{"cells":[{"metadata":{},"cell_type":"markdown","source":"
\n

\n Abnormalities Detection Using YOLOv4

\n\n
"},{"metadata":{},"cell_type":"markdown","source":"# Import required packages"},{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"import os\nimport shutil\nimport threading\nimport time\nimport random\nimport glob\nfrom tqdm import tqdm\nimport numpy as np\nimport pandas as pd\nimport seaborn as sns\nimport matplotlib.pyplot as plt\nimport cv2\n\nimport tensorflow as tf\nfrom tensorflow import keras\nprint('Tensorflow version: %s' % tf.__version__)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"> Copy **dataset** to working folder for easy compiling"},{"metadata":{"trusted":true},"cell_type":"code","source":"src = '../input/vinbigdata-512-image-dataset'\ndst = './vinbigdata-512-image-dataset'\n\n \n# Get total number of files from src folder\nsrc_files_count = 0\nfor (_, _, files) in os.walk(src):\n src_files_count += len(files)\n\n \ndef copied_check(src_files_count, dst):\n dst_files_count = 0\n for (_, _, files) in os.walk(dst):\n dst_files_count += len(files)\n return src_files_count == dst\n\ndef check_on_progress(src_files_count, dst):\n dst_files_count = 0\n logged = 0\n while(dst_files_count < src_files_count):\n dst_files_count = 0\n for (_, _, files) in os.walk(dst):\n dst_files_count += len(files)\n percentage = int(dst_files_count/src_files_count * 100)\n if percentage % 10 == 0 and percentage > logged:\n print('Percentage: {:d}%'.format(percentage))\n logged = percentage\n \ndef copy_dir(src, dst):\n print('Start copying')\n shutil.copytree(src, dst)\n time.sleep(0.1)\n print('Done!')\n print(dst)\n\nif not copied_check(src_files_count, dst):\n if os.path.exists(dst):\n shutil.rmtree(dst)\n # Start the copying procedure on a separate thread\n cp = threading.Thread(name='copy', target=copy_dir, args=(src, dst))\n cp.start()\n # Start the checking on a separate thread\n ch = threading.Thread(name='check', target=check_on_progress, args=(src_files_count, dst))\n ch.start()\nelse:\n print('Dataset has been already copied!')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"> Copy **darknet** to working folder for writing files, making backup, config, etc."},{"metadata":{"trusted":true},"cell_type":"code","source":"darknet_src = '../input/darknetyolo'\ndarknet_dst = './darknetyolo'\n\n# Get total number of files from src folder\nd_src_files_count = 0\nfor (_, _, files) in os.walk(darknet_src):\n d_src_files_count += len(files)\n \nif not copied_check(d_src_files_count, darknet_dst):\n if os.path.exists(darknet_dst):\n shutil.rmtree(darknet_dst)\n # Start the copying procedure on a separate thread\n cp_d = threading.Thread(name='copy_darknet', target=copy_dir, args=(darknet_src, darknet_dst))\n cp_d.start()\n # Start the checking on a separate thread\n ch_d = threading.Thread(name='check_darknet', target=check_on_progress, args=(d_src_files_count, darknet_dst))\n ch_d.start()\nelse:\n print('Dataset has been already copied!')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"> Sanity check"},{"metadata":{"trusted":true},"cell_type":"code","source":"!du -sh ../input/darknetyolo\n!du -sh ./darknetyolo\n!du -sh ../input/vinbigdata-512-image-dataset\n!du -sh ./vinbigdata-512-image-dataset","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# EDA + Preprocessing data"},{"metadata":{},"cell_type":"markdown","source":"Configuration"},{"metadata":{"trusted":true},"cell_type":"code","source":"IMG_WIDTH = 512\nIMG_HEIGHT = 512\nINPUT_SIZE = (512, 512)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"Check the `.csv` files"},{"metadata":{"trusted":true},"cell_type":"code","source":"TRAIN_DIR = './vinbigdata-512-image-dataset/vinbigdata/train'\nTEST_DIR = './vinbigdata-512-image-dataset/vinbigdata/test'\n\ntrain_df = pd.read_csv('./vinbigdata-512-image-dataset/vinbigdata/train.csv')\ntest_df = pd.read_csv('./vinbigdata-512-image-dataset/vinbigdata/test.csv')\n\ntrain_df.head(10)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"> Drop \"no finding\" images which are no useful for training process"},{"metadata":{"trusted":true},"cell_type":"code","source":"train_df = train_df[train_df['class_id']!=14].reset_index(drop=True)\n\n# Retrieve duplicated image names (which have more than 1 annotations)\ntrain_image_names = pd.unique(train_df['image_id']).tolist()\n\nNUM_TRAIN_FILES = len(train_image_names)\nNUM_TEST_FILES = len(test_df)\nprint(f'Dataset has {len(train_df)} elements after removing normal records.')\nprint(f'Number of training files:\\t{NUM_TRAIN_FILES}')\nprint(f'Number of testing files:\\t{NUM_TEST_FILES}')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Creating LabelMap\nlabel_map = train_df.loc[:, [\"class_name\", \"class_id\"]]\nlabel_map = label_map.drop_duplicates().reset_index(drop = True)\nN_CLASSES = len(label_map)\nlabel_map = label_map.sort_values(by=['class_id']).reset_index()['class_name']\nlabel_map","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"plt.figure(figsize=(10, 10))\nplt.grid(axis='x')\nsns.countplot(data=train_df, y='class_name')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"> Normalize annotations and write to `.txt` files for YOLO training"},{"metadata":{"trusted":true},"cell_type":"code","source":"for name in tqdm(train_image_names):\n element = train_df[train_df['image_id']==name].reset_index()\n \n # Extract and normalize annotations\n class_id = element['class_id']\n x_cen = 1/2 * (element['x_max'] + element['x_min']) / element['width']\n y_cen = 1/2 * (element['y_max'] + element['y_min']) / element['height']\n w = (element['x_max'] - element['x_min']) / element['width']\n h = (element['y_max'] - element['y_min']) / element['height']\n \n \n with open(os.path.join(TRAIN_DIR, name + '.txt'), 'w') as f:\n for i in range(len(element)):\n line = f'{class_id[i]} {x_cen[i]} {y_cen[i]} {w[i]} {h[i]}'\n f.write(line)\n if i < len(element) - 1:\n f.write('\\n')\n \nprint('Done!')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"> Do a sanity check"},{"metadata":{"trusted":true},"cell_type":"code","source":"indices = np.random.randint(NUM_TRAIN_FILES, size=4)\nfor name in train_df['image_id'][indices]:\n element = train_df[train_df['image_id']==name]\n with open(os.path.join(TRAIN_DIR, name + '.txt'), 'r') as f:\n if len(f.readlines()) == len(element):\n print(f'Correctly writing in file `{name}.txt`')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"!cat ./vinbigdata-512-image-dataset/vinbigdata/train/fb929e0efd696fe0f54902da5e7ec57a.txt","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"row = 4\ncol = 4\nindices = np.random.randint(len(train_image_names), size=row*col)\n\nplt.figure(figsize=(20, 20))\nfor i in tqdm(range(row*col)):\n plt.subplot(row, col, i+1)\n img = plt.imread(os.path.join(TRAIN_DIR, train_df['image_id'][indices[i]] + '.png'))\n plt.imshow(img, cmap='gray')\n plt.xticks([])\n plt.yticks([])","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"> Random colors represent for specific classes"},{"metadata":{"trusted":true},"cell_type":"code","source":"random_r = [random.uniform(0, 1) for _ in range(N_CLASSES)]\nrandom_g = [random.uniform(0, 1) for _ in range(N_CLASSES)]\nrandom_b = [random.uniform(0, 1) for _ in range(N_CLASSES)]\n\ncolor_map_with_label = list(zip(random_r, random_g, random_b))\nprint('\\n'.join(map(str, color_map_with_label)))","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"Helper function for plotting annotations"},{"metadata":{"trusted":true},"cell_type":"code","source":"def plot_boxes(img_id, directory, ax=None):\n img_path = os.path.join(directory, img_id + '.png')\n anno_path = os.path.join(directory, img_id + '.txt')\n \n # Read image\n img = plt.imread(img_path)\n \n # Convert to RGB\n img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)\n \n with open(anno_path, 'r') as f:\n box_infos = f.readlines()\n class_id_list = []\n for box_info in box_infos:\n class_id, x_cen, y_cen, w, h = list(map(float, box_info.split()))\n class_id = int(class_id)\n class_id_list.append(class_id)\n xmin = int((x_cen - w/2)*IMG_WIDTH)\n ymin = int((y_cen - h/2)*IMG_HEIGHT)\n xmax = int((x_cen + w/2)*IMG_WIDTH)\n ymax = int((y_cen + h/2)*IMG_HEIGHT)\n cv2.rectangle(\n img, \n pt1=(xmin, ymin), \n pt2=(xmax, ymax), \n color=color_map_with_label[class_id], \n thickness=2\n )\n cv2.putText(\n img, \n label_map[class_id], \n (xmin, ymin-5), \n cv2.FONT_HERSHEY_SIMPLEX, \n 0.5, \n color_map_with_label[class_id], \n 1\n )\n \n if ax:\n ax.imshow(img)\n ax.set_title(f'{len(box_infos)} abnormalities detected belonging to {len(set(class_id_list))} classes')\n else:\n plt.figure(figsize=(10, 10))\n plt.imshow(img)\n plt.axis('off')\n plt.title(f'{len(box_infos)} abnormalities detected belonging to {len(set(class_id_list))} classes')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"indices = random.sample(range(NUM_TRAIN_FILES), 2)\n\nfig, axes = plt.subplots(1, 2, figsize=(20, 10))\nplot_boxes(train_image_names[indices[0]], TRAIN_DIR, axes[0])\nplot_boxes(train_image_names[indices[1]], TRAIN_DIR, axes[1])","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Setup configurations for YOLOv4"},{"metadata":{},"cell_type":"markdown","source":"> Create `train.txt` and `val.txt`"},{"metadata":{"trusted":true},"cell_type":"code","source":"data_paths = [os.path.join(TRAIN_DIR, image_name + '.png') for image_name in train_image_names]\ndarknet_path = '/kaggle/working/darknetyolo'\n# Split into training and validation\nsplit = int(len(data_paths) * 0.8)\ntrain_data_paths = data_paths[:split]\nval_data_paths = data_paths[split:]\n\n# Write to files\nwith open(os.path.join(darknet_path, 'train.txt'), 'w') as f:\n f.write('\\n'.join(train_data_paths))\n \nwith open(os.path.join(darknet_path, 'val.txt'), 'w') as f:\n f.write('\\n'.join(val_data_paths))","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"> Set darknet config folder as read-only for preservation"},{"metadata":{"trusted":true},"cell_type":"code","source":"!chmod 0444 ./darknetyolo/cfg","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"class ModelConfiguration():\n def __init__(self, input_size, class_names=None, darknet_path=None, backup=True):\n self.input_size = input_size\n self.class_names = class_names\n self.n_classes = len(class_names)\n self.config_files = ['obj.names', 'obj.data']\n self.darknet_path = darknet_path\n if backup:\n self.backup_path = os.path.join(self.darknet_path, 'backup')\n try:\n # Create back-up folder\n os.mkdir(self.backup_path)\n except:\n pass\n \n def create_config_files(self):\n config_paths = [os.path.join(self.darknet_path, config_file) for config_file in self.config_files]\n \n # Writing `obj.names`\n with open(config_paths[0], 'w') as f:\n for i in range(self.n_classes):\n f.write(self.class_names[i])\n if i < self.n_classes - 1:\n f.write('\\n')\n \n # Writing `obj.data`\n with open(config_paths[1],'w') as f:\n f.write(f'class={self.n_classes}\\n')\n f.write('train=' + os.path.join(self.darknet_path, 'train.txt') + '\\n')\n f.write('valid=' + os.path.join(self.darknet_path, 'val.txt') + '\\n')\n f.write('names=' + config_paths[0] + '\\n')\n f.write('backup=' + self.backup_path)\n \n def config_model(self, lines_with_contents):\n '''Modify the model configuration file at certain lines\n \n Args: \n `lines_with_contents`: dictionary, in which an element contains number of line and its respectively content\n '''\n sorted_keys = list(lines_with_contents.keys())\n sorted_keys.sort()\n \n standard_yolocfg_path = os.path.join(self.darknet_path, 'cfg', 'yolov4-custom.cfg')\n yolocfg_path = os.path.join(self.darknet_path, f'yolov4-{self.n_classes}c-{self.input_size}.cfg')\n \n with open(standard_yolocfg_path, 'r') as f:\n line = f.readline()\n modified_line = None\n modified_f = open(yolocfg_path, 'w')\n \n num_line = 1\n idx = 0\n while line:\n if idx < len(sorted_keys):\n if num_line == sorted_keys[idx]:\n modified_line = lines_with_contents[sorted_keys[idx]]\n print('Changed `{}` into `{}`'.format(line.strip(), modified_line.strip()))\n idx += 1\n modified_line = line\n modified_f.write(modified_line)\n line = f.readline()\n if line:\n modified_f.write('\\n')\n num_line += 1","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"config = ModelConfiguration(\n INPUT_SIZE, \n class_names=label_map, \n darknet_path='./darknetyolo', \n backup=True\n)\n\n# Create `obj.data` and `obj.names` \nconfig.create_config_files()\n\n# Customize config file for model\nlines_with_contents = {\n # Define number of classes\n 970: f'classes={N_CLASSES}', \n 1058: f'classes={N_CLASSES}', \n 1146: f'classes={N_CLASSES}',\n \n # Batch size and Subdivisions\n 6: f'batch={64}',\n 7: f'subdivisions={16}',\n \n # Input size\n 8: f'width={IMG_WIDTH}',\n 9: f'height={IMG_HEIGHT}',\n \n # Max batches\n 20: f'max_batches={20000}',\n \n # Steps\n 22: f'steps={16000},{18000}',\n \n # Burn-in\n 19: f'burn_in={500}',\n \n # Filters before YOLO blocks\n 963: f'filters={57}',\n 1051: f'filters={57}',\n 1139: f'filters={57}'\n \n \n}\nconfig.config_model(lines_with_contents)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"!sed -n '1139p' darknetyolo/cfg/yolov4-custom.cfg","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Build darknet"},{"metadata":{"trusted":true},"cell_type":"code","source":"%cd /kaggle/working/darknetyolo\n!chmod +x ./darknet","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"!/usr/local/cuda/bin/nvcc --version\n\n!sed -i 's/OPENCV=0/OPENCV=1/' Makefile\n!sed -i 's/GPU=0/GPU=1/' Makefile\n!sed -i 's/CUDNN=0/CUDNN=1/' Makefile\n!sed -i 's/CUDNN_HALF=0/CUDNN_HALF=1/' Makefile","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"> Make darknet"},{"metadata":{"trusted":true},"cell_type":"code","source":"!make","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Train the Object Detector"},{"metadata":{"trusted":true},"cell_type":"code","source":"!ls kaggle/working","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]}],"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat":4,"nbformat_minor":4}