Optimize RemoteSensing doc

* modify label tool

* fix label tools bug

* add RemoteSensing

* optimize data clip and normalize method for remote sensing

* fix label problem using mobile devices

* add README.md

* polish README.md, optimize main.py and transforms.py, add transforms.md

* add demo dataset

* add create_dataset_list tool, requirements.txt

* add channel-by-channel normalize and clip, remove RemoteSensing import

* update vdl to 2.0

* polish README

* add tool to randomly split dataset and generate file list

* open eval_best_metric while traing, fix typo

* remove python2.7 ci

.travis.yml

 language: python
 
 python:
-  - '2.7'
   - '3.5'
   - '3.6'
 

contrib/HumanSeg/models/humanseg.py

@@ -602,7 +602,7 @@ class SegModel(object):
             return_details (bool): 是否返回详细信息。默认False。
 
         Returns:
-            dict: 当return_details为False时，返回dict。包含关键字：'miou'、'categore_iou'、'macc'、
+            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。

contrib/RemoteSensing/README.md

 # 遥感分割（RemoteSensing）
 遥感影像分割是图像分割领域中的重要应用场景，广泛应用于土地测绘、环境监测、城市建设等领域。遥感影像分割的目标多种多样，有诸如积雪、农作物、道路、建筑、水源等地物目标，也有例如云层的空中目标。
 
-PaddleSeg提供了针对遥感专题的语义分割库RemoteSensing，涵盖图像预处理、数据增强、模型训练、预测流程，帮助大家利用深度学习技术解决遥感影像分割问题。
+PaddleSeg提供了针对遥感专题的语义分割库RemoteSensing，涵盖图像预处理、数据增强、模型训练、预测流程，帮助用户利用深度学习技术解决遥感影像分割问题。
 
+## 特点
 针对遥感数据多通道、分布范围大、分布不均的特点，我们支持多通道训练预测，内置一系列多通道预处理和数据增强的策略，可结合实际业务场景进行定制组合，提升模型泛化能力和鲁棒性。
 
 **Note:** 所有命令需要在`PaddleSeg/contrib/RemoteSensing/`目录下执行。
@@ -62,8 +63,8 @@ RemoteSensing               # 根目录
 ```
 其中，相应的文件名可根据需要自行定义。
 
-由于遥感领域图像格式多种多样，不同传感器产生的数据格式可能不同。本分割库目前采用npy格式作为遥感数据的格式，采用png无损压缩格式作为标注图片格式。
-
+遥感领域图像格式多种多样，不同传感器产生的数据格式可能不同。为方便数据加载，本分割库统一采用numpy存储格式`npy`作为原图格式，采用`png`无损压缩格式作为标注图片格式。
+原图的前两维是图像的尺寸，第3维是图像的通道数。
 标注图像为单通道图像，像素值即为对应的类别,像素标注类别需要从0开始递增，
 例如0，1，2，3表示有4种类别，标注类别最多为256类。其中可以指定特定的像素值用于表示该值的像素不参与训练和评估（默认为255）。
 
@@ -173,6 +174,7 @@ model.train(
     pretrain_weights=None,
     optimizer=None,
     learning_rate=lr,
+    use_vdl=True
 )
 ```
 
@@ -233,7 +235,7 @@ export CUDA_VISIBLE_DEVICES=0
 ```
 > 在RemoteSensing目录下运行`predict_demo.py`即可开始训练。
 ```shell script
-python predict_demo.py --data_dir dataset/demo/ --load_model_dir saved_model/unet/
+python predict_demo.py --data_dir dataset/demo/ --load_model_dir saved_model/unet/best_model/
 ```
 
 

contrib/RemoteSensing/models/base.py

@@ -213,6 +213,7 @@ class BaseAPI:
                    train_reader,
                    train_batch_size,
                    eval_reader=None,
+                   eval_best_metric=None,
                    save_interval_epochs=1,
                    log_interval_steps=10,
                    save_dir='output',
@@ -261,11 +262,8 @@ class BaseAPI:
 
         if use_vdl:
             # VisualDL component
-            log_writer = LogWriter(vdl_logdir, sync_cycle=20)
-            train_step_component = OrderedDict()
-            eval_component = OrderedDict()
+            log_writer = LogWriter(vdl_logdir)
 
-        best_accuracy_key = ""
         best_accuracy = -1.0
         best_model_epoch = 1
         for i in range(num_epochs):
@@ -302,14 +300,10 @@ class BaseAPI:
 
                     if use_vdl:
                         for k, v in step_metrics.items():
-                            if k not in train_step_component.keys():
-                                with log_writer.mode('Each_Step_while_Training'
-                                                     ) as step_logger:
-                                    train_step_component[
-                                        k] = step_logger.scalar(
-                                            'Training: {}'.format(k))
-                            train_step_component[k].add_record(num_steps, v)
-
+                            log_writer.add_scalar(
+                                tag="Training: {}".format(k),
+                                value=v,
+                                step=num_steps)
                     logging.info(
                         "[TRAIN] Epoch={}/{}, Step={}/{}, {}, eta={}".format(
                             i + 1, num_epochs, step + 1, total_num_steps,
@@ -336,10 +330,9 @@ class BaseAPI:
                     logging.info('[EVAL] Finished, Epoch={}, {} .'.format(
                         i + 1, dict2str(self.eval_metrics)))
                     # 保存最优模型
-                    best_accuracy_key = list(self.eval_metrics.keys())[0]
-                    current_accuracy = self.eval_metrics[best_accuracy_key]
-                    if current_accuracy > best_accuracy:
-                        best_accuracy = current_accuracy
+                    current_metric = self.eval_metrics[eval_best_metric]
+                    if current_metric > best_accuracy:
+                        best_accuracy = current_metric
                         best_model_epoch = i + 1
                         best_model_dir = osp.join(save_dir, "best_model")
                         self.save_model(save_dir=best_model_dir)
@@ -350,13 +343,11 @@ class BaseAPI:
                             if isinstance(v, np.ndarray):
                                 if v.size > 1:
                                     continue
-                            if k not in eval_component:
-                                with log_writer.mode('Each_Epoch_on_Eval_Data'
-                                                     ) as eval_logger:
-                                    eval_component[k] = eval_logger.scalar(
-                                        'Evaluation: {}'.format(k))
-                            eval_component[k].add_record(i + 1, v)
+                            log_writer.add_scalar(
+                                tag="Evaluation: {}".format(k),
+                                step=i + 1,
+                                value=v)
                 self.save_model(save_dir=current_save_dir)
                 logging.info(
                     'Current evaluated best model in eval_reader is epoch_{}, {}={}'
-                    .format(best_model_epoch, best_accuracy_key, best_accuracy))
+                    .format(best_model_epoch, eval_best_metric, best_accuracy))

contrib/RemoteSensing/models/unet.py

@@ -138,6 +138,7 @@ class UNet(BaseAPI):
               train_reader,
               train_batch_size=2,
               eval_reader=None,
+              eval_best_metric='kappa',
               save_interval_epochs=1,
               log_interval_steps=2,
               save_dir='output',
@@ -154,7 +155,8 @@ class UNet(BaseAPI):
             num_epochs (int): 训练迭代轮数。
             train_reader (readers): 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_reader (readers): 评估数据读取器。
+            eval_reader (readers): 边训边评估的评估数据读取器。
+            eval_best_metric (str): 边训边评估保存最好模型的指标。默认为'kappa'。
             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'。
@@ -202,6 +204,7 @@ class UNet(BaseAPI):
             train_reader=train_reader,
             train_batch_size=train_batch_size,
             eval_reader=eval_reader,
+            eval_best_metric=eval_best_metric,
             save_interval_epochs=save_interval_epochs,
             log_interval_steps=log_interval_steps,
             save_dir=save_dir,
@@ -223,7 +226,7 @@ class UNet(BaseAPI):
             return_details (bool): 是否返回详细信息。默认False。
 
         Returns:
-            dict: 当return_details为False时，返回dict。包含关键字：'miou'、'categore_iou'、'macc'、
+            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。
@@ -271,7 +274,7 @@ class UNet(BaseAPI):
         category_acc, macc = conf_mat.accuracy()
 
         metrics = OrderedDict(
-            zip(['miou', 'categore_iou', 'macc', 'category_acc', 'kappa'],
+            zip(['miou', 'category_iou', 'macc', 'category_acc', 'kappa'],
                 [miou, category_iou, macc, category_acc,
                  conf_mat.kappa()]))
         if return_details:

contrib/RemoteSensing/predict_demo.py

@@ -29,8 +29,8 @@ data_dir = args.data_dir
 load_model_dir = args.load_model_dir
 
 # predict
-model = load_model(osp.join(load_model_dir, 'best_model'))
-pred_dir = osp.join(load_model_dir, 'pred')
+model = load_model(load_model_dir)
+pred_dir = osp.join(load_model_dir, 'predict')
 if not osp.exists(pred_dir):
     os.mkdir(pred_dir)
 

contrib/RemoteSensing/requirements.txt

-pre-commit
-yapf == 0.26.0
-flake8
-pyyaml >= 5.1
-Pillow
-numpy
-six
-opencv-python
-tqdm
\ No newline at end of file
+visualdl >= 2.0.0-alpha.2

contrib/RemoteSensing/tools/create_dataset_list.py

@@ -21,7 +21,9 @@ import warnings
 
 def parse_args():
     parser = argparse.ArgumentParser(
-        description='PaddleSeg generate file list on your customized dataset.')
+        description=
+        'A tool for dividing dataset and generating file lists by file directory structure.'
+    )
     parser.add_argument('dataset_root', help='dataset root directory', type=str)
     parser.add_argument(
         '--separator',

contrib/RemoteSensing/tools/split_dataset_list.py

+# coding: utf8
+# copyright (c) 2019 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.
+
+import glob
+import os.path
+import argparse
+import warnings
+import numpy as np
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description=
+        'A tool for proportionally randomizing dataset to produce file lists.')
+    parser.add_argument('dataset_root', help='the dataset root path', type=str)
+    parser.add_argument('images', help='the directory name of images', type=str)
+    parser.add_argument('labels', help='the directory name of labels', type=str)
+    parser.add_argument(
+        '--split', help='', nargs=3, type=float, default=[0.7, 0.3, 0])
+    parser.add_argument(
+        '--label_class',
+        help='label class names',
+        type=str,
+        nargs='*',
+        default=['__background__', '__foreground__'])
+    parser.add_argument(
+        '--separator',
+        dest='separator',
+        help='file list separator',
+        default=" ",
+        type=str)
+    parser.add_argument(
+        '--format',
+        help='data format of images and labels, e.g. jpg, npy or png.',
+        type=str,
+        nargs=2,
+        default=['npy', 'png'])
+    parser.add_argument(
+        '--postfix',
+        help='postfix of images or labels',
+        type=str,
+        nargs=2,
+        default=['', ''])
+
+    return parser.parse_args()
+
+
+def get_files(path, format, postfix):
+    pattern = '*%s.%s' % (postfix, format)
+
+    search_files = os.path.join(path, pattern)
+    search_files2 = os.path.join(path, "*", pattern)  # 包含子目录
+    search_files3 = os.path.join(path, "*", "*", pattern)  # 包含三级目录
+
+    filenames = glob.glob(search_files)
+    filenames2 = glob.glob(search_files2)
+    filenames3 = glob.glob(search_files3)
+
+    filenames = filenames + filenames2 + filenames3
+
+    return sorted(filenames)
+
+
+def generate_list(args):
+    separator = args.separator
+    dataset_root = args.dataset_root
+    if sum(args.split) != 1.0:
+        raise ValueError("划分比例之和必须为1")
+
+    file_list = os.path.join(dataset_root, 'labels.txt')
+    with open(file_list, "w") as f:
+        for label_class in args.label_class:
+            f.write(label_class + '\n')
+
+    image_dir = os.path.join(dataset_root, args.images)
+    label_dir = os.path.join(dataset_root, args.labels)
+    image_files = get_files(image_dir, args.format[0], args.postfix[0])
+    label_files = get_files(label_dir, args.format[1], args.postfix[1])
+    if not image_files:
+        warnings.warn("No files in {}".format(image_dir))
+    num_images = len(image_files)
+
+    if not label_files:
+        warnings.warn("No files in {}".format(label_dir))
+    num_label = len(label_files)
+
+    if num_images != num_label and num_label > 0:
+        raise Exception("Number of images = {}    number of labels = {} \n"
+                        "Either number of images is equal to number of labels, "
+                        "or number of labels is equal to 0.\n"
+                        "Please check your dataset!".format(
+                            num_images, num_label))
+
+    image_files = np.array(image_files)
+    label_files = np.array(label_files)
+    state = np.random.get_state()
+    np.random.shuffle(image_files)
+    np.random.set_state(state)
+    np.random.shuffle(label_files)
+
+    start = 0
+    num_split = len(args.split)
+    dataset_name = ['train', 'val', 'test']
+    for i in range(num_split):
+        dataset_split = dataset_name[i]
+        print("Creating {}.txt...".format(dataset_split))
+        if args.split[i] > 1.0 or args.split[i] < 0:
+            raise ValueError(
+                "{} dataset percentage should be 0~1.".format(dataset_split))
+
+        file_list = os.path.join(dataset_root, dataset_split + '.txt')
+        with open(file_list, "w") as f:
+            num = round(args.split[i] * num_images)
+            end = start + num
+            if i == num_split - 1:
+                end = num_images
+            for item in range(start, end):
+                left = image_files[item].replace(dataset_root, '')
+                if left[0] == os.path.sep:
+                    left = left.lstrip(os.path.sep)
+
+                try:
+                    right = label_files[item].replace(dataset_root, '')
+                    if right[0] == os.path.sep:
+                        right = right.lstrip(os.path.sep)
+                    line = left + separator + right + '\n'
+                except:
+                    line = left + '\n'
+
+                f.write(line)
+                print(line)
+            start = end
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    generate_list(args)

contrib/RemoteSensing/train_demo.py

@@ -103,4 +103,4 @@ model.train(
     pretrain_weights=None,
     optimizer=None,
     learning_rate=lr,
-)
+    use_vdl=True)

pdseg/tools/jingling2seg.py

@@ -17,10 +17,8 @@ from gray2pseudo_color import get_color_map_list
 
 def parse_args():
     parser = argparse.ArgumentParser(
-        formatter_class=argparse.ArgumentDefaultsHelpFormatter
-    )
-    parser.add_argument('input_dir',
-                        help='input annotated directory')
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('input_dir', help='input annotated directory')
     return parser.parse_args()
 
 
@@ -62,8 +60,7 @@ def main(args):
         print('Generating dataset from:', label_file)
         with open(label_file) as f:
             base = osp.splitext(osp.basename(label_file))[0]
-            out_png_file = osp.join(
-                output_dir, base + '.png')
+            out_png_file = osp.join(output_dir, base + '.png')
 
             data = json.load(f)
 
@@ -77,14 +74,20 @@ def main(args):
                         # convert jingling format to labelme format
                         points = []
                         for i in range(1, int(len(polygon) / 2) + 1):
-                            points.append([polygon['x' + str(i)], polygon['y' + str(i)]])
-                        shape = {'label': name, 'points': points, 'shape_type': 'polygon'}
+                            points.append(
+                                [polygon['x' + str(i)], polygon['y' + str(i)]])
+                        shape = {
+                            'label': name,
+                            'points': points,
+                            'shape_type': 'polygon'
+                        }
                         data_shapes.append(shape)
 
             if 'size' not in data:
                 continue
             data_size = data['size']
-            img_shape = (data_size['height'], data_size['width'], data_size['depth'])
+            img_shape = (data_size['height'], data_size['width'],
+                         data_size['depth'])
 
             lbl = labelme.utils.shapes_to_label(
                 img_shape=img_shape,
@@ -102,8 +105,7 @@ def main(args):
             else:
                 raise ValueError(
                     '[%s] Cannot save the pixel-wise class label as PNG. '
-                    'Please consider using the .npy format.' % out_png_file
-                )
+                    'Please consider using the .npy format.' % out_png_file)
 
 
 if __name__ == '__main__':

pdseg/tools/labelme2seg.py

@@ -17,10 +17,8 @@ from gray2pseudo_color import get_color_map_list
 
 def parse_args():
     parser = argparse.ArgumentParser(
-        formatter_class=argparse.ArgumentDefaultsHelpFormatter
-    )
-    parser.add_argument('input_dir',
-                        help='input annotated directory')
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('input_dir', help='input annotated directory')
     return parser.parse_args()
 
 
@@ -61,8 +59,7 @@ def main(args):
         print('Generating dataset from:', label_file)
         with open(label_file) as f:
             base = osp.splitext(osp.basename(label_file))[0]
-            out_png_file = osp.join(
-                output_dir, base + '.png')
+            out_png_file = osp.join(output_dir, base + '.png')
 
             data = json.load(f)
 
@@ -85,8 +82,7 @@ def main(args):
             else:
                 raise ValueError(
                     '[%s] Cannot save the pixel-wise class label as PNG. '
-                    'Please consider using the .npy format.' % out_png_file
-                )
+                    'Please consider using the .npy format.' % out_png_file)
 
 
 if __name__ == '__main__':