fix some hyper parameters

fluid/PaddleCV/deeplabv3+/README.md

@@ -72,20 +72,21 @@ python train.py --help
 以上命令用于测试训练过程是否正常，仅仅迭代了50次并且使用了1的batch size，如果需要复现
 原论文的实验，请使用以下设置：
 ```
+CUDA_VISIBLE_DEVICES=0 \
 python ./train.py \
-    --batch_size=8 \
+    --batch_size=4 \
     --parallel=True \
     --norm_type=gn \
     --train_crop_size=769 \
-    --total_step=90000 \
+    --total_step=500000 \
     --base_lr=0.001 \
     --init_weights_path=deeplabv3plus_gn_init \
     --save_weights_path=output \
     --dataset_path=$DATASET_PATH
 ```
-如果您的显存不足，可以尝试减小`batch_size`，同时等比例放大`total_step`, 保证相乘的值不变，这得益于Group Norm的特性，改变 `batch_size` 并不会显著影响结果，而且能够节约更多显存, 比如您可以设置`--batch_size=4 --total_step=180000`。
+如果您的显存不足，可以尝试减小`batch_size`，同时等比例放大`total_step`, 缩小`base_lr`, 保证相乘的值不变，这得益于Group Norm的特性，改变 `batch_size` 并不会显著影响结果，而且能够节约更多显存, 比如您可以设置`--batch_size=2 --total_step=1000000 --base_lr=0.0005`。
 
-如果您希望使用多卡进行训练，可以同比增加`batch_size`，减小`total_step`, 比如原来单卡训练是`--batch_size=4 --total_step=180000`，使用4卡训练则是`--batch_size=16 --total_step=45000`
+如果您希望使用多卡进行训练，可以同比增加`batch_size`，减小`total_step`, 增加`base_lr`, 比如原来单卡训练是`--batch_size=4 --total_step=500000`，使用4卡训练则是`--batch_size=16 --total_step=125000 --base_lr=0.004`
 
 ### 测试
 执行以下命令在`Cityscape`测试数据集上进行测试：

fluid/PaddleCV/deeplabv3+/reader.py

@@ -9,7 +9,7 @@ import six
 default_config = {
     "shuffle": True,
     "min_resize": 0.5,
-    "max_resize": 2,
+    "max_resize": 4,
     "crop_size": 769,
 }
 
@@ -73,6 +73,16 @@ class CityscapeDataset:
         if self.index >= len(self.label_files):
             self.reset()
 
+    def color_augmentation(self, img):
+        r = np.random.rand()
+        img = img / 255.0
+        if r < 0.5:
+            img = np.power(img, 1 - r)
+        else:
+            img = np.power(img, 1 + (r - 0.5) * 2)
+        img = img * 255.0
+        return img
+
     def get_img(self):
         shape = self.config["crop_size"]
         while True:
@@ -90,9 +100,22 @@ class CityscapeDataset:
                 break
         if shape == -1:
             return img, label, ln
-        random_scale = np.random.rand(1) * (self.config['max_resize'] -
-                                            self.config['min_resize']
-                                            ) + self.config['min_resize']
+
+        if np.random.rand() > 0.5:
+            range_l = 1
+            range_r = self.config['max_resize']
+        else:
+            range_l = self.config['min_resize']
+            range_r = 1
+
+        if np.random.rand() > 0.5:
+            assert len(img.shape) == 3 and len(
+                label.shape) == 3, "{} {}".format(img.shape, label.shape)
+            img = img[:, :, ::-1]
+            label = label[:, :, ::-1]
+        img = self.color_augmentation(img)
+
+        random_scale = np.random.rand(1) * (range_r - range_l) + range_l
         crop_size = int(shape / random_scale)
         bb = crop_size // 2
 

fluid/PaddleCV/deeplabv3+/train.py

@@ -21,10 +21,10 @@ parser = argparse.ArgumentParser()
 add_arg = lambda *args: utility.add_arguments(*args, argparser=parser)
 
 # yapf: disable
-add_arg('batch_size',           int,    2,      "The number of images in each batch during training.")
+add_arg('batch_size',           int,    4,      "The number of images in each batch during training.")
 add_arg('train_crop_size',      int,    769,    "Image crop size during training.")
-add_arg('base_lr',              float,  0.0001, "The base learning rate for model training.")
-add_arg('total_step',           int,    90000,  "Number of the training step.")
+add_arg('base_lr',              float,  0.001,  "The base learning rate for model training.")
+add_arg('total_step',           int,    500000, "Number of the training step.")
 add_arg('init_weights_path',    str,    None,   "Path of the initial weights in paddlepaddle format.")
 add_arg('save_weights_path',    str,    None,   "Path of the saved weights during training.")
 add_arg('dataset_path',         str,    None,   "Cityscape dataset path.")