python-3.x - Mask rcnn 不适用于高分辨率图像

Question

我用了Mask-Rcnn按照这篇引用文章 Mask rcnn usage 使用 VIA 工具训练图像集(注意 具有高分辨率，例如:2400*1920) .在这里，我编辑了Ballon.py，代码如下:

import os
import sys
import json
import datetime
import numpy as np
import skimage.draw

# Root directory of the project
ROOT_DIR = os.path.abspath("../../")

# Import Mask RCNN
sys.path.append(ROOT_DIR)  # To find local version of the library
from mrcnn.config import Config
from mrcnn import model as modellib, utils

# Path to trained weights file
COCO_WEIGHTS_PATH = os.path.join(ROOT_DIR, "mask_rcnn_coco.h5")

if COCO_WEIGHTS_PATH is None:
    print('weights not available')
else:
    print('weights available')


DEFAULT_LOGS_DIR = os.path.join(ROOT_DIR, "logs")

#  Configurations
class NeuralCodeConfig(Config):
    NAME = "screens"

    # We use a GPU with 12GB memory, which can fit two images.
    # Adjust down if you use a smaller GPU.
    IMAGES_PER_GPU = 1

    # Number of classes (including background)
    NUM_CLASSES = 1 + 10 # Background + other region classes

    # Number of training steps per epoch
    STEPS_PER_EPOCH = 30

    # Skip detections with < 90% confidence
    DETECTION_MIN_CONFIDENCE = 0.9


#  Dataset
class NeuralCodeDataset(utils.Dataset):
    def load_screen(self, dataset_dir, subset):
        """Load a subset of the screens dataset.
        dataset_dir: Root directory of the dataset.
        subset: Subset to load: train or val
        """
        # Add classes.
        self.add_class("screens",1,"logo")
        self.add_class("screens",2,"slider")
        self.add_class("screens",3,"navigation")
        self.add_class("screens",4,"forms")
        self.add_class("screens",5,"social_media_icons")
        self.add_class("screens",6,"video")
        self.add_class("screens",7,"map")
        self.add_class("screens",8,"pagination")
        self.add_class("screens",9,"pricing_table_block")
        self.add_class("screens",10,"gallery")

        
        # Train or validation dataset?
        assert subset in ["train", "val"]
        dataset_dir = os.path.join(dataset_dir, subset)
    
         # Load annotations
        # VGG Image Annotator saves each image in the form:
        # { 'filename': '28503151_5b5b7ec140_b.jpg',
        #   'regions': {
        #       '0': {
        #           'region_attributes': {},
        #           'shape_attributes': {
        #               'all_points_x': [...],
        #               'all_points_y': [...],
        #               'name': 'polygon'}},
        #       ... more regions ...
        #   },
        #   'size': 100202
        # }
        # We mostly care about the x and y coordinates of each region
        annotations = json.load(open(os.path.join(dataset_dir, "via_region_data.json")))
        if annotations is None:
            print ("region data json not loaded")
        else:
            print("region data json loaded")
        # print(annotations)
        annotations = list(annotations.values())  # don't need the dict keys

        # The VIA tool saves images in the JSON even if they don't have any
        # annotations. Skip unannotated images.
        annotations = [a for a in annotations if a['regions']]

        # Add images
        for a in annotations:
            # Get the x, y coordinaets of points of the polygons that make up
            # the outline of each object instance. There are stores in the
            # shape_attributes and region_attributes (see json format above)
            polygons = [r['shape_attributes'] for r in a['regions']]
            screens = [r['region_attributes']for r in a['regions']]
            #getting the filename by spliting 
            class_name = screens[0]['html']
            file_name = a['filename'].split("/")
            file_name = file_name[len(file_name)-1]

            #getting class_ids with file_name
            class_ids = class_name+"_"+file_name
            # #getting width an height of the images
            # height = [h['height'] for h in polygons]
            # width = [w['width'] for w in polygons]
            # print(height,'height')
            # print('polygons',polygons)

            # load_mask() needs the image size to convert polygons to masks.
            # Unfortunately, VIA doesn't include it in JSON, so we must readpath
            # the image. This is only managable since the dataset is tiny.
            image_path = os.path.join(dataset_dir,file_name)
            image = skimage.io.imread(image_path)
             #resizing images
            # image = utils.resize_image(image, min_dim=800, max_dim=1000, min_scale=None, mode="square")
            # print('image',image)
            height,width = image.shape[:2]
            # print('height',height)
            # print('width',width)
            # height = 800
            # width = 800
            
   
            self.add_image(
                "screens",
                image_id=file_name,  # use file name as a unique image id
                path=image_path,
                width=width, height=height,
                polygons=polygons,
                class_ids=class_ids)

    def load_mask(self, image_id):
        """Generate instance masks for an image.
       Returns:
        masks: A bool array of shape [height, width, instance count] with
            one mask per instance.
        class_ids: a 1D array of class IDs of the instance masks.
        """
        # If not a screens dataset image, delegate to parent class.
        image_info = self.image_info[image_id]
        if image_info["source"] != "screens":
            return super(self.__class__, self).load_mask(image_id)

        # Convert polygons to a bitmap mask of shape
        # [height, width, instance_count]
        info = self.image_info[image_id]
        mask = np.zeros([info["height"], info["width"], len(info["polygons"])],
                        dtype=np.uint8)
        for i, p in enumerate(info["polygons"]):
            # Get indexes of pixels inside the polygon and set them to 1
            rr, cc = skimage.draw.polygon(p['y'], p['x'])
            mask[rr, cc, i] = 1

        # Return mask, and array of class IDs of each instance. Since we have
        # one class ID only, we return an array of 1s
        # return mask.astype(np.bool), np.ones([mask.shape[-1]], dtype=np.int32)
        # class_ids = np.array(class_ids,dtype=np.int32)
        return mask,class_ids

    def image_reference(self, image_id):
        """Return the path of the image."""
        info = self.image_info[image_id]
        if info["source"] == "screens":
            return info["path"]
        else:
            super(self.__class__, self).image_reference(image_id)



def train(model):
    # Train the model.
    # Training dataset.

    dataset_train = NeuralCodeDataset()
    dataset_train.load_screen(args.dataset, "train")
    dataset_train.prepare()

    # Validation dataset
    dataset_val = NeuralCodeDataset()
    dataset_val.load_screen(args.dataset, "val")
    dataset_val.prepare()

    # *** This training schedule is an example. Update to your needs ***
    # Since we're using a very small dataset, and starting from
    # COCO trained weights, we don't need to train too long. Also,
    # no need to train all layers, just the heads should do it.
    print("Training network heads")
    model.train(dataset_train, dataset_val,
                learning_rate=config.LEARNING_RATE,
                epochs=30,
                layers='heads')

#  Training
if __name__ == '__main__':
    import argparse

# Parse command line arguments
parser = argparse.ArgumentParser(
    description='Train Mask R-CNN to detect screens.')
parser.add_argument("command",
                    metavar="<command>",
                    help="'train' or 'splash'")
parser.add_argument('--dataset', required='True',
                    metavar="../../datasets/screens",
                    help='Directory of the screens dataset')
parser.add_argument('--weights', required=True,
                    metavar="/weights.h5",
                    help="Path to weights .h5 file or 'coco'")
parser.add_argument('--logs', required=False,
                    default=DEFAULT_LOGS_DIR,
                    metavar="../../logs/",
                    help='Logs and checkpoints directory (default=logs/)')
parser.add_argument('--image', required=False,
                    metavar="path or URL to image",
                    help='Image to apply the color splash effect on')
parser.add_argument('--video', required=False,
                    metavar="path or URL to video",
                    help='Video to apply the color splash effect on')
args = parser.parse_args()

# Validate arguments
if args.command == "train":
    assert args.dataset, "Argument --dataset is required for training"
elif args.command == "splash":
    assert args.image or args.video,\
           "Provide --image or --video to apply color splash"

print("Weights: ", args.weights)
print("Dataset: ", args.dataset)
print("Logs: ", args.logs)

# Configurations
if args.command == "train":
    config = NeuralCodeConfig()
else:
    class InferenceConfig(NeuralCodeConfig):
        # Set batch size to 1 since we'll be running inference on
        # one image at a time. Batch size = GPU_COUNT * IMAGES_PER_GPU
        GPU_COUNT = 1
        IMAGES_PER_GPU = 1
    config = InferenceConfig()
config.display()

# Create model
if args.command == "train":
    model = modellib.MaskRCNN(mode="training", config=config,
                              model_dir=args.logs)
else:
    model = modellib.MaskRCNN(mode="inference", config=config,
                              model_dir=args.logs)

# Select weights file to load
if args.weights.lower() == "coco":
    weights_path = COCO_WEIGHTS_PATH
    # Download weights file
    if not os.path.exists(weights_path):
        utils.download_trained_weights(weights_path)
elif args.weights.lower() == "last":
    # Find last trained weights
    weights_path = model.find_last()
elif args.weights.lower() == "imagenet":
    # Start from ImageNet trained weights
    weights_path = model.get_imagenet_weights()
else:
    weights_path = args.weights

# Load weights
print("Loading weights ", weights_path)
if args.weights.lower() == "coco":
    # Exclude the last layers because they require a matching
    # number of classes
    model.load_weights(weights_path, by_name=True, exclude=[
        "mrcnn_class_logits", "mrcnn_bbox_fc",
        "mrcnn_bbox", "mrcnn_mask"])
else:
    model.load_weights(weights_path, by_name=True)

# Train or evaluate
if args.command == "train":
    train(model)
# elif args.command == "splash":
#     detect_and_color_splash(model, image_path=args.image,
#                             video_path=args.video)
else:
    print("'{}' is not recognized. "
          "Use 'train' or 'splash'".format(args.command))

使用预训练的 COCO 数据集训练数据集时出现以下错误:

UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
  "Converting sparse IndexedSlices to a dense Tensor of unknown shape. "
2018-08-09 13:52:27.993239: W tensorflow/core/framework/allocator.cc:108] Allocation of 51380224 exceeds 10% of system memory.
2018-08-09 13:52:28.037704: W tensorflow/core/framework/allocator.cc:108] Allocation of 51380224 exceeds 10% of system memory.
/home/scit/anaconda3/lib/python3.6/site-packages/keras/engine/training.py:2022: UserWarning: Using a generator with use_multiprocessing=True` and multiple workers may duplicate your data. Please consider using the`keras.utils.Sequence class.
  UserWarning('Using a generator with `use_multiprocessing=True`'

ERROR:root:Error processing image {'id': '487.jpg', 'source': 'screens', 'path': '../../datasets/screens/train/487.jpg', 'width': 1920, 'height': 7007, 'polygons': [{'name': 'rect', 'x': 384, 'y': 5, 'width': 116, 'height': 64}, {'name': 'rect', 'x': 989, 'y': 17, 'width': 516, 'height': 42}, {'name': 'rect', 'x': 984, 'y': 5933, 'width': 565, 'height': 273}, {'name': 'rect', 'x': 837, 'y': 6793, 'width': 238, 'height': 50}], 'class_ids': 'logo_487.jpg'}
    Traceback (most recent call last):
      File "/home/scit/Desktop/My_work/object_detection/mask_rcnn/mrcnn/model.py", line 1717, in data_generator
        use_mini_mask=config.USE_MINI_MASK)
      File "/home/scit/Desktop/My_work/object_detection/mask_rcnn/mrcnn/model.py", line 1219, in load_image_gt
        mask, class_ids = dataset.load_mask(image_id)
      File "neural_code.py", line 235, in load_mask
        rr, cc = skimage.draw.polygon(p['y'], p['x'])
      File "/home/scit/anaconda3/lib/python3.6/site-packages/skimage/draw/draw.py", line 441, in polygon
        return _polygon(r, c, shape)
      File "skimage/draw/_draw.pyx", line 217, in skimage.draw._draw._polygon (skimage/draw/_draw.c:4402)
    OverflowError: Python int too large to convert to C ssize_t

我的笔记本电脑显卡规范如下:

具有 250 个 CUDA 内核的 Nvidia GeForce 830M (2 GB)

CPU 规范:

英特尔酷睿 i5(第 4 代)，8 GB 内存

这里可能是什么情况？是图像的分辨率还是我的 GPU 的无能为力。我应该继续处理 CPU 吗？

Answer 1

我在训练我的自定义数据集时与 Mask RCNN 分享我的观察结果。

我的数据集包含各种尺寸的图像(即最小图像大约有 1700 x 1600 像素，最大图像大约有 8500 x 4600 像素)。

我在 nVIDIA RTX 2080Ti、32 GB DDR4 RAM 上进行训练，在训练时我收到以下警告；但训练过程完成。

UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory. "Converting sparse IndexedSlices to a dense Tensor of unknown shape. "

2019-05-23 15:25:23.433774: W T:\src\github\tensorflow\tensorflow\core\common_runtime\bfc_allocator.cc:219] Allocator (GPU_0_bfc) ran out of memory trying to allocate 3.14GiB. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory were available.

几个月前，我尝试了 Matterport Splash of Color Example在我的笔记本电脑上，它有 12 GB RAM 和 nVIDIA 920M (2GB GPU)；并遇到过类似的内存错误。

因此，我们可以怀疑 GPU 内存的大小是导致此错误的一个因素。

此外，批量大小是另一个影响因素；但我看到您设置了 IMAGE_PER_GPU=1。如果您在 mrcnn 文件夹中的 config.py 文件中搜索 BATCH_SIZE，您会发现 –

self.BATCH_SIZE = self.IMAGES_PER_GPU * self.GPU_COUNT

因此，在您的情况下，batch_size 为 1。

总而言之，我建议您在更强大的 GPU 上尝试相同的代码。