Merge pull request #9 from PaddlePaddle/develop

00

deploy/cpp/demo/detector.cpp

@@ -66,7 +66,7 @@ int main(int argc, char** argv) {
         std::cout << "image file: " << image_path
                   << ", predict label: " << result.boxes[i].category
                   << ", label_id:" << result.boxes[i].category_id
-                  << ", score: " << result.boxes[i].score << ", box:("
+                  << ", score: " << result.boxes[i].score << ", box(xmin, ymin, w, h):("
                   << result.boxes[i].coordinate[0] << ", "
                   << result.boxes[i].coordinate[1] << ", "
                   << result.boxes[i].coordinate[2] << ", "
@@ -89,7 +89,7 @@ int main(int argc, char** argv) {
     for (int i = 0; i < result.boxes.size(); ++i) {
       std::cout << ", predict label: " << result.boxes[i].category
                 << ", label_id:" << result.boxes[i].category_id
-                << ", score: " << result.boxes[i].score << ", box:("
+                << ", score: " << result.boxes[i].score << ", box(xmin, ymin, w, h):("
                 << result.boxes[i].coordinate[0] << ", "
                 << result.boxes[i].coordinate[1] << ", "
                 << result.boxes[i].coordinate[2] << ", "

docs/apis/models/classification.md

@@ -35,7 +35,7 @@ train(self, num_epochs, train_dataset, train_batch_size=64, eval_dataset=None, s
 > > - **use_vdl** (bool): 是否使用VisualDL进行可视化。默认值为False。
 > > - **sensitivities_file** (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
 > > - **eval_metric_loss** (float): 可容忍的精度损失。默认为0.05。
-> > - **early_stop** (float): 是否使用提前终止训练策略。默认值为False。
+> > - **early_stop** (bool): 是否使用提前终止训练策略。默认值为False。
 > > - **early_stop_patience** (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内连续下降或持平，则终止训练。默认值为5。
 > > - **resume_checkpoint** (str): 恢复训练时指定上次训练保存的模型路径。若为None，则不会恢复训练。默认值为None。
 
@@ -186,3 +186,7 @@ paddlex.cls.DenseNet161(num_classes=1000)
 paddlex.cls.DenseNet201(num_classes=1000)
 ```
 
+### HRNet_W18
+```python
+paddlex.cls.HRNet_W18(num_classes=1000)
+```

docs/apis/models/detection.md

@@ -53,7 +53,7 @@ train(self, num_epochs, train_dataset, train_batch_size=8, eval_dataset=None, sa
 > > - **use_vdl** (bool): 是否使用VisualDL进行可视化。默认值为False。
 > > - **sensitivities_file** (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，则自动下载在PascalVOC数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
 > > - **eval_metric_loss** (float): 可容忍的精度损失。默认为0.05。
-> > - **early_stop** (float): 是否使用提前终止训练策略。默认值为False。
+> > - **early_stop** (bool): 是否使用提前终止训练策略。默认值为False。
 > > - **early_stop_patience** (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内连续下降或持平，则终止训练。默认值为5。
 > > - **resume_checkpoint** (str): 恢复训练时指定上次训练保存的模型路径。若为None，则不会恢复训练。默认值为None。
 
@@ -107,7 +107,7 @@ paddlex.det.FasterRCNN(num_classes=81, backbone='ResNet50', with_fpn=True, aspec
 > **参数**
 
 > > - **num_classes** (int): 包含了背景类的类别数。默认为81。
-> > - **backbone** (str): FasterRCNN的backbone网络，取值范围为['ResNet18', 'ResNet50', 'ResNet50_vd', 'ResNet101', 'ResNet101_vd']。默认为'ResNet50'。
+> > - **backbone** (str): FasterRCNN的backbone网络，取值范围为['ResNet18', 'ResNet50', 'ResNet50_vd', 'ResNet101', 'ResNet101_vd', 'HRNet_W18']。默认为'ResNet50'。
 > > - **with_fpn** (bool): 是否使用FPN结构。默认为True。
 > > - **aspect_ratios** (list): 生成anchor高宽比的可选值。默认为[0.5, 1.0, 2.0]。
 > > - **anchor_sizes** (list): 生成anchor大小的可选值。默认为[32, 64, 128, 256, 512]。

docs/apis/models/instance_segmentation.md

@@ -12,7 +12,7 @@ paddlex.det.MaskRCNN(num_classes=81, backbone='ResNet50', with_fpn=True, aspect_
 > **参数**
 
 > > - **num_classes** (int): 包含了背景类的类别数。默认为81。
-> > - **backbone** (str): MaskRCNN的backbone网络，取值范围为['ResNet18', 'ResNet50', 'ResNet50_vd', 'ResNet101', 'ResNet101_vd']。默认为'ResNet50'。
+> > - **backbone** (str): MaskRCNN的backbone网络，取值范围为['ResNet18', 'ResNet50', 'ResNet50_vd', 'ResNet101', 'ResNet101_vd', 'HRNet_W18']。默认为'ResNet50'。
 > > - **with_fpn** (bool): 是否使用FPN结构。默认为True。
 > > - **aspect_ratios** (list): 生成anchor高宽比的可选值。默认为[0.5, 1.0, 2.0]。
 > > - **anchor_sizes** (list): 生成anchor大小的可选值。默认为[32, 64, 128, 256, 512]。
@@ -82,4 +82,4 @@ predict(self, img_file, transforms=None)
 >
 > **返回值**
 >
-> > - **list**: 预测结果列表，列表中每个元素均为一个dict，key'bbox', 'mask', 'category', 'category_id', 'score'，分别表示每个预测目标的框坐标信息、Mask信息，类别、类别id、置信度，其中框坐标信息为[xmin, ymin, w, h]，即左上角x, y坐标和框的宽和高。
+> > - **list**: 预测结果列表，列表中每个元素均为一个dict，key'bbox', 'mask', 'category', 'category_id', 'score'，分别表示每个预测目标的框坐标信息、Mask信息，类别、类别id、置信度。其中框坐标信息为[xmin, ymin, w, h]，即左上角x, y坐标和框的宽和高。Mask信息为原图大小的二值图，1表示像素点属于预测类别，0表示像素点是背景。

docs/apis/models/semantic_segmentation.md

@@ -47,7 +47,7 @@ train(self, num_epochs, train_dataset, train_batch_size=2, eval_dataset=None, ev
 > > - **use_vdl** (bool): 是否使用VisualDL进行可视化。默认False。
 > > - **sensitivities_file** (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
 > > - **eval_metric_loss** (float): 可容忍的精度损失。默认为0.05。
-> > - **early_stop** (float): 是否使用提前终止训练策略。默认值为False。
+> > - **early_stop** (bool): 是否使用提前终止训练策略。默认值为False。
 > > - **early_stop_patience** (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内连续下降或持平，则终止训练。默认值为5。
 > > - **resume_checkpoint** (str): 恢复训练时指定上次训练保存的模型路径。若为None，则不会恢复训练。默认值为None。
 
@@ -124,7 +124,7 @@ train(self, num_epochs, train_dataset, train_batch_size=2, eval_dataset=None, ev
 > > - **save_interval_epochs** (int): 模型保存间隔（单位：迭代轮数）。默认为1。
 > > - **log_interval_steps** (int): 训练日志输出间隔（单位：迭代次数）。默认为2。
 > > - **save_dir** (str): 模型保存路径。默认'output'
-> > - **pretrain_weights** (str): 若指定为路径时，则加载路径下预训练模型；若为字符串'IMAGENET'，则自动下载在COCO图片数据上预训练的模型权重；若为None，则不使用预训练模型。默认'COCO'。
+> > - **pretrain_weights** (str): 若指定为路径时，则加载路径下预训练模型；若为字符串'COCO'，则自动下载在COCO图片数据上预训练的模型权重；若为None，则不使用预训练模型。默认'COCO'。
 > > - **optimizer** (paddle.fluid.optimizer): 优化器。当该参数为None时，使用默认的优化器：使用fluid.optimizer.Momentum优化方法，polynomial的学习率衰减策略。
 > > - **learning_rate** (float): 默认优化器的初始学习率。默认0.01。
 > > - **lr_decay_power** (float): 默认优化器学习率衰减指数。默认0.9。
@@ -173,3 +173,88 @@ predict(self, im_file, transforms=None):
 > **返回值**
 > >
 > > - **dict**: 包含关键字'label_map'和'score_map', 'label_map'存储预测结果灰度图，像素值表示对应的类别，'score_map'存储各类别的概率，shape=(h, w, num_classes)。
+
+
+## HRNet类
+
+```python
+paddlex.seg.HRNet(num_classes=2, width=18, use_bce_loss=False, use_dice_loss=False, class_weight=None, ignore_index=255)
+```
+
+> 构建HRNet分割器。
+
+> **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]。
+> > - **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。
+> > - **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
+
+```python
+train(self, num_epochs, train_dataset, train_batch_size=2, eval_dataset=None, eval_batch_size=1, save_interval_epochs=1, log_interval_steps=2, save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=0.01, lr_decay_power=0.9, use_vdl=False, sensitivities_file=None, eval_metric_loss=0.05, early_stop=False, early_stop_patience=5, resume_checkpoint=None):
+```
+
+> HRNet模型训练接口。
+
+> **参数**
+> >
+> > - **num_epochs** (int): 训练迭代轮数。
+> > - **train_dataset** (paddlex.datasets): 训练数据读取器。
+> > - **train_batch_size** (int): 训练数据batch大小。同时作为验证数据batch大小。默认2。
+> > - **eval_dataset** (paddlex.datasets): 评估数据读取器。
+> > - **save_interval_epochs** (int): 模型保存间隔（单位：迭代轮数）。默认为1。
+> > - **log_interval_steps** (int): 训练日志输出间隔（单位：迭代次数）。默认为2。
+> > - **save_dir** (str): 模型保存路径。默认'output'
+> > - **pretrain_weights** (str): 若指定为路径时，则加载路径下预训练模型；若为字符串'IMAGENET'，则自动下载在ImageNet数据集上预训练的模型权重；若为None，则不使用预训练模型。默认'IMAGENET'。
+> > - **optimizer** (paddle.fluid.optimizer): 优化器。当该参数为None时，使用默认的优化器：使用fluid.optimizer.Momentum优化方法，polynomial的学习率衰减策略。
+> > - **learning_rate** (float): 默认优化器的初始学习率。默认0.01。
+> > - **lr_decay_power** (float): 默认优化器学习率衰减指数。默认0.9。
+> > - **use_vdl** (bool): 是否使用VisualDL进行可视化。默认False。
+> > - **sensitivities_file** (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
+> > - **eval_metric_loss** (float): 可容忍的精度损失。默认为0.05。
+> > - **early_stop** (float): 是否使用提前终止训练策略。默认值为False。
+> > - **early_stop_patience** (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内连续下降或持平，则终止训练。默认值为5。
+> > - **resume_checkpoint** (str): 恢复训练时指定上次训练保存的模型路径。若为None，则不会恢复训练。默认值为None。
+
+#### evaluate 评估接口
+
+```
+evaluate(self, eval_dataset, batch_size=1, epoch_id=None, return_details=False):
+```
+
+> HRNet模型评估接口。
+
+> **参数**
+> >
+> > - **eval_dataset** (paddlex.datasets): 评估数据读取器。
+> > - **batch_size** (int): 评估时的batch大小。默认1。
+> > - **epoch_id** (int): 当前评估模型所在的训练轮数。
+> > - **return_details** (bool): 是否返回详细信息。默认False。
+
+> **返回值**
+> >
+> > - **dict**: 当return_details为False时，返回dict。包含关键字：'miou'、'category_iou'、'macc'、
+> >   'category_acc'和'kappa'，分别表示平均iou、各类别iou、平均准确率、各类别准确率和kappa系数。
+> > - **tuple** (metrics, eval_details)：当return_details为True时，增加返回dict (eval_details)，
+> >   包含关键字：'confusion_matrix'，表示评估的混淆矩阵。
+
+#### predict 预测接口
+
+```
+predict(self, im_file, transforms=None):
+```
+
+> HRNet模型预测接口。需要注意的是，只有在训练过程中定义了eval_dataset，模型在保存时才会将预测时的图像处理流程保存在`UNet.test_transforms`和`UNet.eval_transforms`中。如未在训练时定义eval_dataset，那在调用预测`predict`接口时，用户需要再重新定义test_transforms传入给`predict`接口。
+
+> **参数**
+> >
+> > - **img_file** (str): 预测图像路径。
+> > - **transforms** (paddlex.seg.transforms): 数据预处理操作。
+
+> **返回值**
+> >
+> > - **dict**: 包含关键字'label_map'和'score_map', 'label_map'存储预测结果灰度图，像素值表示对应的类别，'score_map'存储各类别的概率，shape=(h, w, num_classes)。

docs/apis/transforms/augment.md

@@ -10,7 +10,7 @@ PaddleX对于图像分类、目标检测、实例分割和语义分割内置了
 | :------- | :------------|
 | 图像分类 | [RandomCrop](cls_transforms.html#randomcrop)、[RandomHorizontalFlip](cls_transforms.html#randomhorizontalflip)、[RandomVerticalFlip](cls_transforms.html#randomverticalflip)、 <br> [RandomRotate](cls_transforms.html#randomratate)、 [RandomDistort](cls_transforms.html#randomdistort) |
 |目标检测<br>实例分割| [RandomHorizontalFlip](det_transforms.html#randomhorizontalflip)、[RandomDistort](det_transforms.html#randomdistort)、[RandomCrop](det_transforms.html#randomcrop)、<br> [MixupImage](det_transforms.html#mixupimage)(仅支持YOLOv3模型)、[RandomExpand](det_transforms.html#randomexpand) |
-|语义分割  | [RandomHorizontalFlip](seg_transforms.html#randomhorizontalflip)、[RandomVerticalFlip](seg_transforms.html#randomverticalflip)、[RandomRangeScaling](seg_transforms.html#randomrangescaling)、<br> [RandomStepScaling](seg_transforms.html#randomstepscaling)、[RandomPaddingCrop](seg_transforms.html#randompaddingcrop)、 [RandomBlur](seg_transforms.html#randomblur)、<br> [RandomRotation](seg_transforms.html#randomrotation)、[RandomScaleAspect](seg_transforms.html#randomscaleaspect)、[RandomDistort](seg_transforms.html#randomdistort) |
+|语义分割  | [RandomHorizontalFlip](seg_transforms.html#randomhorizontalflip)、[RandomVerticalFlip](seg_transforms.html#randomverticalflip)、[RandomRangeScaling](seg_transforms.html#randomrangescaling)、<br> [RandomStepScaling](seg_transforms.html#randomstepscaling)、[RandomPaddingCrop](seg_transforms.html#randompaddingcrop)、 [RandomBlur](seg_transforms.html#randomblur)、<br> [RandomRotate](seg_transforms.html#randomrotate)、[RandomScaleAspect](seg_transforms.html#randomscaleaspect)、[RandomDistort](seg_transforms.html#randomdistort) |
 
 ## imgaug增强库的支持
 

docs/apis/transforms/seg_transforms.md

@@ -120,7 +120,7 @@ paddlex.seg.transforms.RandomBlur(prob=0.1)
 * **prob** (float): 图像模糊概率。默认为0.1。
 
 
-## RandomRotation类
+## RandomRotate类
 ```python
 paddlex.seg.transforms.RandomRotate(rotate_range=15, im_padding_value=[127.5, 127.5, 127.5], label_padding_value=255)
 ```

docs/appendix/model_zoo.md

@@ -27,6 +27,7 @@
 | DenseNet161|116.3MB  | 8.863       | 78.6     | 94.1     |
 | DenseNet201|  84.6MB   | 8.173       | 77.6     | 93.7     |
 | ShuffleNetV2 | 9.0MB   | 10.941        | 68.8     | 88.5     |
+| HRNet_W18 | 21.29MB | 7.368 (V100 GPU) | 76.9 | 93.4 |
 
 ## 目标检测模型
 
@@ -41,6 +42,7 @@
 |FasterRCNN-ResNet50_vd-FPN|168.7MB | 45.773 | 38.9 |
 |FasterRCNN-ResNet101-FPN| 251.7MB | 55.782 | 38.7 |
 |FasterRCNN-ResNet101_vd-FPN |252MB | 58.785 | 40.5 |
+|FasterRCNN-HRNet_W18-FPN |115.5MB | 57.11 | 36 |
 |YOLOv3-DarkNet53|252.4MB | 21.944 | 38.9 |
 |YOLOv3-MobileNetv1 |101.2MB | 12.771 | 29.3 |
 |YOLOv3-MobileNetv3|94.6MB | - | 31.6 |
@@ -49,4 +51,3 @@
 ## 实例分割模型
 
 > 表中模型相关指标均为在MSCOCO数据集上测试得到。
-

docs/slim/index.rst

+模型压缩
+============================
+
+.. toctree::
+   :maxdepth: 2
+
+   prune.md
+   quant.md

docs/slim/prune.md

+# 模型裁剪
+
+## 原理介绍
+
+模型裁剪用于减小模型的计算量和体积，可以加快模型部署后的预测速度，是一种减小模型大小和降低模型计算复杂度的常用方式，通过裁剪卷积层中Kernel输出通道的大小及其关联层参数大小来实现，其关联裁剪的原理可参见[PaddleSlim相关文档](https://paddlepaddle.github.io/PaddleSlim/algo/algo.html#id16)。**一般而言，在同等模型精度前提下，数据复杂度越低，模型可以被裁剪的比例就越高**。
+
+## 裁剪方法
+PaddleX提供了两种方式:
+
+**1.用户自行计算裁剪配置(推荐)，整体流程包含三个步骤,**  
+> **第一步**： 使用数据集训练原始模型  
+> **第二步**：利用第一步训练好的模型，在验证数据集上计算模型中各个参数的敏感度，并将敏感度信息存储至本地文件  
+> **第三步**：使用数据集训练裁剪模型（与第一步差异在于需要在`train`接口中，将第二步计算得到的敏感信息文件传给接口的`sensitivities_file`参数）  
+
+> 在如上三个步骤中，**相当于模型共需要训练两遍**，分别对应第一步和第三步，但其中第三步训练的是裁剪后的模型，因此训练速度较第一步会更快。  
+> 第二步会遍历模型中的部分裁剪参数，分别计算各个参数裁剪后对于模型在验证集上效果的影响，**因此会反复在验证集上评估多次**。  
+
+**2.使用PaddleX内置的裁剪方案**  
+> PaddleX内置的模型裁剪方案是**基于标准数据集**上计算得到的参数敏感度信息，由于不同数据集特征分布会有较大差异，所以该方案相较于第1种方案训练得到的模型**精度一般而言会更低**（**且用户自定义数据集与标准数据集特征分布差异越大，导致训练的模型精度会越低**），仅在用户想节省时间的前提下可以参考使用，使用方式只需一步，  
+
+> **一步**： 使用数据集训练裁剪模型，在训练调用`train`接口时，将接口中的`sensitivities_file`参数设置为'DEFAULT'字符串
+
+> 注：各模型内置的裁剪方案分别依据的数据集为： 图像分类——ImageNet数据集、目标检测——PascalVOC数据集、语义分割——CityScape数据集
+
+## 裁剪实验
+基于上述两种方案，我们在PaddleX上使用样例数据进行了实验，在Tesla P40上实验指标如下所示，
+
+### 图像分类
+实验背景：使用MobileNetV2模型，数据集为蔬菜分类示例数据，见[使用教程-模型压缩-图像分类](../tutorials/compress/classification.md)
+
+| 模型 | 裁剪情况 | 模型大小 | Top1准确率(%) |GPU预测速度 | CPU预测速度 |
+| :-----| :--------| :-------- | :---------- |:---------- |:----------|
+|MobileNetV2 | 无裁剪（原模型）| 13.0M | 97.50|6.47ms |47.44ms |
+|MobileNetV2 | 方案一(eval_metric_loss=0.10) | 2.1M | 99.58 |5.03ms |20.22ms |
+|MobileNetV2 | 方案二(eval_metric_loss=0.10) | 6.0M | 99.58 |5.42ms |29.06ms |
+
+### 目标检测
+实验背景：使用YOLOv3-MobileNetV1模型，数据集为昆虫检测示例数据，见[使用教程-模型压缩-目标检测](../tutorials/compress/detection.md)
+
+
+| 模型 | 裁剪情况 | 模型大小 | MAP(%) |GPU预测速度 | CPU预测速度 |
+| :-----| :--------| :-------- | :---------- |:---------- | :---------|
+|YOLOv3-MobileNetV1 | 无裁剪（原模型）| 139M | 67.57| 14.88ms |976.42ms |
+|YOLOv3-MobileNetV1 | 方案一(eval_metric_loss=0.10) | 34M | 75.49 |10.60ms |558.49ms |
+|YOLOv3-MobileNetV1 | 方案二(eval_metric_loss=0.05) | 29M | 50.27| 9.43ms |360.46ms |
+
+### 语义分割
+实验背景：使用UNet模型，数据集为视盘分割示例数据, 见[使用教程-模型压缩-语义分割](../tutorials/compress/segmentation.md)
+
+| 模型 | 裁剪情况 | 模型大小 | mIOU(%) |GPU预测速度 | CPU预测速度 |
+| :-----| :--------| :-------- | :---------- |:---------- | :---------|
+|UNet | 无裁剪（原模型）| 77M | 91.22 |33.28ms |9523.55ms |
+|UNet | 方案一(eval_metric_loss=0.10) |26M | 90.37 |21.04ms |3936.20ms |
+|UNet | 方案二(eval_metric_loss=0.10) |23M | 91.21 |18.61ms |3447.75ms |

docs/slim/quant.md

+# 模型量化
+
+## 原理介绍
+为了满足低内存带宽、低功耗、低计算资源占用以及低模型存储等需求，定点量化被提出。为此我们提供了训练后量化，该量化使用KL散度确定量化比例因子，将FP32模型转成INT8模型，且不需要重新训练，可以快速得到量化模型。
+
+
+## 使用PaddleX量化模型
+PaddleX提供了`export_quant_model`接口，让用户以接口的形式完成模型以post_quantization方式量化并导出。点击查看[量化接口使用文档](../apis/slim.md)。
+
+## 量化性能对比
+模型量化后的性能对比指标请查阅[PaddleSlim模型库](https://paddlepaddle.github.io/PaddleSlim/model_zoo.html)

docs/tutorials/deploy/deploy_server/deploy_cpp/deploy_cpp_linux.md

@@ -96,6 +96,17 @@ cmake .. \
 make
 
 ```
+**注意：** linux环境下编译会自动下载OPENCV, PaddleX-Encryption和YAML，如果编译环境无法访问外网，可手动下载：
+
+- [opencv3gcc4.8.tar.bz2](https://paddleseg.bj.bcebos.com/deploy/docker/opencv3gcc4.8.tar.bz2)
+- [paddlex-encryption.zip](https://bj.bcebos.com/paddlex/tools/paddlex-encryption.zip)
+- [yaml-cpp.zip](https://bj.bcebos.com/paddlex/deploy/deps/yaml-cpp.zip)
+
+opencv3gcc4.8.tar.bz2文件下载后解压，然后在script/build.sh中指定`OPENCE_DIR`为解压后的路径。
+
+paddlex-encryption.zip文件下载后解压，然后在script/build.sh中指定`ENCRYPTION_DIR`为解压后的路径。
+
+yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://bj.bcebos.com/paddlex/deploy/deps/yaml-cpp.zip` 中的网址，改为下载文件的路径。
 
 修改脚本设置好主要参数后，执行`build`脚本：
  ```shell
@@ -104,8 +115,9 @@ make
 
 ### Step5: 预测及可视化
 
-参考[导出inference模型](../../deploy_python.html#inference)将模型导出为inference格式模型。
-**注意：由于PaddleX代码的持续更新，版本低于1.0.0的模型暂时无法直接用于预测部署，参考[模型版本升级](../../upgrade_version.md)对模型版本进行升级。**
+**在加载模型前，请检查你的模型目录中文件应该包括`model.yml`、`__model__`和`__params__`三个文件。如若不满足这个条件，请参考[模型导出为Inference文档](../deploy_python.html#inference)将模型导出为部署格式。**  
+
+> **注意：由于PaddleX代码的持续更新，版本低于1.0.0的模型（模型版本可查看model.yml文件中的version字段）暂时无法直接用于预测部署，参考[模型版本升级](../../upgrade_version.md)对模型版本进行升级。**  
 
 编译成功后，预测demo的可执行程序分别为`build/demo/detector`，`build/demo/classifer`，`build/demo/segmenter`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
 
@@ -117,7 +129,7 @@ make
 | use_gpu  | 是否使用 GPU 预测, 支持值为0或1(默认值为0) |
 | use_trt  | 是否使用 TensorTr 预测, 支持值为0或1(默认值为0) |
 | gpu_id  | GPU 设备ID, 默认值为0 |
-| save_dir | 保存可视化结果的路径, 默认值为"output"，classfier无该参数 |
+| save_dir | 保存可视化结果的路径, 默认值为"output"，**classfier无该参数** |
 
 ## 样例
 

docs/tutorials/deploy/deploy_server/deploy_cpp/deploy_cpp_win_vs2019.md

@@ -86,7 +86,14 @@ PaddlePaddle C++ 预测库针对不同的`CPU`，`CUDA`，以及是否支持Tens
 | OPENCV_DIR  | OpenCV的安装路径， |
 | PADDLE_DIR | Paddle c++预测库的路径 |
 
-**注意：** 1. 使用`CPU`版预测库，请把`WITH_GPU`的`值`去掉勾 2. 如果使用的是`openblas`版本，请把`WITH_MKL`的`值`去掉勾
+**注意：**
+1. 使用`CPU`版预测库，请把`WITH_GPU`的`值`去掉勾
+
+2. 如果使用的是`openblas`版本，请把`WITH_MKL`的`值`去掉勾
+
+3. Windows环境下编译会自动下载YAML，如果编译环境无法访问外网，可手动下载： [yaml-cpp.zip](https://bj.bcebos.com/paddlex/deploy/deps/yaml-cpp.zip)
+
+yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://bj.bcebos.com/paddlex/deploy/deps/yaml-cpp.zip` 中的网址，改为下载文件的路径。
 
 ![step4](../../images/vs2019_step5.png)
 
@@ -99,8 +106,10 @@ PaddlePaddle C++ 预测库针对不同的`CPU`，`CUDA`，以及是否支持Tens
 
 ### Step5: 预测及可视化
 
-参考[导出inference模型](../deploy_python.html#inference)将模型导出为inference格式模型。
-**注意：由于PaddleX代码的持续更新，版本低于1.0.0的模型暂时无法直接用于预测部署，参考[模型版本升级](../../upgrade_version.md)对模型版本进行升级。**
+
+**在加载模型前，请检查你的模型目录中文件应该包括`model.yml`、`__model__`和`__params__`三个文件。如若不满足这个条件，请参考[模型导出为Inference文档](../deploy_python.html#inference)将模型导出为部署格式。**   
+
+**注意：由于PaddleX代码的持续更新，版本低于1.0.0的模型（模型版本可查看model.yml文件中的version字段）暂时无法直接用于预测部署，参考[模型版本升级](../../upgrade_version.md)对模型版本进行升级。**
 
 上述`Visual Studio 2019`编译产出的可执行文件在`out\build\x64-Release`目录下，打开`cmd`，并切换到该目录：
 

paddlex/__init__.py

@@ -53,4 +53,4 @@ log_level = 2
 
 from . import interpret
 
-__version__ = '1.0.2.github'
+__version__ = '1.0.4'

paddlex/cls.py

@@ -36,5 +36,6 @@ DenseNet121 = cv.models.DenseNet121
 DenseNet161 = cv.models.DenseNet161
 DenseNet201 = cv.models.DenseNet201
 ShuffleNetV2 = cv.models.ShuffleNetV2
+HRNet_W18 = cv.models.HRNet_W18
 
 transforms = cv.transforms.cls_transforms

paddlex/cv/datasets/dataset.py

@@ -209,8 +209,8 @@ def GenerateMiniBatch(batch_data):
     padding_batch = []
     for data in batch_data:
         im_c, im_h, im_w = data[0].shape[:]
-        padding_im = np.zeros((im_c, max_shape[1], max_shape[2]),
-                              dtype=np.float32)
+        padding_im = np.zeros(
+            (im_c, max_shape[1], max_shape[2]), dtype=np.float32)
         padding_im[:, :im_h, :im_w] = data[0]
         padding_batch.append((padding_im, ) + data[1:])
     return padding_batch
@@ -226,8 +226,8 @@ class Dataset:
         if num_workers == 'auto':
             import multiprocessing as mp
             num_workers = mp.cpu_count() // 2 if mp.cpu_count() // 2 < 8 else 8
-        if platform.platform().startswith(
-                "Darwin") or platform.platform().startswith("Windows"):
+        if platform.platform().startswith("Darwin") or platform.platform(
+        ).startswith("Windows"):
             parallel_method = 'thread'
         if transforms is None:
             raise Exception("transform should be defined.")

paddlex/cv/models/__init__.py

@@ -34,11 +34,13 @@ from .classifier import DenseNet121
 from .classifier import DenseNet161
 from .classifier import DenseNet201
 from .classifier import ShuffleNetV2
+from .classifier import HRNet_W18
 from .base import BaseAPI
 from .yolo_v3 import YOLOv3
 from .faster_rcnn import FasterRCNN
 from .mask_rcnn import MaskRCNN
 from .unet import UNet
 from .deeplabv3p import DeepLabv3p
+from .hrnet import HRNet
 from .load_model import load_model
 from .slim import prune

paddlex/cv/models/base.py

@@ -79,9 +79,9 @@ class BaseAPI:
             return int(batch_size // len(self.places))
         else:
             raise Exception("Please support correct batch_size, \
-                            which can be divided by available cards({}) in {}".
-                            format(paddlex.env_info['num'],
-                                   paddlex.env_info['place']))
+                            which can be divided by available cards({}) in {}"
+                            .format(paddlex.env_info['num'], paddlex.env_info[
+                                'place']))
 
     def build_program(self):
         # 构建训练网络
@@ -198,6 +198,8 @@ class BaseAPI:
                 backbone = self.backbone
             else:
                 backbone = self.__class__.__name__
+                if backbone == "HRNet":
+                    backbone = backbone + "_W{}".format(self.width)
             pretrain_weights = get_pretrain_weights(
                 pretrain_weights, self.model_type, backbone, pretrain_dir)
         if startup_prog is None:
@@ -210,8 +212,8 @@ class BaseAPI:
             paddlex.utils.utils.load_pretrain_weights(
                 self.exe, self.train_prog, resume_checkpoint, resume=True)
             if not osp.exists(osp.join(resume_checkpoint, "model.yml")):
-                raise Exception(
-                    "There's not model.yml in {}".format(resume_checkpoint))
+                raise Exception("There's not model.yml in {}".format(
+                    resume_checkpoint))
             with open(osp.join(resume_checkpoint, "model.yml")) as f:
                 info = yaml.load(f.read(), Loader=yaml.Loader)
                 self.completed_epochs = info['completed_epochs']
@@ -269,13 +271,13 @@ class BaseAPI:
         except:
             pass
 
+        if hasattr(self, 'test_transforms'):
             if hasattr(self.test_transforms, 'to_rgb'):
                 if self.test_transforms.to_rgb:
                     info['TransformsMode'] = 'RGB'
                 else:
                     info['TransformsMode'] = 'BGR'
 
-        if hasattr(self, 'test_transforms'):
             if self.test_transforms is not None:
                 info['Transforms'] = list()
                 for op in self.test_transforms.transforms:
@@ -362,8 +364,8 @@ class BaseAPI:
 
         # 模型保存成功的标志
         open(osp.join(save_dir, '.success'), 'w').close()
-        logging.info(
-            "Model for inference deploy saved in {}.".format(save_dir))
+        logging.info("Model for inference deploy saved in {}.".format(
+            save_dir))
 
     def train_loop(self,
                    num_epochs,
@@ -377,7 +379,8 @@ class BaseAPI:
                    early_stop=False,
                    early_stop_patience=5):
         if train_dataset.num_samples < train_batch_size:
-            raise Exception('The amount of training datset must be larger than batch size.')
+            raise Exception(
+                'The amount of training datset must be larger than batch size.')
         if not osp.isdir(save_dir):
             if osp.exists(save_dir):
                 os.remove(save_dir)
@@ -415,8 +418,8 @@ class BaseAPI:
                     build_strategy=build_strategy,
                     exec_strategy=exec_strategy)
 
-        total_num_steps = math.floor(
-            train_dataset.num_samples / train_batch_size)
+        total_num_steps = math.floor(train_dataset.num_samples /
+                                     train_batch_size)
         num_steps = 0
         time_stat = list()
         time_train_one_epoch = None
@@ -430,8 +433,8 @@ class BaseAPI:
         if self.model_type == 'detector':
             eval_batch_size = self._get_single_card_bs(train_batch_size)
         if eval_dataset is not None:
-            total_num_steps_eval = math.ceil(
-                eval_dataset.num_samples / eval_batch_size)
+            total_num_steps_eval = math.ceil(eval_dataset.num_samples /
+                                             eval_batch_size)
 
         if use_vdl:
             # VisualDL component
@@ -473,7 +476,9 @@ class BaseAPI:
 
                     if use_vdl:
                         for k, v in step_metrics.items():
-                            log_writer.add_scalar('Metrics/Training(Step): {}'.format(k), v, num_steps)
+                            log_writer.add_scalar(
+                                'Metrics/Training(Step): {}'.format(k), v,
+                                num_steps)
 
                     # 估算剩余时间
                     avg_step_time = np.mean(time_stat)
@@ -481,11 +486,12 @@ class BaseAPI:
                         eta = (num_epochs - i - 1) * time_train_one_epoch + (
                             total_num_steps - step - 1) * avg_step_time
                     else:
-                        eta = ((num_epochs - i) * total_num_steps - step -
-                               1) * avg_step_time
+                        eta = ((num_epochs - i) * total_num_steps - step - 1
+                               ) * avg_step_time
                     if time_eval_one_epoch is not None:
-                        eval_eta = (total_eval_times - i //
-                                    save_interval_epochs) * time_eval_one_epoch
+                        eval_eta = (
+                            total_eval_times - i // save_interval_epochs
+                        ) * time_eval_one_epoch
                     else:
                         eval_eta = (
                             total_eval_times - i // save_interval_epochs
@@ -495,10 +501,11 @@ class BaseAPI:
                     logging.info(
                         "[TRAIN] Epoch={}/{}, Step={}/{}, {}, time_each_step={}s, eta={}"
                         .format(i + 1, num_epochs, step + 1, total_num_steps,
-                                dict2str(step_metrics), round(
-                                    avg_step_time, 2), eta_str))
+                                dict2str(step_metrics),
+                                round(avg_step_time, 2), eta_str))
             train_metrics = OrderedDict(
-                zip(list(self.train_outputs.keys()), np.mean(records, axis=0)))
+                zip(list(self.train_outputs.keys()), np.mean(
+                    records, axis=0)))
             logging.info('[TRAIN] Epoch {} finished, {} .'.format(
                 i + 1, dict2str(train_metrics)))
             time_train_one_epoch = time.time() - epoch_start_time
@@ -534,7 +541,8 @@ class BaseAPI:
                             if isinstance(v, np.ndarray):
                                 if v.size > 1:
                                     continue
-                            log_writer.add_scalar("Metrics/Eval(Epoch): {}".format(k), v, i+1)
+                            log_writer.add_scalar(
+                                "Metrics/Eval(Epoch): {}".format(k), v, i + 1)
                 self.save_model(save_dir=current_save_dir)
                 time_eval_one_epoch = time.time() - eval_epoch_start_time
                 eval_epoch_start_time = time.time()

paddlex/cv/models/classifier.py

@@ -40,8 +40,8 @@ class BaseClassifier(BaseAPI):
         self.init_params = locals()
         super(BaseClassifier, self).__init__('classifier')
         if not hasattr(paddlex.cv.nets, str.lower(model_name)):
-            raise Exception(
-                "ERROR: There's no model named {}.".format(model_name))
+            raise Exception("ERROR: There's no model named {}.".format(
+                model_name))
         self.model_name = model_name
         self.labels = None
         self.num_classes = num_classes
@@ -218,15 +218,14 @@ class BaseClassifier(BaseAPI):
                 num_pad_samples = batch_size - num_samples
                 pad_images = np.tile(images[0:1], (num_pad_samples, 1, 1, 1))
                 images = np.concatenate([images, pad_images])
-            outputs = self.exe.run(
-                self.parallel_test_prog,
+            outputs = self.exe.run(self.parallel_test_prog,
                                    feed={'image': images},
                                    fetch_list=list(self.test_outputs.values()))
             outputs = [outputs[0][:num_samples]]
             true_labels.extend(labels)
             pred_scores.extend(outputs[0].tolist())
-            logging.debug("[EVAL] Epoch={}, Step={}/{}".format(
-                epoch_id, step + 1, total_steps))
+            logging.debug("[EVAL] Epoch={}, Step={}/{}".format(epoch_id, step +
+                                                               1, total_steps))
 
         pred_top1_label = np.argsort(pred_scores)[:, -1]
         pred_topk_label = np.argsort(pred_scores)[:, -k:]
@@ -263,8 +262,7 @@ class BaseClassifier(BaseAPI):
             self.arrange_transforms(
                 transforms=self.test_transforms, mode='test')
             im = self.test_transforms(img_file)
-        result = self.exe.run(
-            self.test_prog,
+        result = self.exe.run(self.test_prog,
                               feed={'image': im},
                               fetch_list=list(self.test_outputs.values()))
         pred_label = np.argsort(result[0][0])[::-1][:true_topk]
@@ -400,3 +398,9 @@ class ShuffleNetV2(BaseClassifier):
     def __init__(self, num_classes=1000):
         super(ShuffleNetV2, self).__init__(
             model_name='ShuffleNetV2', num_classes=num_classes)
+
+
+class HRNet_W18(BaseClassifier):
+    def __init__(self, num_classes=1000):
+        super(HRNet_W18, self).__init__(
+            model_name='HRNet_W18', num_classes=num_classes)

paddlex/cv/models/deeplabv3p.py

@@ -190,11 +190,6 @@ class DeepLabv3p(BaseAPI):
         if mode == 'train':
             self.optimizer.minimize(model_out)
             outputs['loss'] = model_out
-        elif mode == 'eval':
-            outputs['loss'] = model_out[0]
-            outputs['pred'] = model_out[1]
-            outputs['label'] = model_out[2]
-            outputs['mask'] = model_out[3]
         else:
             outputs['pred'] = model_out[0]
             outputs['logit'] = model_out[1]
@@ -336,15 +331,23 @@ class DeepLabv3p(BaseAPI):
         for step, data in tqdm.tqdm(
                 enumerate(data_generator()), total=total_steps):
             images = np.array([d[0] for d in data])
-            labels = np.array([d[1] for d in data])
+
+            _, _, im_h, im_w = images.shape
+            labels = list()
+            for d in data:
+                padding_label = np.zeros(
+                    (1, im_h, im_w)).astype('int64') + self.ignore_index
+                padding_label[:, :im_h, :im_w] = d[1]
+                labels.append(padding_label)
+            labels = np.array(labels)
+
             num_samples = images.shape[0]
             if num_samples < batch_size:
                 num_pad_samples = batch_size - num_samples
                 pad_images = np.tile(images[0:1], (num_pad_samples, 1, 1, 1))
                 images = np.concatenate([images, pad_images])
             feed_data = {'image': images}
-            outputs = self.exe.run(
-                self.parallel_test_prog,
+            outputs = self.exe.run(self.parallel_test_prog,
                                    feed=feed_data,
                                    fetch_list=list(self.test_outputs.values()),
                                    return_numpy=True)
@@ -364,8 +367,7 @@ class DeepLabv3p(BaseAPI):
 
         metrics = OrderedDict(
             zip(['miou', 'category_iou', 'macc', 'category_acc', 'kappa'],
-                [miou, category_iou, macc, category_acc,
-                 conf_mat.kappa()]))
+                [miou, category_iou, macc, category_acc, conf_mat.kappa()]))
         if return_details:
             eval_details = {
                 'confusion_matrix': conf_mat.confusion_matrix.tolist()
@@ -394,8 +396,7 @@ class DeepLabv3p(BaseAPI):
                 transforms=self.test_transforms, mode='test')
             im, im_info = self.test_transforms(im_file)
         im = np.expand_dims(im, axis=0)
-        result = self.exe.run(
-            self.test_prog,
+        result = self.exe.run(self.test_prog,
                               feed={'image': im},
                               fetch_list=list(self.test_outputs.values()))
         pred = result[0]
@@ -413,6 +414,6 @@ class DeepLabv3p(BaseAPI):
                 pred = pred[0:h, 0:w]
                 logit = logit[0:h, 0:w, :]
             else:
-                raise Exception("Unexpected info '{}' in im_info".format(
-                    info[0]))
+                raise Exception("Unexpected info '{}' in im_info".format(info[
+                    0]))
         return {'label_map': pred, 'score_map': logit}

paddlex/cv/models/faster_rcnn.py

@@ -32,7 +32,7 @@ class FasterRCNN(BaseAPI):
     Args:
         num_classes (int): 包含了背景类的类别数。默认为81。
         backbone (str): FasterRCNN的backbone网络，取值范围为['ResNet18', 'ResNet50',
-            'ResNet50_vd', 'ResNet101', 'ResNet101_vd']。默认为'ResNet50'。
+            'ResNet50_vd', 'ResNet101', 'ResNet101_vd', 'HRNet_W18']。默认为'ResNet50'。
         with_fpn (bool): 是否使用FPN结构。默认为True。
         aspect_ratios (list): 生成anchor高宽比的可选值。默认为[0.5, 1.0, 2.0]。
         anchor_sizes (list): 生成anchor大小的可选值。默认为[32, 64, 128, 256, 512]。
@@ -47,7 +47,8 @@ class FasterRCNN(BaseAPI):
         self.init_params = locals()
         super(FasterRCNN, self).__init__('detector')
         backbones = [
-            'ResNet18', 'ResNet50', 'ResNet50_vd', 'ResNet101', 'ResNet101_vd'
+            'ResNet18', 'ResNet50', 'ResNet50_vd', 'ResNet101', 'ResNet101_vd',
+            'HRNet_W18'
         ]
         assert backbone in backbones, "backbone should be one of {}".format(
             backbones)
@@ -79,6 +80,12 @@ class FasterRCNN(BaseAPI):
             layers = 101
             variant = 'd'
             norm_type = 'affine_channel'
+        elif backbone_name == 'HRNet_W18':
+            backbone = paddlex.cv.nets.hrnet.HRNet(
+                width=18, freeze_norm=True, norm_decay=0., freeze_at=0)
+            if self.with_fpn is False:
+                self.with_fpn = True
+            return backbone
         if self.with_fpn:
             backbone = paddlex.cv.nets.resnet.ResNet(
                 norm_type='bn' if norm_type is None else norm_type,
@@ -117,12 +124,12 @@ class FasterRCNN(BaseAPI):
             model_out = model.build_net(inputs)
             loss = model_out['loss']
             self.optimizer.minimize(loss)
-            outputs = OrderedDict([('loss', model_out['loss']),
+            outputs = OrderedDict(
+                [('loss', model_out['loss']),
                  ('loss_cls', model_out['loss_cls']),
                  ('loss_bbox', model_out['loss_bbox']),
-                                   ('loss_rpn_cls', model_out['loss_rpn_cls']),
-                                   ('loss_rpn_bbox',
-                                    model_out['loss_rpn_bbox'])])
+                 ('loss_rpn_cls', model_out['loss_rpn_cls']), (
+                     'loss_rpn_bbox', model_out['loss_rpn_bbox'])])
         else:
             outputs = model.build_net(inputs)
         return inputs, outputs
@@ -227,7 +234,9 @@ class FasterRCNN(BaseAPI):
         # 构建训练、验证、测试网络
         self.build_program()
         fuse_bn = True
-        if self.with_fpn and self.backbone in ['ResNet18', 'ResNet50']:
+        if self.with_fpn and self.backbone in [
+                'ResNet18', 'ResNet50', 'HRNet_W18'
+        ]:
             fuse_bn = False
         self.net_initialize(
             startup_prog=fluid.default_startup_program(),
@@ -310,8 +319,7 @@ class FasterRCNN(BaseAPI):
                 'im_info': im_infos,
                 'im_shape': im_shapes,
             }
-            outputs = self.exe.run(
-                self.test_prog,
+            outputs = self.exe.run(self.test_prog,
                                    feed=[feed_data],
                                    fetch_list=list(self.test_outputs.values()),
                                    return_numpy=False)
@@ -339,13 +347,13 @@ class FasterRCNN(BaseAPI):
                 res['is_difficult'] = (np.array(res_is_difficult),
                                        [res_is_difficult_lod])
             results.append(res)
-            logging.debug("[EVAL] Epoch={}, Step={}/{}".format(
-                epoch_id, step + 1, total_steps))
+            logging.debug("[EVAL] Epoch={}, Step={}/{}".format(epoch_id, step +
+                                                               1, total_steps))
         box_ap_stats, eval_details = eval_results(
             results, metric, eval_dataset.coco_gt, with_background=True)
         metrics = OrderedDict(
-            zip(['bbox_mmap' if metric == 'COCO' else 'bbox_map'],
-                box_ap_stats))
+            zip(['bbox_mmap'
+                 if metric == 'COCO' else 'bbox_map'], box_ap_stats))
         if return_details:
             return metrics, eval_details
         return metrics
@@ -359,7 +367,8 @@ class FasterRCNN(BaseAPI):
 
         Returns:
             list: 预测结果列表，每个预测结果由预测框类别标签、
-              预测框类别名称、预测框坐标、预测框得分组成。
+              预测框类别名称、预测框坐标(坐标格式为[xmin, ymin, w, h]）、
+              预测框得分组成。
         """
         if transforms is None and not hasattr(self, 'test_transforms'):
             raise Exception("transforms need to be defined, now is None.")
@@ -373,8 +382,7 @@ class FasterRCNN(BaseAPI):
         im = np.expand_dims(im, axis=0)
         im_resize_info = np.expand_dims(im_resize_info, axis=0)
         im_shape = np.expand_dims(im_shape, axis=0)
-        outputs = self.exe.run(
-            self.test_prog,
+        outputs = self.exe.run(self.test_prog,
                                feed={
                                    'image': im,
                                    'im_info': im_resize_info,

paddlex/cv/models/hrnet.py

+#copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+#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.
+
+from __future__ import absolute_import
+import paddle.fluid as fluid
+import paddlex
+from collections import OrderedDict
+from .deeplabv3p import DeepLabv3p
+
+
+class HRNet(DeepLabv3p):
+    """实现HRNet网络的构建并进行训练、评估、预测和模型导出。
+
+    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]。
+        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。
+        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。
+
+    Raises:
+        ValueError: use_bce_loss或use_dice_loss为真且num_calsses > 2。
+        ValueError: class_weight为list, 但长度不等于num_class。
+            class_weight为str, 但class_weight.low()不等于dynamic。
+        TypeError: class_weight不为None时，其类型不是list或str。
+    """
+
+    def __init__(self,
+                 num_classes=2,
+                 width=18,
+                 use_bce_loss=False,
+                 use_dice_loss=False,
+                 class_weight=None,
+                 ignore_index=255):
+        self.init_params = locals()
+        super(DeepLabv3p, self).__init__('segmenter')
+        # dice_loss或bce_loss只适用两类分割中
+        if num_classes > 2 and (use_bce_loss or use_dice_loss):
+            raise ValueError(
+                "dice loss and bce loss is only applicable to binary classfication"
+            )
+
+        if class_weight is not None:
+            if isinstance(class_weight, list):
+                if len(class_weight) != num_classes:
+                    raise ValueError(
+                        "Length of class_weight should be equal to number of classes"
+                    )
+            elif isinstance(class_weight, str):
+                if class_weight.lower() != 'dynamic':
+                    raise ValueError(
+                        "if class_weight is string, must be dynamic!")
+            else:
+                raise TypeError(
+                    'Expect class_weight is a list or string but receive {}'.
+                    format(type(class_weight)))
+        self.num_classes = num_classes
+        self.width = width
+        self.use_bce_loss = use_bce_loss
+        self.use_dice_loss = use_dice_loss
+        self.class_weight = class_weight
+        self.ignore_index = ignore_index
+        self.labels = None
+
+    def build_net(self, mode='train'):
+        model = paddlex.cv.nets.segmentation.HRNet(
+            self.num_classes,
+            width=self.width,
+            mode=mode,
+            use_bce_loss=self.use_bce_loss,
+            use_dice_loss=self.use_dice_loss,
+            class_weight=self.class_weight,
+            ignore_index=self.ignore_index)
+        inputs = model.generate_inputs()
+        model_out = model.build_net(inputs)
+        outputs = OrderedDict()
+        if mode == 'train':
+            self.optimizer.minimize(model_out)
+            outputs['loss'] = model_out
+        elif mode == 'eval':
+            outputs['loss'] = model_out[0]
+            outputs['pred'] = model_out[1]
+            outputs['label'] = model_out[2]
+            outputs['mask'] = model_out[3]
+        else:
+            outputs['pred'] = model_out[0]
+            outputs['logit'] = model_out[1]
+        return inputs, outputs
+
+    def default_optimizer(self,
+                          learning_rate,
+                          num_epochs,
+                          num_steps_each_epoch,
+                          lr_decay_power=0.9):
+        decay_step = num_epochs * num_steps_each_epoch
+        lr_decay = fluid.layers.polynomial_decay(
+            learning_rate,
+            decay_step,
+            end_learning_rate=0,
+            power=lr_decay_power)
+        optimizer = fluid.optimizer.Momentum(
+            lr_decay,
+            momentum=0.9,
+            regularization=fluid.regularizer.L2Decay(
+                regularization_coeff=5e-04))
+        return optimizer
+
+    def train(self,
+              num_epochs,
+              train_dataset,
+              train_batch_size=2,
+              eval_dataset=None,
+              save_interval_epochs=1,
+              log_interval_steps=2,
+              save_dir='output',
+              pretrain_weights='IMAGENET',
+              optimizer=None,
+              learning_rate=0.01,
+              lr_decay_power=0.9,
+              use_vdl=False,
+              sensitivities_file=None,
+              eval_metric_loss=0.05,
+              early_stop=False,
+              early_stop_patience=5,
+              resume_checkpoint=None):
+        """训练。
+
+        Args:
+            num_epochs (int): 训练迭代轮数。
+            train_dataset (paddlex.datasets): 训练数据读取器。
+            train_batch_size (int): 训练数据batch大小。同时作为验证数据batch大小。默认2。
+            eval_dataset (paddlex.datasets): 评估数据读取器。
+            save_interval_epochs (int): 模型保存间隔（单位：迭代轮数）。默认为1。
+            log_interval_steps (int): 训练日志输出间隔（单位：迭代次数）。默认为2。
+            save_dir (str): 模型保存路径。默认'output'。
+            pretrain_weights (str): 若指定为路径时，则加载路径下预训练模型；若为字符串'IMAGENET'，
+                则自动下载在IMAGENET图片数据上预训练的模型权重；若为None，则不使用预训练模型。默认为'IMAGENET'。
+            optimizer (paddle.fluid.optimizer): 优化器。当改参数为None时，使用默认的优化器：使用
+                fluid.optimizer.Momentum优化方法，polynomial的学习率衰减策略。
+            learning_rate (float): 默认优化器的初始学习率。默认0.01。
+            lr_decay_power (float): 默认优化器学习率多项式衰减系数。默认0.9。
+            use_vdl (bool): 是否使用VisualDL进行可视化。默认False。
+            sensitivities_file (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，
+                则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
+            eval_metric_loss (float): 可容忍的精度损失。默认为0.05。
+            early_stop (bool): 是否使用提前终止训练策略。默认值为False。
+            early_stop_patience (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内
+                连续下降或持平，则终止训练。默认值为5。
+            resume_checkpoint (str): 恢复训练时指定上次训练保存的模型路径。若为None，则不会恢复训练。默认值为None。
+
+        Raises:
+            ValueError: 模型从inference model进行加载。
+        """
+        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,
+            early_stop_patience, resume_checkpoint)

paddlex/cv/models/mask_rcnn.py

@@ -32,7 +32,7 @@ class MaskRCNN(FasterRCNN):
     Args:
         num_classes (int): 包含了背景类的类别数。默认为81。
         backbone (str): MaskRCNN的backbone网络，取值范围为['ResNet18', 'ResNet50',
-            'ResNet50_vd', 'ResNet101', 'ResNet101_vd']。默认为'ResNet50'。
+            'ResNet50_vd', 'ResNet101', 'ResNet101_vd', 'HRNet_W18']。默认为'ResNet50'。
         with_fpn (bool): 是否使用FPN结构。默认为True。
         aspect_ratios (list): 生成anchor高宽比的可选值。默认为[0.5, 1.0, 2.0]。
         anchor_sizes (list): 生成anchor大小的可选值。默认为[32, 64, 128, 256, 512]。
@@ -46,7 +46,8 @@ class MaskRCNN(FasterRCNN):
                  anchor_sizes=[32, 64, 128, 256, 512]):
         self.init_params = locals()
         backbones = [
-            'ResNet18', 'ResNet50', 'ResNet50_vd', 'ResNet101', 'ResNet101_vd'
+            'ResNet18', 'ResNet50', 'ResNet50_vd', 'ResNet101', 'ResNet101_vd',
+            'HRNet_W18'
         ]
         assert backbone in backbones, "backbone should be one of {}".format(
             backbones)
@@ -81,13 +82,13 @@ class MaskRCNN(FasterRCNN):
             model_out = model.build_net(inputs)
             loss = model_out['loss']
             self.optimizer.minimize(loss)
-            outputs = OrderedDict([('loss', model_out['loss']),
+            outputs = OrderedDict(
+                [('loss', model_out['loss']),
                  ('loss_cls', model_out['loss_cls']),
                  ('loss_bbox', model_out['loss_bbox']),
                  ('loss_mask', model_out['loss_mask']),
-                                   ('loss_rpn_cls', model_out['loss_rpn_cls']),
-                                   ('loss_rpn_bbox',
-                                    model_out['loss_rpn_bbox'])])
+                 ('loss_rpn_cls', model_out['loss_rpn_cls']), (
+                     'loss_rpn_bbox', model_out['loss_rpn_bbox'])])
         else:
             outputs = model.build_net(inputs)
         return inputs, outputs
@@ -194,7 +195,9 @@ class MaskRCNN(FasterRCNN):
         # 构建训练、验证、测试网络
         self.build_program()
         fuse_bn = True
-        if self.with_fpn and self.backbone in ['ResNet18', 'ResNet50']:
+        if self.with_fpn and self.backbone in [
+                'ResNet18', 'ResNet50', 'HRNet_W18'
+        ]:
             fuse_bn = False
         self.net_initialize(
             startup_prog=fluid.default_startup_program(),
@@ -276,8 +279,7 @@ class MaskRCNN(FasterRCNN):
                 'im_info': im_infos,
                 'im_shape': im_shapes,
             }
-            outputs = self.exe.run(
-                self.test_prog,
+            outputs = self.exe.run(self.test_prog,
                                    feed=[feed_data],
                                    fetch_list=list(self.test_outputs.values()),
                                    return_numpy=False)
@@ -292,8 +294,8 @@ class MaskRCNN(FasterRCNN):
             res['im_shape'] = (im_shapes, [])
             res['im_id'] = (np.array(res_im_id), [])
             results.append(res)
-            logging.debug("[EVAL] Epoch={}, Step={}/{}".format(
-                epoch_id, step + 1, total_steps))
+            logging.debug("[EVAL] Epoch={}, Step={}/{}".format(epoch_id, step +
+                                                               1, total_steps))
 
         ap_stats, eval_details = eval_results(
             results,
@@ -302,8 +304,8 @@ class MaskRCNN(FasterRCNN):
             with_background=True,
             resolution=self.mask_head_resolution)
         if metric == 'VOC':
-            if isinstance(ap_stats[0], np.ndarray) and isinstance(
-                    ap_stats[1], np.ndarray):
+            if isinstance(ap_stats[0], np.ndarray) and isinstance(ap_stats[1],
+                                                                  np.ndarray):
                 metrics = OrderedDict(
                     zip(['bbox_map', 'segm_map'],
                         [ap_stats[0][1], ap_stats[1][1]]))
@@ -311,8 +313,8 @@ class MaskRCNN(FasterRCNN):
                 metrics = OrderedDict(
                     zip(['bbox_map', 'segm_map'], [0.0, 0.0]))
         elif metric == 'COCO':
-            if isinstance(ap_stats[0], np.ndarray) and isinstance(
-                    ap_stats[1], np.ndarray):
+            if isinstance(ap_stats[0], np.ndarray) and isinstance(ap_stats[1],
+                                                                  np.ndarray):
                 metrics = OrderedDict(
                     zip(['bbox_mmap', 'segm_mmap'],
                         [ap_stats[0][0], ap_stats[1][0]]))
@@ -331,7 +333,9 @@ class MaskRCNN(FasterRCNN):
             transforms (paddlex.det.transforms): 数据预处理操作。
 
         Returns:
-            dict: 预测结果列表，每个预测结果由预测框类别标签、预测框类别名称、预测框坐标、预测框内的二值图、
+            dict: 预测结果列表，每个预测结果由预测框类别标签、预测框类别名称、
+                  预测框坐标(坐标格式为[xmin, ymin, w, h]）、
+                  原图大小的预测二值图（1表示预测框类别，0表示背景类）、
                   预测框得分组成。
         """
         if transforms is None and not hasattr(self, 'test_transforms'):
@@ -346,8 +350,7 @@ class MaskRCNN(FasterRCNN):
         im = np.expand_dims(im, axis=0)
         im_resize_info = np.expand_dims(im_resize_info, axis=0)
         im_shape = np.expand_dims(im_shape, axis=0)
-        outputs = self.exe.run(
-            self.test_prog,
+        outputs = self.exe.run(self.test_prog,
                                feed={
                                    'image': im,
                                    'im_info': im_resize_info,
@@ -368,8 +371,8 @@ class MaskRCNN(FasterRCNN):
         import pycocotools.mask as mask_util
         for index, xywh_res in enumerate(xywh_results):
             del xywh_res['image_id']
-            xywh_res['mask'] = mask_util.decode(
-                segm_results[index]['segmentation'])
+            xywh_res['mask'] = mask_util.decode(segm_results[index][
+                'segmentation'])
             xywh_res['category'] = self.labels[xywh_res['category_id']]
             results.append(xywh_res)
         return results

paddlex/cv/models/utils/pretrain_weights.py

@@ -56,6 +56,20 @@ image_pretrain = {
     'https://paddle-imagenet-models-name.bj.bcebos.com/Xception65_deeplab_pretrained.tar',
     'ShuffleNetV2':
     'https://paddle-imagenet-models-name.bj.bcebos.com/ShuffleNetV2_pretrained.tar',
+    'HRNet_W18':
+    'https://paddle-imagenet-models-name.bj.bcebos.com/HRNet_W18_C_pretrained.tar',
+    'HRNet_W30':
+    'https://paddle-imagenet-models-name.bj.bcebos.com/HRNet_W30_C_pretrained.tar',
+    'HRNet_W32':
+    'https://paddle-imagenet-models-name.bj.bcebos.com/HRNet_W32_C_pretrained.tar',
+    'HRNet_W40':
+    'https://paddle-imagenet-models-name.bj.bcebos.com/HRNet_W40_C_pretrained.tar',
+    'HRNet_W48':
+    'https://paddle-imagenet-models-name.bj.bcebos.com/HRNet_W48_C_pretrained.tar',
+    'HRNet_W60':
+    'https://paddle-imagenet-models-name.bj.bcebos.com/HRNet_W60_C_pretrained.tar',
+    'HRNet_W64':
+    'https://paddle-imagenet-models-name.bj.bcebos.com/HRNet_W64_C_pretrained.tar',
 }
 
 coco_pretrain = {

paddlex/cv/models/yolo_v3.py

@@ -306,8 +306,7 @@ class YOLOv3(BaseAPI):
             images = np.array([d[0] for d in data])
             im_sizes = np.array([d[1] for d in data])
             feed_data = {'image': images, 'im_size': im_sizes}
-            outputs = self.exe.run(
-                self.test_prog,
+            outputs = self.exe.run(self.test_prog,
                                    feed=[feed_data],
                                    fetch_list=list(self.test_outputs.values()),
                                    return_numpy=False)
@@ -326,13 +325,13 @@ class YOLOv3(BaseAPI):
                 res['gt_label'] = (res_gt_label, [])
                 res['is_difficult'] = (res_is_difficult, [])
             results.append(res)
-            logging.debug("[EVAL] Epoch={}, Step={}/{}".format(
-                epoch_id, step + 1, total_steps))
+            logging.debug("[EVAL] Epoch={}, Step={}/{}".format(epoch_id, step +
+                                                               1, total_steps))
         box_ap_stats, eval_details = eval_results(
             results, metric, eval_dataset.coco_gt, with_background=False)
         evaluate_metrics = OrderedDict(
-            zip(['bbox_mmap' if metric == 'COCO' else 'bbox_map'],
-                box_ap_stats))
+            zip(['bbox_mmap'
+                 if metric == 'COCO' else 'bbox_map'], box_ap_stats))
         if return_details:
             return evaluate_metrics, eval_details
         return evaluate_metrics
@@ -346,7 +345,8 @@ class YOLOv3(BaseAPI):
 
         Returns:
             list: 预测结果列表，每个预测结果由预测框类别标签、
-              预测框类别名称、预测框坐标、预测框得分组成。
+              预测框类别名称、预测框坐标(坐标格式为[xmin, ymin, w, h]）、
+              预测框得分组成。
         """
         if transforms is None and not hasattr(self, 'test_transforms'):
             raise Exception("transforms need to be defined, now is None.")
@@ -359,12 +359,9 @@ class YOLOv3(BaseAPI):
             im, im_size = self.test_transforms(img_file)
         im = np.expand_dims(im, axis=0)
         im_size = np.expand_dims(im_size, axis=0)
-        outputs = self.exe.run(
-            self.test_prog,
-            feed={
-                'image': im,
-                'im_size': im_size
-            },
+        outputs = self.exe.run(self.test_prog,
+                               feed={'image': im,
+                                     'im_size': im_size},
                                fetch_list=list(self.test_outputs.values()),
                                return_numpy=False)
         res = {

paddlex/cv/nets/__init__.py

@@ -23,6 +23,7 @@ from .segmentation import DeepLabv3p
 from .xception import Xception
 from .densenet import DenseNet
 from .shufflenet_v2 import ShuffleNetV2
+from .hrnet import HRNet
 
 
 def resnet18(input, num_classes=1000):
@@ -51,14 +52,20 @@ def resnet50_vd(input, num_classes=1000):
 
 
 def resnet50_vd_ssld(input, num_classes=1000):
-    model = ResNet(layers=50, num_classes=num_classes, 
-                   variant='d', lr_mult_list=[1.0, 0.1, 0.2, 0.2, 0.3])
+    model = ResNet(
+        layers=50,
+        num_classes=num_classes,
+        variant='d',
+        lr_mult_list=[1.0, 0.1, 0.2, 0.2, 0.3])
     return model(input)
 
 
 def resnet101_vd_ssld(input, num_classes=1000):
-    model = ResNet(layers=101, num_classes=num_classes, 
-                   variant='d', lr_mult_list=[1.0, 0.1, 0.2, 0.2, 0.3])
+    model = ResNet(
+        layers=101,
+        num_classes=num_classes,
+        variant='d',
+        lr_mult_list=[1.0, 0.1, 0.2, 0.2, 0.3])
     return model(input)
 
 
@@ -93,13 +100,17 @@ def mobilenetv3_large(input, num_classes=1000):
 
 
 def mobilenetv3_small_ssld(input, num_classes=1000):
-    model = MobileNetV3(num_classes=num_classes, model_name='small',
+    model = MobileNetV3(
+        num_classes=num_classes,
+        model_name='small',
         lr_mult_list=[0.25, 0.25, 0.5, 0.5, 0.75])
     return model(input)
 
 
 def mobilenetv3_large_ssld(input, num_classes=1000):
-    model = MobileNetV3(num_classes=num_classes, model_name='large',
+    model = MobileNetV3(
+        num_classes=num_classes,
+        model_name='large',
         lr_mult_list=[0.25, 0.25, 0.5, 0.5, 0.75])
     return model(input)
 
@@ -133,6 +144,12 @@ def densenet201(input, num_classes=1000):
     model = DenseNet(layers=201, num_classes=num_classes)
     return model(input)
 
+
 def shufflenetv2(input, num_classes=1000):
     model = ShuffleNetV2(num_classes=num_classes)
     return model(input)
+
+
+def hrnet_w18(input, num_classes=1000):
+    model = HRNet(width=18, num_classes=num_classes)
+    return model(input)

paddlex/cv/nets/darknet.py

@@ -68,13 +68,14 @@ class DarkNet(object):
             bias_attr=False)
 
         bn_name = name + ".bn"
-
+        if self.num_classes:
+            regularizer = None
+        else:
+            regularizer = L2Decay(float(self.norm_decay))
         bn_param_attr = ParamAttr(
-            regularizer=L2Decay(float(self.norm_decay)),
-            name=bn_name + '.scale')
+            regularizer=regularizer, name=bn_name + '.scale')
         bn_bias_attr = ParamAttr(
-            regularizer=L2Decay(float(self.norm_decay)),
-            name=bn_name + '.offset')
+            regularizer=regularizer, name=bn_name + '.offset')
 
         out = fluid.layers.batch_norm(
             input=conv,

paddlex/cv/nets/detection/faster_rcnn.py

@@ -21,7 +21,7 @@ import copy
 
 from paddle import fluid
 
-from .fpn import FPN
+from .fpn import (FPN, HRFPN)
 from .rpn_head import (RPNHead, FPNRPNHead)
 from .roi_extractor import (RoIAlign, FPNRoIAlign)
 from .bbox_head import (BBoxHead, TwoFCHead)
@@ -82,6 +82,11 @@ class FasterRCNN(object):
         self.backbone = backbone
         self.mode = mode
         if with_fpn and fpn is None:
+            if self.backbone.__class__.__name__.startswith('HRNet'):
+                fpn = HRFPN()
+                fpn.min_level = 2
+                fpn.max_level = 6
+            else:
                 fpn = FPN()
         self.fpn = fpn
         self.num_classes = num_classes

paddlex/cv/nets/detection/fpn.py

@@ -23,7 +23,7 @@ from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.initializer import Xavier
 from paddle.fluid.regularizer import L2Decay
 
-__all__ = ['FPN']
+__all__ = ['FPN', 'HRFPN']
 
 
 def ConvNorm(input,
@@ -219,8 +219,8 @@ class FPN(object):
             body_name = body_name_list[i]
             body_input = body_dict[body_name]
             top_output = self.fpn_inner_output[i - 1]
-            fpn_inner_single = self._add_topdown_lateral(
-                body_name, body_input, top_output)
+            fpn_inner_single = self._add_topdown_lateral(body_name, body_input,
+                                                         top_output)
             self.fpn_inner_output[i] = fpn_inner_single
         fpn_dict = {}
         fpn_name_list = []
@@ -293,3 +293,107 @@ class FPN(object):
                 spatial_scale.insert(0, spatial_scale[0] * 0.5)
         res_dict = OrderedDict([(k, fpn_dict[k]) for k in fpn_name_list])
         return res_dict, spatial_scale
+
+
+class HRFPN(object):
+    """
+    HRNet, see https://arxiv.org/abs/1908.07919
+
+    Args:
+        num_chan (int): number of feature channels
+        pooling_type (str): pooling type of downsampling
+        share_conv (bool): whethet to share conv for different layers' reduction
+        spatial_scale (list): feature map scaling factor
+    """
+
+    def __init__(
+            self,
+            num_chan=256,
+            pooling_type="avg",
+            share_conv=False,
+            spatial_scale=[1. / 64, 1. / 32, 1. / 16, 1. / 8, 1. / 4], ):
+        self.num_chan = num_chan
+        self.pooling_type = pooling_type
+        self.share_conv = share_conv
+        self.spatial_scale = spatial_scale
+
+    def get_output(self, body_dict):
+        num_out = len(self.spatial_scale)
+        body_name_list = list(body_dict.keys())
+
+        num_backbone_stages = len(body_name_list)
+
+        outs = []
+        outs.append(body_dict[body_name_list[0]])
+
+        # resize
+        for i in range(1, len(body_dict)):
+            resized = self.resize_input_tensor(body_dict[body_name_list[i]],
+                                               outs[0], 2**i)
+            outs.append(resized)
+
+        # concat
+        out = fluid.layers.concat(outs, axis=1)
+
+        # reduction
+        out = fluid.layers.conv2d(
+            input=out,
+            num_filters=self.num_chan,
+            filter_size=1,
+            stride=1,
+            padding=0,
+            param_attr=ParamAttr(name='hrfpn_reduction_weights'),
+            bias_attr=False)
+
+        # conv
+        outs = [out]
+        for i in range(1, num_out):
+            outs.append(
+                self.pooling(
+                    out,
+                    size=2**i,
+                    stride=2**i,
+                    pooling_type=self.pooling_type))
+        outputs = []
+
+        for i in range(num_out):
+            conv_name = "shared_fpn_conv" if self.share_conv else "shared_fpn_conv_" + str(
+                i)
+            conv = fluid.layers.conv2d(
+                input=outs[i],
+                num_filters=self.num_chan,
+                filter_size=3,
+                stride=1,
+                padding=1,
+                param_attr=ParamAttr(name=conv_name + "_weights"),
+                bias_attr=False)
+            outputs.append(conv)
+
+        for idx in range(0, num_out - len(body_name_list)):
+            body_name_list.append("fpn_res5_sum_subsampled_{}x".format(2**(
+                idx + 1)))
+
+        outputs = outputs[::-1]
+        body_name_list = body_name_list[::-1]
+
+        res_dict = OrderedDict([(body_name_list[k], outputs[k])
+                                for k in range(len(body_name_list))])
+        return res_dict, self.spatial_scale
+
+    def resize_input_tensor(self, body_input, ref_output, scale):
+        shape = fluid.layers.shape(ref_output)
+        shape_hw = fluid.layers.slice(shape, axes=[0], starts=[2], ends=[4])
+        out_shape_ = shape_hw
+        out_shape = fluid.layers.cast(out_shape_, dtype='int32')
+        out_shape.stop_gradient = True
+        body_output = fluid.layers.resize_bilinear(
+            body_input, scale=scale, out_shape=out_shape)
+        return body_output
+
+    def pooling(self, input, size, stride, pooling_type):
+        pool = fluid.layers.pool2d(
+            input=input,
+            pool_size=size,
+            pool_stride=stride,
+            pool_type=pooling_type)
+        return pool

paddlex/cv/nets/detection/mask_rcnn.py

@@ -21,7 +21,7 @@ import copy
 
 import paddle.fluid as fluid
 
-from .fpn import FPN
+from .fpn import (FPN, HRFPN)
 from .rpn_head import (RPNHead, FPNRPNHead)
 from .roi_extractor import (RoIAlign, FPNRoIAlign)
 from .bbox_head import (BBoxHead, TwoFCHead)
@@ -92,8 +92,12 @@ class MaskRCNN(object):
         self.backbone = backbone
         self.mode = mode
         if with_fpn and fpn is None:
-            fpn = FPN(
-                num_chan=num_chan,
+            if self.backbone.__class__.__name__.startswith('HRNet'):
+                fpn = HRFPN()
+                fpn.min_level = 2
+                fpn.max_level = 6
+            else:
+                fpn = FPN(num_chan=num_chan,
                           min_level=min_level,
                           max_level=max_level,
                           spatial_scale=spatial_scale)

paddlex/cv/nets/hrnet.py

+# 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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from collections import OrderedDict
+
+from paddle import fluid
+from paddle.fluid.param_attr import ParamAttr
+from paddle.fluid.framework import Variable
+from paddle.fluid.regularizer import L2Decay
+
+from numbers import Integral
+from paddle.fluid.initializer import MSRA
+import math
+
+__all__ = ['HRNet']
+
+
+class HRNet(object):
+    def __init__(self,
+                 width=40,
+                 has_se=False,
+                 freeze_at=0,
+                 norm_type='bn',
+                 freeze_norm=False,
+                 norm_decay=0.,
+                 feature_maps=[2, 3, 4, 5],
+                 num_classes=None):
+        super(HRNet, self).__init__()
+
+        if isinstance(feature_maps, Integral):
+            feature_maps = [feature_maps]
+
+        assert 0 <= freeze_at <= 4, "freeze_at should be 0, 1, 2, 3 or 4"
+        assert len(feature_maps) > 0, "need one or more feature maps"
+        assert norm_type in ['bn', 'sync_bn']
+
+        self.width = width
+        self.has_se = has_se
+        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]],
+        }
+
+        self.freeze_at = freeze_at
+        self.norm_type = norm_type
+        self.norm_decay = norm_decay
+        self.freeze_norm = freeze_norm
+        self.feature_maps = feature_maps
+        self.num_classes = num_classes
+        self.end_points = []
+        return
+
+    def net(self, input, class_dim=1000):
+        width = self.width
+        channels_2, channels_3, channels_4 = self.channels[width]
+        num_modules_2, num_modules_3, num_modules_4 = 1, 4, 3
+
+        x = self.conv_bn_layer(
+            input=input,
+            filter_size=3,
+            num_filters=64,
+            stride=2,
+            if_act=True,
+            name='layer1_1')
+        x = self.conv_bn_layer(
+            input=x,
+            filter_size=3,
+            num_filters=64,
+            stride=2,
+            if_act=True,
+            name='layer1_2')
+
+        la1 = self.layer1(x, name='layer2')
+        tr1 = self.transition_layer([la1], [256], channels_2, name='tr1')
+        st2 = self.stage(tr1, num_modules_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')
+        tr3 = self.transition_layer(st3, channels_3, channels_4, name='tr3')
+        st4 = self.stage(tr3, num_modules_4, channels_4, name='st4')
+
+        # classification
+        if self.num_classes:
+            last_cls = self.last_cls_out(x=st4, name='cls_head')
+            y = last_cls[0]
+            last_num_filters = [256, 512, 1024]
+            for i in range(3):
+                y = fluid.layers.elementwise_add(
+                    last_cls[i + 1],
+                    self.conv_bn_layer(
+                        input=y,
+                        filter_size=3,
+                        num_filters=last_num_filters[i],
+                        stride=2,
+                        name='cls_head_add' + str(i + 1)))
+
+            y = self.conv_bn_layer(
+                input=y,
+                filter_size=1,
+                num_filters=2048,
+                stride=1,
+                name='cls_head_last_conv')
+            pool = fluid.layers.pool2d(
+                input=y, pool_type='avg', global_pooling=True)
+            stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
+            out = fluid.layers.fc(
+                input=pool,
+                size=class_dim,
+                param_attr=ParamAttr(
+                    name='fc_weights',
+                    initializer=fluid.initializer.Uniform(-stdv, stdv)),
+                bias_attr=ParamAttr(name='fc_offset'))
+            return out
+
+        # segmentation
+        if self.feature_maps == "stage4":
+            return st4
+
+        self.end_points = st4
+        return st4[-1]
+
+    def layer1(self, input, name=None):
+        conv = input
+        for i in range(4):
+            conv = self.bottleneck_block(
+                conv,
+                num_filters=64,
+                downsample=True if i == 0 else False,
+                name=name + '_' + str(i + 1))
+        return conv
+
+    def transition_layer(self, x, in_channels, out_channels, name=None):
+        num_in = len(in_channels)
+        num_out = len(out_channels)
+        out = []
+        for i in range(num_out):
+            if i < num_in:
+                if in_channels[i] != out_channels[i]:
+                    residual = self.conv_bn_layer(
+                        x[i],
+                        filter_size=3,
+                        num_filters=out_channels[i],
+                        name=name + '_layer_' + str(i + 1))
+                    out.append(residual)
+                else:
+                    out.append(x[i])
+            else:
+                residual = self.conv_bn_layer(
+                    x[-1],
+                    filter_size=3,
+                    num_filters=out_channels[i],
+                    stride=2,
+                    name=name + '_layer_' + str(i + 1))
+                out.append(residual)
+        return out
+
+    def branches(self, x, block_num, channels, name=None):
+        out = []
+        for i in range(len(channels)):
+            residual = x[i]
+            for j in range(block_num):
+                residual = self.basic_block(
+                    residual,
+                    channels[i],
+                    name=name + '_branch_layer_' + str(i + 1) + '_' +
+                    str(j + 1))
+            out.append(residual)
+        return out
+
+    def fuse_layers(self, x, channels, multi_scale_output=True, name=None):
+        out = []
+        for i in range(len(channels) if multi_scale_output else 1):
+            residual = x[i]
+            if self.feature_maps == "stage4":
+                shape = fluid.layers.shape(residual)
+                width = shape[-1]
+                height = shape[-2]
+            for j in range(len(channels)):
+                if j > i:
+                    y = self.conv_bn_layer(
+                        x[j],
+                        filter_size=1,
+                        num_filters=channels[i],
+                        if_act=False,
+                        name=name + '_layer_' + str(i + 1) + '_' + str(j + 1))
+                    if self.feature_maps == "stage4":
+                        y = fluid.layers.resize_bilinear(
+                            input=y, out_shape=[height, width])
+                    else:
+                        y = fluid.layers.resize_nearest(
+                            input=y, scale=2**(j - i))
+                    residual = fluid.layers.elementwise_add(
+                        x=residual, y=y, act=None)
+                elif j < i:
+                    y = x[j]
+                    for k in range(i - j):
+                        if k == i - j - 1:
+                            y = self.conv_bn_layer(
+                                y,
+                                filter_size=3,
+                                num_filters=channels[i],
+                                stride=2,
+                                if_act=False,
+                                name=name + '_layer_' + str(i + 1) + '_' +
+                                str(j + 1) + '_' + str(k + 1))
+                        else:
+                            y = self.conv_bn_layer(
+                                y,
+                                filter_size=3,
+                                num_filters=channels[j],
+                                stride=2,
+                                name=name + '_layer_' + str(i + 1) + '_' +
+                                str(j + 1) + '_' + str(k + 1))
+                    residual = fluid.layers.elementwise_add(
+                        x=residual, y=y, act=None)
+
+            residual = fluid.layers.relu(residual)
+            out.append(residual)
+        return out
+
+    def high_resolution_module(self,
+                               x,
+                               channels,
+                               multi_scale_output=True,
+                               name=None):
+        residual = self.branches(x, 4, channels, name=name)
+        out = self.fuse_layers(
+            residual,
+            channels,
+            multi_scale_output=multi_scale_output,
+            name=name)
+        return out
+
+    def stage(self,
+              x,
+              num_modules,
+              channels,
+              multi_scale_output=True,
+              name=None):
+        out = x
+        for i in range(num_modules):
+            if i == num_modules - 1 and multi_scale_output == False:
+                out = self.high_resolution_module(
+                    out,
+                    channels,
+                    multi_scale_output=False,
+                    name=name + '_' + str(i + 1))
+            else:
+                out = self.high_resolution_module(
+                    out, channels, name=name + '_' + str(i + 1))
+
+        return out
+
+    def last_cls_out(self, x, name=None):
+        out = []
+        num_filters_list = [32, 64, 128, 256]
+        for i in range(len(x)):
+            out.append(
+                self.bottleneck_block(
+                    input=x[i],
+                    num_filters=num_filters_list[i],
+                    name=name + 'conv_' + str(i + 1),
+                    downsample=True))
+        return out
+
+    def basic_block(self,
+                    input,
+                    num_filters,
+                    stride=1,
+                    downsample=False,
+                    name=None):
+        residual = input
+        conv = self.conv_bn_layer(
+            input=input,
+            filter_size=3,
+            num_filters=num_filters,
+            stride=stride,
+            name=name + '_conv1')
+        conv = self.conv_bn_layer(
+            input=conv,
+            filter_size=3,
+            num_filters=num_filters,
+            if_act=False,
+            name=name + '_conv2')
+        if downsample:
+            residual = self.conv_bn_layer(
+                input=input,
+                filter_size=1,
+                num_filters=num_filters,
+                if_act=False,
+                name=name + '_downsample')
+        if self.has_se:
+            conv = self.squeeze_excitation(
+                input=conv,
+                num_channels=num_filters,
+                reduction_ratio=16,
+                name=name + '_fc')
+        return fluid.layers.elementwise_add(x=residual, y=conv, act='relu')
+
+    def bottleneck_block(self,
+                         input,
+                         num_filters,
+                         stride=1,
+                         downsample=False,
+                         name=None):
+        residual = input
+        conv = self.conv_bn_layer(
+            input=input,
+            filter_size=1,
+            num_filters=num_filters,
+            name=name + '_conv1')
+        conv = self.conv_bn_layer(
+            input=conv,
+            filter_size=3,
+            num_filters=num_filters,
+            stride=stride,
+            name=name + '_conv2')
+        conv = self.conv_bn_layer(
+            input=conv,
+            filter_size=1,
+            num_filters=num_filters * 4,
+            if_act=False,
+            name=name + '_conv3')
+        if downsample:
+            residual = self.conv_bn_layer(
+                input=input,
+                filter_size=1,
+                num_filters=num_filters * 4,
+                if_act=False,
+                name=name + '_downsample')
+        if self.has_se:
+            conv = self.squeeze_excitation(
+                input=conv,
+                num_channels=num_filters * 4,
+                reduction_ratio=16,
+                name=name + '_fc')
+        return fluid.layers.elementwise_add(x=residual, y=conv, act='relu')
+
+    def squeeze_excitation(self,
+                           input,
+                           num_channels,
+                           reduction_ratio,
+                           name=None):
+        pool = fluid.layers.pool2d(
+            input=input, pool_size=0, pool_type='avg', global_pooling=True)
+        stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
+        squeeze = fluid.layers.fc(
+            input=pool,
+            size=num_channels / reduction_ratio,
+            act='relu',
+            param_attr=fluid.param_attr.ParamAttr(
+                initializer=fluid.initializer.Uniform(-stdv, stdv),
+                name=name + '_sqz_weights'),
+            bias_attr=ParamAttr(name=name + '_sqz_offset'))
+        stdv = 1.0 / math.sqrt(squeeze.shape[1] * 1.0)
+        excitation = fluid.layers.fc(
+            input=squeeze,
+            size=num_channels,
+            act='sigmoid',
+            param_attr=fluid.param_attr.ParamAttr(
+                initializer=fluid.initializer.Uniform(-stdv, stdv),
+                name=name + '_exc_weights'),
+            bias_attr=ParamAttr(name=name + '_exc_offset'))
+        scale = fluid.layers.elementwise_mul(x=input, y=excitation, axis=0)
+        return scale
+
+    def conv_bn_layer(self,
+                      input,
+                      filter_size,
+                      num_filters,
+                      stride=1,
+                      padding=1,
+                      num_groups=1,
+                      if_act=True,
+                      name=None):
+        conv = fluid.layers.conv2d(
+            input=input,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=(filter_size - 1) // 2,
+            groups=num_groups,
+            act=None,
+            param_attr=ParamAttr(
+                initializer=MSRA(), name=name + '_weights'),
+            bias_attr=False)
+        bn_name = name + '_bn'
+        bn = self._bn(input=conv, bn_name=bn_name)
+        if if_act:
+            bn = fluid.layers.relu(bn)
+        return bn
+
+    def _bn(self, input, act=None, bn_name=None):
+        norm_lr = 0. if self.freeze_norm else 1.
+        norm_decay = self.norm_decay
+        if self.num_classes or self.feature_maps == "stage4":
+            regularizer = None
+            pattr_initializer = fluid.initializer.Constant(1.0)
+            battr_initializer = fluid.initializer.Constant(0.0)
+        else:
+            regularizer = L2Decay(norm_decay)
+            pattr_initializer = None
+            battr_initializer = None
+        pattr = ParamAttr(
+            name=bn_name + '_scale',
+            learning_rate=norm_lr,
+            regularizer=regularizer,
+            initializer=pattr_initializer)
+        battr = ParamAttr(
+            name=bn_name + '_offset',
+            learning_rate=norm_lr,
+            regularizer=regularizer,
+            initializer=battr_initializer)
+
+        global_stats = True if self.freeze_norm else False
+        out = fluid.layers.batch_norm(
+            input=input,
+            act=act,
+            name=bn_name + '.output.1',
+            param_attr=pattr,
+            bias_attr=battr,
+            moving_mean_name=bn_name + '_mean',
+            moving_variance_name=bn_name + '_variance',
+            use_global_stats=global_stats)
+        scale = fluid.framework._get_var(pattr.name)
+        bias = fluid.framework._get_var(battr.name)
+        if self.freeze_norm:
+            scale.stop_gradient = True
+            bias.stop_gradient = True
+        return out
+
+    def __call__(self, input):
+        assert isinstance(input, Variable)
+        if isinstance(self.feature_maps, (list, tuple)):
+            assert not (set(self.feature_maps) - set([2, 3, 4, 5])), \
+                "feature maps {} not in [2, 3, 4, 5]".format(self.feature_maps)
+
+        res_endpoints = []
+
+        res = input
+        feature_maps = self.feature_maps
+        out = self.net(input)
+        if self.num_classes or self.feature_maps == "stage4":
+            return out
+
+        for i in feature_maps:
+            res = self.end_points[i - 2]
+            if i in self.feature_maps:
+                res_endpoints.append(res)
+            if self.freeze_at >= i:
+                res.stop_gradient = True
+
+        return OrderedDict([('res{}_sum'.format(self.feature_maps[idx]), feat)
+                            for idx, feat in enumerate(res_endpoints)])

paddlex/cv/nets/mobilenet_v1.py

@@ -79,10 +79,14 @@ class MobileNetV1(object):
 
         bn_name = name + "_bn"
         norm_decay = self.norm_decay
+        if self.num_classes:
+            regularizer = None
+        else:
+            regularizer = L2Decay(norm_decay)
         bn_param_attr = ParamAttr(
-            regularizer=L2Decay(norm_decay), name=bn_name + '_scale')
+            regularizer=regularizer, name=bn_name + '_scale')
         bn_bias_attr = ParamAttr(
-            regularizer=L2Decay(norm_decay), name=bn_name + '_offset')
+            regularizer=regularizer, name=bn_name + '_offset')
         return fluid.layers.batch_norm(
             input=conv,
             act=act,
@@ -189,11 +193,11 @@ class MobileNetV1(object):
         if self.num_classes:
             out = fluid.layers.pool2d(
                 input=out, pool_type='avg', global_pooling=True)
-            output = fluid.layers.fc(
-                input=out,
+            output = fluid.layers.fc(input=out,
                                      size=self.num_classes,
                                      param_attr=ParamAttr(
-                    initializer=fluid.initializer.MSRA(), name="fc7_weights"),
+                                         initializer=fluid.initializer.MSRA(),
+                                         name="fc7_weights"),
                                      bias_attr=ParamAttr(name="fc7_offset"))
             return output
 

paddlex/cv/nets/mobilenet_v3.py

@@ -31,6 +31,7 @@ class MobileNetV3():
         with_extra_blocks (bool): if extra blocks should be added.
         extra_block_filters (list): number of filter for each extra block.
     """
+
     def __init__(self,
                  scale=1.0,
                  model_name='small',
@@ -113,10 +114,16 @@ class MobileNetV3():
         lr_idx = self.curr_stage // self.lr_interval
         lr_idx = min(lr_idx, len(self.lr_mult_list) - 1)
         lr_mult = self.lr_mult_list[lr_idx]
-        conv_param_attr = ParamAttr(name=name + '_weights',
+        if self.num_classes:
+            regularizer = None
+        else:
+            regularizer = L2Decay(self.conv_decay)
+        conv_param_attr = ParamAttr(
+            name=name + '_weights',
             learning_rate=lr_mult,
-                                    regularizer=L2Decay(self.conv_decay))
-        conv = fluid.layers.conv2d(input=input,
+            regularizer=regularizer)
+        conv = fluid.layers.conv2d(
+            input=input,
             num_filters=num_filters,
             filter_size=filter_size,
             stride=stride,
@@ -127,11 +134,12 @@ class MobileNetV3():
             param_attr=conv_param_attr,
             bias_attr=False)
         bn_name = name + '_bn'
-        bn_param_attr = ParamAttr(name=bn_name + "_scale",
-                                  regularizer=L2Decay(self.norm_decay))
-        bn_bias_attr = ParamAttr(name=bn_name + "_offset",
-                                 regularizer=L2Decay(self.norm_decay))
-        bn = fluid.layers.batch_norm(input=conv,
+        bn_param_attr = ParamAttr(
+            name=bn_name + "_scale", regularizer=L2Decay(self.norm_decay))
+        bn_bias_attr = ParamAttr(
+            name=bn_name + "_offset", regularizer=L2Decay(self.norm_decay))
+        bn = fluid.layers.batch_norm(
+            input=conv,
             param_attr=bn_param_attr,
             bias_attr=bn_bias_attr,
             moving_mean_name=bn_name + '_mean',
@@ -154,10 +162,8 @@ class MobileNetV3():
         lr_mult = self.lr_mult_list[lr_idx]
 
         num_mid_filter = int(num_out_filter // ratio)
-        pool = fluid.layers.pool2d(input=input,
-                                   pool_type='avg',
-                                   global_pooling=True,
-                                   use_cudnn=False)
+        pool = fluid.layers.pool2d(
+            input=input, pool_type='avg', global_pooling=True, use_cudnn=False)
         conv1 = fluid.layers.conv2d(
             input=pool,
             filter_size=1,
@@ -191,7 +197,8 @@ class MobileNetV3():
                        use_se=False,
                        name=None):
         input_data = input
-        conv0 = self._conv_bn_layer(input=input,
+        conv0 = self._conv_bn_layer(
+            input=input,
             filter_size=1,
             num_filters=num_mid_filter,
             stride=1,
@@ -201,7 +208,8 @@ class MobileNetV3():
             name=name + '_expand')
         if self.block_stride == 16 and stride == 2:
             self.end_points.append(conv0)
-        conv1 = self._conv_bn_layer(input=conv0,
+        conv1 = self._conv_bn_layer(
+            input=conv0,
             filter_size=filter_size,
             num_filters=num_mid_filter,
             stride=stride,
@@ -213,11 +221,11 @@ class MobileNetV3():
             name=name + '_depthwise')
 
         if use_se:
-            conv1 = self._se_block(input=conv1,
-                                   num_out_filter=num_mid_filter,
-                                   name=name + '_se')
+            conv1 = self._se_block(
+                input=conv1, num_out_filter=num_mid_filter, name=name + '_se')
 
-        conv2 = self._conv_bn_layer(input=conv1,
+        conv2 = self._conv_bn_layer(
+            input=conv1,
             filter_size=1,
             num_filters=num_out_filter,
             stride=1,
@@ -227,7 +235,8 @@ class MobileNetV3():
         if num_in_filter != num_out_filter or stride != 1:
             return conv2
         else:
-            return fluid.layers.elementwise_add(x=input_data, y=conv2, act=None)
+            return fluid.layers.elementwise_add(
+                x=input_data, y=conv2, act=None)
 
     def _extra_block_dw(self,
                         input,
@@ -235,14 +244,16 @@ class MobileNetV3():
                         num_filters2,
                         stride,
                         name=None):
-        pointwise_conv = self._conv_bn_layer(input=input,
+        pointwise_conv = self._conv_bn_layer(
+            input=input,
             filter_size=1,
             num_filters=int(num_filters1),
             stride=1,
             padding="SAME",
             act='relu6',
             name=name + "_extra1")
-        depthwise_conv = self._conv_bn_layer(input=pointwise_conv,
+        depthwise_conv = self._conv_bn_layer(
+            input=pointwise_conv,
             filter_size=3,
             num_filters=int(num_filters2),
             stride=stride,
@@ -251,7 +262,8 @@ class MobileNetV3():
             act='relu6',
             use_cudnn=False,
             name=name + "_extra2_dw")
-        normal_conv = self._conv_bn_layer(input=depthwise_conv,
+        normal_conv = self._conv_bn_layer(
+            input=depthwise_conv,
             filter_size=1,
             num_filters=int(num_filters2),
             stride=1,
@@ -282,7 +294,8 @@ class MobileNetV3():
             self.block_stride *= layer_cfg[5]
             if layer_cfg[5] == 2:
                 blocks.append(conv)
-            conv = self._residual_unit(input=conv,
+            conv = self._residual_unit(
+                input=conv,
                 num_in_filter=inplanes,
                 num_mid_filter=int(scale * layer_cfg[1]),
                 num_out_filter=int(scale * layer_cfg[2]),
@@ -298,7 +311,8 @@ class MobileNetV3():
         blocks.append(conv)
 
         if self.num_classes:
-            conv = self._conv_bn_layer(input=conv,
+            conv = self._conv_bn_layer(
+                input=conv,
                 filter_size=1,
                 num_filters=int(scale * self.cls_ch_squeeze),
                 stride=1,
@@ -308,7 +322,8 @@ class MobileNetV3():
                 act='hard_swish',
                 name='conv_last')
 
-            conv = fluid.layers.pool2d(input=conv,
+            conv = fluid.layers.pool2d(
+                input=conv,
                 pool_type='avg',
                 global_pooling=True,
                 use_cudnn=False)
@@ -333,7 +348,8 @@ class MobileNetV3():
             return blocks
 
         # extra block
-        conv_extra = self._conv_bn_layer(conv,
+        conv_extra = self._conv_bn_layer(
+            conv,
             filter_size=1,
             num_filters=int(scale * cfg[-1][1]),
             stride=1,

paddlex/cv/nets/resnet.py

@@ -135,8 +135,10 @@ class ResNet(object):
             filter_size=filter_size,
             stride=stride,
             padding=padding,
-            param_attr=ParamAttr(initializer=Constant(0.0), name=name + ".w_0"),
-            bias_attr=ParamAttr(initializer=Constant(0.0), name=name + ".b_0"),
+            param_attr=ParamAttr(
+                initializer=Constant(0.0), name=name + ".w_0"),
+            bias_attr=ParamAttr(
+                initializer=Constant(0.0), name=name + ".b_0"),
             act=act,
             name=name)
         return out
@@ -151,7 +153,8 @@ class ResNet(object):
                    name=None,
                    dcn_v2=False,
                    use_lr_mult_list=False):
-        lr_mult = self.lr_mult_list[self.curr_stage] if use_lr_mult_list else 1.0
+        lr_mult = self.lr_mult_list[
+            self.curr_stage] if use_lr_mult_list else 1.0
         _name = self.prefix_name + name if self.prefix_name != '' else name
         if not dcn_v2:
             conv = fluid.layers.conv2d(
@@ -162,8 +165,8 @@ class ResNet(object):
                 padding=(filter_size - 1) // 2,
                 groups=groups,
                 act=None,
-                param_attr=ParamAttr(name=_name + "_weights",
-                                     learning_rate=lr_mult),
+                param_attr=ParamAttr(
+                    name=_name + "_weights", learning_rate=lr_mult),
                 bias_attr=False,
                 name=_name + '.conv2d.output.1')
         else:
@@ -202,14 +205,18 @@ class ResNet(object):
 
         norm_lr = 0. if self.freeze_norm else lr_mult
         norm_decay = self.norm_decay
+        if self.num_classes:
+            regularizer = None
+        else:
+            regularizer = L2Decay(norm_decay)
         pattr = ParamAttr(
             name=bn_name + '_scale',
             learning_rate=norm_lr,
-            regularizer=L2Decay(norm_decay))
+            regularizer=regularizer)
         battr = ParamAttr(
             name=bn_name + '_offset',
             learning_rate=norm_lr,
-            regularizer=L2Decay(norm_decay))
+            regularizer=regularizer)
 
         if self.norm_type in ['bn', 'sync_bn']:
             global_stats = True if self.freeze_norm else False
@@ -262,8 +269,8 @@ class ResNet(object):
                     pool_padding=0,
                     ceil_mode=True,
                     pool_type='avg')
-                return self._conv_norm(input, ch_out, 1, 1, name=name,
-                                      use_lr_mult_list=True)
+                return self._conv_norm(
+                    input, ch_out, 1, 1, name=name, use_lr_mult_list=True)
             return self._conv_norm(input, ch_out, 1, stride, name=name)
         else:
             return input

paddlex/cv/nets/segmentation/__init__.py

@@ -14,5 +14,6 @@
 
 from .unet import UNet
 from .deeplabv3p import DeepLabv3p
+from .hrnet import HRNet
 from .model_utils import libs
 from .model_utils import loss

paddlex/cv/nets/segmentation/deeplabv3p.py

@@ -135,7 +135,8 @@ class DeepLabv3p(object):
         param_attr = fluid.ParamAttr(
             name=name_scope + 'weights',
             regularizer=None,
-            initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.06))
+            initializer=fluid.initializer.TruncatedNormal(
+                loc=0.0, scale=0.06))
         with scope('encoder'):
             channel = 256
             with scope("image_pool"):
@@ -151,8 +152,8 @@ class DeepLabv3p(object):
                         padding=0,
                         param_attr=param_attr))
                 input_shape = fluid.layers.shape(input)
-                image_avg = fluid.layers.resize_bilinear(
-                    image_avg, input_shape[2:])
+                image_avg = fluid.layers.resize_bilinear(image_avg,
+                                                         input_shape[2:])
 
             with scope("aspp0"):
                 aspp0 = bn_relu(
@@ -244,7 +245,8 @@ class DeepLabv3p(object):
         param_attr = fluid.ParamAttr(
             name=name_scope + 'weights',
             regularizer=None,
-            initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.06))
+            initializer=fluid.initializer.TruncatedNormal(
+                loc=0.0, scale=0.06))
         with scope('decoder'):
             with scope('concat'):
                 decode_shortcut = bn_relu(
@@ -326,9 +328,6 @@ class DeepLabv3p(object):
         if self.mode == 'train':
             inputs['label'] = fluid.data(
                 dtype='int32', shape=[None, 1, None, None], name='label')
-        elif self.mode == 'eval':
-            inputs['label'] = fluid.data(
-                dtype='int32', shape=[None, 1, None, None], name='label')
         return inputs
 
     def build_net(self, inputs):
@@ -351,7 +350,8 @@ class DeepLabv3p(object):
             name=name_scope + 'weights',
             regularizer=fluid.regularizer.L2DecayRegularizer(
                 regularization_coeff=0.0),
-            initializer=fluid.initializer.TruncatedNormal(loc=0.0, scale=0.01))
+            initializer=fluid.initializer.TruncatedNormal(
+                loc=0.0, scale=0.01))
         with scope('logit'):
             with fluid.name_scope('last_conv'):
                 logit = conv(

paddlex/cv/nets/segmentation/hrnet.py

+# coding: utf8
+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from collections import OrderedDict
+
+import paddle.fluid as fluid
+from paddle.fluid.initializer import MSRA
+from paddle.fluid.param_attr import ParamAttr
+from .model_utils.libs import sigmoid_to_softmax
+from .model_utils.loss import softmax_with_loss
+from .model_utils.loss import dice_loss
+from .model_utils.loss import bce_loss
+import paddlex
+import paddlex.utils.logging as logging
+
+
+class HRNet(object):
+    def __init__(self,
+                 num_classes,
+                 mode='train',
+                 width=18,
+                 use_bce_loss=False,
+                 use_dice_loss=False,
+                 class_weight=None,
+                 ignore_index=255):
+        # dice_loss或bce_loss只适用两类分割中
+        if num_classes > 2 and (use_bce_loss or use_dice_loss):
+            raise ValueError(
+                "dice loss and bce loss is only applicable to binary classfication"
+            )
+
+        if class_weight is not None:
+            if isinstance(class_weight, list):
+                if len(class_weight) != num_classes:
+                    raise ValueError(
+                        "Length of class_weight should be equal to number of classes"
+                    )
+            elif isinstance(class_weight, str):
+                if class_weight.lower() != 'dynamic':
+                    raise ValueError(
+                        "if class_weight is string, must be dynamic!")
+            else:
+                raise TypeError(
+                    'Expect class_weight is a list or string but receive {}'.
+                    format(type(class_weight)))
+
+        self.num_classes = num_classes
+        self.mode = mode
+        self.use_bce_loss = use_bce_loss
+        self.use_dice_loss = use_dice_loss
+        self.class_weight = class_weight
+        self.ignore_index = ignore_index
+        self.backbone = paddlex.cv.nets.hrnet.HRNet(
+            width=width, feature_maps="stage4")
+
+    def build_net(self, inputs):
+        if self.use_dice_loss or self.use_bce_loss:
+            self.num_classes = 1
+        image = inputs['image']
+        st4 = self.backbone(image)
+        # upsample
+        shape = fluid.layers.shape(st4[0])[-2:]
+        st4[1] = fluid.layers.resize_bilinear(st4[1], out_shape=shape)
+        st4[2] = fluid.layers.resize_bilinear(st4[2], out_shape=shape)
+        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])
+
+        out = self._conv_bn_layer(
+            input=out,
+            filter_size=1,
+            num_filters=last_channels,
+            stride=1,
+            if_act=True,
+            name='conv-2')
+        out = fluid.layers.conv2d(
+            input=out,
+            num_filters=self.num_classes,
+            filter_size=1,
+            stride=1,
+            padding=0,
+            act=None,
+            param_attr=ParamAttr(
+                initializer=MSRA(), name='conv-1_weights'),
+            bias_attr=False)
+
+        input_shape = fluid.layers.shape(image)[-2:]
+        logit = fluid.layers.resize_bilinear(out, input_shape)
+
+        if self.num_classes == 1:
+            out = sigmoid_to_softmax(logit)
+            out = fluid.layers.transpose(out, [0, 2, 3, 1])
+        else:
+            out = fluid.layers.transpose(logit, [0, 2, 3, 1])
+
+        pred = fluid.layers.argmax(out, axis=3)
+        pred = fluid.layers.unsqueeze(pred, axes=[3])
+
+        if self.mode == 'train':
+            label = inputs['label']
+            mask = label != self.ignore_index
+            return self._get_loss(logit, label, mask)
+        elif self.mode == 'eval':
+            label = inputs['label']
+            mask = label != self.ignore_index
+            loss = self._get_loss(logit, label, mask)
+            return loss, pred, label, mask
+        else:
+            if self.num_classes == 1:
+                logit = sigmoid_to_softmax(logit)
+            else:
+                logit = fluid.layers.softmax(logit, axis=1)
+            return pred, logit
+
+    def generate_inputs(self):
+        inputs = OrderedDict()
+        inputs['image'] = fluid.data(
+            dtype='float32', shape=[None, 3, None, None], name='image')
+        if self.mode == 'train':
+            inputs['label'] = fluid.data(
+                dtype='int32', shape=[None, 1, None, None], name='label')
+        elif self.mode == 'eval':
+            inputs['label'] = fluid.data(
+                dtype='int32', shape=[None, 1, None, None], name='label')
+        return inputs
+
+    def _get_loss(self, logit, label, mask):
+        avg_loss = 0
+        if not (self.use_dice_loss or self.use_bce_loss):
+            avg_loss += softmax_with_loss(
+                logit,
+                label,
+                mask,
+                num_classes=self.num_classes,
+                weight=self.class_weight,
+                ignore_index=self.ignore_index)
+        else:
+            if self.use_dice_loss:
+                avg_loss += dice_loss(logit, label, mask)
+            if self.use_bce_loss:
+                avg_loss += bce_loss(
+                    logit, label, mask, ignore_index=self.ignore_index)
+
+        return avg_loss
+
+    def _conv_bn_layer(self,
+                       input,
+                       filter_size,
+                       num_filters,
+                       stride=1,
+                       padding=1,
+                       num_groups=1,
+                       if_act=True,
+                       name=None):
+        conv = fluid.layers.conv2d(
+            input=input,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            stride=stride,
+            padding=(filter_size - 1) // 2,
+            groups=num_groups,
+            act=None,
+            param_attr=ParamAttr(
+                initializer=MSRA(), name=name + '_weights'),
+            bias_attr=False)
+        bn_name = name + '_bn'
+        bn = fluid.layers.batch_norm(
+            input=conv,
+            param_attr=ParamAttr(
+                name=bn_name + "_scale",
+                initializer=fluid.initializer.Constant(1.0)),
+            bias_attr=ParamAttr(
+                name=bn_name + "_offset",
+                initializer=fluid.initializer.Constant(0.0)),
+            moving_mean_name=bn_name + '_mean',
+            moving_variance_name=bn_name + '_variance')
+        if if_act:
+            bn = fluid.layers.relu(bn)
+        return bn

paddlex/cv/transforms/cls_transforms.py

@@ -92,6 +92,12 @@ class Compose(ClsTransform):
                     outputs = (im, label)
         return outputs
 
+    def add_augmenters(self, augmenters):
+        if not isinstance(augmenters, list):
+            raise Exception(
+                "augmenters should be list type in func add_augmenters()")
+        self.transforms = augmenters + self.transforms.transforms
+
 
 class RandomCrop(ClsTransform):
     """对图像进行随机剪裁，模型训练时的数据增强操作。
@@ -461,3 +467,56 @@ class ArrangeClassifier(ClsTransform):
         else:
             outputs = (im, )
         return outputs
+
+
+class ComposedClsTransforms(Compose):
+    """ 分类模型的基础Transforms流程，具体如下
+        训练阶段：
+        1. 随机从图像中crop一块子图，并resize成crop_size大小
+        2. 将1的输出按0.5的概率随机进行水平翻转
+        3. 将图像进行归一化
+        验证/预测阶段：
+        1. 将图像按比例Resize，使得最小边长度为crop_size[0] * 1.14
+        2. 从图像中心crop出一个大小为crop_size的图像
+        3. 将图像进行归一化
+
+        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'
+            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
+    """
+
+    def __init__(self,
+                 mode,
+                 crop_size=[224, 224],
+                 mean=[0.485, 0.456, 0.406],
+                 std=[0.229, 0.224, 0.225]):
+        width = crop_size
+        if isinstance(crop_size, list):
+            if crop_size[0] != crop_size[1]:
+                raise Exception(
+                    "In classifier model, width and height should be equal, please modify your parameter `crop_size`"
+                )
+            width = crop_size[0]
+        if width % 32 != 0:
+            raise Exception(
+                "In classifier model, width and height should be multiple of 32, e.g 224、256、320...., please modify your parameter `crop_size`"
+            )
+
+        if mode == 'train':
+            # 训练时的transforms，包含数据增强
+            transforms = [
+                RandomCrop(crop_size=width), RandomHorizontalFlip(prob=0.5),
+                Normalize(
+                    mean=mean, std=std)
+            ]
+        else:
+            # 验证/预测时的transforms
+            transforms = [
+                ResizeByShort(short_size=int(width * 1.14)),
+                CenterCrop(crop_size=width), Normalize(
+                    mean=mean, std=std)
+            ]
+
+        super(ComposedClsTransforms, self).__init__(transforms)

paddlex/cv/transforms/det_transforms.py

@@ -152,6 +152,12 @@ class Compose(DetTransform):
                     outputs = (im, im_info)
         return outputs
 
+    def add_augmenters(self, augmenters):
+        if not isinstance(augmenters, list):
+            raise Exception(
+                "augmenters should be list type in func add_augmenters()")
+        self.transforms = augmenters + self.transforms.transforms
+
 
 class ResizeByShort(DetTransform):
     """根据图像的短边调整图像大小（resize）。
@@ -1227,3 +1233,108 @@ class ArrangeYOLOv3(DetTransform):
             im_shape = im_info['image_shape']
             outputs = (im, im_shape)
         return outputs
+
+
+class ComposedRCNNTransforms(Compose):
+    """ RCNN模型(faster-rcnn/mask-rcnn)图像处理流程，具体如下，
+        训练阶段：
+        1. 随机以0.5的概率将图像水平翻转
+        2. 图像归一化
+        3. 图像按比例Resize，scale计算方式如下
+            scale = min_max_size[0] / short_size_of_image
+            if max_size_of_image * scale > min_max_size[1]:
+                scale = min_max_size[1] / max_size_of_image
+        4. 将3步骤的长宽进行padding，使得长宽为32的倍数
+        验证阶段：
+        1. 图像归一化
+        2. 图像按比例Resize，scale计算方式同上训练阶段
+        3. 将2步骤的长宽进行padding，使得长宽为32的倍数
+
+        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'
+            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
+    """
+
+    def __init__(self,
+                 mode,
+                 min_max_size=[800, 1333],
+                 mean=[0.485, 0.456, 0.406],
+                 std=[0.229, 0.224, 0.225]):
+        if mode == 'train':
+            # 训练时的transforms，包含数据增强
+            transforms = [
+                RandomHorizontalFlip(prob=0.5), Normalize(
+                    mean=mean, std=std), ResizeByShort(
+                        short_size=min_max_size[0], max_size=min_max_size[1]),
+                Padding(coarsest_stride=32)
+            ]
+        else:
+            # 验证/预测时的transforms
+            transforms = [
+                Normalize(
+                    mean=mean, std=std), ResizeByShort(
+                        short_size=min_max_size[0], max_size=min_max_size[1]),
+                Padding(coarsest_stride=32)
+            ]
+
+        super(ComposedRCNNTransforms, self).__init__(transforms)
+
+
+class ComposedYOLOTransforms(Compose):
+    """YOLOv3模型的图像预处理流程，具体如下，
+        训练阶段：
+        1. 在前mixup_epoch轮迭代中，使用MixupImage策略，见https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/det_transforms.html#mixupimage
+        2. 对图像进行随机扰动，包括亮度，对比度，饱和度和色调
+        3. 随机扩充图像，见https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/det_transforms.html#randomexpand
+        4. 随机裁剪图像
+        5. 将4步骤的输出图像Resize成shape参数的大小
+        6. 随机0.5的概率水平翻转图像
+        7. 图像归一化
+        验证/预测阶段：
+        1. 将图像Resize成shape参数大小
+        2. 图像归一化
+
+        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
+            mean(list): 图像均值
+            std(list): 图像方差
+    """
+
+    def __init__(self,
+                 mode,
+                 shape=[608, 608],
+                 mixup_epoch=250,
+                 mean=[0.485, 0.456, 0.406],
+                 std=[0.229, 0.224, 0.225]):
+        width = shape
+        if isinstance(shape, list):
+            if shape[0] != shape[1]:
+                raise Exception(
+                    "In YOLOv3 model, width and height should be equal")
+            width = shape[0]
+        if width % 32 != 0:
+            raise Exception(
+                "In YOLOv3 model, width and height should be multiple of 32, e.g 224、256、320...."
+            )
+
+        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(
+                        mean=mean, std=std)
+            ]
+        else:
+            # 验证/预测时的transforms
+            transforms = [
+                Resize(
+                    target_size=width, interp='CUBIC'), Normalize(
+                        mean=mean, std=std)
+            ]
+        super(ComposedYOLOTransforms, self).__init__(transforms)

paddlex/cv/transforms/seg_transforms.py

@@ -108,6 +108,12 @@ class Compose(SegTransform):
                     outputs = (im, im_info)
         return outputs
 
+    def add_augmenters(self, augmenters):
+        if not isinstance(augmenters, list):
+            raise Exception(
+                "augmenters should be list type in func add_augmenters()")
+        self.transforms = augmenters + self.transforms.transforms
+
 
 class RandomHorizontalFlip(SegTransform):
     """以一定的概率对图像进行水平翻转。当存在标注图像时，则同步进行翻转。
@@ -1088,3 +1094,39 @@ class ArrangeSegmenter(SegTransform):
             return (im, im_info)
         else:
             return (im, )
+
+
+class ComposedSegTransforms(Compose):
+    """ 语义分割模型(UNet/DeepLabv3p)的图像处理流程，具体如下
+        训练阶段：
+        1. 随机对图像以0.5的概率水平翻转
+        2. 按不同的比例随机Resize原图
+        3. 从原图中随机crop出大小为train_crop_size大小的子图，如若crop出来的图小于train_crop_size，则会将图padding到对应大小
+        4. 图像归一化
+        预测阶段：
+        1. 图像归一化
+
+        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
+            mean(list): 图像均值
+            std(list): 图像方差
+    """
+
+    def __init__(self,
+                 mode,
+                 train_crop_size=[769, 769],
+                 mean=[0.5, 0.5, 0.5],
+                 std=[0.5, 0.5, 0.5]):
+        if mode == 'train':
+            # 训练时的transforms，包含数据增强
+            transforms = [
+                RandomHorizontalFlip(prob=0.5), ResizeStepScaling(),
+                RandomPaddingCrop(crop_size=train_crop_size), Normalize(
+                    mean=mean, std=std)
+            ]
+        else:
+            # 验证/预测时的transforms
+            transforms = [Resize(512), Normalize(mean=mean, std=std)]
+
+        super(ComposedSegTransforms, self).__init__(transforms)

paddlex/deploy.py

@@ -97,8 +97,6 @@ class Predictor:
             config.disable_glog_info()
         if memory_optimize:
             config.enable_memory_optim()
-        else:
-            config.diable_memory_optim()
 
         # 开启计算图分析优化，包括OP融合等
         config.switch_ir_optim(True)

paddlex/seg.py

@@ -17,5 +17,6 @@ from . import cv
 
 UNet = cv.models.UNet
 DeepLabv3p = cv.models.DeepLabv3p
+HRNet = cv.models.HRNet
 transforms = cv.transforms.seg_transforms
 visualize = cv.models.utils.visualize.visualize_segmentation

paddlex/tools/x2coco.py

@@ -110,7 +110,7 @@ class LabelMe2COCO(X2COCO):
         annotation["segmentation"] = [list(np.asarray(points).flatten())]
         annotation["iscrowd"] = 0
         annotation["image_id"] = image_id + 1
-        annotation["bbox"] = list(map(float, get_bbox(height, width, points)))
+        annotation["bbox"] = list(map(float, self.get_bbox(height, width, points)))
         annotation["area"] = annotation["bbox"][2] * annotation["bbox"][3]
         annotation["category_id"] = label_to_num[label]
         annotation["id"] = object_id + 1

setup.py

@@ -19,7 +19,7 @@ long_description = "PaddleX. A end-to-end deeplearning model development toolkit
 
 setuptools.setup(
     name="paddlex",
-    version='1.0.2',
+    version='1.0.4',
     author="paddlex",
     author_email="paddlex@baidu.com",
     description=long_description,

tutorials/train/segmentation/hrnet.py

+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
+train_transforms = transforms.Compose([
+    transforms.RandomHorizontalFlip(), transforms.ResizeRangeScaling(),
+    transforms.RandomPaddingCrop(crop_size=512), transforms.Normalize()
+])
+
+eval_transforms = transforms.Compose([
+    transforms.ResizeByLong(long_size=512),
+    transforms.Padding(target_size=512), transforms.Normalize()
+])
+
+# 定义训练和验证所用的数据集
+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
+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)