前回に引き続き
以下のコンペをもとに勉強していきます。
今回は学習済モデル(Xception_model)を活用して転移学習を行ったものです。
XceptionとはImageNetを学習したモデルでInceptinの改良版みたいです。
XceptionはInceptionと違い、空間方向の畳み込みとチャンネル方向の畳み込みを完全分離しても良い
という考えで作られたモデルみたいです。
データの前加工に関しては以前の記事でも紹介した
マスクによるセグメンテーションとぼかし効果を使ったノイズ除去をメインで使っています。
以下コードです。
#必要なライブラリの読み込み import tensorflow as tf from keras.applications import xception from keras.preprocessing import image import keras.preprocessing.image from sklearn.linear_model import LogisticRegression from sklearn.metrics import accuracy_score, confusion_matrix import sklearn.ensemble import numpy as np import pandas as pd import datetime as dt from tqdm import tqdm from glob import glob import cv2 import os import seaborn as sns import mpl_toolkits.axes_grid1 import matplotlib import matplotlib.pyplot as plt import datetime %matplotlib inline plt.rcParams['figure.figsize'] = [16, 10] plt.rcParams['font.size'] = 16 global_start = datetime.datetime.now() #テストデータのサイズ指定 valid_set_size_percentage = 10 running_on_kaggle = True # train data take_only_samples_of_train_data = True; # set False to train on all train data num_samples_of_train_data_per_species = 200 # < 221, ignored if take_only_samples_of_train_data = True load_bf_of_train_data = False # set True to load bottleneck features from file # test data take_only_samples_of_test_data = False; # set False to predict on all test data num_samples_of_test_data = 200 # < 794, ignored if take_only_samples_of_test_data = True load_bf_of_test_data = False # set True to load bottleneck features from file # augmented images use_aug_data = False # set True to use augmented images load_bf_of_aug_data = True # set True to load bottleneck features from file # show plots show_plots = True #----------------------------------------------------------------------------------------------------- #データの読み込み cw_dir = os.getcwd() data_dir = '' train_dir = os.path.join(data_dir, 'train') test_dir = os.path.join(data_dir, 'test') #カテゴリをリストに格納 species = ['Black-grass', 'Charlock', 'Cleavers', 'Common Chickweed', 'Common wheat', 'Fat Hen', 'Loose Silky-bent', 'Maize', 'Scentless Mayweed', 'Shepherds Purse', 'Small-flowered Cranesbill', 'Sugar beet'] num_species = len(species) #それぞれのカテゴリのデータ数確認 for sp in species: print("{} images of {}".format(len(os.listdir(os.path.join(train_dir + "/" +sp))),sp)) #トレーニングデータの読み込み,トレーニングデータのサイズ確認 train = [] for species_id, sp in enumerate(species): for file in os.listdir(os.path.join(train_dir, sp)): train.append(['train/{}/{}'.format(sp, file), file, species_id, sp]) train_df = pd.DataFrame(train, columns=['filepath', 'file', 'species_id', 'species']) print('') print('train_df.shape = ', train_df.shape) #テストデータの読み込み test = [] for file in os.listdir(test_dir): test.append(['test/{}'.format(file), file]) test_df = pd.DataFrame(test, columns=['filepath', 'file']) print('test_df.shape = ', test_df.shape) #データ（画像）の読み取り用関数作成 def read_image(filepath , target_size = None): #画像読み込み img = cv2.imread(os.path.join(data_dir, filepath), cv2.IMREAD_COLOR) #画像サイズの変更する場合、ターゲットサイズの引数に指定 img = cv2.resize(img.copy(), target_size, interpolation = cv2.INTER_AREA) #img = image.load_img(os.path.join(data_dir, filepath),target_size=target_size) #img = image.img_to_array(img) return img #全ての画像サイズの取得（時間かかる） if False: train_df['image_heigth'] = 0 train_df['image_width'] = 0 train_df['image_channel'] = 0 #get all image shapes for i in range(len(train_df)): img = read_img(train_df.filepath.values[i]) train_df.loc[i,'image_heigth'] = img.shape[0] train_df.loc[i,'image_width'] = img.shape[1] train_df.loc[i,'image_channel'] = img.shape[2] test_df['image_heigth'] = 0 test_df['image_width'] = 0 test_df['image_channel'] = 0 # get all image shapes for i in range(len(test_df)): img = read_img(test_df.filepath.values[i]) test_df.loc[i,'image_heigth'] = img.shape[0] test_df.loc[i,'image_width'] = img.shape[1] test_df.loc[i,'image_channel'] = img.shape[2] # train data #print(train_df.describe()) #今回は最初の5つを表示 train_df.head() #全てのカテゴリの画像を1枚ずつプロットしてみる if show_plots: fig = plt.figure(1, figsize=(num_species, num_species)) grid = mpl_toolkits.axes_grid1.ImageGrid(fig, 111, nrows_ncols=(num_species, num_species), axes_pad=0.05) i = 0 for species_id, sp in enumerate(species): for filepath in train_df[train_df['species'] == sp]['filepath'].values[:num_species]: ax = grid[i] img = read_image(filepath, (224, 224)) ax.imshow(img.astype(np.uint8)) ax.axis('off') if i % num_species == num_species - 1: ax.text(250, 112, sp, verticalalignment='center') i += 1 plt.show(); #テストイメージをプロットしてみる if show_plots: fig = plt.figure(1, figsize=(10, 10)) grid = mpl_toolkits.axes_grid1.ImageGrid(fig, 111, nrows_ncols=(5, 10), axes_pad=0.05) i = 0 for j in range(5): for filepath in test_df['filepath'].values[j*5:j*5+10]: ax = grid[i] img = read_image(filepath, (224, 224)) ax.imshow(img.astype(np.uint8)) ax.axis('off') i += 1 plt.show(); #----------------------------------------------------------------------------------------------------- #データ前加工 #①植物を検出してセグメント化する #②(299,299,3)を(1,-1)の範囲に整形 #③画像をnumpy配列に変換 #試しにデータの一部のファイルを使用してみる if take_only_samples_of_train_data: train_df = pd.concat([train_df[train_df['species'] == sp][:num_samples_of_train_data_per_species] for sp in species]) train_df.index = np.arange(len(train_df)) if take_only_samples_of_test_data: test_df = test_df[:num_samples_of_test_data] #OpenCVを使ったセグメント検出用のマスクを作成 def create_mask_for_plant(image): #RGBからHSVに変換 image_hsv = cv2.cvtColor(image,cv2.COLOR_BGR2HSV) sensitivity = 35 #下限となるhsv色相の設定 lower_hsv = np.array([60 - sensitivity , 100 , 50]) #上限となるhsv色相の設定 upper_hsv = np.array([60 + sensitivity , 255 , 255]) #上で設定した値でマスク作成 mask = cv2.inRange(image_hsv , lower_hsv,upper_hsv) kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(11,11)) mask = cv2.morphologyEx(mask , cv2.MORPH_CLOSE,kernel) return mask #オリジナル画像とマスクを合成 def segment_plant(image): mask = create_mask_for_plant(image) output = cv2.bitwise_and(image , image , mask = mask) return output #画像のノイズ除去 def sharpen_image(image): #画像にぼかしを入れる image_blurred = cv2.GaussianBlur(image , (0,0),3) #ぼかした画像とオリジナル画像を合成 image_sharp = cv2.addWeighted(image , 1.5 , image_blurred,-0.5 ,0) return image_sharp #上記処理後の画像の読み込み def read_segmented_image(filepath,img_size): #オリジナル画像の読み込み img = cv2.imread(os.path.join(data_dir, filepath), cv2.IMREAD_COLOR) #画像のサイズ変更 img = cv2.resize(img.copy(), img_size, interpolation = cv2.INTER_AREA) #マスクの作成 image_mask = create_mask_for_plant(img) #マスクとオリジナル画像合成（セグメント検出） image_segmented = segment_plant(img) #画像にぼかしを入れる image_sharpen = sharpen_image(image_segmented) return img, image_mask, image_segmented, image_sharpen #加工後の写真をいくつか見てみる if show_plots: for i in range(4): img, image_mask, image_segmented, image_sharpen = read_segmented_image( train_df.loc[i,'filepath'],(224,224)) fig, axs = plt.subplots(1, 4, figsize=(20, 20)) axs[0].imshow(img.astype(np.uint8)) axs[1].imshow(image_mask.astype(np.uint8)) axs[2].imshow(image_segmented.astype(np.uint8)) axs[3].imshow(image_sharpen.astype(np.uint8)) #加工後のデータとしてすべての画像の読み込み def preprocess_image(img): img /= 255. img -= 0.5 img *= 2 return img target_image_size = 299 print('read and preprocess training and validation images') #トレーニング、バリデーションデータの前加工・読み込み x_train_valid = np.zeros((len(train_df), target_image_size, target_image_size, 3), dtype='float32') y_train_valid = train_df.loc[:, 'species_id'].values for i, filepath in tqdm(enumerate(train_df['filepath'])): _,_,_,img = read_segmented_image(filepath, (target_image_size, target_image_size)) # 正則化 x_train_valid[i] = preprocess_image(np.expand_dims(img.copy().astype(np.float), axis=0)) print('read and preprocess test images') #テストデータの前加工・読み込み x_test = np.zeros((len(test_df), target_image_size, target_image_size, 3), dtype='float32') for i, filepath in tqdm(enumerate(test_df['filepath'])): _,_,_,img = read_segmented_image(filepath, (target_image_size, target_image_size)) #正則化 x_test[i] = preprocess_image(np.expand_dims(img.copy().astype(np.float), axis=0)) print('x_train_valid.shape = ', x_train_valid.shape) print('x_test.shape = ', x_test.shape) #----------------------------------------------------------------------------------------------------- #前加工後のデータの複製 def generate_images(imgs): imgs_len = len(imgs) image_generator = keras.preprocessing.image.ImageDataGenerator( rotation_range = 10, width_shift_range = 0.1 , height_shift_range = 0.1, horizontal_flip = False, vertical_flip = False, zoom_range = 0.1) imgs = image_generator.flow(imgs.copy(), np.zeros(imgs_len), batch_size=imgs_len, shuffle = False).next() return imgs[0] #いくつかのデータをプロットしてみる if show_plots: imgs = (((x_train_valid[0:4]+1.)/2.)*255.) # transform pixels into range [0,255] imgs_generated = imgs fig, axs = plt.subplots(4, 8, figsize=(20, 10)) for i in range(8): axs[0,i].imshow(imgs_generated[0].astype(np.uint8)) axs[1,i].imshow(imgs_generated[1].astype(np.uint8)) axs[2,i].imshow(imgs_generated[2].astype(np.uint8)) axs[3,i].imshow(imgs_generated[3].astype(np.uint8)) imgs_generated = generate_images(imgs) #----------------------------------------------------------------------------------------------------- #学習済モデルの学習済特徴量を使用する if running_on_kaggle: # キャッシュとモデル用ディレクトリの作成 cache_dir = os.path.expanduser(os.path.join('~', '.keras')) if not os.path.exists(cache_dir): os.makedirs(cache_dir) models_dir = os.path.join(cache_dir, 'models') if not os.path.exists(models_dir): os.makedirs(models_dir) print('') print('use xception models') if not load_bf_of_train_data: print('x_train_valid.shape = ', x_train_valid.shape) print('y_train_valid.shape = ', y_train_valid.shape) print('') print('compute bottleneck features from Xception network') local_start = datetime.datetime.now() # xception base modelを読み込んで、最後の層（出力層?）で予測を行う base_model = xception.Xception(weights='imagenet', include_top=False, pooling='avg') x_train_valid_bf = base_model.predict(x_train_valid, batch_size=32, verbose=1) print('running time: ', datetime.datetime.now()-local_start) print('') print('x_train_valid_bf.shape = ', x_train_valid_bf.shape) print('') print('save bottleneck features and labels for later ') np.save(os.path.join(os.getcwd(),'x_train_valid_bf.npy'), x_train_valid_bf) np.save(os.path.join(os.getcwd(),'y_train_valid.npy'), y_train_valid) else: # ボトルネック特徴とラベルの読み込み print('load bottleneck features and labels') x_train_valid_bf = np.load(os.path.join(os.getcwd(),'x_train_valid_bf_of_segmented_images.npy')) y_train_valid = np.load(os.path.join(os.getcwd(),'y_train_valid_of_segmented_images.npy')) print('x_train_valid_bf.shape = ', x_train_valid_bf.shape) print('y_train_valid.shape = ', y_train_valid.shape) # テストイメージに対して予測を行う if not load_bf_of_test_data: local_start = datetime.datetime.now() base_model = xception.Xception(weights='imagenet', include_top=False, pooling='avg') x_test_bf = base_model.predict(x_test, batch_size=32, verbose=1) print('running time: ', datetime.datetime.now()-local_start) print('') print('x_test_bf = ',x_test_bf.shape) print('save bottleneck features ') np.save(os.path.join(os.getcwd(),'x_test_bf.npy'), x_test_bf) else: print('load bottleneck features') x_test_bf = np.load(os.path.join(os.getcwd(),'x_test_bf_of_segmented_images.npy')) print('x_test_bf.shape = ', x_test_bf.shape) ## 複製したデータに対して2048のボトルネック特徴を計算する if use_aug_data: if not load_bf_of_aug_data: for i in range(2): x_aug_tmp = generate_images(x_train_valid) y_aug_tmp = y_train_valid print('compute bottleneck features from Xception network') print('x_aug_tmp.shape = ', x_aug_tmp) print('y_aug_tmp.shape = ', y_aug_tmp) local_start = datetime.datetime.now() # load xception model and predict the last layer having 2048 neurons per image base_model = xception.Xception(weights='imagenet', include_top=False, pooling='avg') x_aug_tmp_bf = base_model.predict(x_aug_tmp, batch_size=32, verbose=1) print('running time: ', datetime.datetime.now()-local_start) print('') print('x_aug_tmp_bf.shape = ', x_aug_tmp_bf.shape) if i==0: x_aug = x_aug_tmp_bf y_aug = y_aug_tmp else: x_aug = np.concatenate([x_aug,x_aug_tmp_bf]) y_aug = np.concatenate([y_aug,y_aug_tmp]) print('') print('save bottleneck features and labels for later ') np.save(os.path.join(os.getcwd(),'x_aug.npy'), x_aug) np.save(os.path.join(os.getcwd(),'y_aug.npy'), y_aug) else: print('load bottleneck features') x_aug_bf = np.load(os.path.join(os.getcwd(), 'x_aug_bf_of_segmented_images.npy')) y_aug = np.load(os.path.join(os.getcwd(), 'y_aug_of_segmented_images.npy')) print('x_aug_bf.shape = ', x_aug_bf.shape) print('y_aug.shape = ', y_aug.shape) #ファイルの結合 if False: x_aug_bf_1 = np.load(os.path.join(os.getcwd(), 'x_aug_bf_of_segmented_images.npy')) y_aug_1 = np.load(os.path.join(os.getcwd(), 'y_aug_of_segmented_images.npy')) print(x_aug_bf_1.shape, y_aug_1.shape) x_aug_bf_2 = np.load(os.path.join(os.getcwd(), 'x_aug_bf.npy')) y_aug_2 = np.load(os.path.join(os.getcwd(), 'y_aug.npy')) print(x_aug_bf_2.shape, y_aug_2.shape) x_aug_bf = np.concatenate([x_aug_bf_1,x_aug_bf_2]) y_aug = np.concatenate([y_aug_1,y_aug_2]) print(x_aug_bf.shape,y_aug.shape) #----------------------------------------------------------------------------------------------------- #モデルの学習と評価 #ラベルをワンホットエンコーディングする def dense_to_one_hot(labels_dense, num_classes): num_labels = labels_dense.shape[0] index_offset = np.arange(num_labels) * num_classes labels_one_hot = np.zeros((num_labels, num_classes)) labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1 return labels_one_hot def one_hot_to_dense(labels_one_hot): num_labels = labels_one_hot.shape[0] num_classes = labels_one_hot.shape[1] labels_dense = np.where(labels_one_hot == 1)[1] return labels_dense #トレーニングデータとバリデーションデータをシャッフルする def shuffle_train_valid_data(): print('shuffle train and validation data') #オリジナルデータを学習・評価用でシャッフル perm_array = np.arange(len(x_train_valid_bf)) np.random.shuffle(perm_array) #オリジナルデータを分割 x_train_bf = x_train_valid_bf[perm_array[:train_set_size]] y_train = dense_to_one_hot(y_train_valid[perm_array[:train_set_size]], num_species) x_valid_bf = x_train_valid_bf[perm_array[-valid_set_size:]] y_valid = dense_to_one_hot(y_train_valid[perm_array[-valid_set_size:]], num_species) # オリジナルデータと複製データを結合 if use_aug_data: x_train_bf = np.concatenate([x_train_bf, x_aug_bf]) y_train = np.concatenate([y_train, dense_to_one_hot(y_aug, num_species)]) # データをシャッフル perm_array = np.arange(len(x_train_bf)) np.random.shuffle(perm_array) x_train_bf = x_train_bf[perm_array] y_train = y_train[perm_array] return x_train_bf, y_train, x_valid_bf, y_valid if valid_set_size_percentage > 0: #データの分割 valid_set_size = int(len(x_train_valid_bf) * valid_set_size_percentage/100); train_set_size = len(x_train_valid_bf) - valid_set_size; else: valid_set_size = int(len(x_train_valid_bf) * 0.1); train_set_size = len(x_train_valid_bf) x_train_bf, y_train, x_valid_bf, y_valid = shuffle_train_valid_data() print('x_train_bf.shape = ', x_train_bf.shape) print('y_train.shape = ', y_train.shape) print('x_valid_bf.shape = ', x_valid_bf.shape) print('y_valid.shape = ', y_valid.shape) shuffle train and validation data x_train_bf.shape = (2160, 2048) y_train.shape = (2160, 12) x_valid_bf.shape = (240, 2048) y_valid.shape = (240, 12) #線形回帰モデル cv_num = 1 acc_logreg_train = acc_logreg_valid = 0 y_test_pred_proba_logreg = 0 for i in range(cv_num): shuffle_train_valid_data() # shuffle data logreg = LogisticRegression(multi_class='multinomial', solver='lbfgs') logreg.fit(x_train_bf, one_hot_to_dense(y_train)) acc_logreg_train += logreg.score(x_train_bf, one_hot_to_dense(y_train)) acc_logreg_valid += logreg.score(x_valid_bf, one_hot_to_dense(y_valid)) y_test_pred_proba_logreg += logreg.predict_proba(x_test_bf) acc_logreg_train /= cv_num acc_logreg_valid /= cv_num y_test_pred_proba_logreg /= cv_num y_test_pred_class_logreg = np.argmax(y_test_pred_proba_logreg, axis = 1) print('Logistic Regression') print('Accuracy train/valid = %.4f/%.4f'%(acc_logreg_train, acc_logreg_valid)) print('y_test_pred_class_logreg.shape = ', y_test_pred_class_logreg.shape) shuffle train and validation data Logistic Regression Accuracy train/valid = 0.9986/0.9208 y_test_pred_class_logreg.shape = (794,) ## random forest random_forest = sklearn.ensemble.RandomForestClassifier(n_estimators=10) random_forest.fit(x_train_bf, one_hot_to_dense(y_train)) acc_train_random_forest = random_forest.score(x_train_bf, one_hot_to_dense(y_train)) acc_valid_random_forest = random_forest.score(x_valid_bf, one_hot_to_dense(y_valid)) print('Random Forest') print('Accuracy train/valid = %.4f/%.4f'%(acc_train_random_forest, acc_valid_random_forest)) Random Forest Accuracy train/valid = 0.9958/0.7208 ## tensorflowを使ったニューラルネットワーク perm_array_train = np.arange(len(x_train_bf)) index_in_epoch = 0 #見にバッチ取得の関数 def get_next_batch(batch_size): global index_in_epoch, perm_array_train start = index_in_epoch index_in_epoch += batch_size if index_in_epoch > train_set_size: np.random.shuffle(perm_array_train) start = 0 index_in_epoch = batch_size end = index_in_epoch return x_train_bf[perm_array_train[start:end]], y_train[perm_array_train[start:end]] x_size = x_train_bf.shape[1] y_size = num_species n_n_fc1 = 1024 n_n_fc2 = num_species x_data = tf.placeholder('float', shape=[None, x_size]) y_data = tf.placeholder('float', shape=[None, y_size]) # 1.layer: fully connected W_fc1 = tf.Variable(tf.truncated_normal(shape = [x_size, n_n_fc1], stddev = 0.1)) b_fc1 = tf.Variable(tf.constant(0.1, shape = [n_n_fc1])) h_fc1 = tf.nn.relu(tf.matmul(x_data, W_fc1) + b_fc1) # add dropout tf_keep_prob = tf.placeholder('float') h_fc1_drop = tf.nn.dropout(h_fc1, tf_keep_prob) # 3.layer: fully connected W_fc2 = tf.Variable(tf.truncated_normal(shape = [n_n_fc1, n_n_fc2], stddev = 0.1)) b_fc2 = tf.Variable(tf.constant(0.1, shape = [n_n_fc2])) z_pred = tf.matmul(h_fc1_drop, W_fc2) + b_fc2 # cost function cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_data, logits=z_pred)); # optimisation function tf_learn_rate = tf.placeholder(dtype='float', name="tf_learn_rate") train_step = tf.train.AdamOptimizer(tf_learn_rate).minimize(cross_entropy) # evaluation y_pred = tf.cast(tf.nn.softmax(z_pred), dtype = tf.float32); y_pred_class = tf.cast(tf.argmax(y_pred,1), tf.int32) y_data_class = tf.cast(tf.argmax(y_data,1), tf.int32) accuracy = tf.reduce_mean(tf.cast(tf.equal(y_pred_class, y_data_class), tf.float32)) # parameters cv_num = 1 # number of cross validations n_epoch = 15 # number of epochs batch_size = 50 keep_prob = 0.33 # dropout regularization with keeping probability learn_rate_range = [0.01,0.005,0.0025,0.001,0.001,0.001,0.00075,0.0005,0.00025,0.0001, 0.0001,0.0001,0.0001,0.0001,0.0001,0.0001,0.0001,0.0001,0.0001]; learn_rate_step = 3 # in terms of epochs acc_train_DNN = 0 acc_valid_DNN = 0 loss_train_DNN = 0 loss_valid_DNN = 0 y_test_pred_proba_DNN = 0 y_valid_pred_proba = 0 # use cross validation for j in range(cv_num): # start TensorFlow session and initialize global variables sess = tf.InteractiveSession() sess.run(tf.global_variables_initializer()) # shuffle train/validation splits shuffle_train_valid_data() n_step = -1; # training model for i in range(int(n_epoch*train_set_size/batch_size)): if i%int(learn_rate_step*train_set_size/batch_size) == 0: n_step += 1; learn_rate = learn_rate_range[n_step]; print('learnrate = ', learn_rate) x_batch, y_batch = get_next_batch(batch_size) sess.run(train_step, feed_dict={x_data: x_batch, y_data: y_batch, tf_keep_prob: keep_prob, tf_learn_rate: learn_rate}) if i%int(0.25*train_set_size/batch_size) == 0: train_loss = sess.run(cross_entropy, feed_dict={x_data: x_train_bf[:valid_set_size], y_data: y_train[:valid_set_size], tf_keep_prob: 1.0}) train_acc = accuracy.eval(feed_dict={x_data: x_train_bf[:valid_set_size], y_data: y_train[:valid_set_size], tf_keep_prob: 1.0}) valid_loss = sess.run(cross_entropy,feed_dict={x_data: x_valid_bf, y_data: y_valid, tf_keep_prob: 1.0}) valid_acc = accuracy.eval(feed_dict={x_data: x_valid_bf, y_data: y_valid, tf_keep_prob: 1.0}) print('%.2f epoch: train/val loss = %.4f/%.4f, train/val acc = %.4f/%.4f'%( (i+1)*batch_size/train_set_size, train_loss, valid_loss, train_acc, valid_acc)) acc_train_DNN += train_acc acc_valid_DNN += valid_acc loss_train_DNN += train_loss loss_valid_DNN += valid_loss y_valid_pred_proba += y_pred.eval(feed_dict={x_data: x_valid_bf, tf_keep_prob: 1.0}) y_test_pred_proba_DNN += y_pred.eval(feed_dict={x_data: x_test_bf, tf_keep_prob: 1.0}) sess.close() acc_train_DNN /= float(cv_num) acc_valid_DNN /= float(cv_num) loss_train_DNN /= float(cv_num) loss_valid_DNN /= float(cv_num) # final validation prediction y_valid_pred_proba /= float(cv_num) y_valid_pred_class = np.argmax(y_valid_pred_proba, axis = 1) # final test prediction y_test_pred_proba_DNN /= float(cv_num) y_test_pred_class_DNN = np.argmax(y_test_pred_proba_DNN, axis = 1) # final loss and accuracy print('') print('final: train/val loss = %.4f/%.4f, train/val acc = %.4f/%.4f'%(loss_train_DNN, loss_valid_DNN, acc_train_DNN, acc_valid_DNN)) #----------------------------------------------------------------------------------------------------- #コンフュージョンマトリックスの作成 if show_plots: cnf_matrix = confusion_matrix(one_hot_to_dense(y_valid), y_valid_pred_class) abbreviation = ['BG', 'Ch', 'Cl', 'CC', 'CW', 'FH', 'LSB', 'M', 'SM', 'SP', 'SFC', 'SB'] pd.DataFrame({'class': species, 'abbreviation': abbreviation}) fig, ax = plt.subplots(1) ax = sns.heatmap(cnf_matrix, ax=ax, cmap=plt.cm.Greens, annot=True) ax.set_xticklabels(abbreviation) ax.set_yticklabels(abbreviation) plt.title('Confusion matrix of validation set') plt.ylabel('True species') plt.xlabel('Predicted species') plt.show(); #----------------------------------------------------------------------------------------------------- #推論と提出用ファイルの作成 take the results from the best model or a combination write the submission file y_test_pred_class = y_test_pred_class_logreg test_df['species_id'] = y_test_pred_class test_df['species'] = [species[sp] for sp in y_test_pred_class] test_df[['file', 'species']].to_csv('submission.csv', index=False) print('total running time: ', datetime.datetime.now()-global_start