Add PortraitNet.

configs/portraitnet/README.md

@@ -0,0 +1,20 @@
+# PortraitNet: Real-time Portrait Segmentation Network for Mobile Device
+
+## Reference
+
+> Song-Hai Zhanga, Xin Donga, Jia Lib, Ruilong Lia, Yong-Liang Yang. "PortraitNet: Real-time portrait segmentation network for mobile device". @ CAD&Graphics 2019.
+
+## Usage
+[https://aistudio.baidu.com/aistudio/projectdetail/1754799](https://aistudio.baidu.com/aistudio/projectdetail/1754799)
+
+## Dataset
+[https://pan.baidu.com/s/15uBpR7zFF2zpUccoq5pQYg](https://pan.baidu.com/s/15uBpR7zFF2zpUccoq5pQYg)
+
+password: ajcs
+
+## Performance
+
+| Model | Backbone | Dataset | Resolution | Training Iters | mIoU |
+|-|-|-|-|-|-|-|-|
+|PortraitNet|MobileNetV2|EG1800|224x224|46000|96.92%|
+|PortraitNet|MobileNetV2|Supervise.ly|224x224|60000|93.94%|

configs/portraitnet/portraitnet_eg1800_224x224_46k.yml

@@ -0,0 +1,78 @@
+batch_size: 64
+iters: 46000
+
+train_dataset:
+  type: EG1800
+  dataset_root: data/EG1800
+  common_transforms:
+    - type: RandomAffine
+      max_rotation: 45
+      min_scale_factor: 0.5
+      max_scale_factor: 1.5
+      size: [ 224, 224 ]
+      translation_offset: 56
+    - type: RandomHorizontalFlip
+  transforms1:
+    - type: Normalize
+      mean: [0.485, 0.458, 0.408]
+      std: [0.23, 0.23, 0.23]
+  transforms2:
+    - type: RandomDistort
+      brightness_range: 0.6
+      contrast_range: 0.4
+      saturation_range: 0.6
+      hue_prob: 0.0
+      sharpness_range: 0.2
+      sharpness_prob: 0.5
+    - type: RandomBlur
+      prob: 0.5
+      blur_type: random
+    - type: RandomNoise
+    - type: Normalize
+      mean: [ 0.485, 0.458, 0.408 ]
+      std: [ 0.23, 0.23, 0.23 ]
+  mode: train
+
+val_dataset:
+  type: EG1800
+  dataset_root: data/EG1800
+  common_transforms:
+    - type: ScalePadding
+      target_size: [ 224, 224 ]
+      im_padding_value: [127.5, 127.5, 127.5]
+      label_padding_value: 0
+    - type: Normalize
+      mean: [0.485, 0.458, 0.408]
+      std: [0.23, 0.23, 0.23]
+  transforms1: null
+  transforms2: null
+  mode: val
+
+optimizer:
+  type: adam
+  weight_decay: 5.0e-4
+
+lr_scheduler:
+  type: StepDecay
+  learning_rate: 0.001
+  step_size: 460
+  gamma: 0.95
+
+loss:
+  types:
+    - type: CrossEntropyLoss
+    - type: CrossEntropyLoss
+    - type: FocalLoss
+      edge_label: True
+    - type: KLLoss
+  coef: [1, 1, 0.3, 2]
+
+model:
+  type: PortraitNet
+  backbone:
+    type: MobileNetV2
+    channel_ratio: 1.0
+    min_channel: 16
+    pretrained: null
+  add_edge: True
+  num_classes: 2

configs/portraitnet/portraitnet_supervisely_224x224_60k.yml

@@ -0,0 +1,50 @@
+_base_: './portraitnet_eg1800_224x224_46k.yml'
+batch_size: 64
+iters: 60000
+
+train_dataset:
+  type: SUPERVISELY
+  dataset_root: data/Supervisely_face
+  common_transforms:
+    - type: RandomAffine
+      max_rotation: 45
+      min_scale_factor: 0.5
+      max_scale_factor: 1.5
+      size: [ 224, 224 ]
+      translation_offset: 56
+    - type: RandomHorizontalFlip
+  transforms1:
+    - type: Normalize
+      mean: [0.485, 0.458, 0.408]
+      std: [0.23, 0.23, 0.23]
+  transforms2:
+    - type: RandomDistort
+      brightness_range: 0.6
+      contrast_range: 0.4
+      saturation_range: 0.6
+      hue_prob: 0.0
+      sharpness_range: 0.2
+      sharpness_prob: 0.5
+    - type: RandomBlur
+      prob: 0.5
+      blur_type: random
+    - type: RandomNoise
+    - type: Normalize
+      mean: [ 0.485, 0.458, 0.408 ]
+      std: [ 0.23, 0.23, 0.23 ]
+  mode: train
+
+val_dataset:
+  type: SUPERVISELY
+  dataset_root: data/Supervisely_face
+  common_transforms:
+    - type: ScalePadding
+      target_size: [ 224, 224 ]
+      im_padding_value: [127.5, 127.5, 127.5]
+      label_padding_value: 0
+    - type: Normalize
+      mean: [0.485, 0.458, 0.408]
+      std: [0.23, 0.23, 0.23]
+  transforms1: null
+  transforms2: null
+  mode: val

paddleseg/core/train.py

@@ -40,8 +40,12 @@ def loss_computation(logits_list, labels, losses, edges=None):
         logits = logits_list[i]
         loss_i = losses['types'][i]
         # Whether to use edges as labels According to loss type.
-        if loss_i.__class__.__name__ in ('BCELoss', ) and loss_i.edge_label:
+        if loss_i.__class__.__name__ in ('BCELoss',
+                                         'FocalLoss') and loss_i.edge_label:
             loss_list.append(losses['coef'][i] * loss_i(logits, edges))
+        elif loss_i.__class__.__name__ in ("KLLoss", ):
+            loss_list.append(losses['coef'][i] * loss_i(
+                logits_list[0], logits_list[1].detach()))
         else:
             loss_list.append(losses['coef'][i] * loss_i(logits, labels))
     return loss_list

paddleseg/cvlibs/config.py

@@ -176,6 +176,9 @@ def learning_rate(self) -> paddle.optimizer.lr.LRScheduler:
         elif decay_type == 'piecewise':
             values = _learning_rate
             return paddle.optimizer.lr.PiecewiseDecay(values=values, **args)
+        elif decay_type == 'stepdecay':
+            lr = _learning_rate
+            return paddle.optimizer.lr.StepDecay(lr, **args)
         else:
             raise RuntimeError('Only poly and piecewise decay support.')
 

paddleseg/datasets/__init__.py

@@ -19,3 +19,5 @@
 from .optic_disc_seg import OpticDiscSeg
 from .pascal_context import PascalContext
 from .mini_deep_globe_road_extraction import MiniDeepGlobeRoadExtraction
+from .eg1800 import EG1800
+from .supervisely import SUPERVISELY

paddleseg/datasets/eg1800.py

@@ -0,0 +1,117 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import copy
+
+import cv2
+import numpy as np
+
+from paddleseg.datasets import Dataset
+from paddleseg.cvlibs import manager
+from paddleseg.transforms import Compose
+import paddleseg.transforms.functional as F
+
+
+@manager.DATASETS.add_component
+class EG1800(Dataset):
+    """
+    EG1800 dataset `http://xiaoyongshen.me/webpage_portrait/index.html`.
+
+    Args:
+        common_transforms (list): A list of common image transformations for two inputs of portrait net.
+        transforms1 (list): A list of image transformations for the first input of portrait net.
+        transforms2 (list): A list of image transformations for the second input of portrait net.
+        dataset_root (str, optional): The EG1800 dataset directory. Default: None.
+        mode (str, optional): A subset of the entire dataset. It should be one of ('train', 'val'). Default: 'train'.
+        edge (bool, optional): Whether to compute edge while training. Default: False
+    """
+    NUM_CLASSES = 2
+
+    def __init__(self,
+                 common_transforms,
+                 transforms1,
+                 transforms2,
+                 dataset_root=None,
+                 mode='train',
+                 edge=False):
+        self.dataset_root = dataset_root
+        self.common_transforms = Compose(common_transforms)
+        self.transforms = self.common_transforms
+        if transforms1 is not None:
+            self.transforms1 = Compose(transforms1, to_rgb=False)
+        if transforms2 is not None:
+            self.transforms2 = Compose(transforms2, to_rgb=False)
+        mode = mode.lower()
+        self.ignore_index = 255
+        self.mode = mode
+        self.num_classes = self.NUM_CLASSES
+        self.input_width = 224
+        self.input_height = 224
+
+        if mode == 'train':
+            path = os.path.join(dataset_root, 'eg1800_train.txt')
+        else:
+            path = os.path.join(dataset_root, 'eg1800_test.txt')
+        with open(path, 'r') as f:
+            files = f.readlines()
+        img_files = [
+            os.path.join(dataset_root, 'Images', file).strip() for file in files
+        ]
+        label_files = [
+            os.path.join(dataset_root, 'Labels', file).strip() for file in files
+        ]
+
+        self.file_list = [[
+            img_path, label_path
+        ] for img_path, label_path in zip(img_files, label_files)]
+        pass
+
+    def __getitem__(self, item):
+        image_path, label_path = self.file_list[item]
+        im = cv2.imread(image_path)
+        label = cv2.imread(label_path, 0)
+        label[label > 1] = 0
+
+        if self.mode == "val":
+            common_im, label = self.common_transforms(im=im, label=label)
+            im = np.float32(common_im[::-1, :, :])  # RGB => BGR
+            im_aug = copy.deepcopy(im)
+        else:
+            common_im, label = self.common_transforms(im=im, label=label)
+            common_im = np.transpose(common_im, [1, 2, 0])
+            # add augmentation
+            im, _ = self.transforms1(common_im)
+            im_aug, _ = self.transforms2(common_im)
+
+            im = np.float32(im[::-1, :, :])  # RGB => BGR
+            im_aug = np.float32(im_aug[::-1, :, :])  # RGB => BGR
+
+        label = cv2.resize(
+            np.uint8(label), (self.input_width, self.input_height),
+            interpolation=cv2.INTER_NEAREST)
+
+        # add mask blur
+        label = np.uint8(cv2.blur(label, (5, 5)))
+        label[label >= 0.5] = 1
+        label[label < 0.5] = 0
+
+        edge_mask = F.mask_to_binary_edge(
+            label, radius=4, num_classes=self.num_classes)
+        edge_mask = np.transpose(edge_mask, [1, 2, 0]).squeeze(axis=-1)
+        im = np.concatenate([im_aug, im])
+        if self.mode == "train":
+            return im, label, edge_mask
+        else:
+            return im, label