{"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_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"Welcome to the world where fashion meets computer vision! This is a starter kernel that applies Mask R-CNN with COCO pretrained weights to the task of [iMaterialist (Fashion) 2019 at FGVC6](https://www.kaggle.com/c/imaterialist-fashion-2019-FGVC6).","metadata":{}},{"cell_type":"code","source":"import os\nimport gc\nimport sys\nimport json\nimport glob\nimport random\nfrom pathlib import Path\n\nimport cv2\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\nimport itertools\nfrom tqdm import tqdm\n\nfrom imgaug import augmenters as iaa\nfrom sklearn.model_selection import StratifiedKFold, KFold","metadata":{"execution":{"iopub.status.busy":"2023-01-31T10:06:58.08167Z","iopub.execute_input":"2023-01-31T10:06:58.082219Z","iopub.status.idle":"2023-01-31T10:07:00.257757Z","shell.execute_reply.started":"2023-01-31T10:06:58.082164Z","shell.execute_reply":"2023-01-31T10:07:00.256975Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"DATA_DIR = Path('/kaggle/input')\nROOT_DIR = Path('/kaggle/working')\n\n# For demonstration purpose, the classification ignores attributes (only categories),\n# and the image size is set to 512, which is the same as the size of submission masks\nNUM_CATS = 46\nIMAGE_SIZE = 512","metadata":{"execution":{"iopub.status.busy":"2023-01-31T10:08:04.850683Z","iopub.execute_input":"2023-01-31T10:08:04.85104Z","iopub.status.idle":"2023-01-31T10:08:04.856014Z","shell.execute_reply.started":"2023-01-31T10:08:04.850963Z","shell.execute_reply":"2023-01-31T10:08:04.855028Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Dowload Libraries and Pretrained Weights","metadata":{}},{"cell_type":"code","source":"#!git clone https://www.github.com/akTwelve/Mask_RCN\n!git clone https://www.github.com/matterport/Mask_RCNN.git\nos.chdir('Mask_RCNN')\n\n!rm -rf .git # to prevent an error when the kernel is committed\n!rm -rf images assets # to prevent displaying images at the bottom of a kernel","metadata":{"_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","execution":{"iopub.status.busy":"2023-01-29T08:36:08.231301Z","iopub.execute_input":"2023-01-29T08:36:08.234384Z","iopub.status.idle":"2023-01-29T08:36:17.891566Z","shell.execute_reply.started":"2023-01-29T08:36:08.234332Z","shell.execute_reply":"2023-01-29T08:36:17.890414Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!pwd","metadata":{"execution":{"iopub.status.busy":"2023-01-29T08:36:17.896223Z","iopub.execute_input":"2023-01-29T08:36:17.898221Z","iopub.status.idle":"2023-01-29T08:36:18.897446Z","shell.execute_reply.started":"2023-01-29T08:36:17.898161Z","shell.execute_reply":"2023-01-29T08:36:18.896548Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"os.chdir('Mask_RCNN')\nsys.path.append(ROOT_DIR/'Mask_RCNN')\nfrom mrcnn.config import Config\nfrom mrcnn import utils\nimport mrcnn.model as modellib\nfrom mrcnn import visualize\nfrom mrcnn.model import log","metadata":{"_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","execution":{"iopub.status.busy":"2023-01-31T10:08:18.186074Z","iopub.execute_input":"2023-01-31T10:08:18.186557Z","iopub.status.idle":"2023-01-31T10:08:19.208721Z","shell.execute_reply.started":"2023-01-31T10:08:18.186497Z","shell.execute_reply":"2023-01-31T10:08:19.207215Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!wget --quiet https://github.com/matterport/Mask_RCNN/releases/download/v2.0/mask_rcnn_coco.h5\n!ls -lh mask_rcnn_coco.h5\n\nCOCO_WEIGHTS_PATH = 'mask_rcnn_coco.h5'","metadata":{"execution":{"iopub.status.busy":"2023-01-31T10:08:26.538176Z","iopub.execute_input":"2023-01-31T10:08:26.53857Z","iopub.status.idle":"2023-01-31T10:08:43.124183Z","shell.execute_reply.started":"2023-01-31T10:08:26.538524Z","shell.execute_reply":"2023-01-31T10:08:43.123216Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Set Config","metadata":{}},{"cell_type":"markdown","source":"Mask R-CNN has a load of hyperparameters. I only adjust some of them.","metadata":{}},{"cell_type":"code","source":"class FashionConfig(Config):\n    NAME = \"fashion\"\n    NUM_CLASSES = NUM_CATS + 1 # +1 for the background class\n    \n    GPU_COUNT = 1\n    IMAGES_PER_GPU = 4 # a memory error occurs when IMAGES_PER_GPU is too high\n    \n    BACKBONE = 'resnet50'\n    \n    IMAGE_MIN_DIM = IMAGE_SIZE\n    IMAGE_MAX_DIM = IMAGE_SIZE    \n    IMAGE_RESIZE_MODE = 'none'\n    \n    RPN_ANCHOR_SCALES = (16, 32, 64, 128, 256)\n    #DETECTION_NMS_THRESHOLD = 0.0\n    \n    # STEPS_PER_EPOCH should be the number of instances \n    # divided by (GPU_COUNT*IMAGES_PER_GPU), and so should VALIDATION_STEPS;\n    # however, due to the time limit, I set them so that this kernel can be run in 9 hours\n    STEPS_PER_EPOCH = 1000\n    VALIDATION_STEPS = 200\n    \nconfig = FashionConfig()\nconfig.display()","metadata":{"execution":{"iopub.status.busy":"2023-01-31T10:08:45.31493Z","iopub.execute_input":"2023-01-31T10:08:45.315271Z","iopub.status.idle":"2023-01-31T10:08:45.326675Z","shell.execute_reply.started":"2023-01-31T10:08:45.315206Z","shell.execute_reply":"2023-01-31T10:08:45.323537Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Make Datasets","metadata":{}},{"cell_type":"code","source":"with open(DATA_DIR/\"label_descriptions.json\") as f:\n    label_descriptions = json.load(f)\n\nlabel_names = [x['name'] for x in label_descriptions['categories']]","metadata":{"execution":{"iopub.status.busy":"2023-01-31T10:09:14.496642Z","iopub.execute_input":"2023-01-31T10:09:14.497005Z","iopub.status.idle":"2023-01-31T10:09:14.508496Z","shell.execute_reply.started":"2023-01-31T10:09:14.496926Z","shell.execute_reply":"2023-01-31T10:09:14.507782Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"segment_df = pd.read_csv(DATA_DIR/\"train.csv\")\n\nmultilabel_percent = len(segment_df[segment_df['ClassId'].str.contains('_')])/len(segment_df)*100\nprint(f\"Segments that have attributes: {multilabel_percent:.2f}%\")","metadata":{"execution":{"iopub.status.busy":"2023-01-31T10:09:16.810379Z","iopub.execute_input":"2023-01-31T10:09:16.81068Z","iopub.status.idle":"2023-01-31T10:09:57.315273Z","shell.execute_reply.started":"2023-01-31T10:09:16.810638Z","shell.execute_reply":"2023-01-31T10:09:57.314074Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Segments that contain attributes are only 3.46% of data, and [according to the host](https://www.kaggle.com/c/imaterialist-fashion-2019-FGVC6/discussion/90643#523135), 80% of images have no attribute. So, in the first step, we can only deal with categories to reduce the complexity of the task.","metadata":{}},{"cell_type":"code","source":"segment_df['CategoryId'] = segment_df['ClassId'].str.split('_').str[0]\n\nprint(\"Total segments: \", len(segment_df))\nsegment_df.head()","metadata":{"execution":{"iopub.status.busy":"2023-01-31T10:11:28.580641Z","iopub.execute_input":"2023-01-31T10:11:28.58112Z","iopub.status.idle":"2023-01-31T10:11:29.276471Z","shell.execute_reply.started":"2023-01-31T10:11:28.580895Z","shell.execute_reply":"2023-01-31T10:11:29.275028Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Rows with the same image are grouped together because the subsequent operations perform in an image level.","metadata":{}},{"cell_type":"code","source":"image_df = segment_df.groupby('ImageId')['EncodedPixels', 'CategoryId'].agg(lambda x: list(x))\nsize_df = segment_df.groupby('ImageId')['Height', 'Width'].mean()\nimage_df = image_df.join(size_df, on='ImageId')\n\nprint(\"Total images: \", len(image_df))\nimage_df.head()","metadata":{"execution":{"iopub.status.busy":"2023-01-31T10:24:12.266047Z","iopub.execute_input":"2023-01-31T10:24:12.266922Z","iopub.status.idle":"2023-01-31T10:24:20.106614Z","shell.execute_reply.started":"2023-01-31T10:24:12.266862Z","shell.execute_reply":"2023-01-31T10:24:20.105218Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Here is the custom function that resizes an image.","metadata":{}},{"cell_type":"code","source":"def resize_image(image_path):\n    img = cv2.imread(image_path)\n    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n    img = cv2.resize(img, (IMAGE_SIZE, IMAGE_SIZE), interpolation=cv2.INTER_AREA)  \n    return img","metadata":{"execution":{"iopub.status.busy":"2023-01-31T10:11:39.077114Z","iopub.execute_input":"2023-01-31T10:11:39.077411Z","iopub.status.idle":"2023-01-31T10:11:39.083904Z","shell.execute_reply.started":"2023-01-31T10:11:39.07735Z","shell.execute_reply":"2023-01-31T10:11:39.082641Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"The crucial part is to create a dataset for this task.","metadata":{}},{"cell_type":"code","source":"class FashionDataset(utils.Dataset):\n\n    def __init__(self, df):\n        super().__init__(self)\n        \n        # Add classes\n        for i, name in enumerate(label_names):\n            self.add_class(\"fashion\", i+1, name)\n        \n        # Add images \n        for i, row in df.iterrows():\n            self.add_image(\"fashion\", \n                           image_id=row.name, \n                           path=str(DATA_DIR/'train'/row.name), \n                           labels=row['CategoryId'],\n                           annotations=row['EncodedPixels'], \n                           height=row['Height'], width=row['Width'])\n\n    def image_reference(self, image_id):\n        info = self.image_info[image_id]\n        return info['path'], [label_names[int(x)] for x in info['labels']]\n    \n    def load_image(self, image_id):\n        return resize_image(self.image_info[image_id]['path'])\n\n    def load_mask(self, image_id):\n        info = self.image_info[image_id]\n                \n        mask = np.zeros((IMAGE_SIZE, IMAGE_SIZE, len(info['annotations'])), dtype=np.uint8)\n        labels = []\n        \n        for m, (annotation, label) in enumerate(zip(info['annotations'], info['labels'])):\n            sub_mask = np.full(info['height']*info['width'], 0, dtype=np.uint8)\n            annotation = [int(x) for x in annotation.split(' ')]\n            \n            for i, start_pixel in enumerate(annotation[::2]):\n                sub_mask[start_pixel: start_pixel+annotation[2*i+1]] = 1\n\n            sub_mask = sub_mask.reshape((info['height'], info['width']), order='F')\n            sub_mask = cv2.resize(sub_mask, (IMAGE_SIZE, IMAGE_SIZE), interpolation=cv2.INTER_NEAREST)\n            \n            mask[:, :, m] = sub_mask\n            labels.append(int(label)+1)\n            \n        return mask, np.array(labels)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:22:41.288707Z","iopub.execute_input":"2023-01-31T07:22:41.289Z","iopub.status.idle":"2023-01-31T07:22:41.298878Z","shell.execute_reply.started":"2023-01-31T07:22:41.288949Z","shell.execute_reply":"2023-01-31T07:22:41.298061Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Let's visualize some random images and their masks.","metadata":{}},{"cell_type":"code","source":"dataset = FashionDataset(image_df)\ndataset.prepare()\n\nfor i in range(6):\n    image_id = random.choice(dataset.image_ids)\n    print(dataset.image_reference(image_id))\n    \n    image = dataset.load_image(image_id)\n    mask, class_ids = dataset.load_mask(image_id)\n    visualize.display_top_masks(image, mask, class_ids, dataset.class_names, limit=4)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:22:44.376089Z","iopub.execute_input":"2023-01-31T07:22:44.376413Z","iopub.status.idle":"2023-01-31T07:22:57.229734Z","shell.execute_reply.started":"2023-01-31T07:22:44.376355Z","shell.execute_reply":"2023-01-31T07:22:57.229159Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Now, the data are partitioned into train and validation sets.","metadata":{}},{"cell_type":"code","source":"# This code partially supports k-fold training, \n# you can specify the fold to train and the total number of folds here\nFOLD = 0\nN_FOLDS = 5\n\nkf = KFold(n_splits=N_FOLDS, random_state=42, shuffle=True)\nsplits = kf.split(image_df) # ideally, this should be multilabel stratification\n\ndef get_fold():    \n    for i, (train_index, valid_index) in enumerate(splits):\n        if i == FOLD:\n            return image_df.iloc[train_index], image_df.iloc[valid_index]\n        \ntrain_df, valid_df = get_fold()\n\ntrain_dataset = FashionDataset(train_df)\ntrain_dataset.prepare()\n\nvalid_dataset = FashionDataset(valid_df)\nvalid_dataset.prepare()","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:23:09.094646Z","iopub.execute_input":"2023-01-31T07:23:09.094935Z","iopub.status.idle":"2023-01-31T07:23:15.139845Z","shell.execute_reply.started":"2023-01-31T07:23:09.094881Z","shell.execute_reply":"2023-01-31T07:23:15.139008Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Let's visualize class distributions of the train and validation data.","metadata":{}},{"cell_type":"code","source":"train_segments = np.concatenate(train_df['CategoryId'].values).astype(int)\nprint(\"Total train images: \", len(train_df))\nprint(\"Total train segments: \", len(train_segments))\n\nplt.figure(figsize=(12, 3))\nvalues, counts = np.unique(train_segments, return_counts=True)\nplt.bar(values, counts)\nplt.xticks(values, label_names, rotation='vertical')\nplt.show()\n\nvalid_segments = np.concatenate(valid_df['CategoryId'].values).astype(int)\nprint(\"Total train images: \", len(valid_df))\nprint(\"Total validation segments: \", len(valid_segments))\n\nplt.figure(figsize=(12, 3))\nvalues, counts = np.unique(valid_segments, return_counts=True)\nplt.bar(values, counts)\nplt.xticks(values, label_names, rotation='vertical')\nplt.show()","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:23:17.592048Z","iopub.execute_input":"2023-01-31T07:23:17.592403Z","iopub.status.idle":"2023-01-31T07:23:19.07187Z","shell.execute_reply.started":"2023-01-31T07:23:17.592342Z","shell.execute_reply":"2023-01-31T07:23:19.071154Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Train","metadata":{}},{"cell_type":"code","source":"# Note that any hyperparameters here, such as LR, may still not be optimal\nLR = 1e-4\nEPOCHS = [2, 6, 8]\n\nimport warnings \nwarnings.filterwarnings(\"ignore\")","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:23:23.496208Z","iopub.execute_input":"2023-01-31T07:23:23.496508Z","iopub.status.idle":"2023-01-31T07:23:23.500611Z","shell.execute_reply.started":"2023-01-31T07:23:23.496457Z","shell.execute_reply":"2023-01-31T07:23:23.499829Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"This section creates a Mask R-CNN model and specifies augmentations to be used.","metadata":{}},{"cell_type":"code","source":"model = modellib.MaskRCNN(mode='training', config=config, model_dir=ROOT_DIR)\n\nmodel.load_weights(COCO_WEIGHTS_PATH, by_name=True, exclude=[\n    'mrcnn_class_logits', 'mrcnn_bbox_fc', 'mrcnn_bbox', 'mrcnn_mask'])","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:09.804892Z","iopub.execute_input":"2023-01-31T07:24:09.805235Z","iopub.status.idle":"2023-01-31T07:24:21.73648Z","shell.execute_reply.started":"2023-01-31T07:24:09.805168Z","shell.execute_reply":"2023-01-31T07:24:21.735589Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"augmentation = iaa.Sequential([\n    iaa.Fliplr(0.5) # only horizontal flip here\n])","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:32.198778Z","iopub.execute_input":"2023-01-31T07:24:32.199073Z","iopub.status.idle":"2023-01-31T07:24:32.204384Z","shell.execute_reply.started":"2023-01-31T07:24:32.199017Z","shell.execute_reply":"2023-01-31T07:24:32.20329Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"**ПРИХОДИТСЯ ЧИСТИТЬ ДАТАСЕТ**","metadata":{}},{"cell_type":"code","source":"#Список изображений которые есть на данный момент\nspisok=os.listdir(str(DATA_DIR/'test'))\nlen(spisok)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:36.26439Z","iopub.execute_input":"2023-01-31T07:24:36.264688Z","iopub.status.idle":"2023-01-31T07:24:36.660794Z","shell.execute_reply.started":"2023-01-31T07:24:36.264626Z","shell.execute_reply":"2023-01-31T07:24:36.660069Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sample_df = pd.read_csv(DATA_DIR/\"sample_submission.csv\")\nsample_df.head()","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:37.711724Z","iopub.execute_input":"2023-01-31T07:24:37.712016Z","iopub.status.idle":"2023-01-31T07:24:37.739172Z","shell.execute_reply.started":"2023-01-31T07:24:37.711959Z","shell.execute_reply":"2023-01-31T07:24:37.738205Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"len(sample_df)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:39.327666Z","iopub.execute_input":"2023-01-31T07:24:39.327965Z","iopub.status.idle":"2023-01-31T07:24:39.333565Z","shell.execute_reply.started":"2023-01-31T07:24:39.327909Z","shell.execute_reply":"2023-01-31T07:24:39.332779Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"list2=[]\nfor i in sample_df['ImageId']:\n  if i in spisok:\n    continue\n  else:\n    list2.append(i)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:40.567959Z","iopub.execute_input":"2023-01-31T07:24:40.568319Z","iopub.status.idle":"2023-01-31T07:24:40.656517Z","shell.execute_reply.started":"2023-01-31T07:24:40.568246Z","shell.execute_reply":"2023-01-31T07:24:40.655737Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"list2","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:42.304544Z","iopub.execute_input":"2023-01-31T07:24:42.304841Z","iopub.status.idle":"2023-01-31T07:24:42.313862Z","shell.execute_reply.started":"2023-01-31T07:24:42.304786Z","shell.execute_reply":"2023-01-31T07:24:42.313168Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sample_df=sample_df.reset_index(drop=True)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:43.711733Z","iopub.execute_input":"2023-01-31T07:24:43.712031Z","iopub.status.idle":"2023-01-31T07:24:43.719024Z","shell.execute_reply.started":"2023-01-31T07:24:43.711973Z","shell.execute_reply":"2023-01-31T07:24:43.71825Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"for i in range(len(sample_df)-len(list2)):\n    for d in list2:\n      if sample_df['ImageId'][i]==d:\n        sample_df=sample_df.drop(i)\n        sample_df=sample_df.reset_index(drop=True)\n\n#sample_df.drop('12fea2852daf1832ce9646dd4554ada2.jpg', axis=0)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:45.575683Z","iopub.execute_input":"2023-01-31T07:24:45.575982Z","iopub.status.idle":"2023-01-31T07:24:45.856113Z","shell.execute_reply.started":"2023-01-31T07:24:45.575928Z","shell.execute_reply":"2023-01-31T07:24:45.854362Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"len(sample_df)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:48.295844Z","iopub.execute_input":"2023-01-31T07:24:48.296178Z","iopub.status.idle":"2023-01-31T07:24:48.303Z","shell.execute_reply.started":"2023-01-31T07:24:48.296098Z","shell.execute_reply":"2023-01-31T07:24:48.301974Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"len(spisok)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:24:49.430347Z","iopub.execute_input":"2023-01-31T07:24:49.430641Z","iopub.status.idle":"2023-01-31T07:24:49.435914Z","shell.execute_reply.started":"2023-01-31T07:24:49.430586Z","shell.execute_reply":"2023-01-31T07:24:49.435089Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"**ВОТ ДО СЮДА. Готово. Из списка предсказаний убраны отсутствующие файлы**","metadata":{}},{"cell_type":"markdown","source":"Вот те,что удалены","metadata":{}},{"cell_type":"markdown","source":"First, we train only the heads.","metadata":{}},{"cell_type":"code","source":"%%time\nmodel.train(train_dataset, valid_dataset,\n            learning_rate=LR*2,\n# Сначала 2 эпохи для обучения только \"головы\" с высоким LR\n            epochs=EPOCHS[0],\n            layers='heads',\n            augmentation=None)\n\nhistory = model.keras_model.history.history","metadata":{"execution":{"iopub.status.busy":"2023-01-29T08:38:08.953647Z","iopub.execute_input":"2023-01-29T08:38:08.954067Z","iopub.status.idle":"2023-01-29T09:55:55.433897Z","shell.execute_reply.started":"2023-01-29T08:38:08.954018Z","shell.execute_reply":"2023-01-29T09:55:55.432534Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Then, all layers are trained.","metadata":{}},{"cell_type":"code","source":"%%time\nmodel.train(train_dataset, valid_dataset,\n            learning_rate=LR,\n            epochs=EPOCHS[1],\n            layers='all',\n            augmentation=augmentation)\n\nnew_history = model.keras_model.history.history\nfor k in new_history: history[k] = history[k] + new_history[k]","metadata":{"execution":{"iopub.status.busy":"2023-01-29T09:55:55.436564Z","iopub.execute_input":"2023-01-29T09:55:55.436883Z","iopub.status.idle":"2023-01-29T12:30:15.64936Z","shell.execute_reply.started":"2023-01-29T09:55:55.436816Z","shell.execute_reply":"2023-01-29T12:30:15.64744Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Afterwards, we reduce LR and train again.","metadata":{}},{"cell_type":"code","source":"%%time\nmodel.train(train_dataset, valid_dataset,\n            learning_rate=LR/5,\n            epochs=EPOCHS[2],\n            layers='all',\n            augmentation=augmentation)\n\nnew_history = model.keras_model.history.history\nfor k in new_history: history[k] = history[k] + new_history[k]","metadata":{"execution":{"iopub.status.busy":"2023-01-29T12:30:15.651322Z","iopub.execute_input":"2023-01-29T12:30:15.655294Z","iopub.status.idle":"2023-01-29T13:54:06.763787Z","shell.execute_reply.started":"2023-01-29T12:30:15.655204Z","shell.execute_reply":"2023-01-29T13:54:06.759606Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Let's visualize training history and choose the best epoch.","metadata":{}},{"cell_type":"code","source":"epochs = range(EPOCHS[-1])\n\nplt.figure(figsize=(18, 6))\n\nplt.subplot(131)\nplt.plot(epochs, history['loss'], label=\"train loss\")\nplt.plot(epochs, history['val_loss'], label=\"valid loss\")\nplt.legend()\nplt.subplot(132)\nplt.plot(epochs, history['mrcnn_class_loss'], label=\"train class loss\")\nplt.plot(epochs, history['val_mrcnn_class_loss'], label=\"valid class loss\")\nplt.legend()\nplt.subplot(133)\nplt.plot(epochs, history['mrcnn_mask_loss'], label=\"train mask loss\")\nplt.plot(epochs, history['val_mrcnn_mask_loss'], label=\"valid mask loss\")\nplt.legend()\n\nplt.show()","metadata":{"execution":{"iopub.status.busy":"2023-01-29T13:54:06.76665Z","iopub.execute_input":"2023-01-29T13:54:06.771827Z","iopub.status.idle":"2023-01-29T13:54:10.602786Z","shell.execute_reply.started":"2023-01-29T13:54:06.768698Z","shell.execute_reply":"2023-01-29T13:54:10.600597Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"best_epoch = np.argmin(history[\"val_loss\"]) + 1\nprint(\"Best epoch: \", best_epoch)\nprint(\"Valid loss: \", history[\"val_loss\"][best_epoch-1])","metadata":{"execution":{"iopub.status.busy":"2023-01-29T13:54:10.604876Z","iopub.execute_input":"2023-01-29T13:54:10.609923Z","iopub.status.idle":"2023-01-29T13:54:10.629212Z","shell.execute_reply.started":"2023-01-29T13:54:10.605143Z","shell.execute_reply":"2023-01-29T13:54:10.62333Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Predict","metadata":{}},{"cell_type":"markdown","source":"The final step is to use our model to predict test data.","metadata":{}},{"cell_type":"code","source":"glob_list = glob.glob(f'/kaggle/working/fashion*/mask_rcnn_fashion_{best_epoch:04d}.h5')\nmodel_path = glob_list[0] if glob_list else ''","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:28:27.18332Z","iopub.execute_input":"2023-01-31T07:28:27.183603Z","iopub.status.idle":"2023-01-31T07:28:27.20074Z","shell.execute_reply.started":"2023-01-31T07:28:27.183553Z","shell.execute_reply":"2023-01-31T07:28:27.199553Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"This cell defines InferenceConfig and loads the best trained model.","metadata":{}},{"cell_type":"code","source":"class InferenceConfig(FashionConfig):\n    GPU_COUNT = 1\n    IMAGES_PER_GPU = 1\n\ninference_config = InferenceConfig()\n\nmodel = modellib.MaskRCNN(mode='inference', \n                          config=inference_config,\n                          model_dir=ROOT_DIR)\n\nassert model_path != '', \"Provide path to trained weights\"\nprint(\"Loading weights from \", model_path)\nmodel.load_weights(model_path, by_name=True)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:28:22.541374Z","iopub.execute_input":"2023-01-31T07:28:22.54167Z","iopub.status.idle":"2023-01-31T07:28:25.589308Z","shell.execute_reply.started":"2023-01-31T07:28:22.541608Z","shell.execute_reply":"2023-01-31T07:28:25.588321Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Then, load the submission data.","metadata":{}},{"cell_type":"markdown","source":"## ","metadata":{}},{"cell_type":"markdown","source":"Here is the main prediction steps, along with some helper functions.","metadata":{}},{"cell_type":"code","source":"# Convert data to run-length encoding\ndef to_rle(bits):\n    rle = []\n    pos = 0\n    for bit, group in itertools.groupby(bits):\n        group_list = list(group)\n        if bit:\n            rle.extend([pos, sum(group_list)])\n        pos += len(group_list)\n    return rle","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:28:19.557242Z","iopub.execute_input":"2023-01-31T07:28:19.557533Z","iopub.status.idle":"2023-01-31T07:28:19.562921Z","shell.execute_reply.started":"2023-01-31T07:28:19.557482Z","shell.execute_reply":"2023-01-31T07:28:19.561573Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Since the submission system does not permit overlapped masks, we have to fix them\ndef refine_masks(masks, rois):\n    areas = np.sum(masks.reshape(-1, masks.shape[-1]), axis=0)\n    mask_index = np.argsort(areas)\n    union_mask = np.zeros(masks.shape[:-1], dtype=bool)\n    for m in mask_index:\n        masks[:, :, m] = np.logical_and(masks[:, :, m], np.logical_not(union_mask))\n        union_mask = np.logical_or(masks[:, :, m], union_mask)\n    for m in range(masks.shape[-1]):\n        mask_pos = np.where(masks[:, :, m]==True)\n        if np.any(mask_pos):\n            y1, x1 = np.min(mask_pos, axis=1)\n            y2, x2 = np.max(mask_pos, axis=1)\n            rois[m, :] = [y1, x1, y2, x2]\n    return masks, rois","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:28:18.158649Z","iopub.execute_input":"2023-01-31T07:28:18.158951Z","iopub.status.idle":"2023-01-31T07:28:18.165451Z","shell.execute_reply.started":"2023-01-31T07:28:18.158896Z","shell.execute_reply":"2023-01-31T07:28:18.16458Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"%%time\nsub_list = []\nmissing_count = 0\nfor i, row in tqdm(sample_df.iterrows(), total=len(sample_df)):\n    image = resize_image(str(DATA_DIR/'test'/row['ImageId']))\n    result = model.detect([image])[0]\n    if result['masks'].size > 0:\n        masks, _ = refine_masks(result['masks'], result['rois'])\n        for m in range(masks.shape[-1]):\n            mask = masks[:, :, m].ravel(order='F')\n            rle = to_rle(mask)\n            label = result['class_ids'][m] - 1\n            sub_list.append([row['ImageId'], ' '.join(list(map(str, rle))), label])\n    else:\n        # The system does not allow missing ids, this is an easy way to fill them \n        sub_list.append([row['ImageId'], '1 1', 23])\n        missing_count += 1","metadata":{"_kg_hide-input":true,"_kg_hide-output":true,"execution":{"iopub.status.busy":"2023-01-31T07:28:09.60695Z","iopub.execute_input":"2023-01-31T07:28:09.607271Z","iopub.status.idle":"2023-01-31T07:28:09.660015Z","shell.execute_reply.started":"2023-01-31T07:28:09.607211Z","shell.execute_reply":"2023-01-31T07:28:09.659268Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!pwd","metadata":{"execution":{"iopub.status.busy":"2023-01-29T14:18:35.849635Z","iopub.execute_input":"2023-01-29T14:18:35.850158Z","iopub.status.idle":"2023-01-29T14:18:36.933901Z","shell.execute_reply.started":"2023-01-29T14:18:35.850086Z","shell.execute_reply":"2023-01-29T14:18:36.932964Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"The submission file is created, when all predictions are ready.","metadata":{}},{"cell_type":"code","source":"submission_df = pd.DataFrame(sub_list, columns=sample_df.columns.values)\nprint(\"Total image results: \", submission_df['ImageId'].nunique())\nprint(\"Missing Images: \", missing_count)\nsubmission_df.head()","metadata":{"execution":{"iopub.status.busy":"2023-01-29T14:18:36.937077Z","iopub.execute_input":"2023-01-29T14:18:36.937386Z","iopub.status.idle":"2023-01-29T14:18:37.028244Z","shell.execute_reply.started":"2023-01-29T14:18:36.937327Z","shell.execute_reply":"2023-01-29T14:18:37.027463Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"submission_df.to_csv(\"submission.csv\", index=False)","metadata":{"execution":{"iopub.status.busy":"2023-01-29T14:18:37.029378Z","iopub.execute_input":"2023-01-29T14:18:37.029648Z","iopub.status.idle":"2023-01-29T14:18:37.495801Z","shell.execute_reply.started":"2023-01-29T14:18:37.029603Z","shell.execute_reply":"2023-01-29T14:18:37.494973Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Finally, it's pleasing to visualize the results! Sample images contain both fashion models and predictions from the Mask R-CNN model.","metadata":{}},{"cell_type":"code","source":"for i in range(9):\n    image_id = sample_df.sample()['ImageId'].values[0]\n    image_path = str(DATA_DIR/'test'/image_id)\n    \n    img = cv2.imread(image_path)\n    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n    \n    result = model.detect([resize_image(image_path)])\n    r = result[0]\n    \n    if r['masks'].size > 0:\n        masks = np.zeros((img.shape[0], img.shape[1], r['masks'].shape[-1]), dtype=np.uint8)\n        for m in range(r['masks'].shape[-1]):\n            masks[:, :, m] = cv2.resize(r['masks'][:, :, m].astype('uint8'), \n                                        (img.shape[1], img.shape[0]), interpolation=cv2.INTER_NEAREST)\n        \n        y_scale = img.shape[0]/IMAGE_SIZE\n        x_scale = img.shape[1]/IMAGE_SIZE\n        rois = (r['rois'] * [y_scale, x_scale, y_scale, x_scale]).astype(int)\n        \n        masks, rois = refine_masks(masks, rois)\n    else:\n        masks, rois = r['masks'], r['rois']\n        \n    visualize.display_instances(img, rois, masks, r['class_ids'], \n                                ['bg']+label_names, r['scores'],\n                                title=image_id, figsize=(12, 12))","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:28:39.960196Z","iopub.execute_input":"2023-01-31T07:28:39.960486Z","iopub.status.idle":"2023-01-31T07:29:52.30508Z","shell.execute_reply.started":"2023-01-31T07:28:39.960434Z","shell.execute_reply":"2023-01-31T07:29:52.303565Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import tensorflow as tf\nfrom tensorflow import keras","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:25:59.064644Z","iopub.execute_input":"2023-01-31T07:25:59.06496Z","iopub.status.idle":"2023-01-31T07:25:59.069605Z","shell.execute_reply.started":"2023-01-31T07:25:59.064905Z","shell.execute_reply":"2023-01-31T07:25:59.068663Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#model.keras_model.save_weights(ROOT_DIR/'my_model')","metadata":{"execution":{"iopub.status.busy":"2023-01-29T14:18:56.186768Z","iopub.execute_input":"2023-01-29T14:18:56.18724Z","iopub.status.idle":"2023-01-29T14:20:23.505631Z","shell.execute_reply.started":"2023-01-29T14:18:56.187186Z","shell.execute_reply":"2023-01-29T14:20:23.504758Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"model.keras_model.load_weights(ROOT_DIR/'my_model', by_name=True)","metadata":{"execution":{"iopub.status.busy":"2023-01-31T07:27:41.522313Z","iopub.execute_input":"2023-01-31T07:27:41.522618Z","iopub.status.idle":"2023-01-31T07:27:42.794745Z","shell.execute_reply.started":"2023-01-31T07:27:41.522564Z","shell.execute_reply":"2023-01-31T07:27:42.793891Z"},"trusted":true},"execution_count":null,"outputs":[]}]}