update train test for doc_seg

Author: ndcuong91

configs/_base_/datasets/doc_seg.py

@@ -0,0 +1,54 @@
+# dataset settings
+dataset_type = 'doc_segDataset'
+data_root = '/data_backup/cuongnd/mmseg/doc_seg'
+img_norm_cfg = dict(
+    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
+crop_size = (640, 640)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', reduce_zero_label=False),
+    dict(type='Resize', img_scale=(1270, 900), ratio_range=(0.9, 1.1)),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(1270, 900),
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='imgs/train',
+        ann_dir='anno/train',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='imgs/val',
+        ann_dir='anno/val',
+        pipeline=test_pipeline),
+    test=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='imgs/val',
+        ann_dir='anno/val',
+        pipeline=test_pipeline))

configs/_base_/default_runtime.py

@@ -1,6 +1,6 @@
 # yapf:disable
 log_config = dict(
-    interval=200,
+    interval=100,
     hooks=[
         dict(type='TextLoggerHook', by_epoch=False),
         # dict(type='TensorboardLoggerHook')

configs/_base_/schedules/schedule_80k_new.py

@@ -6,4 +6,4 @@
 # runtime settings
 total_iters = 80000
 checkpoint_config = dict(by_epoch=False, interval=8000)
-evaluation = dict(interval=80000, metric='mIoU')
+evaluation = dict(interval=800, metric='mIoU')

configs/fastscnn/fast_scnn_4x8_20k_lr0.12_doc_seg.py

@@ -0,0 +1,10 @@
+_base_ = [
+    '../_base_/models/fast_scnn.py', '../_base_/datasets/doc_seg.py',
+    '../_base_/default_runtime.py', '../_base_/schedules/schedule_20k.py'
+]
+
+# Re-config the data sampler.
+data = dict(samples_per_gpu=4, workers_per_gpu=4)
+
+# Re-config the optimizer.
+optimizer = dict(type='SGD', lr=0.12, momentum=0.9, weight_decay=4e-5)

configs/fastscnn/fast_scnn_4x8_80k_lr0.12_doc_seg.py

@@ -0,0 +1,10 @@
+_base_ = [
+    '../_base_/models/fast_scnn.py', '../_base_/datasets/doc_seg.py',
+    '../_base_/default_runtime.py', '../_base_/schedules/schedule_80k_new.py'
+]
+
+# Re-config the data sampler.
+data = dict(samples_per_gpu=8, workers_per_gpu=4)
+
+# Re-config the optimizer.
+optimizer = dict(type='SGD', lr=0.12, momentum=0.9, weight_decay=4e-5)

mmseg/apis/train.py

@@ -118,7 +118,7 @@ def train_segmentor(model,
         eval_cfg = cfg.get('evaluation', {})
         eval_cfg['by_epoch'] = cfg.runner['type'] != 'IterBasedRunner'
         eval_hook = DistEvalHook if distributed else EvalHook
-        runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
+        runner.register_hook(eval_hook(val_dataloader, **eval_cfg), priority='LOW')
 
     if cfg.resume_from:
         runner.resume(cfg.resume_from)

mmseg/datasets/__init__.py

@@ -14,11 +14,12 @@
 from .publaynet_split1 import publaynet_split1Dataset
 from .doc_structure1 import doc_structure1Dataset
 from .popular_doc import popular_docDataset
+from .doc_seg import doc_segDataset
 
 __all__ = [
     'CustomDataset', 'build_dataloader', 'ConcatDataset', 'RepeatDataset',
     'DATASETS', 'build_dataset', 'PIPELINES', 'CityscapesDataset',
     'PascalVOCDataset', 'ADE20KDataset', 'PascalContextDataset',
     'PascalContextDataset59', 'ChaseDB1Dataset', 'DRIVEDataset', 'HRFDataset',
-    'STAREDataset','popular_docDataset'
+    'STAREDataset','popular_docDataset','doc_segDataset'
 ]

mmseg/datasets/doc_seg.py

@@ -0,0 +1,16 @@
+from .builder import DATASETS
+from .custom import CustomDataset
+
+
+@DATASETS.register_module()
+class doc_segDataset(CustomDataset):
+    """doc_segDataset
+    """
+    CLASSES = ('background','doc')
+    PALETTE = [[120,120,120],[255, 0, 0]]
+    def __init__(self, **kwargs):
+        super(doc_segDataset, self).__init__(
+            img_suffix='.jpg',
+            seg_map_suffix='.png',
+            reduce_zero_label=False,
+            **kwargs)

tools/labelme2normal.py

@@ -0,0 +1,74 @@
+import os, cv2
+import numpy as np
+import json
+
+def get_list_file_in_folder(dir, ext=['jpg', 'png', 'JPG', 'PNG']):
+    included_extensions = ext
+    file_names = [fn for fn in os.listdir(dir)
+                  if any(fn.endswith(ext) for ext in included_extensions)]
+    return file_names
+
+def convert_labelme_label_to_normal_format(src_anno_dir, src_img_dir, dst_anno_dir, label_list, debug=False):
+    print('convert_voc_label_to_normal_format')
+    print('src_anno_dir',src_anno_dir)
+    print('dst_anno_dir',dst_anno_dir)
+
+    list_imgs = get_list_file_in_folder(src_img_dir)
+    list_imgs = sorted(list_imgs)
+
+    count_samples ={}
+    for label in label_list:
+        count_samples[label]=0
+
+    for idx, img_name in enumerate(list_imgs):
+        base_name = img_name.split('.')[0]
+        if idx < 0:
+            continue
+        print(idx, 'labelme2normal. Convert', base_name)
+
+        json_path = os.path.join(src_anno_dir, base_name+'.json')
+        img = cv2.imread(os.path.join(src_img_dir, img_name))
+
+        segment_img = np.zeros((img.shape[0], img.shape[1]), dtype=np.uint8)
+        with open(json_path) as json_file:
+            data = json.load(json_file)
+            shapes = data["shapes"]
+            for shape in shapes:
+                point = shape["points"]
+                label = shape["label"]
+                pts = np.asarray(point,np.int32)
+                # label_idx = label_list.index(label)
+                label_idx = 1 # chỉ có vùng giấy tờ với idx  =1
+                if label not in count_samples.keys():
+                    count_samples[label] = 0
+                count_samples[label] +=1
+
+                # color = int(20*label_idx)
+                color = label_idx
+
+                cv2.fillPoly(segment_img, pts=[pts], color=color)
+        if debug:
+            cv2.imshow('origin' ,img)
+            cv2.imshow('mask' ,segment_img)
+            cv2.waitKey(0)
+
+
+        output_anno_path = os.path.join(dst_anno_dir, base_name+'.png')
+        cv2.imwrite(output_anno_path, segment_img)
+
+    print('Number of samples', count_samples)
+
+
+if __name__ == "__main__":
+    # test = cv2.imread('/home/duycuong/PycharmProjects/ocr/others/conversion_tools/segmentation/00004.png')
+
+    src_anno_dir ='/data_backup/cuongnd/mmseg/doc_seg/anno/bhyt'
+    src_img_dir = '/data_backup/cuongnd/Viettel_freeform/MAFC/BHYT_origin/imgs/clean'
+    dst_anno_dir ='/data_backup/cuongnd/mmseg/doc_seg/anno/bhyt_imgs'
+    label_list = ['background','cccd','cccd_back','cmnd_new','cmnd_old','cmnd_old_back',
+                    'driverlicense_new','driverlicense_new_back','driverlicense_old','driverlicense_old_back']
+    convert_labelme_label_to_normal_format(src_anno_dir,
+                                           src_img_dir,
+                                           dst_anno_dir,
+                                           label_list,
+                                           debug = False)

tools/prepare_segmentation_data.py

@@ -57,26 +57,6 @@ def convert_anno_detection_to_segmentation(img_dir, anno_det_dir, output_anno_se
             cv2.rectangle(anno_mask,(int(left)-extend,int(top)-extend),(int(right)+extend,int(bottom)+extend),1,-1)
         cv2.imwrite(os.path.join(output_anno_segment_dir,img_name),anno_mask)
 
-def convert_anno_objective2_to_segmentation(img_dir, anno_det_dir, output_anno_segment_dir, extend=-1, format_anno_det='icdar', class_list=dict()):
-    list_images = get_list_file_in_folder(img_dir)
-    list_images = sorted(list_images)
-    for idx, img_name in enumerate(list_images):
-        print(idx, img_name)
-        img_path=os.path.join(img_dir,img_name)
-        img = cv2.imread(img_path)
-        anno_mask =  np.zeros((img.shape[0], img.shape[1]), np.uint8)
-        anno_file = os.path.join(anno_det_dir,img_name.replace('.jpg','.json').replace('.png','.json'))
-
-        import json
-        with open(anno_file, "r") as anno:
-            anno_str = json.load(anno)
-
-        for i, line in enumerate(anno_str['cellboxes']):
-            left, top, right, bottom = line[0], line[1], line[2], line[3]
-            cv2.rectangle(anno_mask,(int(left)-extend,int(top)-extend),(int(right)+extend,int(bottom)+extend),1,-1)
-        cv2.imwrite(os.path.join(output_anno_segment_dir,img_name),anno_mask)
-        print('ok')
-
 def split_dataset(img_dir, ann_dir, img_dst_dir, ann_dst_dir, ratio=0.5):
     list_images = get_list_file_in_folder(img_dir)
     random.shuffle(list_images)
@@ -260,41 +240,20 @@ def visualize_normal_format_dataset(img_dir, ann_dir):
 
 
 if __name__=='__main__':
-    #img=cv2.imread('/home/cuongnd/PycharmProjects/aicr/source/mmsegmentation/data/ade/ADEChallengeData2016/annotations/validation/ADE_val_00000012.png', cv2.IMREAD_GRAYSCALE)
-
-    data_dir='/data20.04/data/table recognition/from_Korea/201012_172754_pubtabnet_valid_sample_objective#2'
-    img_dir= data_dir + '/images'
-    anno_det_dir=data_dir + '/annots'
-    output_anno_segment_dir=data_dir + '/annot_seg'
-
-
-    #convert_anno_objective2_to_segmentation(img_dir, anno_det_dir, output_anno_segment_dir)
-    #
     # split_dataset(img_dir='/data4T/cuongnd/dataset/publaynet_split1/img_dir/train',
     #               ann_dir='/data4T/cuongnd/dataset/publaynet_split1/ann_dir/train_3classes',
     #               img_dst_dir='/data4T/cuongnd/dataset/doc_structure1/img_dir/train',
     #               ann_dst_dir='/data4T/cuongnd/dataset/doc_structure1/ann_dir/train',
     #               ratio=0.002)
 
-    # del_dataset(img_dir='/data20.04/data/doc_structure/publaynet/img_dir/train',
-    #             ann_dir='/data20.04/data/doc_structure/publaynet/ann_dir/train')
-
-    src_anno_dir='/data4T/cuongnd/dataset/publaynet_split1/ann_dir/val'
-    dst_anno_dir='/data4T/cuongnd/dataset/publaynet_split1/ann_dir/val_3classes'
-    # refactor_classes_of_dataset(src_anno_dir, dst_anno_dir,
-    #                             src_classes=[1, 2, 3, 4, 5], #('text', 'title', 'list', 'table', 'figure')
-    #                             dst_classes=[1, 1, 3, 2, 1])
-
+    src_anno_dir='/data_backup/cuongnd/Viettel_freeform/MAFC/BHYT_origin/imgs/clean'
+    dst_anno_dir='/data_backup/cuongnd/mmseg/doc_seg/imgs/bhyt'
+    convert_all_imgs_to_jpg(src_anno_dir,dst_anno_dir)
 
-    #onvert_voc_label_to_normal_format(src_anno_dir,dst_anno_dir)
 
-    #convert_all_imgs_to_jpg(src_anno_dir,dst_anno_dir)
-    #
-    # refine_dataset(img_dir='/data4T/ntanh/publaynet/train',
-    #             ann_dir='/data4T/ntanh/publaynet_gen_gt_oct2.1/train/label')
-    img_dir='/home/duycuong/home_data/mmlab/mmseg/popular_doc/images/train'
-    ann_dir='/home/duycuong/home_data/mmlab/mmseg/popular_doc/annotations/train'
-    visualize_normal_format_dataset(img_dir=img_dir,
-                                    ann_dir=ann_dir)
+    # img_dir='/data_backup/cuongnd/mmseg/doc_seg_data/imgs/train'
+    # ann_dir='/data_backup/cuongnd/mmseg/doc_seg_data/anno/train'
+    # visualize_normal_format_dataset(img_dir=img_dir,
+    #                                 ann_dir=ann_dir)