{"cells":[{"metadata":{},"cell_type":"markdown","source":"# 1. Base\n## 1-1. Model Parameters"},{"metadata":{"trusted":true},"cell_type":"code","source":"BATCH_SIZE = 48\nEDGE_CROP = 16\nGAUSSIAN_NOISE = 0.1\nUPSAMPLE_MODE = 'SIMPLE'\n# downsampling inside the network\nNET_SCALING = (1, 1)\n# downsampling in preprocessing\nIMG_SCALING = (3, 3)\n# number of validation images to use\nVALID_IMG_COUNT = 900\n# maximum number of steps_per_epoch in training\nMAX_TRAIN_STEPS = 9\nMAX_TRAIN_EPOCHS = 99\nAUGMENT_BRIGHTNESS = False","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 1-2. Import Modules"},{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"import numpy as np\nimport pandas as pd\nimport os\nimport matplotlib.pyplot as plt\nfrom skimage.io import imread\nfrom skimage.morphology import binary_opening, disk, label","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 1-3. Read Data"},{"metadata":{"trusted":true},"cell_type":"code","source":"BASE_DIR = '/kaggle/input/airbus-ship-detection/'\nTRAIN_DIR = BASE_DIR + '/train_v2/'\nTEST_DIR = BASE_DIR + '/test_v2/'","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train = os.listdir(TRAIN_DIR)\ntest = os.listdir(TEST_DIR)\n\nprint(f\"Train files: {len(train)}. ---> {train[:3]}\")\nprint(f\"Test files :  {len(test)}. ---> {test[:3]}\")","execution_count":null,"outputs":[]},{"metadata":{"_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","trusted":true},"cell_type":"code","source":"from PIL import Image\n\nImage.open(TRAIN_DIR+train[0])","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"masks = pd.read_csv(os.path.join(BASE_DIR, 'train_ship_segmentations_v2.csv'))\nnot_empty = pd.notna(masks.EncodedPixels)\nprint(not_empty.sum(), 'masks in', masks[not_empty].ImageId.nunique(), 'images')\nprint((~not_empty).sum(), 'empty images in', masks.ImageId.nunique(), 'total images')\nmasks.head()","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 1-4. Utility functions"},{"metadata":{"trusted":true},"cell_type":"code","source":"def multi_rle_encode(img, **kwargs):\n    '''\n    Encode connected regions as separated masks\n    '''\n    labels = label(img)\n    if img.ndim > 2:\n        return [rle_encode(np.sum(labels==k, axis=2), **kwargs) for k in np.unique(labels[labels>0])]\n    else:\n        return [rle_encode(labels==k, **kwargs) for k in np.unique(labels[labels>0])]\n\n# ref: https://www.kaggle.com/paulorzp/run-length-encode-and-decode\ndef rle_encode(img, min_max_threshold=1e-3, max_mean_threshold=None):\n    '''\n    img: numpy array, 1 - mask, 0 - background\n    Returns run length as string formated\n    '''\n    if np.max(img) < min_max_threshold:\n        return '' ## no need to encode if it's all zeros\n    if max_mean_threshold and np.mean(img) > max_mean_threshold:\n        return '' ## ignore overfilled mask\n    pixels = img.T.flatten()\n    pixels = np.concatenate([[0], pixels, [0]])\n    runs = np.where(pixels[1:] != pixels[:-1])[0] + 1\n    runs[1::2] -= runs[::2]\n    return ' '.join(str(x) for x in runs)\n\ndef rle_decode(mask_rle, shape=(768, 768)):\n    '''\n    mask_rle: run-length as string formated (start length)\n    shape: (height,width) of array to return \n    Returns numpy array, 1 - mask, 0 - background\n    '''\n    s = mask_rle.split()\n    starts, lengths = [np.asarray(x, dtype=int) for x in (s[0:][::2], s[1:][::2])]\n    starts -= 1\n    ends = starts + lengths\n    img = np.zeros(shape[0]*shape[1], dtype=np.uint8)\n    for lo, hi in zip(starts, ends):\n        img[lo:hi] = 1\n    return img.reshape(shape).T  # Needed to align to RLE direction\n\ndef masks_as_image(in_mask_list):\n    # Take the individual ship masks and create a single mask array for all ships\n    all_masks = np.zeros((768, 768), dtype = np.uint8)\n    for mask in in_mask_list:\n        if isinstance(mask, str):\n            all_masks |= rle_decode(mask)\n    return all_masks\n\ndef masks_as_color(in_mask_list):\n    # Take the individual ship masks and create a color mask array for each ships\n    all_masks = np.zeros((768, 768), dtype = np.float)\n    scale = lambda x: (len(in_mask_list)+x+1) / (len(in_mask_list)*2) ## scale the heatmap image to shift \n    for i,mask in enumerate(in_mask_list):\n        if isinstance(mask, str):\n            all_masks[:,:] += scale(i) * rle_decode(mask)\n    return all_masks\n\ndef showImage(image_name):\n    fig, (ax1, ax2, ax3, ax4) = plt.subplots(1, 4, figsize = (16, 5))\n    rle_0 = masks.query('ImageId==\"'+image_name+'\"')['EncodedPixels']\n    img_0 = masks_as_image(rle_0)\n    ax1.imshow(img_0)\n    ax1.set_title('Mask as image')\n    rle_1 = multi_rle_encode(img_0)\n    img_1 = masks_as_image(rle_1)\n    ax2.imshow(img_1)\n    ax2.set_title('Re-encoded')\n    img_c = masks_as_color(rle_0)\n    ax3.imshow(img_c)\n    ax3.set_title('Masks in colors')\n    img_c = masks_as_color(rle_1)\n    ax4.imshow(img_c)\n    ax4.set_title('Re-encoded in colors')\n    print('Check Decoding->Encoding',\n          'RLE_0:', len(rle_0), '->',\n          'RLE_1:', len(rle_1))\n    print(np.sum(img_0 - img_1), 'error')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 1-5. Data Preparation Tools"},{"metadata":{"trusted":true},"cell_type":"code","source":"masks['ships'] = masks['EncodedPixels'].map(lambda c_row: 1 if isinstance(c_row, str) else 0)\nunique_img_ids = masks.groupby('ImageId').agg({'ships': 'sum'}).reset_index()\nunique_img_ids['has_ship'] = unique_img_ids['ships'].map(lambda x: 1.0 if x>0 else 0.0)\nmasks.drop(['ships'], axis=1, inplace=True)\nprint(unique_img_ids.loc[unique_img_ids.ships>=2].head())\nshowImage(unique_img_ids.loc[unique_img_ids.ships>=2].iloc[0].ImageId)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"unique_img_ids['ships'].hist(bins=unique_img_ids['ships'].max())\nprint('Max of ships : ',unique_img_ids['ships'].max())\nprint('Avg of ships : ',unique_img_ids['ships'].mean())","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# 2. Make dataset"},{"metadata":{},"cell_type":"markdown","source":"## 2-1. Undersample Empty Images"},{"metadata":{},"cell_type":"markdown","source":"Divide groups by the number of ships in the image to determine the maximum number of elements in each group."},{"metadata":{"trusted":true},"cell_type":"code","source":"SAMPLES_PER_GROUP = 4000\nbalanced_train_df = unique_img_ids.groupby('ships').apply(lambda x: x.sample(SAMPLES_PER_GROUP) if len(x) > SAMPLES_PER_GROUP else x)\nbalanced_train_df['ships'].hist(bins=balanced_train_df['ships'].max()+1)\nprint(balanced_train_df.shape[0], 'masks')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 2-2. Split train & test set"},{"metadata":{"trusted":true},"cell_type":"code","source":"from sklearn.model_selection import train_test_split\ntrain_ids, valid_ids = train_test_split(balanced_train_df, \n                 test_size = 0.2, \n                 stratify = balanced_train_df['ships'])\ntrain_df = pd.merge(masks, train_ids)\nvalid_df = pd.merge(masks, valid_ids)\nprint(train_df.shape[0], 'training masks')\nprint(valid_df.shape[0], 'validation masks')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"Use generator for memory saving."},{"metadata":{},"cell_type":"markdown","source":"## 2-3. Image generator"},{"metadata":{"trusted":true},"cell_type":"code","source":"BATCH_SIZE = 100\ndef make_image_gen(in_df, batch_size = BATCH_SIZE):\n    all_batches = list(in_df.groupby('ImageId'))\n    out_rgb = []\n    out_mask = []\n    while True:\n        np.random.shuffle(all_batches)\n        for c_img_id, c_masks in all_batches:\n            rgb_path = os.path.join(TRAIN_DIR, c_img_id)\n            c_img = imread(rgb_path)\n            c_mask = np.expand_dims(masks_as_image(c_masks['EncodedPixels'].values), -1)\n            if IMG_SCALING is not None:\n                c_img = c_img[::IMG_SCALING[0], ::IMG_SCALING[1]]\n                c_mask = c_mask[::IMG_SCALING[0], ::IMG_SCALING[1]]\n            out_rgb += [c_img]\n            out_mask += [c_mask]\n            if len(out_rgb)>=batch_size:\n                yield np.stack(out_rgb, 0)/255.0, np.stack(out_mask, 0)\n                out_rgb, out_mask=[], []","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train_gen = make_image_gen(train_df)\ntrain_x, train_y = next(train_gen)\nprint('x', train_x.shape, train_x.min(), train_x.max())\nprint('y', train_y.shape, train_y.min(), train_y.max())","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 2-4. Make validatation set"},{"metadata":{"trusted":true},"cell_type":"code","source":"valid_x, valid_y = next(make_image_gen(valid_df, VALID_IMG_COUNT))\nprint(valid_x.shape, valid_y.shape)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 2-5. Data Augmentation"},{"metadata":{"trusted":true},"cell_type":"code","source":"from keras.preprocessing.image import ImageDataGenerator\ndg_args = dict(featurewise_center = False, \n                  samplewise_center = False,\n                  rotation_range = 45, \n                  width_shift_range = 0.1, \n                  height_shift_range = 0.1, \n                  shear_range = 0.01,\n                  zoom_range = [0.9, 1.25],  \n                  horizontal_flip = True, \n                  vertical_flip = True,\n                  fill_mode = 'reflect',\n                   data_format = 'channels_last')\n# brightness can be problematic since it seems to change the labels differently from the images \nif AUGMENT_BRIGHTNESS:\n    dg_args[' brightness_range'] = [0.5, 1.5]\nimage_gen = ImageDataGenerator(**dg_args)\n\nif AUGMENT_BRIGHTNESS:\n    dg_args.pop('brightness_range')\nlabel_gen = ImageDataGenerator(**dg_args)\n\ndef create_aug_gen(in_gen, seed = None):\n    np.random.seed(seed if seed is not None else np.random.choice(range(9999)))\n    for in_x, in_y in in_gen:\n        seed = np.random.choice(range(9999))\n        # keep the seeds syncronized otherwise the augmentation to the images is different from the masks\n        g_x = image_gen.flow(255*in_x, \n                             batch_size = in_x.shape[0], \n                             seed = seed, \n                             shuffle=True)\n        g_y = label_gen.flow(in_y, \n                             batch_size = in_x.shape[0], \n                             seed = seed, \n                             shuffle=True)\n\n        yield next(g_x)/255.0, next(g_y)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"cur_gen = create_aug_gen(train_gen)\nt_x, t_y = next(cur_gen)\nprint('x', t_x.shape, t_x.dtype, t_x.min(), t_x.max())\nprint('y', t_y.shape, t_y.dtype, t_y.min(), t_y.max())","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"import gc; gc.enable() \n\ngc.collect()","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# 3. Modeling"},{"metadata":{},"cell_type":"markdown","source":"## 3-1. Base Model"},{"metadata":{"trusted":true},"cell_type":"code","source":"from keras import models, layers\n# Build U-Net model\ndef upsample_conv(filters, kernel_size, strides, padding):\n    return layers.Conv2DTranspose(filters, kernel_size, strides=strides, padding=padding)\ndef upsample_simple(filters, kernel_size, strides, padding):\n    return layers.UpSampling2D(strides)\n\nif UPSAMPLE_MODE=='DECONV':\n    upsample=upsample_conv\nelse:\n    upsample=upsample_simple\n    \ninput_img = layers.Input(t_x.shape[1:], name = 'RGB_Input')\npp_in_layer = input_img\n\nif NET_SCALING is not None:\n    pp_in_layer = layers.AvgPool2D(NET_SCALING)(pp_in_layer)\n    \npp_in_layer = layers.GaussianNoise(GAUSSIAN_NOISE)(pp_in_layer)\npp_in_layer = layers.BatchNormalization()(pp_in_layer)\n\nc1 = layers.Conv2D(8, (3, 3), activation='relu', padding='same') (pp_in_layer)\nc1 = layers.Conv2D(8, (3, 3), activation='relu', padding='same') (c1)\np1 = layers.MaxPooling2D((2, 2)) (c1)\n\nc2 = layers.Conv2D(16, (3, 3), activation='relu', padding='same') (p1)\nc2 = layers.Conv2D(16, (3, 3), activation='relu', padding='same') (c2)\np2 = layers.MaxPooling2D((2, 2)) (c2)\n\nc3 = layers.Conv2D(32, (3, 3), activation='relu', padding='same') (p2)\nc3 = layers.Conv2D(32, (3, 3), activation='relu', padding='same') (c3)\np3 = layers.MaxPooling2D((2, 2)) (c3)\n\nc4 = layers.Conv2D(64, (3, 3), activation='relu', padding='same') (p3)\nc4 = layers.Conv2D(64, (3, 3), activation='relu', padding='same') (c4)\np4 = layers.MaxPooling2D(pool_size=(2, 2)) (c4)\n\n\nc5 = layers.Conv2D(128, (3, 3), activation='relu', padding='same') (p4)\nc5 = layers.Conv2D(128, (3, 3), activation='relu', padding='same') (c5)\n\nu6 = upsample(64, (2, 2), strides=(2, 2), padding='same') (c5)\nu6 = layers.concatenate([u6, c4])\nc6 = layers.Conv2D(64, (3, 3), activation='relu', padding='same') (u6)\nc6 = layers.Conv2D(64, (3, 3), activation='relu', padding='same') (c6)\n\nu7 = upsample(32, (2, 2), strides=(2, 2), padding='same') (c6)\nu7 = layers.concatenate([u7, c3])\nc7 = layers.Conv2D(32, (3, 3), activation='relu', padding='same') (u7)\nc7 = layers.Conv2D(32, (3, 3), activation='relu', padding='same') (c7)\n\nu8 = upsample(16, (2, 2), strides=(2, 2), padding='same') (c7)\nu8 = layers.concatenate([u8, c2])\nc8 = layers.Conv2D(16, (3, 3), activation='relu', padding='same') (u8)\nc8 = layers.Conv2D(16, (3, 3), activation='relu', padding='same') (c8)\n\nu9 = upsample(8, (2, 2), strides=(2, 2), padding='same') (c8)\nu9 = layers.concatenate([u9, c1], axis=3)\nc9 = layers.Conv2D(8, (3, 3), activation='relu', padding='same') (u9)\nc9 = layers.Conv2D(8, (3, 3), activation='relu', padding='same') (c9)\n\nd = layers.Conv2D(1, (1, 1), activation='sigmoid') (c9)\n# d = layers.Cropping2D((EDGE_CROP, EDGE_CROP))(d)\n# d = layers.ZeroPadding2D((EDGE_CROP, EDGE_CROP))(d)\nif NET_SCALING is not None:\n    d = layers.UpSampling2D(NET_SCALING)(d)\n\nseg_model = models.Model(inputs=[input_img], outputs=[d])\nseg_model.summary()","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 3-2. loss function"},{"metadata":{"trusted":true},"cell_type":"code","source":"import keras.backend as K\nfrom keras.optimizers import Adam\nfrom keras.losses import binary_crossentropy\n\n## intersection over union\ndef IoU(y_true, y_pred, eps=1e-6):\n    intersection = K.sum(y_true * y_pred, axis=[1,2,3])\n    union = K.sum(y_true, axis=[1,2,3]) + K.sum(y_pred, axis=[1,2,3]) - intersection\n    return -K.mean( (intersection + eps) / (union + eps), axis=0)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 3-3. callbacks setting"},{"metadata":{"trusted":true},"cell_type":"code","source":"from keras.callbacks import ModelCheckpoint, LearningRateScheduler, EarlyStopping, ReduceLROnPlateau\nweight_path=\"{}_weights.best.hdf5\".format('seg_model')\n\ncheckpoint = ModelCheckpoint(weight_path, monitor='val_loss', verbose=1, save_best_only=True, mode='min', save_weights_only=True)\n\nreduceLROnPlat = ReduceLROnPlateau(monitor='val_loss', factor=0.33,\n                                   patience=1, verbose=1, mode='min',\n                                   min_delta=0.0001, cooldown=0, min_lr=1e-8)\n\nearly = EarlyStopping(monitor=\"val_loss\", mode=\"min\", verbose=2,\n                      patience=20) # probably needs to be more patient, but kaggle time is limited\n\ncallbacks_list = [checkpoint, early, reduceLROnPlat]","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 3-4. Training"},{"metadata":{"trusted":true},"cell_type":"code","source":"def fit():\n    seg_model.compile(optimizer=Adam(1e-3, decay=1e-6), loss=IoU, metrics=['binary_accuracy'])\n    \n    step_count = min(MAX_TRAIN_STEPS, train_df.shape[0]//BATCH_SIZE)\n    aug_gen = create_aug_gen(make_image_gen(train_df))\n    loss_history = [seg_model.fit_generator(aug_gen,\n                                 steps_per_epoch=step_count,\n                                 epochs=MAX_TRAIN_EPOCHS,\n                                 validation_data=(valid_x, valid_y),\n                                 callbacks=callbacks_list,\n                                workers=1 # the generator is not very thread safe\n                                           )]\n    return loss_history\n\nwhile True:\n    loss_history = fit()\n    if np.min([mh.history['val_loss'] for mh in loss_history]) < -0.2:\n        break","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## 3-5. Full Resolution Model"},{"metadata":{"trusted":true},"cell_type":"code","source":"if IMG_SCALING is not None:\n    fullres_model = models.Sequential()\n    fullres_model.add(layers.AvgPool2D(IMG_SCALING, input_shape = (None, None, 3)))\n    fullres_model.add(seg_model)\n    fullres_model.add(layers.UpSampling2D(IMG_SCALING))\nelse:\n    fullres_model = seg_model\nfullres_model.save('fullres_model.h5')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def raw_prediction(img, path=TEST_DIR):\n    img = imread(os.path.join(path, img))\n    img = np.expand_dims(img, 0)/255\n    seg = fullres_model.predict(img)[0]\n    return seg, img[0]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"from skimage.morphology import binary_opening, disk\n\ndef smooth(seg):\n    return binary_opening(seg>0.99, np.expand_dims(disk(2), -1))","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def predict(img, path=TEST_DIR):\n    seg, img = raw_prediction(img, path=path)\n    return smooth(seg), img","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"test_paths = np.array(os.listdir(TEST_DIR))\nprint('{} test images found'.format(len(test_paths)))","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"from tqdm import tqdm_notebook\n\ndef pred_encode(img, **kwargs):\n    cur_seg, _ = predict(img)\n    cur_rles = multi_rle_encode(cur_seg, **kwargs)\n    return [[img, rle] for rle in cur_rles if rle is not None]\n\nout_pred_rows = []\nfor c_img_name in tqdm_notebook(test_paths[:30000]): ## only a subset as it takes too long to run\n    out_pred_rows += pred_encode(c_img_name, min_max_threshold=1.0)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"sub = pd.DataFrame(out_pred_rows)\nsub.columns = ['ImageId', 'EncodedPixels']\nsub = sub[sub.EncodedPixels.notnull()]\nsub.head()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"TOP_PREDICTIONS=5\nfig, m_axs = plt.subplots(TOP_PREDICTIONS, 2, figsize = (9, TOP_PREDICTIONS*5))\n[c_ax.axis('off') for c_ax in m_axs.flatten()]\n\nfor (ax1, ax2), c_img_name in zip(m_axs, sub.ImageId.unique()[:TOP_PREDICTIONS]):\n    c_img = imread(os.path.join(TEST_DIR, c_img_name))\n    c_img = np.expand_dims(c_img, 0)/255.0\n    ax1.imshow(c_img[0])\n    ax1.set_title('Image: ' + c_img_name)\n    ax2.imshow(masks_as_color(sub.query('ImageId==\"{}\"'.format(c_img_name))['EncodedPixels']))\n    ax2.set_title('Prediction')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"sub1 = pd.read_csv(BASE_DIR+'sample_submission_v2.csv')\nsub1 = pd.DataFrame(np.setdiff1d(sub1['ImageId'].unique(), sub['ImageId'].unique(), assume_unique=True), columns=['ImageId'])\nsub1['EncodedPixels'] = None\nprint(len(sub1), len(sub))\n\nsub = pd.concat([sub, sub1])\nprint(len(sub))\nsub.to_csv('submission.csv', index=False)\nsub.head()","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":1}