_vdl=True)
new_tutorials/train/segmentation/unet.py 已删除 100644 → 0
浏览文件 @ 2bbff977
import os
# 选择使用0号卡
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
import paddlex as pdx
from paddlex.seg import transforms
# 下载和解压视盘分割数据集
optic_dataset = 'https://bj.bcebos.com/paddlex/datasets/optic_disc_seg.tar.gz'
pdx.utils.download_and_decompress(optic_dataset, path='./')
# 定义训练和验证时的transforms
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/seg_transforms.html#composedsegtransforms
train_transforms = transforms.ComposedSegTransforms(mode='train', train_crop_size=[769, 769])
eval_transforms = transforms.ComposedSegTransforms(mode='eval')
# 定义训练和验证所用的数据集
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/datasets/semantic_segmentation.html#segdataset
train_dataset = pdx.datasets.SegDataset(
data_dir='optic_disc_seg',
file_list='optic_disc_seg/train_list.txt',
label_list='optic_disc_seg/labels.txt',
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.SegDataset(
data_dir='optic_disc_seg',
file_list='optic_disc_seg/val_list.txt',
label_list='optic_disc_seg/labels.txt',
transforms=eval_transforms)
# 初始化模型，并进行训练
# 可使用VisualDL查看训练指标
# VisualDL启动方式: visualdl --logdir output/unet/vdl_log --port 8001
# 浏览器打开 https://0.0.0.0:8001即可
# 其中0.0.0.0为本机访问，如为远程服务, 改成相应机器IP
# API说明: https://paddlex.readthedocs.io/zh_CN/latest/apis/models/semantic_segmentation.html#unet
num_classes = len(train_dataset.labels)
model = pdx.seg.UNet(num_classes=num_classes)
model.train(
num_epochs=20,
train_dataset=train_dataset,
train_batch_size=4,
eval_dataset=eval_dataset,
learning_rate=0.01,
save_dir='output/unet',
use_vdl=True)
paddlex/__init__.py
浏览文件 @ 4c631565
......@@ -53,4 +53,4 @@ log_level = 2
from . import interpret
__version__ = '1.0.6'
__version__ = '1.0.7'
paddlex/command.py
浏览文件 @ 4c631565
......@@ -15,6 +15,7 @@
from six import text_type as _text_type
import argparse
import sys
import paddlex.utils.logging as logging
def arg_parser():
......@@ -94,15 +95,15 @@ def main():
if args.export_onnx:
assert args.model_dir is not None, "--model_dir should be defined while exporting onnx model"
assert args.save_dir is not None, "--save_dir should be defined to create onnx model"
assert args.fixed_input_shape is not None, "--fixed_input_shape should be defined [w,h] to create onnx model, such as [224,224]"
fixed_input_shape = []
if args.fixed_input_shape is not None:
fixed_input_shape = eval(args.fixed_input_shape)
assert len(
fixed_input_shape
) == 2, "len of fixed input shape must == 2, such as [224,224]"
model = pdx.load_model(args.model_dir, fixed_input_shape)
model = pdx.load_model(args.model_dir)
if model.status == "Normal" or model.status == "Prune":
logging.error(
"Only support inference model, try to export model first as below,",
exit=False)
logging.error(
"paddlex --export_inference --model_dir model_path --save_dir infer_model"
)
pdx.convertor.export_onnx_model(model, args.save_dir)
......
paddlex/convertor.py
浏览文件 @ 4c631565
......@@ -30,119 +30,17 @@ def export_onnx(model_dir, save_dir, fixed_input_shape):
def export_onnx_model(model, save_dir):
support_list = [
'ResNet18', 'ResNet34', 'ResNet50', 'ResNet101', 'ResNet50_vd',
'ResNet101_vd', 'ResNet50_vd_ssld', 'ResNet101_vd_ssld', 'DarkNet53',
'MobileNetV1', 'MobileNetV2', 'DenseNet121', 'DenseNet161',
'DenseNet201'
]
if model.__class__.__name__ not in support_list:
raise Exception("Model: {} unsupport export to ONNX".format(
model.__class__.__name__))
try:
from fluid.utils import op_io_info, init_name_prefix
from onnx import helper, checker
import fluid_onnx.ops as ops
from fluid_onnx.variables import paddle_variable_to_onnx_tensor, paddle_onnx_weight
from debug.model_check import debug_model, Tracker
except Exception as e:
if model.model_type == "detector" or model.__class__.__name__ == "FastSCNN":
logging.error(
"Import Module Failed! Please install paddle2onnx. Related requirements see https://github.com/PaddlePaddle/paddle2onnx."
"Only image classifier models and semantic segmentation models(except FastSCNN) are supported to export to ONNX"
)
raise e
place = fluid.CPUPlace()
exe = fluid.Executor(place)
inference_scope = fluid.global_scope()
with fluid.scope_guard(inference_scope):
test_input_names = [
var.name for var in list(model.test_inputs.values())
]
inputs_outputs_list = ["fetch", "feed"]
weights, weights_value_info = [], []
global_block = model.test_prog.global_block()
for var_name in global_block.vars:
var = global_block.var(var_name)
if var_name not in test_input_names\
and var.persistable:
weight, val_info = paddle_onnx_weight(
var=var, scope=inference_scope)
weights.append(weight)
weights_value_info.append(val_info)
# Create inputs
inputs = [
paddle_variable_to_onnx_tensor(v, global_block)
for v in test_input_names
]
logging.INFO("load the model parameter done.")
onnx_nodes = []
op_check_list = []
op_trackers = []
nms_first_index = -1
nms_outputs = []
for block in model.test_prog.blocks:
for op in block.ops:
if op.type in ops.node_maker:
# TODO: deal with the corner case that vars in
# different blocks have the same name
node_proto = ops.node_maker[str(op.type)](
operator=op, block=block)
op_outputs = []
last_node = None
if isinstance(node_proto, tuple):
onnx_nodes.extend(list(node_proto))
last_node = list(node_proto)
else:
onnx_nodes.append(node_proto)
last_node = [node_proto]
tracker = Tracker(str(op.type), last_node)
op_trackers.append(tracker)
op_check_list.append(str(op.type))
if op.type == "multiclass_nms" and nms_first_index < 0:
nms_first_index = 0
if nms_first_index >= 0:
_, _, output_op = op_io_info(op)
for output in output_op:
nms_outputs.extend(output_op[output])
else:
if op.type not in ['feed', 'fetch']:
op_check_list.append(op.type)
logging.info('The operator sets to run test case.')
logging.info(set(op_check_list))
# Create outputs
# Get the new names for outputs if they've been renamed in nodes' making
renamed_outputs = op_io_info.get_all_renamed_outputs()
test_outputs = list(model.test_outputs.values())
test_outputs_names = [var.name for var in model.test_outputs.values()]
test_outputs_names = [
name if name not in renamed_outputs else renamed_outputs[name]
for name in test_outputs_names
]
outputs = [
paddle_variable_to_onnx_tensor(v, global_block)
for v in test_outputs_names
]
# Make graph
onnx_name = 'paddlex.onnx'
onnx_graph = helper.make_graph(
nodes=onnx_nodes,
name=onnx_name,
initializer=weights,
inputs=inputs + weights_value_info,
outputs=outputs)
# Make model
onnx_model = helper.make_model(
onnx_graph, producer_name='PaddlePaddle')
# Model check
checker.check_model(onnx_model)
if onnx_model is not None:
onnx_model_file = os.path.join(save_dir, onnx_name)
if not os.path.exists(save_dir):
os.mkdir(save_dir)
with open(onnx_model_file, 'wb') as f:
f.write(onnx_model.SerializeToString())
logging.info("Saved converted model to path: %s" % onnx_model_file)
try:
import x2paddle
if x2paddle.__version__ < '0.7.4':
logging.error("You need to upgrade x2paddle >= 0.7.4")
except:
logging.error(
"You need to install x2paddle first, pip install x2paddle>=0.7.4")
from x2paddle.op_mapper.paddle_op_mapper import PaddleOpMapper
mapper = PaddleOpMapper()
mapper.convert(model.test_prog, save_dir)
paddlex/cv/datasets/coco.py
浏览文件 @ 4c631565
......@@ -100,7 +100,7 @@ class CocoDetection(VOCDetection):
gt_score = np.ones((num_bbox, 1), dtype=np.float32)
is_crowd = np.zeros((num_bbox, 1), dtype=np.int32)
difficult = np.zeros((num_bbox, 1), dtype=np.int32)
gt_poly = None
gt_poly = [None] * num_bbox
for i, box in enumerate(bboxes):
catid = box['category_id']
......@@ -108,8 +108,6 @@ class CocoDetection(VOCDetection):
gt_bbox[i, :] = box['clean_bbox']
is_crowd[i][0] = box['iscrowd']
if 'segmentation' in box:
if gt_poly is None:
gt_poly = [None] * num_bbox
gt_poly[i] = box['segmentation']
im_info = {
......@@ -121,10 +119,9 @@ class CocoDetection(VOCDetection):
'gt_class': gt_class,
'gt_bbox': gt_bbox,
'gt_score': gt_score,
'gt_poly': gt_poly,
'difficult': difficult
}
if gt_poly is not None:
label_info['gt_poly'] = gt_poly
coco_rec = (im_info, label_info)
self.file_list.append([im_fname, coco_rec])
......
paddlex/cv/datasets/easydata_cls.py
浏览文件 @ 4c631565
......@@ -46,7 +46,7 @@ class EasyDataCls(ImageNet):
label_list,
transforms=None,
num_workers='auto',
buffer_size=100,
buffer_size=8,
parallel_method='process',
shuffle=False):
super(ImageNet, self).__init__(
......@@ -73,8 +73,8 @@ class EasyDataCls(ImageNet):
if not osp.isfile(json_file):
continue
if not osp.exists(img_file):
raise IOError(
'The image file {} is not exist!'.format(img_file))
raise IOError('The image file {} is not exist!'.format(
img_file))
with open(json_file, mode='r', \
encoding=get_encoding(json_file)) as j:
json_info = json.load(j)
......@@ -83,4 +83,3 @@ class EasyDataCls(ImageNet):
self.num_samples = len(self.file_list)
logging.info("{} samples in file {}".format(
len(self.file_list), file_list))
\ No newline at end of file
paddlex/cv/datasets/imagenet.py
浏览文件 @ 4c631565
......@@ -45,7 +45,7 @@ class ImageNet(Dataset):
label_list,
transforms=None,
num_workers='auto',
buffer_size=100,
buffer_size=8,
parallel_method='process',
shuffle=False):
super(ImageNet, self).__init__(
......@@ -70,8 +70,8 @@ class ImageNet(Dataset):
continue
full_path = osp.join(data_dir, items[0])
if not osp.exists(full_path):
raise IOError(
'The image file {} is not exist!'.format(full_path))
raise IOError('The image file {} is not exist!'.format(
full_path))
self.file_list.append([full_path, int(items[1])])
self.num_samples = len(self.file_list)
logging.info("{} samples in file {}".format(
......
paddlex/cv/datasets/seg_dataset.py
浏览文件 @ 4c631565
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
......@@ -28,7 +28,7 @@ class SegDataset(Dataset):
Args:
data_dir (str): 数据集所在的目录路径。
file_list (str): 描述数据集图片文件和对应标注文件的文件路径（文本内每行路径为相对data_dir的相对路）。
label_list (str): 描述数据集包含的类别信息文件路径。
label_list (str): 描述数据集包含的类别信息文件路径。默认值为None。
transforms (list): 数据集中每个样本的预处理/增强算子。
num_workers (int): 数据集中样本在预处理过程中的线程或进程数。默认为4。
buffer_size (int): 数据集中样本在预处理过程中队列的缓存长度，以样本数为单位。默认为100。
......@@ -40,7 +40,7 @@ class SegDataset(Dataset):
def __init__(self,
data_dir,
file_list,
label_list,
label_list=None,
transforms=None,
num_workers='auto',
buffer_size=100,
......@@ -56,6 +56,7 @@ class SegDataset(Dataset):
self.labels = list()
self._epoch = 0
if label_list is not None:
with open(label_list, encoding=get_encoding(label_list)) as f:
for line in f:
item = line.strip()
......@@ -69,8 +70,8 @@ class SegDataset(Dataset):
full_path_im = osp.join(data_dir, items[0])
full_path_label = osp.join(data_dir, items[1])
if not osp.exists(full_path_im):
raise IOError(
'The image file {} is not exist!'.format(full_path_im))
raise IOError('The image file {} is not exist!'.format(
full_path_im))
if not osp.exists(full_path_label):
raise IOError('The image file {} is not exist!'.format(
full_path_label))
......
paddlex/cv/datasets/voc.py
浏览文件 @ 4c631565
......@@ -17,6 +17,7 @@ import copy
import os
import os.path as osp
import random
import re
import numpy as np
from collections import OrderedDict
import xml.etree.ElementTree as ET
......@@ -104,23 +105,60 @@ class VOCDetection(Dataset):
else:
ct = int(tree.find('id').text)
im_id = np.array([int(tree.find('id').text)])
objs = tree.findall('object')
im_w = float(tree.find('size').find('width').text)
im_h = float(tree.find('size').find('height').text)
pattern = re.compile('<object>', re.IGNORECASE)
obj_tag = pattern.findall(
str(ET.tostringlist(tree.getroot())))[0][1:-1]
objs = tree.findall(obj_tag)
pattern = re.compile('<size>', re.IGNORECASE)
size_tag = pattern.findall(
str(ET.tostringlist(tree.getroot())))[0][1:-1]
size_element = tree.find(size_tag)
pattern = re.compile('<width>', re.IGNORECASE)
width_tag = pattern.findall(
str(ET.tostringlist(size_element)))[0][1:-1]
im_w = float(size_element.find(width_tag).text)
pattern = re.compile('<height>', re.IGNORECASE)
height_tag = pattern.findall(
str(ET.tostringlist(size_element)))[0][1:-1]
im_h = float(size_element.find(height_tag).text)
gt_bbox = np.zeros((len(objs), 4), dtype=np.float32)
gt_class = np.zeros((len(objs), 1), dtype=np.int32)
gt_score = np.ones((len(objs), 1), dtype=np.float32)
is_crowd = np.zeros((len(objs), 1), dtype=np.int32)
difficult = np.zeros((len(objs), 1), dtype=np.int32)
for i, obj in enumerate(objs):
cname = obj.find('name').text.strip()
pattern = re.compile('<name>', re.IGNORECASE)
name_tag = pattern.findall(str(ET.tostringlist(obj)))[0][
1:-1]
cname = obj.find(name_tag).text.strip()
gt_class[i][0] = cname2cid[cname]
_difficult = int(obj.find('difficult').text)
x1 = float(obj.find('bndbox').find('xmin').text)
y1 = float(obj.find('bndbox').find('ymin').text)
x2 = float(obj.find('bndbox').find('xmax').text)
y2 = float(obj.find('bndbox').find('ymax').text)
pattern = re.compile('<difficult>', re.IGNORECASE)
diff_tag = pattern.findall(str(ET.tostringlist(obj)))[0][
1:-1]
try:
_difficult = int(obj.find(diff_tag).text)
except Exception:
_difficult = 0
pattern = re.compile('<bndbox>', re.IGNORECASE)
box_tag = pattern.findall(str(ET.tostringlist(obj)))[0][1:
-1]
box_element = obj.find(box_tag)
pattern = re.compile('<xmin>', re.IGNORECASE)
xmin_tag = pattern.findall(
str(ET.tostringlist(box_element)))[0][1:-1]
x1 = float(box_element.find(xmin_tag).text)
pattern = re.compile('<ymin>', re.IGNORECASE)
ymin_tag = pattern.findall(
str(ET.tostringlist(box_element)))[0][1:-1]
y1 = float(box_element.find(ymin_tag).text)
pattern = re.compile('<xmax>', re.IGNORECASE)
xmax_tag = pattern.findall(
str(ET.tostringlist(box_element)))[0][1:-1]
x2 = float(box_element.find(xmax_tag).text)
pattern = re.compile('<ymax>', re.IGNORECASE)
ymax_tag = pattern.findall(
str(ET.tostringlist(box_element)))[0][1:-1]
y2 = float(box_element.find(ymax_tag).text)
x1 = max(0, x1)
y1 = max(0, y1)
if im_w > 0.5 and im_h > 0.5:
......@@ -149,6 +187,7 @@ class VOCDetection(Dataset):
'gt_class': gt_class,
'gt_bbox': gt_bbox,
'gt_score': gt_score,
'gt_poly': [],
'difficult': difficult
}
voc_rec = (im_info, label_info)
......
paddlex/cv/models/hrnet.py
浏览文件 @ 4c631565
......@@ -24,11 +24,12 @@ class HRNet(DeepLabv3p):
Args:
num_classes (int): 类别数。
width (int): 高分辨率分支中特征层的通道数量。默认值为18。可选择取值为[18, 30, 32, 40, 44, 48, 60, 64]。
width (int|str): 高分辨率分支中特征层的通道数量。默认值为18。可选择取值为[18, 30, 32, 40, 44, 48, 60, 64, '18_small_v1']。
'18_small_v1'是18的轻量级版本。
use_bce_loss (bool): 是否使用bce loss作为网络的损失函数，只能用于两类分割。可与dice loss同时使用。默认False。
use_dice_loss (bool): 是否使用dice loss作为网络的损失函数，只能用于两类分割，可与bce loss同时使用。
当use_bce_loss和use_dice_loss都为False时，使用交叉熵损失函数。默认False。
class_weight (list/str): 交叉熵损失函数各类损失的权重。当class_weight为list的时候，长度应为
class_weight (list|str): 交叉熵损失函数各类损失的权重。当class_weight为list的时候，长度应为
num_classes。当class_weight为str时， weight.lower()应为'dynamic'，这时会根据每一轮各类像素的比重
自行计算相应的权重，每一类的权重为：每类的比例 * num_classes。class_weight取默认值None是，各类的权重1，
即平时使用的交叉熵损失函数。
......@@ -168,6 +169,6 @@ class HRNet(DeepLabv3p):
return super(HRNet, self).train(
num_epochs, train_dataset, train_batch_size, eval_dataset,
save_interval_epochs, log_interval_steps, save_dir,
pretrain_weights, optimizer, learning_rate, lr_decay_power, use_vdl,
sensitivities_file, eval_metric_loss, early_stop,
pretrain_weights, optimizer, learning_rate, lr_decay_power,
use_vdl, sensitivities_file, eval_metric_loss, early_stop,
early_stop_patience, resume_checkpoint)
paddlex/cv/models/load_model.py
浏览文件 @ 4c631565
......@@ -108,6 +108,7 @@ def load_model(model_dir, fixed_input_shape=None):
logging.info("Model[{}] loaded.".format(info['Model']))
model.trainable = False
model.status = status
return model
......
paddlex/cv/models/slim/prune.py
浏览文件 @ 4c631565
......@@ -158,6 +158,7 @@ def prune_program(model, prune_params_ratios=None):
prune_params_ratios (dict): 由裁剪参数名和裁剪率组成的字典，当为None时
使用默认裁剪参数名和裁剪率。默认为None。
"""
assert model.status == 'Normal', 'Only the models saved while training are supported!'
place = model.places[0]
train_prog = model.train_prog
eval_prog = model.test_prog
......@@ -235,6 +236,7 @@ def cal_params_sensitivities(model, save_file, eval_dataset, batch_size=8):
其中``weight_0``是卷积Kernel名；``sensitivities['weight_0']``是一个字典，key是裁剪率，value是敏感度。
"""
assert model.status == 'Normal', 'Only the models saved while training are supported!'
if os.path.exists(save_file):
os.remove(save_file)
......
paddlex/cv/models/slim/prune_config.py
浏览文件 @ 4c631565
......@@ -19,6 +19,8 @@ import paddle.fluid as fluid
import paddlex
sensitivities_data = {
'AlexNet':
'https://bj.bcebos.com/paddlex/slim_prune/alexnet_sensitivities.data',
'ResNet18':
'https://bj.bcebos.com/paddlex/slim_prune/resnet18.sensitivities',
'ResNet34':
......@@ -41,6 +43,10 @@ sensitivities_data = {
'https://bj.bcebos.com/paddlex/slim_prune/mobilenetv3_large.sensitivities',
'MobileNetV3_small':
'https://bj.bcebos.com/paddlex/slim_prune/mobilenetv3_small.sensitivities',
'MobileNetV3_large_ssld':
'https://bj.bcebos.com/paddlex/slim_prune/mobilenetv3_large_ssld_sensitivities.data',
'MobileNetV3_small_ssld':
'https://bj.bcebos.com/paddlex/slim_prune/mobilenetv3_small_ssld_sensitivities.data',
'DenseNet121':
'https://bj.bcebos.com/paddlex/slim_prune/densenet121.sensitivities',
'DenseNet161':
......@@ -51,6 +57,8 @@ sensitivities_data = {
'https://bj.bcebos.com/paddlex/slim_prune/xception41.sensitivities',
'Xception65':
'https://bj.bcebos.com/paddlex/slim_prune/xception65.sensitivities',
'ShuffleNetV2':
'https://bj.bcebos.com/paddlex/slim_prune/shufflenetv2_sensitivities.data',
'YOLOv3_MobileNetV1':
'https://bj.bcebos.com/paddlex/slim_prune/yolov3_mobilenetv1.sensitivities',
'YOLOv3_MobileNetV3_large':
......@@ -143,7 +151,8 @@ def get_prune_params(model):
if model_type.startswith('ResNet') or \
model_type.startswith('DenseNet') or \
model_type.startswith('DarkNet') or \
model_type.startswith('AlexNet'):
model_type.startswith('AlexNet') or \
model_type.startswith('ShuffleNetV2'):
for block in program.blocks:
for param in block.all_parameters():
pd_var = fluid.global_scope().find_var(param.name)
......@@ -152,6 +161,28 @@ def get_prune_params(model):
prune_names.append(param.name)
if model_type == 'AlexNet':
prune_names.remove('conv5_weights')
if model_type == 'ShuffleNetV2':
not_prune_names = ['stage_2_1_conv5_weights',
'stage_2_1_conv3_weights',
'stage_2_2_conv3_weights',
'stage_2_3_conv3_weights',
'stage_2_4_conv3_weights',
'stage_3_1_conv5_weights',
'stage_3_1_conv3_weights',
'stage_3_2_conv3_weights',
'stage_3_3_conv3_weights',
'stage_3_4_conv3_weights',
'stage_3_5_conv3_weights',
'stage_3_6_conv3_weights',
'stage_3_7_conv3_weights',
'stage_3_8_conv3_weights',
'stage_4_1_conv5_weights',
'stage_4_1_conv3_weights',
'stage_4_2_conv3_weights',
'stage_4_3_conv3_weights',
'stage_4_4_conv3_weights',]
for name in not_prune_names:
prune_names.remove(name)
elif model_type == "MobileNetV1":
prune_names.append("conv1_weights")
for param in program.global_block().all_parameters():
......
paddlex/cv/models/utils/pretrain_weights.py
浏览文件 @ 4c631565
......@@ -81,7 +81,7 @@ coco_pretrain = {
'YOLOv3_MobileNetV1_COCO':
'https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar',
'YOLOv3_MobileNetV3_large_COCO':
'https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v3.pdparams',
'https://bj.bcebos.com/paddlex/models/yolov3_mobilenet_v3.tar',
'YOLOv3_ResNet34_COCO':
'https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar',
'YOLOv3_ResNet50_vd_COCO':
......
paddlex/cv/nets/hrnet.py
浏览文件 @ 4c631565
......@@ -51,15 +51,38 @@ class HRNet(object):
self.width = width
self.has_se = has_se
self.num_modules = {
'18_small_v1': [1, 1, 1, 1],
'18': [1, 1, 4, 3],
'30': [1, 1, 4, 3],
'32': [1, 1, 4, 3],
'40': [1, 1, 4, 3],
'44': [1, 1, 4, 3],
'48': [1, 1, 4, 3],
'60': [1, 1, 4, 3],
'64': [1, 1, 4, 3]
}
self.num_blocks = {
'18_small_v1': [[1], [2, 2], [2, 2, 2], [2, 2, 2, 2]],
'18': [[4], [4, 4], [4, 4, 4], [4, 4, 4, 4]],
'30': [[4], [4, 4], [4, 4, 4], [4, 4, 4, 4]],
'32': [[4], [4, 4], [4, 4, 4], [4, 4, 4, 4]],
'40': [[4], [4, 4], [4, 4, 4], [4, 4, 4, 4]],
'44': [[4], [4, 4], [4, 4, 4], [4, 4, 4, 4]],
'48': [[4], [4, 4], [4, 4, 4], [4, 4, 4, 4]],
'60': [[4], [4, 4], [4, 4, 4], [4, 4, 4, 4]],
'64': [[4], [4, 4], [4, 4, 4], [4, 4, 4, 4]]
}
self.channels = {
18: [[18, 36], [18, 36, 72], [18, 36, 72, 144]],
30: [[30, 60], [30, 60, 120], [30, 60, 120, 240]],
32: [[32, 64], [32, 64, 128], [32, 64, 128, 256]],
40: [[40, 80], [40, 80, 160], [40, 80, 160, 320]],
44: [[44, 88], [44, 88, 176], [44, 88, 176, 352]],
48: [[48, 96], [48, 96, 192], [48, 96, 192, 384]],
60: [[60, 120], [60, 120, 240], [60, 120, 240, 480]],
64: [[64, 128], [64, 128, 256], [64, 128, 256, 512]],
'18_small_v1': [[32], [16, 32], [16, 32, 64], [16, 32, 64, 128]],
'18': [[64], [18, 36], [18, 36, 72], [18, 36, 72, 144]],
'30': [[64], [30, 60], [30, 60, 120], [30, 60, 120, 240]],
'32': [[64], [32, 64], [32, 64, 128], [32, 64, 128, 256]],
'40': [[64], [40, 80], [40, 80, 160], [40, 80, 160, 320]],
'44': [[64], [44, 88], [44, 88, 176], [44, 88, 176, 352]],
'48': [[64], [48, 96], [48, 96, 192], [48, 96, 192, 384]],
'60': [[64], [60, 120], [60, 120, 240], [60, 120, 240, 480]],
'64': [[64], [64, 128], [64, 128, 256], [64, 128, 256, 512]],
}
self.freeze_at = freeze_at
......@@ -73,31 +96,38 @@ class HRNet(object):
def net(self, input):
width = self.width
channels_2, channels_3, channels_4 = self.channels[width]
num_modules_2, num_modules_3, num_modules_4 = 1, 4, 3
channels_1, channels_2, channels_3, channels_4 = self.channels[str(
width)]
num_modules_1, num_modules_2, num_modules_3, num_modules_4 = self.num_modules[
str(width)]
num_blocks_1, num_blocks_2, num_blocks_3, num_blocks_4 = self.num_blocks[
str(width)]
x = self.conv_bn_layer(
input=input,
filter_size=3,
num_filters=64,
num_filters=channels_1[0],
stride=2,
if_act=True,
name='layer1_1')
x = self.conv_bn_layer(
input=x,
filter_size=3,
num_filters=64,
num_filters=channels_1[0],
stride=2,
if_act=True,
name='layer1_2')
la1 = self.layer1(x, name='layer2')
la1 = self.layer1(x, num_blocks_1, channels_1, name='layer2')
tr1 = self.transition_layer([la1], [256], channels_2, name='tr1')
st2 = self.stage(tr1, num_modules_2, channels_2, name='st2')
st2 = self.stage(
tr1, num_modules_2, num_blocks_2, channels_2, name='st2')
tr2 = self.transition_layer(st2, channels_2, channels_3, name='tr2')
st3 = self.stage(tr2, num_modules_3, channels_3, name='st3')
st3 = self.stage(
tr2, num_modules_3, num_blocks_3, channels_3, name='st3')
tr3 = self.transition_layer(st3, channels_3, channels_4, name='tr3')
st4 = self.stage(tr3, num_modules_4, channels_4, name='st4')
st4 = self.stage(
tr3, num_modules_4, num_blocks_4, channels_4, name='st4')
# classification
if self.num_classes:
......@@ -139,12 +169,12 @@ class HRNet(object):
self.end_points = st4
return st4[-1]
def layer1(self, input, name=None):
def layer1(self, input, num_blocks, channels, name=None):
conv = input
for i in range(4):
for i in range(num_blocks[0]):
conv = self.bottleneck_block(
conv,
num_filters=64,
num_filters=channels[0],
downsample=True if i == 0 else False,
name=name + '_' + str(i + 1))
return conv
......@@ -178,7 +208,7 @@ class HRNet(object):
out = []
for i in range(len(channels)):
residual = x[i]
for j in range(block_num):
for j in range(block_num[i]):
residual = self.basic_block(
residual,
channels[i],
......@@ -240,10 +270,11 @@ class HRNet(object):
def high_resolution_module(self,
x,
num_blocks,
channels,
multi_scale_output=True,
name=None):
residual = self.branches(x, 4, channels, name=name)
residual = self.branches(x, num_blocks, channels, name=name)
out = self.fuse_layers(
residual,
channels,
......@@ -254,6 +285,7 @@ class HRNet(object):
def stage(self,
x,
num_modules,
num_blocks,
channels,
multi_scale_output=True,
name=None):
......@@ -262,12 +294,13 @@ class HRNet(object):
if i == num_modules - 1 and multi_scale_output == False:
out = self.high_resolution_module(
out,
num_blocks,
channels,
multi_scale_output=False,
name=name + '_' + str(i + 1))
else:
out = self.high_resolution_module(
out, channels, name=name + '_' + str(i + 1))
out, num_blocks, channels, name=name + '_' + str(i + 1))
return out
......
paddlex/cv/nets/segmentation/hrnet.py
浏览文件 @ 4c631565
......@@ -82,7 +82,8 @@ class HRNet(object):
st4[3] = fluid.layers.resize_bilinear(st4[3], out_shape=shape)
out = fluid.layers.concat(st4, axis=1)
last_channels = sum(self.backbone.channels[self.backbone.width][-1])
last_channels = sum(self.backbone.channels[str(self.backbone.width)][
-1])
out = self._conv_bn_layer(
input=out,
......
paddlex/cv/transforms/cls_transforms.py
浏览文件 @ 4c631565
......@@ -70,8 +70,8 @@ class Compose(ClsTransform):
if isinstance(im, np.ndarray):
if len(im.shape) != 3:
raise Exception(
"im should be 3-dimension, but now is {}-dimensions".
format(len(im.shape)))
"im should be 3-dimension, but now is {}-dimensions".format(
len(im.shape)))
else:
try:
im = cv2.imread(im).astype('float32')
......@@ -100,7 +100,9 @@ class Compose(ClsTransform):
transform_names = [type(x).__name__ for x in self.transforms]
for aug in augmenters:
if type(aug).__name__ in transform_names:
logging.error("{} is already in ComposedTransforms, need to remove it from add_augmenters().".format(type(aug).__name__))
logging.error(
"{} is already in ComposedTransforms, need to remove it from add_augmenters().".
format(type(aug).__name__))
self.transforms = augmenters + self.transforms
......@@ -139,8 +141,8 @@ class RandomCrop(ClsTransform):
tuple: 当label为空时，返回的tuple为(im, )，对应图像np.ndarray数据；
当label不为空时，返回的tuple为(im, label)，分别对应图像np.ndarray数据、图像类别id。
"""
im = random_crop(im, self.crop_size, self.lower_scale,
self.lower_ratio, self.upper_ratio)
im = random_crop(im, self.crop_size, self.lower_scale, self.lower_ratio,
self.upper_ratio)
if label is None:
return (im, )
else:
......@@ -270,14 +272,12 @@ class ResizeByShort(ClsTransform):
im_short_size = min(im.shape[0], im.shape[1])
im_long_size = max(im.shape[0], im.shape[1])
scale = float(self.short_size) / im_short_size
if self.max_size > 0 and np.round(scale *
im_long_size) > self.max_size:
if self.max_size > 0 and np.round(scale * im_long_size) > self.max_size:
scale = float(self.max_size) / float(im_long_size)
resized_width = int(round(im.shape[1] * scale))
resized_height = int(round(im.shape[0] * scale))
im = cv2.resize(
im, (resized_width, resized_height),
interpolation=cv2.INTER_LINEAR)
im, (resized_width, resized_height), interpolation=cv2.INTER_LINEAR)
if label is None:
return (im, )
......@@ -490,13 +490,15 @@ class ComposedClsTransforms(Compose):
crop_size(int|list): 输入模型里的图像大小
mean(list): 图像均值
std(list): 图像方差
random_horizontal_flip(bool): 是否以0.5的概率使用随机水平翻转增强，该仅在mode为`train`时生效，默认为True
"""
def __init__(self,
mode,
crop_size=[224, 224],
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]):
std=[0.229, 0.224, 0.225],
random_horizontal_flip=True):
width = crop_size
if isinstance(crop_size, list):
if crop_size[0] != crop_size[1]:
......@@ -512,10 +514,11 @@ class ComposedClsTransforms(Compose):
if mode == 'train':
# 训练时的transforms，包含数据增强
transforms = [
RandomCrop(crop_size=width), RandomHorizontalFlip(prob=0.5),
Normalize(
RandomCrop(crop_size=width), Normalize(
mean=mean, std=std)
]
if random_horizontal_flip:
transforms.insert(0, RandomHorizontalFlip())
else:
# 验证/预测时的transforms
transforms = [
......
paddlex/cv/transforms/det_transforms.py
浏览文件 @ 4c631565
......@@ -160,7 +160,9 @@ class Compose(DetTransform):
transform_names = [type(x).__name__ for x in self.transforms]
for aug in augmenters:
if type(aug).__name__ in transform_names:
logging.error("{} is already in ComposedTransforms, need to remove it from add_augmenters().".format(type(aug).__name__))
logging.error(
"{} is already in ComposedTransforms, need to remove it from add_augmenters().".
format(type(aug).__name__))
self.transforms = augmenters + self.transforms
......@@ -220,15 +222,13 @@ class ResizeByShort(DetTransform):
im_short_size = min(im.shape[0], im.shape[1])
im_long_size = max(im.shape[0], im.shape[1])
scale = float(self.short_size) / im_short_size
if self.max_size > 0 and np.round(scale *
im_long_size) > self.max_size:
if self.max_size > 0 and np.round(scale * im_long_size) > self.max_size:
scale = float(self.max_size) / float(im_long_size)
resized_width = int(round(im.shape[1] * scale))
resized_height = int(round(im.shape[0] * scale))
im_resize_info = [resized_height, resized_width, scale]
im = cv2.resize(
im, (resized_width, resized_height),
interpolation=cv2.INTER_LINEAR)
im, (resized_width, resized_height), interpolation=cv2.INTER_LINEAR)
im_info['im_resize_info'] = np.array(im_resize_info).astype(np.float32)
if label_info is None:
return (im, im_info)
......@@ -268,8 +268,7 @@ class Padding(DetTransform):
if not isinstance(target_size, tuple) and not isinstance(
target_size, list):
raise TypeError(
"Padding: Type of target_size must in (int|list|tuple)."
)
"Padding: Type of target_size must in (int|list|tuple).")
elif len(target_size) != 2:
raise ValueError(
"Padding: Length of target_size must equal 2.")
......@@ -454,8 +453,7 @@ class RandomHorizontalFlip(DetTransform):
ValueError: 数据长度不匹配。
"""
if not isinstance(im, np.ndarray):
raise TypeError(
"RandomHorizontalFlip: image is not a numpy array.")
raise TypeError("RandomHorizontalFlip: image is not a numpy array.")
if len(im.shape) != 3:
raise ValueError(
"RandomHorizontalFlip: image is not 3-dimensional.")
......@@ -785,9 +783,7 @@ class RandomExpand(DetTransform):
fill_value (list): 扩张图像的初始填充值（0-255）。默认为[123.675, 116.28, 103.53]。
"""
def __init__(self,
ratio=4.,
prob=0.5,
def __init__(self, ratio=4., prob=0.5,
fill_value=[123.675, 116.28, 103.53]):
super(RandomExpand, self).__init__()
assert ratio > 1.01, "expand ratio must be larger than 1.01"
......@@ -1281,21 +1277,25 @@ class ComposedRCNNTransforms(Compose):
min_max_size(list): 图像在缩放时，最小边和最大边的约束条件
mean(list): 图像均值
std(list): 图像方差
random_horizontal_flip(bool): 是否以0.5的概率使用随机水平翻转增强，该仅在mode为`train`时生效，默认为True
"""
def __init__(self,
mode,
min_max_size=[800, 1333],
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]):
std=[0.229, 0.224, 0.225],
random_horizontal_flip=True):
if mode == 'train':
# 训练时的transforms，包含数据增强
transforms = [
RandomHorizontalFlip(prob=0.5), Normalize(
Normalize(
mean=mean, std=std), ResizeByShort(
short_size=min_max_size[0], max_size=min_max_size[1]),
Padding(coarsest_stride=32)
]
if random_horizontal_flip:
transforms.insert(0, RandomHorizontalFlip())
else:
# 验证/预测时的transforms
transforms = [
......@@ -1325,9 +1325,14 @@ class ComposedYOLOv3Transforms(Compose):
Args:
mode(str): 图像处理流程所处阶段，训练/验证/预测，分别对应'train', 'eval', 'test'
shape(list): 输入模型中图像的大小，输入模型的图像会被Resize成此大小
mixup_epoch(int): 模型训练过程中，前mixup_epoch会使用mixup策略
mixup_epoch(int): 模型训练过程中，前mixup_epoch会使用mixup策略, 若设为-1，则表示不使用该策略
mean(list): 图像均值
std(list): 图像方差
random_distort(bool): 数据增强方式，参数仅在mode为`train`时生效，表示是否在训练过程中随机扰动图像，默认为True
random_expand(bool): 数据增强方式，参数仅在mode为`train`时生效，表示是否在训练过程中随机扩张图像，默认为True
random_crop(bool): 数据增强方式，参数仅在mode为`train`时生效，表示是否在训练过程中随机裁剪图像，默认为True
random_horizontal_flip(bool): 数据增强方式，参数仅在mode为`train`时生效，表示是否在训练过程中随机水平翻转图像，默认为True
"""
def __init__(self,
......@@ -1335,7 +1340,11 @@ class ComposedYOLOv3Transforms(Compose):
shape=[608, 608],
mixup_epoch=250,
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]):
std=[0.229, 0.224, 0.225],
random_distort=True,
random_expand=True,
random_crop=True,
random_horizontal_flip=True):
width = shape
if isinstance(shape, list):
if shape[0] != shape[1]:
......@@ -1350,12 +1359,18 @@ class ComposedYOLOv3Transforms(Compose):
if mode == 'train':
# 训练时的transforms，包含数据增强
transforms = [
MixupImage(mixup_epoch=mixup_epoch), RandomDistort(),
RandomExpand(), RandomCrop(), Resize(
target_size=width,
interp='RANDOM'), RandomHorizontalFlip(), Normalize(
MixupImage(mixup_epoch=mixup_epoch), Resize(
target_size=width, interp='RANDOM'), Normalize(
mean=mean, std=std)
]
if random_horizontal_flip:
transforms.insert(1, RandomHorizontalFlip())
if random_crop:
transforms.insert(1, RandomCrop())
if random_expand:
transforms.insert(1, RandomExpand())
if random_distort:
transforms.insert(1, RandomDistort())
else:
# 验证/预测时的transforms
transforms = [
......
paddlex/cv/transforms/seg_transforms.py
浏览文件 @ 4c631565
......@@ -116,7 +116,9 @@ class Compose(SegTransform):
transform_names = [type(x).__name__ for x in self.transforms]
for aug in augmenters:
if type(aug).__name__ in transform_names:
logging.error("{} is already in ComposedTransforms, need to remove it from add_augmenters().".format(type(aug).__name__))
logging.error(
"{} is already in ComposedTransforms, need to remove it from add_augmenters().".
format(type(aug).__name__))
self.transforms = augmenters + self.transforms
......@@ -401,8 +403,7 @@ class ResizeByShort(SegTransform):
im_short_size = min(im.shape[0], im.shape[1])
im_long_size = max(im.shape[0], im.shape[1])
scale = float(self.short_size) / im_short_size
if self.max_size > 0 and np.round(scale *
im_long_size) > self.max_size:
if self.max_size > 0 and np.round(scale * im_long_size) > self.max_size:
scale = float(self.max_size) / float(im_long_size)
resized_width = int(round(im.shape[1] * scale))
resized_height = int(round(im.shape[0] * scale))
......@@ -1113,25 +1114,35 @@ class ComposedSegTransforms(Compose):
Args:
mode(str): 图像处理所处阶段，训练/验证/预测，分别对应'train', 'eval', 'test'
train_crop_size(list): 模型训练阶段，随机从原图crop的大小
min_max_size(list): 训练过程中，图像的最长边会随机resize至此区间（短边按比例相应resize)；预测阶段，图像最长边会resize至此区间中间值，即(min_size+max_size)/2。默认为[400, 600]
train_crop_size(list): 仅在mode为'train`时生效，训练过程中，随机从图像中裁剪出对应大小的子图（如若原图小于此大小，则会padding到此大小)，默认为[400, 600]
mean(list): 图像均值
std(list): 图像方差
random_horizontal_flip(bool): 数据增强方式，仅在mode为`train`时生效，表示训练过程是否随机水平翻转图像，默认为True
"""
def __init__(self,
mode,
train_crop_size=[769, 769],
min_max_size=[400, 600],
train_crop_size=[512, 512],
mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5]):
std=[0.5, 0.5, 0.5],
random_horizontal_flip=True):
if mode == 'train':
# 训练时的transforms，包含数据增强
transforms = [
RandomHorizontalFlip(prob=0.5), ResizeStepScaling(),
ResizeRangeScaling(
min_value=min(min_max_size), max_value=max(min_max_size)),
RandomPaddingCrop(crop_size=train_crop_size), Normalize(
mean=mean, std=std)
]
if random_horizontal_flip:
transforms.insert(0, RandomHorizontalFlip())
else:
# 验证/预测时的transforms
transforms = [Normalize(mean=mean, std=std)]
long_size = (min(min_max_size) + max(min_max_size)) // 2
transforms = [
ResizeByLong(long_size=long_size), Normalize(
mean=mean, std=std)
]
super(ComposedSegTransforms, self).__init__(transforms)
paddlex/interpret/visualize.py
浏览文件 @ 4c631565
......@@ -70,8 +70,10 @@ def normlime(img_file,
normlime_weights_file=None):
"""使用NormLIME算法将模型预测结果的可解释性可视化。
NormLIME是利用一定数量的样本来出一个全局的解释。NormLIME会提前计算一定数量的测
试样本的LIME结果，然后对相同的特征进行权重的归一化，这样来得到一个全局的输入和输出的关系。
NormLIME是利用一定数量的样本来出一个全局的解释。由于NormLIME计算量较大，此处采用一种简化的方式：
使用一定数量的测试样本（目前默认使用所有测试样本），对每个样本进行特征提取，映射到同一个特征空间；
然后以此特征做为输入，以模型输出做为输出，使用线性回归对其进行拟合，得到一个全局的输入和输出的关系。
之后，对一测试样本进行解释时，使用NormLIME全局的解释，来对LIME的结果进行滤波，使最终的可视化结果更加稳定。
注意1：dataset读取的是一个数据集，该数据集不宜过大，否则计算时间会较长，但应包含所有类别的数据。
注意2：NormLIME可解释性结果可视化目前只支持分类模型。
......
setup.py
浏览文件 @ 4c631565
......@@ -15,11 +15,11 @@
import setuptools
import sys
long_description = "PaddleX. A end-to-end deeplearning model development toolkit base on PaddlePaddle\n\n"
long_description = "PaddlePaddle Entire Process Development Toolkit"
setuptools.setup(
name="paddlex",
version='1.0.6',
version='1.0.7',
author="paddlex",
author_email="paddlex@baidu.com",
description=long_description,
......
new_tutorials/train/segmentation/fast_scnn.py → tutorials/train/segmentation/fast_scnn.py
浏览文件 @ 4c631565
......@@ -35,7 +35,7 @@ eval_dataset = pdx.datasets.SegDataset(
# 浏览器打开 https://0.0.0.0:8001即可
# 其中0.0.0.0为本机访问，如为远程服务, 改成相应机器IP
# https://paddlex.readthedocs.io/zh_CN/latest/apis/models/semantic_segmentation.html#hrnet
# https://paddlex.readthedocs.io/zh_CN/latest/apis/models/semantic_segmentation.html#fastscnn
num_classes = len(train_dataset.labels)
model = pdx.seg.FastSCNN(num_classes=num_classes)
model.train(
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或