Update configs

configs/cityscape.yaml

@@ -10,18 +10,6 @@ AUG:
     MIN_SCALE_FACTOR: 0.5  # for stepscaling
     SCALE_STEP_SIZE: 0.25  # for stepscaling
     MIRROR: True
-    RICH_CROP:
-        ENABLE: False
-        ASPECT_RATIO: 0.33
-        BLUR: True
-        BLUR_RATIO: 0.1
-        FLIP: True
-        FLIP_RATIO: 0.2
-        MAX_ROTATION: 15
-        MIN_AREA_RATIO: 0.5
-        BRIGHTNESS_JITTER_RATIO: 0.5
-        CONTRAST_JITTER_RATIO: 0.5
-        SATURATION_JITTER_RATIO: 0.5
 BATCH_SIZE: 4
 DATASET:
     DATA_DIR: "./dataset/cityscapes/"
@@ -44,8 +32,7 @@ TEST:
     TEST_MODEL: "snapshots/cityscape_v5/final/"
 TRAIN:
     MODEL_SAVE_DIR: "snapshots/cityscape_v7/"
-    PRETRAINED_MODEL: u"pretrain/deeplabv3plus_gn_init"
-    RESUME: False
+    PRETRAINED_MODEL_DIR: "pretrain/deeplabv3plus_gn_init"
     SNAPSHOT_EPOCH: 10
 SOLVER:
     LR: 0.001

configs/unet_pet.yaml

@@ -12,18 +12,6 @@ AUG:
     MIN_SCALE_FACTOR: 0.75  # for stepscaling
     SCALE_STEP_SIZE: 0.25  # for stepscaling
     MIRROR: True
-    RICH_CROP:
-        ENABLE: False
-        ASPECT_RATIO: 0.33
-        BLUR: True
-        BLUR_RATIO: 0.1
-        FLIP: True
-        FLIP_RATIO: 0.2
-        MAX_ROTATION: 15
-        MIN_AREA_RATIO: 0.5
-        BRIGHTNESS_JITTER_RATIO: 0.5
-        CONTRAST_JITTER_RATIO: 0.5
-        SATURATION_JITTER_RATIO: 0.5
 BATCH_SIZE: 4
 DATASET:
     DATA_DIR: "./dataset/mini_pet/"
@@ -45,8 +33,7 @@ TEST:
     TEST_MODEL: "./test/saved_model/unet_pet/final/"
 TRAIN:
     MODEL_SAVE_DIR: "./test/saved_models/unet_pet/"
-    PRETRAINED_MODEL: "./test/models/unet_coco/"
-    RESUME: False
+    PRETRAINED_MODEL_DIR: "./test/models/unet_coco/"
     SNAPSHOT_EPOCH: 10
 SOLVER:
     NUM_EPOCHS: 500

docs/configs/train_group.md

@@ -11,7 +11,7 @@ TRAIN Group存放所有和训练相关的配置
 <br/>
 <br/>
 
-## `PRETRAINED_MODEL`
+## `PRETRAINED_MODEL_DIR`
 预训练模型路径
 
 ## 默认值
@@ -28,19 +28,15 @@ TRAIN Group存放所有和训练相关的配置
 <br/>
 <br/>
 
-## `RESUME`
-是否从预训练模型中恢复参数并继续训练
+## `RESUME_MODEL_DIR`
+从指定路径中恢复参数并继续训练
 
 ## 默认值
-False
+无
 
 ## 注意事项
 
-* 当该字段被置为True且`PRETRAINED_MODEL`不存在时，该选项不生效
-
-* 当该字段被置为True且`PRETRAINED_MODEL`存在时，PaddleSeg会恢复到上一次训练的最近一个epoch，并且恢复训练过程中的临时变量（如已经衰减过的学习率，Optimizer的动量数据等）
-
-* 当该字段被置为True且`PRETRAINED_MODEL`存在时，`PRETRAINED_MODEL`路径的最后一个目录必须为int数值或者字符串final，PaddleSeg会将int数值作为当前起始EPOCH继续训练，若目录为final，则不会继续训练。若目录不满足上述条件，PaddleSeg会抛出错误。
+* 当`RESUME_MODEL_DIR`存在时，PaddleSeg会恢复到上一次训练的最近一个epoch，并且恢复训练过程中的临时变量（如已经衰减过的学习率，Optimizer的动量数据等），`PRETRAINED_MODEL`路径的最后一个目录必须为int数值或者字符串final，PaddleSeg会将int数值作为当前起始EPOCH继续训练，若目录为final，则不会继续训练。若目录不满足上述条件，PaddleSeg会抛出错误。
 
 <br/>
 <br/>
@@ -57,4 +53,4 @@ False
 
 * 仅在GPU多卡训练时该开关有效（Windows不支持多卡训练，因此无需打开该开关）
 
-* GPU多卡训练时，建议开启该开关，可以提升模型的训练效果
+* GPU多卡训练时，建议开启该开关，可以提升模型的训练效果
\ No newline at end of file

docs/data_aug.md

@@ -35,9 +35,6 @@ rich crop是指对图像进行多种变换，保证在训练过程中数据的
 - blur
 图像加模糊，使用开关`AUG.RICH_CROP.BLUR`，为False时该项功能关闭。`AUG.RICH_CROP.BLUR_RATIO`控制加入模糊的概率。
 
-- flip
-图像上下翻转，使用开关`AUG.RICH_CROP.FLIP`，为False时该项功能关闭。`AUG.RICH_CROP.FLIP_RATIO`控制上下翻转的概率。
-
 - rotation
 图像旋转，`AUG.RICH_CROP.MAX_ROTATION`控制最大旋转角度。旋转产生的多余的区域的填充值为均值。
 

docs/usage.md

@@ -76,7 +76,7 @@ python pdseg/train.py --use_gpu \
                       --tb_log_dir train_log \
                       --cfg configs/unet_pet.yaml \
                       BATCH_SIZE 4 \
-                      TRAIN.PRETRAINED_MODEL pretrained_model/unet_bn_coco \
+                      TRAIN.PRETRAINED_MODEL_DIR pretrained_model/unet_bn_coco \
                       TRAIN.SYNC_BATCH_NORM True \
                       SOLVER.LR 5e-5
 ```

pdseg/reader.py

@@ -261,17 +261,18 @@ class SegDataset(object):
                     SATURATION_JITTER_RATIO,
                     contrast_jitter_ratio=cfg.AUG.RICH_CROP.
                     CONTRAST_JITTER_RATIO)
-                if cfg.AUG.RICH_CROP.FLIP:
-                    if cfg.AUG.RICH_CROP.FLIP_RATIO <= 0:
-                        n = 0
-                    elif cfg.AUG.RICH_CROP.FLIP_RATIO >= 1:
-                        n = 1
-                    else:
-                        n = int(1.0 / cfg.AUG.RICH_CROP.FLIP_RATIO)
-                    if n > 0:
-                        if np.random.randint(0, n) == 0:
-                            img = img[::-1, :, :]
-                            grt = grt[::-1, :]
+
+            if cfg.AUG.FLIP:
+                if cfg.AUG.FLIP_RATIO <= 0:
+                    n = 0
+                elif cfg.AUG.FLIP_RATIO >= 1:
+                    n = 1
+                else:
+                    n = int(1.0 / cfg.AUG.FLIP_RATIO)
+                if n > 0:
+                    if np.random.randint(0, n) == 0:
+                        img = img[::-1, :, :]
+                        grt = grt[::-1, :]
 
             if cfg.AUG.MIRROR:
                 if np.random.randint(0, 2) == 1:

pdseg/train.py

@@ -152,15 +152,15 @@ def load_checkpoint(exe, program):
     Load checkpoiont from pretrained model directory for resume training
     """
 
-    print('Resume model training from:', cfg.TRAIN.PRETRAINED_MODEL)
-    if not os.path.exists(cfg.TRAIN.PRETRAINED_MODEL):
+    print('Resume model training from:', cfg.TRAIN.RESUME_MODEL_DIR)
+    if not os.path.exists(cfg.TRAIN.RESUME_MODEL_DIR):
         raise ValueError("TRAIN.PRETRAIN_MODEL {} not exist!".format(
-            cfg.TRAIN.PRETRAINED_MODEL))
+            cfg.TRAIN.RESUME_MODEL_DIR))
 
     fluid.io.load_persistables(
-        exe, cfg.TRAIN.PRETRAINED_MODEL, main_program=program)
+        exe, cfg.TRAIN.RESUME_MODEL_DIR, main_program=program)
 
-    model_path = cfg.TRAIN.PRETRAINED_MODEL
+    model_path = cfg.TRAIN.RESUME_MODEL_DIR
     # Check is path ended by path spearator
     if model_path[-1] == os.sep:
         model_path = model_path[0:-1]
@@ -255,11 +255,11 @@ def train(cfg):
 
     # Resume training
     begin_epoch = cfg.SOLVER.BEGIN_EPOCH
-    if cfg.TRAIN.RESUME:
+    if cfg.TRAIN.RESUME_MODEL_DIR:
         begin_epoch = load_checkpoint(exe, train_prog)
     # Load pretrained model
-    elif os.path.exists(cfg.TRAIN.PRETRAINED_MODEL):
-        print('Pretrained model dir:', cfg.TRAIN.PRETRAINED_MODEL)
+    elif os.path.exists(cfg.TRAIN.PRETRAINED_MODEL_DIR):
+        print('Pretrained model dir:', cfg.TRAIN.PRETRAINED_MODEL_DIR)
         load_vars = []
         load_fail_vars = []
 
@@ -268,10 +268,10 @@ def train(cfg):
             Check whehter persitable variable shape is match with current network
             """
             var_exist = os.path.exists(
-                os.path.join(cfg.TRAIN.PRETRAINED_MODEL, var.name))
+                os.path.join(cfg.TRAIN.PRETRAINED_MODEL_DIR, var.name))
             if var_exist:
                 var_shape = parse_shape_from_file(
-                    os.path.join(cfg.TRAIN.PRETRAINED_MODEL, var.name))
+                    os.path.join(cfg.TRAIN.PRETRAINED_MODEL_DIR, var.name))
                 return var_shape == shape
             return False
 
@@ -285,10 +285,10 @@ def train(cfg):
                     load_fail_vars.append(x)
         if cfg.MODEL.FP16:
             # If open FP16 training mode, load FP16 var separate
-            load_fp16_vars(exe, cfg.TRAIN.PRETRAINED_MODEL, train_prog)
+            load_fp16_vars(exe, cfg.TRAIN.PRETRAINED_MODEL_DIR, train_prog)
         else:
             fluid.io.load_vars(
-                exe, dirname=cfg.TRAIN.PRETRAINED_MODEL, vars=load_vars)
+                exe, dirname=cfg.TRAIN.PRETRAINED_MODEL_DIR, vars=load_vars)
         for var in load_vars:
             print("Parameter[{}] loaded sucessfully!".format(var.name))
         for var in load_fail_vars:
@@ -299,7 +299,7 @@ def train(cfg):
             len(load_vars) + len(load_fail_vars)))
     else:
         print('Pretrained model dir {} not exists, training from scratch...'.
-              format(cfg.TRAIN.PRETRAINED_MODEL))
+              format(cfg.TRAIN.PRETRAINED_MODEL_DIR))
 
     fetch_list = [avg_loss.name, lr.name]
     if args.debug:

pdseg/utils/config.py

@@ -69,6 +69,10 @@ cfg.DATASET.IGNORE_INDEX = 255
 ########################### 数据增强配置 ######################################
 # 图像镜像左右翻转
 cfg.AUG.MIRROR = True
+# 图像上下翻转开关，True/False
+cfg.AUG.FLIP = False
+# 图像启动上下翻转的概率，0-1
+cfg.AUG.FLIP_RATIO = 0.5
 # 图像resize的固定尺寸（宽，高），非负
 cfg.AUG.FIX_RESIZE_SIZE = tuple()
 # 图像resize的方式有三种：
@@ -107,18 +111,14 @@ cfg.AUG.RICH_CROP.CONTRAST_JITTER_RATIO = 0.5
 cfg.AUG.RICH_CROP.BLUR = False
 # 图像启动模糊百分比，0-1
 cfg.AUG.RICH_CROP.BLUR_RATIO = 0.1
-# 图像上下翻转开关，True/False
-cfg.AUG.RICH_CROP.FLIP = False
-# 图像启动上下翻转的概率，0-1
-cfg.AUG.RICH_CROP.FLIP_RATIO = 0.2
 
 ########################### 训练配置 ##########################################
 # 模型保存路径
 cfg.TRAIN.MODEL_SAVE_DIR = ''
 # 预训练模型路径
-cfg.TRAIN.PRETRAINED_MODEL = ''
+cfg.TRAIN.PRETRAINED_MODEL_DIR = ''
 # 是否resume，继续训练
-cfg.TRAIN.RESUME = False
+cfg.TRAIN.RESUME_MODEL_DIR = ''
 # 是否使用多卡间同步BatchNorm均值和方差
 cfg.TRAIN.SYNC_BATCH_NORM = False
 # 模型参数保存的epoch间隔数，可用来继续训练中断的模型

test/configs/deeplabv3p_xception65_cityscapes.yaml

@@ -10,18 +10,6 @@ AUG:
     MIN_SCALE_FACTOR: 0.5  # for stepscaling
     SCALE_STEP_SIZE: 0.25  # for stepscaling
     MIRROR: True
-    RICH_CROP:
-        ENABLE: False
-        ASPECT_RATIO: 0.33
-        BLUR: True
-        BLUR_RATIO: 0.1
-        FLIP: True
-        FLIP_RATIO: 0.2
-        MAX_ROTATION: 15
-        MIN_AREA_RATIO: 0.5
-        BRIGHTNESS_JITTER_RATIO: 0.5
-        CONTRAST_JITTER_RATIO: 0.5
-        SATURATION_JITTER_RATIO: 0.5
 BATCH_SIZE: 4
 DATASET:
     DATA_DIR: "./dataset/cityscapes/"
@@ -46,8 +34,7 @@ TEST:
     TEST_MODEL: "snapshots/cityscape_v5/final/"
 TRAIN:
     MODEL_SAVE_DIR: "snapshots/cityscape_v5/"
-    PRETRAINED_MODEL: "pretrain/deeplabv3plus_gn_init"
-    RESUME: False
+    PRETRAINED_MODEL_DIR: "pretrain/deeplabv3plus_gn_init"
     SNAPSHOT_EPOCH: 10
 SOLVER:
     LR: 0.001

test/configs/unet_pet.yaml

@@ -12,18 +12,6 @@ AUG:
     MIN_SCALE_FACTOR: 0.75  # for stepscaling
     SCALE_STEP_SIZE: 0.25  # for stepscaling
     MIRROR: True
-    RICH_CROP:
-        ENABLE: False
-        ASPECT_RATIO: 0.33
-        BLUR: True
-        BLUR_RATIO: 0.1
-        FLIP: True
-        FLIP_RATIO: 0.2
-        MAX_ROTATION: 15
-        MIN_AREA_RATIO: 0.5
-        BRIGHTNESS_JITTER_RATIO: 0.5
-        CONTRAST_JITTER_RATIO: 0.5
-        SATURATION_JITTER_RATIO: 0.5
 BATCH_SIZE: 6
 DATASET:
     DATA_DIR: "./dataset/pet/"
@@ -45,8 +33,7 @@ TEST:
     TEST_MODEL: "./test/saved_model/unet_pet/final/"
 TRAIN:
     MODEL_SAVE_DIR: "./test/saved_models/unet_pet/"
-    PRETRAINED_MODEL: "./test/models/unet_coco/"
-    RESUME: False
+    PRETRAINED_MODEL_DIR: "./test/models/unet_coco/"
     SNAPSHOT_EPOCH: 10
 SOLVER:
     NUM_EPOCHS: 500

test/local_test_cityscapes.py

@@ -45,7 +45,8 @@ if __name__ == "__main__":
     saved_model = os.path.join(LOCAL_PATH, "saved_model", model_name)
 
     parser = argparse.ArgumentParser(description="PaddleSeg loacl test")
-    parser.add_argument("--devices",
+    parser.add_argument(
+        "--devices",
         dest="devices",
         help="GPU id of running. if more than one, use spacing to separate.",
         nargs="+",
@@ -75,7 +76,7 @@ if __name__ == "__main__":
     train(
         flags=["--cfg", cfg, "--use_gpu", "--log_steps", "10"],
         options=[
-            "SOLVER.NUM_EPOCHS", "1", "TRAIN.PRETRAINED_MODEL", test_model,
+            "SOLVER.NUM_EPOCHS", "1", "TRAIN.PRETRAINED_MODEL_DIR", test_model,
             "TRAIN.MODEL_SAVE_DIR", saved_model
         ],
         devices=devices)

test/local_test_pet.py

@@ -46,7 +46,8 @@ if __name__ == "__main__":
     saved_model = os.path.join(LOCAL_PATH, "saved_model", model_name)
 
     parser = argparse.ArgumentParser(description="PaddleSeg loacl test")
-    parser.add_argument("--devices",
+    parser.add_argument(
+        "--devices",
         dest="devices",
         help="GPU id of running. if more than one, use spacing to separate.",
         nargs="+",
@@ -59,7 +60,7 @@ if __name__ == "__main__":
     train(
         flags=["--cfg", cfg, "--use_gpu", "--log_steps", "10"],
         options=[
-            "SOLVER.NUM_EPOCHS", "1", "TRAIN.PRETRAINED_MODEL", test_model,
+            "SOLVER.NUM_EPOCHS", "1", "TRAIN.PRETRAINED_MODEL_DIR", test_model,
             "TRAIN.MODEL_SAVE_DIR", saved_model, "DATASET.TRAIN_FILE_LIST",
             os.path.join(DATASET_PATH, "mini_pet", "file_list",
                          "train_list.txt"), "DATASET.VAL_FILE_LIST",

turtorial/finetune_deeplabv3plus.md

@@ -67,7 +67,7 @@ MODEL:
     DEEPLAB:
         BACKBONE: "xception_65"
 TRAIN:
-    PRETRAINED_MODEL: "./pretrained_model/deeplabv3p_xception65_bn_pet/"
+    PRETRAINED_MODEL_DIR: "./pretrained_model/deeplabv3p_xception65_bn_pet/"
 
 
 # 其他配置
@@ -79,7 +79,6 @@ AUG:
 BATCH_SIZE: 4
 TRAIN:
     MODEL_SAVE_DIR: "./finetune/deeplabv3p_xception65_bn_pet/"
-    RESUME: False
     SNAPSHOT_EPOCH: 10
 TEST:
     TEST_MODEL: "./finetune/deeplabv3p_xception65_bn_pet/final"

turtorial/finetune_icnet.md

@@ -66,7 +66,7 @@ MODEL:
     DEFAULT_NORM_TYPE: "bn"
     MULTI_LOSS_WEIGHT: "[1.0, 0.4, 0.16]"
 TRAIN:
-    PRETRAINED_MODEL: "./pretrained_model/icnet_bn_cityscapes/"
+    PRETRAINED_MODEL_DIR: "./pretrained_model/icnet_bn_cityscapes/"
 
 
 # 其他配置
@@ -78,7 +78,6 @@ AUG:
 BATCH_SIZE: 4
 TRAIN:
     MODEL_SAVE_DIR: "./finetune/icnet_pet/"
-    RESUME: False
     SNAPSHOT_EPOCH: 10
 TEST:
     TEST_MODEL: "./finetune/icnet_pet/final"
@@ -118,4 +117,4 @@ python pdseg/eval.py --use_gpu --cfg ./configs/test_pet.yaml
 
 |预训练模型名称|BackBone|Norm|数据集|配置|
 |-|-|-|-|-|
-|icnet_bn_cityscapes|-|bn|Cityscapes|MODEL.MODEL_NAME: icnet <br> MODEL.DEFAULT_NORM_TYPE: bn|
+|icnet_bn_cityscapes|-|bn|Cityscapes|MODEL.MODEL_NAME: icnet <br> MODEL.DEFAULT_NORM_TYPE: bn <br> MULTI_LOSS_WEIGHT: [1.0, 0.4, 0.16]|

turtorial/finetune_unet.md

@@ -65,7 +65,7 @@ MODEL:
     MODEL_NAME: "unet"
     DEFAULT_NORM_TYPE: "bn"
 TRAIN:
-    PRETRAINED_MODEL: "./pretrained_model/unet_bn_coco/"
+    PRETRAINED_MODEL_DIR: "./pretrained_model/unet_bn_coco/"
 
 
 # 其他配置
@@ -77,7 +77,6 @@ AUG:
 BATCH_SIZE: 4
 TRAIN:
     MODEL_SAVE_DIR: "./finetune/unet_pet/"
-    RESUME: False
     SNAPSHOT_EPOCH: 10
 TEST:
     TEST_MODEL: "./finetune/unet_pet/final"