__author__ = "n01z3" import matplotlib.pyplot as plt import numpy as np import cv2 import pandas as pd from shapely.wkt import loads as wkt_loads import tifffile as tiff import os import random from keras.models import Model from keras.layers import Input, merge, Convolution2D, MaxPooling2D, UpSampling2D, Reshape, core, Dropout from keras.optimizers import Adam from keras.callbacks import ModelCheckpoint, LearningRateScheduler from keras import backend as K from sklearn.metrics import jaccard_similarity_score from shapely.geometry import MultiPolygon, Polygon import shapely.wkt import shapely.affinity from collections import defaultdict N_Cls = 10 inDir = '../input' DF = pd.read_csv(inDir + '/train_wkt_v4.csv') GS = pd.read_csv(inDir + '/grid_sizes.csv', names=['ImageId', 'Xmax', 'Ymin'], skiprows=1) SB = pd.read_csv(os.path.join(inDir, 'sample_submission.csv')) ISZ = 160 smooth = 1e-12 def _convert_coordinates_to_raster(coords, img_size, xymax): # __author__ = visoft # https://www.kaggle.com/visoft/dstl-satellite-imagery-feature-detection/export-pixel-wise-mask Xmax, Ymax = xymax H, W = img_size W1 = 1.0 * W * W / (W + 1) H1 = 1.0 * H * H / (H + 1) xf = W1 / Xmax yf = H1 / Ymax coords[:, 1] *= yf coords[:, 0] *= xf coords_int = np.round(coords).astype(np.int32) return coords_int def _get_xmax_ymin(grid_sizes_panda, imageId): # __author__ = visoft # https://www.kaggle.com/visoft/dstl-satellite-imagery-feature-detection/export-pixel-wise-mask xmax, ymin = grid_sizes_panda[grid_sizes_panda.ImageId == imageId].iloc[0, 1:].astype(float) return (xmax, ymin) def _get_polygon_list(wkt_list_pandas, imageId, cType): # __author__ = visoft # https://www.kaggle.com/visoft/dstl-satellite-imagery-feature-detection/export-pixel-wise-mask df_image = wkt_list_pandas[wkt_list_pandas.ImageId == imageId] multipoly_def = df_image[df_image.ClassType == cType].MultipolygonWKT polygonList = None if len(multipoly_def) > 0: assert len(multipoly_def) == 1 polygonList = wkt_loads(multipoly_def.values[0]) return polygonList def _get_and_convert_contours(polygonList, raster_img_size, xymax): # __author__ = visoft # https://www.kaggle.com/visoft/dstl-satellite-imagery-feature-detection/export-pixel-wise-mask perim_list = [] interior_list = [] if polygonList is None: return None for k in range(len(polygonList)): poly = polygonList[k] perim = np.array(list(poly.exterior.coords)) perim_c = _convert_coordinates_to_raster(perim, raster_img_size, xymax) perim_list.append(perim_c) for pi in poly.interiors: interior = np.array(list(pi.coords)) interior_c = _convert_coordinates_to_raster(interior, raster_img_size, xymax) interior_list.append(interior_c) return perim_list, interior_list def _plot_mask_from_contours(raster_img_size, contours, class_value=1): # __author__ = visoft # https://www.kaggle.com/visoft/dstl-satellite-imagery-feature-detection/export-pixel-wise-mask img_mask = np.zeros(raster_img_size, np.uint8) if contours is None: return img_mask perim_list, interior_list = contours cv2.fillPoly(img_mask, perim_list, class_value) cv2.fillPoly(img_mask, interior_list, 0) return img_mask def generate_mask_for_image_and_class(raster_size, imageId, class_type, grid_sizes_panda=GS, wkt_list_pandas=DF): # __author__ = visoft # https://www.kaggle.com/visoft/dstl-satellite-imagery-feature-detection/export-pixel-wise-mask xymax = _get_xmax_ymin(grid_sizes_panda, imageId) polygon_list = _get_polygon_list(wkt_list_pandas, imageId, class_type) contours = _get_and_convert_contours(polygon_list, raster_size, xymax) mask = _plot_mask_from_contours(raster_size, contours, 1) return mask def M(image_id): # __author__ = amaia # https://www.kaggle.com/aamaia/dstl-satellite-imagery-feature-detection/rgb-using-m-bands-example filename = os.path.join(inDir, 'sixteen_band', '{}_M.tif'.format(image_id)) img = tiff.imread(filename) img = np.rollaxis(img, 0, 3) return img def stretch_n(bands, lower_percent=5, higher_percent=95): out = np.zeros_like(bands) n = bands.shape[2] for i in range(n): a = 0 # np.min(band) b = 1 # np.max(band) c = np.percentile(bands[:, :, i], lower_percent) d = np.percentile(bands[:, :, i], higher_percent) t = a + (bands[:, :, i] - c) * (b - a) / (d - c) t[t < a] = a t[t > b] = b out[:, :, i] = t return out.astype(np.float32) def jaccard_coef(y_true, y_pred): # __author__ = Vladimir Iglovikov intersection = K.sum(y_true * y_pred, axis=[0, -1, -2]) sum_ = K.sum(y_true + y_pred, axis=[0, -1, -2]) jac = (intersection + smooth) / (sum_ - intersection + smooth) return K.mean(jac) def jaccard_coef_int(y_true, y_pred): # __author__ = Vladimir Iglovikov y_pred_pos = K.round(K.clip(y_pred, 0, 1)) intersection = K.sum(y_true * y_pred_pos, axis=[0, -1, -2]) sum_ = K.sum(y_true + y_pred, axis=[0, -1, -2]) jac = (intersection + smooth) / (sum_ - intersection + smooth) return K.mean(jac) def stick_all_train(): # print "let's stick all imgs together" s = 835 x = np.zeros((5 * s, 5 * s, 8)) y = np.zeros((5 * s, 5 * s, N_Cls)) ids = sorted(DF.ImageId.unique()) print(len(ids)) for i in range(5): for j in range(5): id = ids[5 * i + j] img = M(id) img = stretch_n(img) print(img.shape, id, np.amax(img), np.amin(img)) x[s * i:s * i + s, s * j:s * j + s, :] = img[:s, :s, :] for z in range(N_Cls): y[s * i:s * i + s, s * j:s * j + s, z] = generate_mask_for_image_and_class( (img.shape[0], img.shape[1]), id, z + 1)[:s, :s] print(np.amax(y), np.amin(y)) print(x.shape, y.shape) return x, y # np.save('data/x_trn_%d' % N_Cls, x) # np.save('data/y_trn_%d' % N_Cls, y) def get_patches(img, msk, amt=10000, aug=True): is2 = int(1.0 * ISZ) xm, ym = img.shape[0] - is2, img.shape[1] - is2 x, y = [], [] tr = [0.4, 0.1, 0.1, 0.15, 0.3, 0.95, 0.1, 0.05, 0.001, 0.005] for i in range(amt): xc = random.randint(0, xm) yc = random.randint(0, ym) im = img[xc:xc + is2, yc:yc + is2] ms = msk[xc:xc + is2, yc:yc + is2] for j in range(N_Cls): sm = np.sum(ms[:, :, j]) if 1.0 * sm / is2 ** 2 > tr[j]: if aug: if random.uniform(0, 1) > 0.5: im = im[::-1] ms = ms[::-1] if random.uniform(0, 1) > 0.5: im = im[:, ::-1] ms = ms[:, ::-1] x.append(im) y.append(ms) x, y = 2 * np.transpose(x, (0, 3, 1, 2)) - 1, np.transpose(y, (0, 3, 1, 2)) print (x.shape, y.shape, np.amax(x), np.amin(x), np.amax(y), np.amin(y)) return x, y def make_val(): print( "let's pick some samples for validation") img,msk = stick_all_train()#= np.load('data/x_trn_%d.npy' % N_Cls) # msk = np.load('data/y_trn_%d.npy' % N_Cls) x, y = get_patches(img, msk, amt=3000) # np.save('data/x_tmp_%d' % N_Cls, x) # np.save('data/y_tmp_%d' % N_Cls, y) # def get_unet(): # inputs = Input((8, ISZ, ISZ)) # conv1 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(inputs) # conv1 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(conv1) # pool1 = MaxPooling2D(pool_size=(2, 2))(conv1) # conv2 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(pool1) # conv2 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(conv2) # pool2 = MaxPooling2D(pool_size=(2, 2))(conv2) # conv3 = Convolution2D(128, 3, 3, activation='relu', border_mode='same')(pool2) # conv3 = Convolution2D(128, 3, 3, activation='relu', border_mode='same')(conv3) # pool3 = MaxPooling2D(pool_size=(2, 2))(conv3) # conv4 = Convolution2D(256, 3, 3, activation='relu', border_mode='same')(pool3) # conv4 = Convolution2D(256, 3, 3, activation='relu', border_mode='same')(conv4) # pool4 = MaxPooling2D(pool_size=(2, 2))(conv4) # conv5 = Convolution2D(512, 3, 3, activation='relu', border_mode='same')(pool4) # conv5 = Convolution2D(512, 3, 3, activation='relu', border_mode='same')(conv5) # up6 = merge([UpSampling2D(size=(2, 2))(conv5), conv4], mode='concat', concat_axis=1) # conv6 = Convolution2D(256, 3, 3, activation='relu', border_mode='same')(up6) # conv6 = Convolution2D(256, 3, 3, activation='relu', border_mode='same')(conv6) # up7 = merge([UpSampling2D(size=(2, 2))(conv6), conv3], mode='concat', concat_axis=1) # conv7 = Convolution2D(128, 3, 3, activation='relu', border_mode='same')(up7) # conv7 = Convolution2D(128, 3, 3, activation='relu', border_mode='same')(conv7) # up8 = merge([UpSampling2D(size=(2, 2))(conv7), conv2], mode='concat', concat_axis=1) # conv8 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(up8) # conv8 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(conv8) # up9 = merge([UpSampling2D(size=(2, 2))(conv8), conv1], mode='concat', concat_axis=1) # conv9 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(up9) # conv9 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(conv9) # conv10 = Convolution2D(N_Cls, 1, 1, activation='sigmoid')(conv9) # model = Model(input=inputs, output=conv10) # model.compile(optimizer=Adam(), loss='binary_crossentropy', metrics=[jaccard_coef, jaccard_coef_int, 'accuracy']) # return model # def calc_jacc(model): # img = np.load('data/x_tmp_%d.npy' % N_Cls) # msk = np.load('data/y_tmp_%d.npy' % N_Cls) # prd = model.predict(img, batch_size=4) # print prd.shape, msk.shape # avg, trs = [], [] # for i in range(N_Cls): # t_msk = msk[:, i, :, :] # t_prd = prd[:, i, :, :] # t_msk = t_msk.reshape(msk.shape[0] * msk.shape[2], msk.shape[3]) # t_prd = t_prd.reshape(msk.shape[0] * msk.shape[2], msk.shape[3]) # m, b_tr = 0, 0 # for j in range(10): # tr = j / 10.0 # pred_binary_mask = t_prd > tr # jk = jaccard_similarity_score(t_msk, pred_binary_mask) # if jk > m: # m = jk # b_tr = tr # print i, m, b_tr # avg.append(m) # trs.append(b_tr) # score = sum(avg) / 10.0 # return score, trs # def mask_for_polygons(polygons, im_size): # # __author__ = Konstantin Lopuhin # # https://www.kaggle.com/lopuhin/dstl-satellite-imagery-feature-detection/full-pipeline-demo-poly-pixels-ml-poly # img_mask = np.zeros(im_size, np.uint8) # if not polygons: # return img_mask # int_coords = lambda x: np.array(x).round().astype(np.int32) # exteriors = [int_coords(poly.exterior.coords) for poly in polygons] # interiors = [int_coords(pi.coords) for poly in polygons # for pi in poly.interiors] # cv2.fillPoly(img_mask, exteriors, 1) # cv2.fillPoly(img_mask, interiors, 0) # return img_mask # def mask_to_polygons(mask, epsilon=5, min_area=1.): # # __author__ = Konstantin Lopuhin # # https://www.kaggle.com/lopuhin/dstl-satellite-imagery-feature-detection/full-pipeline-demo-poly-pixels-ml-poly # # first, find contours with cv2: it's much faster than shapely # image, contours, hierarchy = cv2.findContours( # ((mask == 1) * 255).astype(np.uint8), # cv2.RETR_CCOMP, cv2.CHAIN_APPROX_TC89_KCOS) # # create approximate contours to have reasonable submission size # approx_contours = [cv2.approxPolyDP(cnt, epsilon, True) # for cnt in contours] # if not contours: # return MultiPolygon() # # now messy stuff to associate parent and child contours # cnt_children = defaultdict(list) # child_contours = set() # assert hierarchy.shape[0] == 1 # # http://docs.opencv.org/3.1.0/d9/d8b/tutorial_py_contours_hierarchy.html # for idx, (_, _, _, parent_idx) in enumerate(hierarchy[0]): # if parent_idx != -1: # child_contours.add(idx) # cnt_children[parent_idx].append(approx_contours[idx]) # # create actual polygons filtering by area (removes artifacts) # all_polygons = [] # for idx, cnt in enumerate(approx_contours): # if idx not in child_contours and cv2.contourArea(cnt) >= min_area: # assert cnt.shape[1] == 1 # poly = Polygon( # shell=cnt[:, 0, :], # holes=[c[:, 0, :] for c in cnt_children.get(idx, []) # if cv2.contourArea(c) >= min_area]) # all_polygons.append(poly) # # approximating polygons might have created invalid ones, fix them # all_polygons = MultiPolygon(all_polygons) # if not all_polygons.is_valid: # all_polygons = all_polygons.buffer(0) # # Sometimes buffer() converts a simple Multipolygon to just a Polygon, # # need to keep it a Multi throughout # if all_polygons.type == 'Polygon': # all_polygons = MultiPolygon([all_polygons]) # return all_polygons # def get_scalers(im_size, x_max, y_min): # # __author__ = Konstantin Lopuhin # # https://www.kaggle.com/lopuhin/dstl-satellite-imagery-feature-detection/full-pipeline-demo-poly-pixels-ml-poly # h, w = im_size # they are flipped so that mask_for_polygons works correctly # h, w = float(h), float(w) # w_ = 1.0 * w * (w / (w + 1)) # h_ = 1.0 * h * (h / (h + 1)) # return w_ / x_max, h_ / y_min # def train_net(): # print "start train net" # x_val, y_val = np.load('data/x_tmp_%d.npy' % N_Cls), np.load('data/y_tmp_%d.npy' % N_Cls) # img = np.load('data/x_trn_%d.npy' % N_Cls) # msk = np.load('data/y_trn_%d.npy' % N_Cls) # x_trn, y_trn = get_patches(img, msk) # model = get_unet() # model.load_weights('weights/unet_10_jk0.7878') # model_checkpoint = ModelCheckpoint('weights/unet_tmp.hdf5', monitor='loss', save_best_only=True) # for i in range(1): # model.fit(x_trn, y_trn, batch_size=64, nb_epoch=1, verbose=1, shuffle=True, # callbacks=[model_checkpoint], validation_data=(x_val, y_val)) # del x_trn # del y_trn # x_trn, y_trn = get_patches(img, msk) # score, trs = calc_jacc(model) # print 'val jk', score # model.save_weights('weights/unet_10_jk%.4f' % score) # return model # def predict_id(id, model, trs): # img = M(id) # x = stretch_n(img) # cnv = np.zeros((960, 960, 8)).astype(np.float32) # prd = np.zeros((N_Cls, 960, 960)).astype(np.float32) # cnv[:img.shape[0], :img.shape[1], :] = x # for i in range(0, 6): # line = [] # for j in range(0, 6): # line.append(cnv[i * ISZ:(i + 1) * ISZ, j * ISZ:(j + 1) * ISZ]) # x = 2 * np.transpose(line, (0, 3, 1, 2)) - 1 # tmp = model.predict(x, batch_size=4) # for j in range(tmp.shape[0]): # prd[:, i * ISZ:(i + 1) * ISZ, j * ISZ:(j + 1) * ISZ] = tmp[j] # # trs = [0.4, 0.1, 0.4, 0.3, 0.3, 0.5, 0.3, 0.6, 0.1, 0.1] # for i in range(N_Cls): # prd[i] = prd[i] > trs[i] # return prd[:, :img.shape[0], :img.shape[1]] # def predict_test(model, trs): # print "predict test" # for i, id in enumerate(sorted(set(SB['ImageId'].tolist()))): # msk = predict_id(id, model, trs) # np.save('msk/10_%s' % id, msk) # if i % 100 == 0: print i, id # def make_submit(): # print "make submission file" # df = pd.read_csv(os.path.join(inDir, 'sample_submission.csv')) # print df.head() # for idx, row in df.iterrows(): # id = row[0] # kls = row[1] - 1 # msk = np.load('msk/10_%s.npy' % id)[kls] # pred_polygons = mask_to_polygons(msk) # x_max = GS.loc[GS['ImageId'] == id, 'Xmax'].as_matrix()[0] # y_min = GS.loc[GS['ImageId'] == id, 'Ymin'].as_matrix()[0] # x_scaler, y_scaler = get_scalers(msk.shape, x_max, y_min) # scaled_pred_polygons = shapely.affinity.scale(pred_polygons, xfact=1.0 / x_scaler, yfact=1.0 / y_scaler, # origin=(0, 0, 0)) # df.iloc[idx, 2] = shapely.wkt.dumps(scaled_pred_polygons) # if idx % 100 == 0: print idx # print df.head() # df.to_csv('subm/1.csv', index=False) # def check_predict(id='6120_2_3'): # model = get_unet() # model.load_weights('weights/unet_10_jk0.7878') # msk = predict_id(id, model, [0.4, 0.1, 0.4, 0.3, 0.3, 0.5, 0.3, 0.6, 0.1, 0.1]) # img = M(id) # plt.figure() # ax1 = plt.subplot(131) # ax1.set_title('image ID:6120_2_3') # ax1.imshow(img[:, :, 5], cmap=plt.get_cmap('gist_ncar')) # ax2 = plt.subplot(132) # ax2.set_title('predict bldg pixels') # ax2.imshow(msk[0], cmap=plt.get_cmap('gray')) # ax3 = plt.subplot(133) # ax3.set_title('predict bldg polygones') # ax3.imshow(mask_for_polygons(mask_to_polygons(msk[0], epsilon=1), img.shape[:2]), cmap=plt.get_cmap('gray')) # plt.show() if __name__ == '__main__': stick_all_train() make_val() # model = train_net() # score, trs = calc_jacc(model) # predict_test(model, trs) # make_submit() # # bonus # check_predict()