push icnet code

model_zoo/research/cv/ICNet/README.md

@@ -8,6 +8,7 @@
     - [Script and Sample Code](#script-and-sample-code)
     - [Script Parameters](#script-parameters)
     - [Training Process](#training-process)
+        - [Prepare Dataset](#prepare-dataset)
         - [Distributed Training](#distributed-training)
         - [Training Results](#training-results)
     - [Evaluation Process](#evaluation-process)
@@ -41,8 +42,8 @@ It contains 5,000 finely annotated images split into training, validation and te
 - frame:
     - [Mindspore](https://www.mindspore.cn/install)
 - For details, please refer to the following resources:
-    - [MindSpore course](https://www.mindspore.cn/tutorials/zh-CN/master/index.html)
+    - [MindSpore course](https://www.mindspore.cn/tutorial/training/zh-CN/master/index.html)
-    - [MindSpore Python API](https://www.mindspore.cn/docs/api/zh-CN/master/index.html)
+    - [MindSpore Python API](https://www.mindspore.cn/doc/api_python/zh-CN/master/index.html)
 
 # [Scription Description](#Content)
 
@@ -88,7 +89,7 @@ It contains 5,000 finely annotated images split into training, validation and te
 
 ## Script Parameters
 
-Set script parameters in configs/icnet.yaml .
+Set script parameters in src/model_utils/icnet.yaml .
 
 ### Model
 
@@ -112,22 +113,29 @@ weight_decay: 0.0001
 ```bash
 train_batch_size_percard: 4
 valid_batch_size: 1
-cityscapes_root: "/data/cityscapes/"
+cityscapes_root: "/data/cityscapes/" # set dataset path
 epochs: 160
-val_epoch: 1        # run validation every val-epoch
+val_epoch: 1
-ckpt_dir: "./ckpt/" # ckpt and training log will be saved here
+ckpt_dir: "./ckpt/"                  # ckpt and training log will be saved here
-mindrecord_dir: '/root/bigpingping/mindrecord'
+mindrecord_dir: ''                   # set mindrecord path
+pretrained_model_path: '/root/ResNet50V1B-150_625.ckpt' # set the pretrained model path correctly
 save_checkpoint_epochs: 5
 keep_checkpoint_max: 10
 ```
 
-### Valid
+## Training Process
 
-```bash
+### Prepare Datast
-ckpt_path: ""       # set the pretrained model path correctly
+
+- Convert dataset to Mindrecord
+
+```python
+    python cityscapes_mindrecord.py [DATASET_PATH] [MINDRECORD_PATH]
 ```
 
-## Training Process
+- Note:
+
+[MINDRCORD_PATH] in script should be consistent with 'mindrecord_dir' in config file.
 
 ### Distributed Training
 

model_zoo/research/cv/ICNet/eval.py

@@ -47,7 +47,7 @@ class Evaluator:
         self.image_paths, self.mask_paths = _get_city_pairs(config["train"]["cityscapes_root"], "val")
 
         # create network
-        self.model = ICNet(nclass=19, backbone='resnet50', istraining=False)
+        self.model = ICNet(nclass=19, pretrained_path=cfg["train"]["pretrained_model_path"], istraining=False)
 
         # load ckpt
         ckpt_file_name = args_opt.checkpoint_path
@@ -74,6 +74,7 @@ class Evaluator:
             mask = self._mask_transform(mask)  # mask shape: (H,w)
 
             image = Tensor(image)
+            print(image)
 
             expand_dims = ops.ExpandDims()
             image = expand_dims(image, 0)
@@ -163,7 +164,8 @@ if __name__ == '__main__':
     from src.metric import SegmentationMetric
     from src.logger import SetupLogger
     # Set config file
-    config_path = args_opt.project_path + "/src/model_utils/icnet.yaml"
+    config_file = "src/model_utils/icnet.yaml"
+    config_path = os.path.join(args_opt.project_path, config_file)
     with open(config_path, "r") as yaml_file:
         cfg = yaml.load(yaml_file.read())
     logger = SetupLogger(name="semantic_segmentation",

model_zoo/research/cv/ICNet/postprocess.py

@@ -106,7 +106,7 @@ if __name__ == '__main__':
     from src.metric import SegmentationMetric
     from src.logger import SetupLogger
     # Set config file
-    config_file = "/src/model_utils/icnet.yaml"
+    config_file = "src/model_utils/icnet.yaml"
     config_path = os.path.join(args_opt.project_path, config_file)
     with open(config_path, "r") as yaml_file:
         cfg = yaml.load(yaml_file.read())

model_zoo/research/cv/ICNet/scripts/run_distribute_train8p.sh

@@ -17,20 +17,28 @@
 if [ $# != 2 ]
 then
     echo "=============================================================================================================="
-    echo "Usage: bash scripts/run_distribute_train.sh [RANK_TABLE_FILE] [PROJECT_PATH]"
+    echo "Usage: bash scripts/run_distribute_train8p.sh [RANK_TABLE_FILE] [PROJECT_PATH]"
     echo "Please run the script as: "
-    echo "bash scripts/run_distribute_train.sh [RANK_TABLE_FILE] [PROJECT_PATH]"
+    echo "bash scripts/run_distribute_train8p.sh [RANK_TABLE_FILE] [PROJECT_PATH]"
-    echo "for example: bash run_distribute_train.sh /absolute/path/to/RANK_TABLE_FILE /root/ICNet/"
+    echo "for example: bash script/run_distribute_train8p.sh /absolute/path/to/RANK_TABLE_FILE /root/ICNet/"
     echo "=============================================================================================================="
     exit 1
 fi
 
-PATH1=$(get_real_path $1)
+get_real_path(){
+  if [ "${1:0:1}" == "/" ]; then
+    echo "$1"
+  else
+    echo "$(realpath -m $PWD/$1)"
+  fi
+}
+
+PATH2=$(get_real_path $2)
 
 
-if [ ! -d $PATH1 ]
+if [ ! -d $PATH2 ]
 then
-    echo "error: DATASET_PATH=$PATH1 is not a directory"
+    echo "error: PROJECT_PATH=$PATH2 is not a directory"
 exit 1
 fi
 
@@ -51,7 +59,7 @@ do
     echo "start training for rank $i, device $DEVICE_ID"
     env > env.log
 
-    python3 train.py --project_path=$PATH1 > log.txt 2>&1 &
+    python3 train.py --project_path=$PATH2 > log.txt 2>&1 &
 
     cd ../
-done
+done

model_zoo/research/cv/ICNet/scripts/run_eval.sh

@@ -30,6 +30,7 @@ get_real_path(){
 
 PATH1=$(get_real_path $1)
 PATH2=$(get_real_path $2)
+PATH3=$(get_real_path $3)
 
 
 if [ ! -d $PATH1 ]
@@ -44,9 +45,15 @@ then
 exit 1
 fi
 
+if [ ! -d $PATH3 ]
+then
+    echo "error: PROJECT_PATH=$PATH3 is not a directory"
+exit 1
+fi
+
 ulimit -u unlimited
 export DEVICE_NUM=1
-export DEVICE_ID=$5
+export DEVICE_ID=0
 export RANK_SIZE=1
 export RANK_ID=0
 
@@ -61,6 +68,6 @@ cp -r ../src ./eval
 cd ./eval || exit
 env > env.log
 echo "start evaluation for device $DEVICE_ID"
-python eval.py --dataset_path=$PATH1 --checkpoint_path=$PATH2 &> log &
+python eval.py --dataset_path=$PATH1 --checkpoint_path=$PATH2 --project_path=$PATH3 &> log &
 
-cd ..
+cd ..

model_zoo/research/cv/ICNet/src/cityscapes_mindrecord.py

@@ -15,6 +15,7 @@
 """Prepare Cityscapes dataset"""
 import os
 import random
+import argparse
 import numpy as np
 from PIL import Image
 from PIL import ImageOps
@@ -27,6 +28,7 @@ import mindspore.dataset.transforms.py_transforms as tc
 
 def _get_city_pairs(folder, split='train'):
     """Return two path arrays of data set img and mask"""
+
     def get_path_pairs(image_folder, masks_folder):
         image_paths = []
         masks_paths = []
@@ -145,11 +147,10 @@ def _img_mask_transform(img, mask):
     return (img, mask)
 
 
-def data_to_mindrecord_img(prefix='cityscapes.mindrecord', file_num=1,
+def data_to_mindrecord_img(prefix='cityscapes-2975.mindrecord', file_num=1,
-                           root='/data/cityscapes/', split='train'):
+                           root='./', split='train', mindrecord_dir="./"):
     """to mindrecord"""
-    mindrecord_dir = '/data/Mindrecord_cityscapes/'
+    mindrecord_path = os.path.join(mindrecord_dir, prefix)
-    mindrecord_path = os.path.join(mindrecord_dir, prefix + "2975-2")
 
     writter = FileWriter(mindrecord_path, file_num)
 
@@ -202,4 +203,10 @@ def create_icnet_dataset(mindrecord_file, batch_size=16, device_num=1, rank_id=0
 
 
 if __name__ == '__main__':
-    data_to_mindrecord_img()
+    parser = argparse.ArgumentParser(description="dataset_to_mindrecord")
+    parser.add_argument("--dataset_path", type=str, default="/data/cityscapes/", help="dataset path")
+    parser.add_argument("--mindrecord_path", type=str, default="/data/cityscapes_mindrecord/",
+                        help="mindrecord_path")
+
+    args_opt = parser.parse_args()
+    data_to_mindrecord_img(root=args_opt.dataset_path, mindrecord_dir=args_opt.mindrecord_path)

model_zoo/research/cv/ICNet/src/model_utils/icnet.yaml

@@ -19,10 +19,11 @@ train:
   epochs: 160
   val_epoch: 1        # run validation every val-epoch
   ckpt_dir: "./ckpt/" # ckpt and training log will be saved here
-  mindrecord_dir: '/root/ICNet/mindrecord'
+  mindrecord_dir: '/data/cityscapes_mindrecord'
+  pretrained_model_path: '/root/ResNet50V1B-150_625.ckpt'
   save_checkpoint_epochs: 5
   keep_checkpoint_max: 10
 
 ### 4.Valid
 test:
-  ckpt_path: "" # set the pretrained model path correctly
+  ckpt_path: "" # set the pretrained model path correctly

model_zoo/research/cv/ICNet/src/models/__init__.py

@@ -1,4 +1,3 @@
 """__init__"""
 from .icnet import ICNet
 from .icnet_dc import ICNetdc
-from .icnet_1p import ICNet1p

model_zoo/research/cv/ICNet/src/models/icnet.py

@@ -28,7 +28,7 @@ context.set_context(mode=context.GRAPH_MODE, device_target='Ascend')
 class ICNet(nn.Cell):
     """Image Cascade Network"""
 
-    def __init__(self, nclass=19, backbone='resnet50', pretrained_base=True, istraining=True):
+    def __init__(self, nclass=19, backbone='resnet50', pretrained_path='', istraining=True):
         super(ICNet, self).__init__()
         self.conv_sub1 = nn.SequentialCell(
             _ConvBNReLU(3, 32, 3, 2),
@@ -38,7 +38,7 @@ class ICNet(nn.Cell):
         self.istraining = istraining
         self.ppm = PyramidPoolingModule()
 
-        self.backbone = SegBaseModel()
+        self.backbone = SegBaseModel(root=pretrained_path)
 
         self.head = _ICHead(nclass)
 
@@ -232,11 +232,11 @@ class CascadeFeatureFusion24(nn.Cell):
 class SegBaseModel(nn.Cell):
     """Base Model for Semantic Segmentation"""
 
-    def __init__(self, nclass=19, backbone='resnet50', pretrained_base=True, **kwargs):
+    def __init__(self, nclass=19, backbone='resnet50', root=''):
         super(SegBaseModel, self).__init__()
         self.nclass = nclass
         if backbone == 'resnet50':
-            self.pretrained = get_resnet50v1b()
+            self.pretrained = get_resnet50v1b(ckpt_root=root)
 
     def construct(self, x):
         """forwarding pre-trained network"""

model_zoo/research/cv/ICNet/src/models/icnet_dc.py

@@ -28,7 +28,7 @@ context.set_context(mode=context.GRAPH_MODE, device_target='Ascend')
 class ICNetdc(nn.Cell):
     """Image Cascade Network"""
 
-    def __init__(self, nclass=19, backbone='resnet50', pretrained_base=True, istraining=True):
+    def __init__(self, nclass=19, pretrained_path="", istraining=True):
         super(ICNetdc, self).__init__()
         self.conv_sub1 = nn.SequentialCell(
             _ConvBNReLU(3, 32, 3, 2),
@@ -38,7 +38,7 @@ class ICNetdc(nn.Cell):
         self.istraining = istraining
         self.ppm = PyramidPoolingModule()
 
-        self.backbone = SegBaseModel()
+        self.backbone = SegBaseModel(root=pretrained_path)
 
         self.head = _ICHead(nclass)
 
@@ -235,11 +235,11 @@ class CascadeFeatureFusion24(nn.Cell):
 class SegBaseModel(nn.Cell):
     """Base Model for Semantic Segmentation"""
 
-    def __init__(self, nclass=19, backbone='resnet50', pretrained_base=True, **kwargs):
+    def __init__(self, nclass=19, backbone='resnet50', root=""):
         super(SegBaseModel, self).__init__()
         self.nclass = nclass
         if backbone == 'resnet50':
-            self.pretrained = get_resnet50v1b()
+            self.pretrained = get_resnet50v1b(ckpt_root=root)
 
     def construct(self, x):
         """forwarding pre-trained network"""

model_zoo/research/cv/ICNet/src/models/resnet50_v1.py

@@ -258,7 +258,7 @@ class Resnet50v1b(nn.Cell):
         return out
 
 
-def get_resnet50v1b(class_num=1001, ckpt_root='/root/ICNet/ckpt/ResNet50V1B-150_625.ckpt', pretrained=True):
+def get_resnet50v1b(class_num=1001, ckpt_root='', pretrained=True):
     """
     Get SE-ResNet50 neural network.
     Default : GE Theta+ version (best)

model_zoo/research/cv/ICNet/train.py

@@ -48,16 +48,16 @@ def train_net():
                                           parameter_broadcast=True,
                                           gradients_mean=True)
         init()
-    prefix = 'cityscapes.mindrecord'
+    prefix = 'cityscapes-2975.mindrecord'
     mindrecord_dir = cfg['train']["mindrecord_dir"]
-    mindrecord_file = os.path.join(mindrecord_dir, prefix + '2975')
+    mindrecord_file = os.path.join(mindrecord_dir, prefix)
     dataset = create_icnet_dataset(mindrecord_file, batch_size=cfg['train']["train_batch_size_percard"],
                                    device_num=device_num, rank_id=device_id)
 
     train_data_size = dataset.get_dataset_size()
     print("data_size", train_data_size)
     epoch = cfg["train"]["epochs"]
-    network = ICNetdc()  # __init__
+    network = ICNetdc(pretrained_path=cfg["train"]["pretrained_model_path"])  # __init__
 
     iters_per_epoch = train_data_size
     total_train_steps = iters_per_epoch * epoch
@@ -85,7 +85,8 @@ if __name__ == '__main__':
     from src.cityscapes_mindrecord import create_icnet_dataset
     from src.models.icnet_dc import ICNetdc
     from src.lr_scheduler import poly_lr
-    config_path = args_opt.project_path + "src/model_utils/icnet.yaml"
+    config_file = "src/model_utils/icnet.yaml"
+    config_path = os.path.join(args_opt.project_path, config_file)
     with open(config_path, "r") as yaml_file:
         cfg = yaml.load(yaml_file.read())
     logging.basicConfig(level=logging.INFO)