Klasifikace Elektrokardiogramu (EKG) - díl třetí a poslední
9. 2. 2024 0:00
Jiří Raška
Tento článek navazuje na předchozí: Klasifikace Elektrokardiogramu (EKG) – díl druhý.
Kombinované modely
Další oblastí, kterou bych chtěl ověřit, je kombinace výše vytvořených modelů pro samostatné vstupní datové sady do jednoho modelu.
Nemá cenu to nějak blíže rozebírat, rovnou do toho …
Kombinovaný model pro metadata X a EKG křivky Y
Vytvořím model, kdy výstupy z vrstev X a Y jsou spojeny do jedné dimenze ve vrstvě Concatenate :
In [41]:
def create_model04(X_shape, Y_shape, Z_shape):
X_inputs = keras.Input(X_shape[1:], name='X_inputs')
Y_inputs = keras.Input(Y_shape[1:], name='Y_inputs')
X = create_X_model(X_inputs)
Y = create_Y_model(Y_inputs, filters=(64, 128, 256), kernel_size=(7, 3, 3))
X = keras.layers.Concatenate(name='Z_concat')([X, Y])
X = keras.layers.Dense(128, activation='relu', name='Z_dense_1')(X)
X = keras.layers.Dense(128, activation='relu', name='Z_dense_2')(X)
X = keras.layers.Dropout(0.5, name='Z_drop_1')(X)
outputs = keras.layers.Dense(Z_shape[-1], activation='sigmoid', name='Z_outputs')(X)
model = keras.Model(inputs=[X_inputs, Y_inputs], outputs=outputs, name='X_Y_model')
return model
Nyní můžu model přeložit:
In [42]:
model04 = create_model04(X_train.shape, Y_train.shape, Z_train.shape) model04.compile(optimizer='adam', loss='binary_crossentropy', metrics=['binary_accuracy', 'Precision', 'Recall']) model04.summary() Model: "X_Y_model" __________________________________________________________________________________________________ Layer (type) Output Shape Param # Connected to ================================================================================================== Y_inputs (InputLayer) [(None, 1000, 12)] 0 [] Y_norm (Normalization) (None, 1000, 12) 25 ['Y_inputs[0][0]'] Y_conv_1 (Conv1D) (None, 1000, 64) 5440 ['Y_norm[0][0]'] Y_norm_1 (BatchNormalizati (None, 1000, 64) 256 ['Y_conv_1[0][0]'] on) Y_relu_1 (ReLU) (None, 1000, 64) 0 ['Y_norm_1[0][0]'] Y_pool_1 (MaxPooling1D) (None, 500, 64) 0 ['Y_relu_1[0][0]'] Y_conv_2 (Conv1D) (None, 500, 128) 24704 ['Y_pool_1[0][0]'] Y_norm_2 (BatchNormalizati (None, 500, 128) 512 ['Y_conv_2[0][0]'] on) Y_relu_2 (ReLU) (None, 500, 128) 0 ['Y_norm_2[0][0]'] X_inputs (InputLayer) [(None, 7)] 0 [] Y_pool_2 (MaxPooling1D) (None, 250, 128) 0 ['Y_relu_2[0][0]'] X_norm (Normalization) (None, 7) 15 ['X_inputs[0][0]'] Y_conv_3 (Conv1D) (None, 250, 256) 98560 ['Y_pool_2[0][0]'] X_dense_1 (Dense) (None, 32) 256 ['X_norm[0][0]'] Y_norm_3 (BatchNormalizati (None, 250, 256) 1024 ['Y_conv_3[0][0]'] on) X_drop_1 (Dropout) (None, 32) 0 ['X_dense_1[0][0]'] Y_relu_3 (ReLU) (None, 250, 256) 0 ['Y_norm_3[0][0]'] X_dense_2 (Dense) (None, 32) 1056 ['X_drop_1[0][0]'] Y_aver (GlobalAveragePooli (None, 256) 0 ['Y_relu_3[0][0]'] ng1D) X_drop_2 (Dropout) (None, 32) 0 ['X_dense_2[0][0]'] Y_drop (Dropout) (None, 256) 0 ['Y_aver[0][0]'] Z_concat (Concatenate) (None, 288) 0 ['X_drop_2[0][0]', 'Y_drop[0][0]'] Z_dense_1 (Dense) (None, 128) 36992 ['Z_concat[0][0]'] Z_dense_2 (Dense) (None, 128) 16512 ['Z_dense_1[0][0]'] Z_drop_1 (Dropout) (None, 128) 0 ['Z_dense_2[0][0]'] Z_outputs (Dense) (None, 5) 645 ['Z_drop_1[0][0]'] ================================================================================================== Total params: 185997 (726.56 KB) Trainable params: 185061 (722.89 KB) Non-trainable params: 936 (3.66 KB) __________________________________________________________________________________________________
A rovnou jej trénovat:
In [43]:
MODEL_CHECKPOINT = '/kaggle/working/model/model04.ckpt' callbacks_list = [ keras.callbacks.EarlyStopping(monitor='val_binary_accuracy', patience=10), keras.callbacks.ModelCheckpoint(filepath=MODEL_CHECKPOINT, monitor='val_binary_accuracy', save_best_only=True) ] history = model04.fit([X_train, Y_train], Z_train, epochs=100, batch_size=32, callbacks=callbacks_list, validation_data=([X_valid, Y_valid], Z_valid)) model04 = keras.models.load_model(MODEL_CHECKPOINT) model04.save('/kaggle/working/model/model04.keras') Epoch 1/100 546/546 [==============================] - 14s 17ms/step - loss: 0.5135 - binary_accuracy: 0.7760 - precision: 0.5898 - recall: 0.3968 - val_loss: 0.3685 - val_binary_accuracy: 0.8338 - val_precision: 0.7517 - val_recall: 0.5227 Epoch 2/100 546/546 [==============================] - 8s 15ms/step - loss: 0.3815 - binary_accuracy: 0.8339 - precision: 0.7225 - recall: 0.5655 - val_loss: 0.3438 - val_binary_accuracy: 0.8505 - val_precision: 0.7462 - val_recall: 0.6293 Epoch 3/100 546/546 [==============================] - 6s 10ms/step - loss: 0.3520 - binary_accuracy: 0.8516 - precision: 0.7547 - recall: 0.6187 - val_loss: 0.3874 - val_binary_accuracy: 0.8459 - val_precision: 0.7479 - val_recall: 0.5990 Epoch 4/100 546/546 [==============================] - 9s 17ms/step - loss: 0.3346 - binary_accuracy: 0.8622 - precision: 0.7760 - recall: 0.6454 - val_loss: 0.3259 - val_binary_accuracy: 0.8625 - val_precision: 0.7736 - val_recall: 0.6532 Epoch 5/100 546/546 [==============================] - 8s 15ms/step - loss: 0.3222 - binary_accuracy: 0.8681 - precision: 0.7869 - recall: 0.6616 - val_loss: 0.3106 - val_binary_accuracy: 0.8656 - val_precision: 0.7798 - val_recall: 0.6607 Epoch 6/100 546/546 [==============================] - 8s 15ms/step - loss: 0.3122 - binary_accuracy: 0.8732 - precision: 0.7973 - recall: 0.6739 - val_loss: 0.3083 - val_binary_accuracy: 0.8711 - val_precision: 0.7915 - val_recall: 0.6732 Epoch 7/100 546/546 [==============================] - 6s 11ms/step - loss: 0.3009 - binary_accuracy: 0.8787 - precision: 0.8079 - recall: 0.6874 - val_loss: 0.3148 - val_binary_accuracy: 0.8711 - val_precision: 0.8011 - val_recall: 0.6597 Epoch 8/100 546/546 [==============================] - 9s 16ms/step - loss: 0.2972 - binary_accuracy: 0.8790 - precision: 0.8053 - recall: 0.6928 - val_loss: 0.2975 - val_binary_accuracy: 0.8784 - val_precision: 0.7873 - val_recall: 0.7185 Epoch 9/100 546/546 [==============================] - 8s 15ms/step - loss: 0.2931 - binary_accuracy: 0.8819 - precision: 0.8113 - recall: 0.6991 - val_loss: 0.2907 - val_binary_accuracy: 0.8847 - val_precision: 0.7993 - val_recall: 0.7331 Epoch 10/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2901 - binary_accuracy: 0.8836 - precision: 0.8140 - recall: 0.7039 - val_loss: 0.3016 - val_binary_accuracy: 0.8789 - val_precision: 0.7963 - val_recall: 0.7071 Epoch 11/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2828 - binary_accuracy: 0.8865 - precision: 0.8207 - recall: 0.7097 - val_loss: 0.3254 - val_binary_accuracy: 0.8627 - val_precision: 0.7449 - val_recall: 0.7042 Epoch 12/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2825 - binary_accuracy: 0.8870 - precision: 0.8217 - recall: 0.7108 - val_loss: 0.2861 - val_binary_accuracy: 0.8844 - val_precision: 0.7942 - val_recall: 0.7392 Epoch 13/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2783 - binary_accuracy: 0.8888 - precision: 0.8221 - recall: 0.7194 - val_loss: 0.2960 - val_binary_accuracy: 0.8804 - val_precision: 0.7880 - val_recall: 0.7281 Epoch 14/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2737 - binary_accuracy: 0.8901 - precision: 0.8244 - recall: 0.7225 - val_loss: 0.2795 - val_binary_accuracy: 0.8847 - val_precision: 0.8060 - val_recall: 0.7232 Epoch 15/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2725 - binary_accuracy: 0.8906 - precision: 0.8227 - recall: 0.7273 - val_loss: 0.2797 - val_binary_accuracy: 0.8842 - val_precision: 0.8038 - val_recall: 0.7235 Epoch 16/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2704 - binary_accuracy: 0.8914 - precision: 0.8251 - recall: 0.7282 - val_loss: 0.2810 - val_binary_accuracy: 0.8824 - val_precision: 0.8004 - val_recall: 0.7196 Epoch 17/100 546/546 [==============================] - 9s 17ms/step - loss: 0.2689 - binary_accuracy: 0.8917 - precision: 0.8273 - recall: 0.7266 - val_loss: 0.2885 - val_binary_accuracy: 0.8883 - val_precision: 0.8005 - val_recall: 0.7499 Epoch 18/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2666 - binary_accuracy: 0.8931 - precision: 0.8306 - recall: 0.7291 - val_loss: 0.2870 - val_binary_accuracy: 0.8793 - val_precision: 0.7842 - val_recall: 0.7285 Epoch 19/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2640 - binary_accuracy: 0.8943 - precision: 0.8310 - recall: 0.7349 - val_loss: 0.3127 - val_binary_accuracy: 0.8720 - val_precision: 0.7847 - val_recall: 0.6878 Epoch 20/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2588 - binary_accuracy: 0.8959 - precision: 0.8322 - recall: 0.7408 - val_loss: 0.2906 - val_binary_accuracy: 0.8837 - val_precision: 0.7957 - val_recall: 0.7335 Epoch 21/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2559 - binary_accuracy: 0.8984 - precision: 0.8398 - recall: 0.7431 - val_loss: 0.2920 - val_binary_accuracy: 0.8824 - val_precision: 0.8087 - val_recall: 0.7075 Epoch 22/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2560 - binary_accuracy: 0.8974 - precision: 0.8380 - recall: 0.7406 - val_loss: 0.2740 - val_binary_accuracy: 0.8868 - val_precision: 0.8063 - val_recall: 0.7335 Epoch 23/100 546/546 [==============================] - 9s 16ms/step - loss: 0.2505 - binary_accuracy: 0.9002 - precision: 0.8437 - recall: 0.7469 - val_loss: 0.2773 - val_binary_accuracy: 0.8897 - val_precision: 0.8118 - val_recall: 0.7403 Epoch 24/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2498 - binary_accuracy: 0.9013 - precision: 0.8446 - recall: 0.7507 - val_loss: 0.2842 - val_binary_accuracy: 0.8846 - val_precision: 0.8030 - val_recall: 0.7271 Epoch 25/100 546/546 [==============================] - 8s 15ms/step - loss: 0.2488 - binary_accuracy: 0.9009 - precision: 0.8454 - recall: 0.7479 - val_loss: 0.2716 - val_binary_accuracy: 0.8908 - val_precision: 0.8202 - val_recall: 0.7339 Epoch 26/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2451 - binary_accuracy: 0.9033 - precision: 0.8471 - recall: 0.7572 - val_loss: 0.2892 - val_binary_accuracy: 0.8806 - val_precision: 0.7882 - val_recall: 0.7289 Epoch 27/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2444 - binary_accuracy: 0.9035 - precision: 0.8489 - recall: 0.7559 - val_loss: 0.2891 - val_binary_accuracy: 0.8891 - val_precision: 0.8185 - val_recall: 0.7274 Epoch 28/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2417 - binary_accuracy: 0.9031 - precision: 0.8492 - recall: 0.7534 - val_loss: 0.2857 - val_binary_accuracy: 0.8854 - val_precision: 0.8087 - val_recall: 0.7224 Epoch 29/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2411 - binary_accuracy: 0.9043 - precision: 0.8499 - recall: 0.7583 - val_loss: 0.2764 - val_binary_accuracy: 0.8846 - val_precision: 0.8126 - val_recall: 0.7132 Epoch 30/100 546/546 [==============================] - 8s 15ms/step - loss: 0.2393 - binary_accuracy: 0.9059 - precision: 0.8523 - recall: 0.7629 - val_loss: 0.2731 - val_binary_accuracy: 0.8941 - val_precision: 0.8245 - val_recall: 0.7442 Epoch 31/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2392 - binary_accuracy: 0.9050 - precision: 0.8519 - recall: 0.7592 - val_loss: 0.2821 - val_binary_accuracy: 0.8893 - val_precision: 0.8184 - val_recall: 0.7285 Epoch 32/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2359 - binary_accuracy: 0.9073 - precision: 0.8556 - recall: 0.7655 - val_loss: 0.2758 - val_binary_accuracy: 0.8898 - val_precision: 0.8148 - val_recall: 0.7364 Epoch 33/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2349 - binary_accuracy: 0.9065 - precision: 0.8513 - recall: 0.7671 - val_loss: 0.2871 - val_binary_accuracy: 0.8887 - val_precision: 0.8086 - val_recall: 0.7399 Epoch 34/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2340 - binary_accuracy: 0.9074 - precision: 0.8569 - recall: 0.7644 - val_loss: 0.2672 - val_binary_accuracy: 0.8906 - val_precision: 0.8169 - val_recall: 0.7371 Epoch 35/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2325 - binary_accuracy: 0.9076 - precision: 0.8578 - recall: 0.7640 - val_loss: 0.2790 - val_binary_accuracy: 0.8885 - val_precision: 0.8135 - val_recall: 0.7314 Epoch 36/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2312 - binary_accuracy: 0.9076 - precision: 0.8538 - recall: 0.7690 - val_loss: 0.2766 - val_binary_accuracy: 0.8876 - val_precision: 0.8101 - val_recall: 0.7321 Epoch 37/100 546/546 [==============================] - 9s 17ms/step - loss: 0.2272 - binary_accuracy: 0.9096 - precision: 0.8583 - recall: 0.7730 - val_loss: 0.2598 - val_binary_accuracy: 0.8942 - val_precision: 0.8175 - val_recall: 0.7545 Epoch 38/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2274 - binary_accuracy: 0.9102 - precision: 0.8599 - recall: 0.7736 - val_loss: 0.2663 - val_binary_accuracy: 0.8931 - val_precision: 0.8220 - val_recall: 0.7428 Epoch 39/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2275 - binary_accuracy: 0.9103 - precision: 0.8606 - recall: 0.7731 - val_loss: 0.2714 - val_binary_accuracy: 0.8907 - val_precision: 0.8173 - val_recall: 0.7374 Epoch 40/100 546/546 [==============================] - 8s 15ms/step - loss: 0.2252 - binary_accuracy: 0.9110 - precision: 0.8632 - recall: 0.7733 - val_loss: 0.2699 - val_binary_accuracy: 0.8943 - val_precision: 0.8162 - val_recall: 0.7570 Epoch 41/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2234 - binary_accuracy: 0.9105 - precision: 0.8613 - recall: 0.7732 - val_loss: 0.2834 - val_binary_accuracy: 0.8888 - val_precision: 0.8128 - val_recall: 0.7342 Epoch 42/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2233 - binary_accuracy: 0.9117 - precision: 0.8642 - recall: 0.7755 - val_loss: 0.2745 - val_binary_accuracy: 0.8894 - val_precision: 0.7991 - val_recall: 0.7578 Epoch 43/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2197 - binary_accuracy: 0.9123 - precision: 0.8647 - recall: 0.7776 - val_loss: 0.2963 - val_binary_accuracy: 0.8898 - val_precision: 0.8227 - val_recall: 0.7253 Epoch 44/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2189 - binary_accuracy: 0.9140 - precision: 0.8667 - recall: 0.7829 - val_loss: 0.2891 - val_binary_accuracy: 0.8875 - val_precision: 0.8061 - val_recall: 0.7371 Epoch 45/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2194 - binary_accuracy: 0.9130 - precision: 0.8654 - recall: 0.7798 - val_loss: 0.2812 - val_binary_accuracy: 0.8907 - val_precision: 0.8151 - val_recall: 0.7406 Epoch 46/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2186 - binary_accuracy: 0.9131 - precision: 0.8636 - recall: 0.7826 - val_loss: 0.2920 - val_binary_accuracy: 0.8861 - val_precision: 0.8101 - val_recall: 0.7242 Epoch 47/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2157 - binary_accuracy: 0.9142 - precision: 0.8683 - recall: 0.7817 - val_loss: 0.2746 - val_binary_accuracy: 0.8889 - val_precision: 0.8192 - val_recall: 0.7257 Epoch 48/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2153 - binary_accuracy: 0.9145 - precision: 0.8694 - recall: 0.7821 - val_loss: 0.2672 - val_binary_accuracy: 0.8924 - val_precision: 0.8139 - val_recall: 0.7506 Epoch 49/100 546/546 [==============================] - 6s 10ms/step - loss: 0.2145 - binary_accuracy: 0.9159 - precision: 0.8715 - recall: 0.7856 - val_loss: 0.2997 - val_binary_accuracy: 0.8899 - val_precision: 0.8043 - val_recall: 0.7524 Epoch 50/100 546/546 [==============================] - 6s 11ms/step - loss: 0.2123 - binary_accuracy: 0.9152 - precision: 0.8697 - recall: 0.7849 - val_loss: 0.2770 - val_binary_accuracy: 0.8938 - val_precision: 0.8105 - val_recall: 0.7628
Průběh trénování vypadal zhruba takto:
In [44]:
sns.relplot(data=pd.DataFrame(history.history), kind='line', height=4, aspect=4)
plt.show()
Na závěr kapitoly si můžu model vyhodnotit proti testovacím sadám dat pro X a Y:
In [45]:
Z_pred = np.around(model04.predict([X_test, Y_test], verbose=0))
metrics = pd.concat([metrics.drop(model04.name, level=0, errors='ignore'), model_quality_reporting(Z_test, Z_pred, model_name=model04.name)])
precision recall f1-score support
NORM 0.83 0.92 0.87 964
MI 0.89 0.66 0.76 553
STTC 0.79 0.71 0.74 523
CD 0.76 0.72 0.74 498
HYP 0.73 0.45 0.56 263
micro avg 0.81 0.75 0.78 2801
macro avg 0.80 0.69 0.74 2801
weighted avg 0.81 0.75 0.77 2801
samples avg 0.79 0.77 0.76 2801
Posouzení výsledku modelu v porovnání s ostatními se nechám až na závěr tohoto článku.
Kombinovaný model pro metadata X, EKG křivky Y a spektrogramy S
V tomto pokusu zkombinuji všechny tři modely pro vstupní data do jednoho.
In [46]:
def create_model05(X_shape, Y_shape, S_shape, Z_shape):
X_inputs = keras.Input(X_shape[1:], name='X_inputs')
Y_inputs = keras.Input(Y_shape[1:], name='Y_inputs')
S_inputs = keras.Input(S_shape[1:], name='S_inputs')
X = create_X_model(X_inputs)
Y = create_Y_model(Y_inputs, filters=(64, 128, 256), kernel_size=(7, 3, 3))
S = create_S_model(S_inputs, filters=(32, 64, 128), kernel_size=((3, 3), (3, 3), (3, 3)))
X = keras.layers.Concatenate(name='Z_concat')([X, Y, S])
X = keras.layers.Dense(128, activation='relu', name='Z_dense_1')(X)
X = keras.layers.Dense(128, activation='relu', name='Z_dense_2')(X)
X = keras.layers.Dropout(0.5, name='Z_drop_1')(X)
outputs = keras.layers.Dense(Z_shape[-1], activation='sigmoid', name='Z_outputs')(X)
model = keras.Model(inputs=[X_inputs, Y_inputs, S_inputs], outputs=outputs, name='X_Y_S_model')
return model
A nyní kompilace modelu a následně trénování tak, jak jsem to dělal doposud u všech modelů:
In [47]:
model05 = create_model05(X_train.shape, Y_train.shape, S_train.shape, Z_train.shape) model05.compile(optimizer='adam', loss='binary_crossentropy', metrics=['binary_accuracy', 'Precision', 'Recall']) model05.summary() Model: "X_Y_S_model" __________________________________________________________________________________________________ Layer (type) Output Shape Param # Connected to ================================================================================================== S_inputs (InputLayer) [(None, 12, 31, 34, 1)] 0 [] S_norm (Normalization) (None, 12, 31, 34, 1) 3 ['S_inputs[0][0]'] tf.split_1 (TFOpLambda) [(None, 1, 31, 34, 1), 0 ['S_norm[0][0]'] (None, 1, 31, 34, 1), (None, 1, 31, 34, 1), (None, 1, 31, 34, 1), (None, 1, 31, 34, 1), (None, 1, 31, 34, 1), (None, 1, 31, 34, 1), (None, 1, 31, 34, 1), (None, 1, 31, 34, 1), (None, 1, 31, 34, 1), (None, 1, 31, 34, 1), (None, 1, 31, 34, 1)] tf.compat.v1.squeeze_12 (T (None, 31, 34, 1) 0 ['tf.split_1[0][0]'] FOpLambda) tf.compat.v1.squeeze_13 (T (None, 31, 34, 1) 0 ['tf.split_1[0][1]'] FOpLambda) tf.compat.v1.squeeze_14 (T (None, 31, 34, 1) 0 ['tf.split_1[0][2]'] FOpLambda) tf.compat.v1.squeeze_15 (T (None, 31, 34, 1) 0 ['tf.split_1[0][3]'] FOpLambda) tf.compat.v1.squeeze_16 (T (None, 31, 34, 1) 0 ['tf.split_1[0][4]'] FOpLambda) tf.compat.v1.squeeze_17 (T (None, 31, 34, 1) 0 ['tf.split_1[0][5]'] FOpLambda) tf.compat.v1.squeeze_18 (T (None, 31, 34, 1) 0 ['tf.split_1[0][6]'] FOpLambda) tf.compat.v1.squeeze_19 (T (None, 31, 34, 1) 0 ['tf.split_1[0][7]'] FOpLambda) tf.compat.v1.squeeze_20 (T (None, 31, 34, 1) 0 ['tf.split_1[0][8]'] FOpLambda) tf.compat.v1.squeeze_21 (T (None, 31, 34, 1) 0 ['tf.split_1[0][9]'] FOpLambda) tf.compat.v1.squeeze_22 (T (None, 31, 34, 1) 0 ['tf.split_1[0][10]'] FOpLambda) tf.compat.v1.squeeze_23 (T (None, 31, 34, 1) 0 ['tf.split_1[0][11]'] FOpLambda) S_conv_0_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_12[0][0 ]'] S_conv_1_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_13[0][0 ]'] S_conv_2_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_14[0][0 ]'] S_conv_3_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_15[0][0 ]'] S_conv_4_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_16[0][0 ]'] S_conv_5_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_17[0][0 ]'] S_conv_6_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_18[0][0 ]'] S_conv_7_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_19[0][0 ]'] S_conv_8_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_20[0][0 ]'] S_conv_9_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_21[0][0 ]'] S_conv_10_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_22[0][0 ]'] S_conv_11_1 (Conv2D) (None, 31, 34, 32) 320 ['tf.compat.v1.squeeze_23[0][0 ]'] Y_inputs (InputLayer) [(None, 1000, 12)] 0 [] S_norm_0_1 (BatchNormaliza (None, 31, 34, 32) 128 ['S_conv_0_1[0][0]'] tion) S_norm_1_1 (BatchNormaliza (None, 31, 34, 32) 128 ['S_conv_1_1[0][0]'] tion) S_norm_2_1 (BatchNormaliza (None, 31, 34, 32) 128 ['S_conv_2_1[0][0]'] tion) S_norm_3_1 (BatchNormaliza (None, 31, 34, 32) 128 ['S_conv_3_1[0][0]'] tion) S_norm_4_1 (BatchNormaliza (None, 31, 34, 32) 128 ['S_conv_4_1[0][0]'] tion) S_norm_5_1 (BatchNormaliza (None, 31, 34, 32) 128 ['S_conv_5_1[0][0]'] tion) S_norm_6_1 (BatchNormaliza (None, 31, 34, 32) 128 ['S_conv_6_1[0][0]'] tion) S_norm_7_1 (BatchNormaliza (None, 31, 34, 32) 128 ['S_conv_7_1[0][0]'] tion) S_norm_8_1 (BatchNormaliza (None, 31, 34, 32) 128 ['S_conv_8_1[0][0]'] tion) S_norm_9_1 (BatchNormaliza (None, 31, 34, 32) 128 ['S_conv_9_1[0][0]'] tion) S_norm_10_1 (BatchNormaliz (None, 31, 34, 32) 128 ['S_conv_10_1[0][0]'] ation) S_norm_11_1 (BatchNormaliz (None, 31, 34, 32) 128 ['S_conv_11_1[0][0]'] ation) Y_norm (Normalization) (None, 1000, 12) 25 ['Y_inputs[0][0]'] S_relu_0_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_0_1[0][0]'] S_relu_1_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_1_1[0][0]'] S_relu_2_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_2_1[0][0]'] S_relu_3_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_3_1[0][0]'] S_relu_4_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_4_1[0][0]'] S_relu_5_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_5_1[0][0]'] S_relu_6_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_6_1[0][0]'] S_relu_7_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_7_1[0][0]'] S_relu_8_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_8_1[0][0]'] S_relu_9_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_9_1[0][0]'] S_relu_10_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_10_1[0][0]'] S_relu_11_1 (ReLU) (None, 31, 34, 32) 0 ['S_norm_11_1[0][0]'] Y_conv_1 (Conv1D) (None, 1000, 64) 5440 ['Y_norm[0][0]'] S_pool_0_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_0_1[0][0]'] S_pool_1_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_1_1[0][0]'] S_pool_2_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_2_1[0][0]'] S_pool_3_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_3_1[0][0]'] S_pool_4_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_4_1[0][0]'] S_pool_5_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_5_1[0][0]'] S_pool_6_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_6_1[0][0]'] S_pool_7_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_7_1[0][0]'] S_pool_8_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_8_1[0][0]'] S_pool_9_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_9_1[0][0]'] S_pool_10_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_10_1[0][0]'] S_pool_11_1 (MaxPooling2D) (None, 15, 17, 32) 0 ['S_relu_11_1[0][0]'] Y_norm_1 (BatchNormalizati (None, 1000, 64) 256 ['Y_conv_1[0][0]'] on) S_conv_0_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_0_1[0][0]'] S_conv_1_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_1_1[0][0]'] S_conv_2_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_2_1[0][0]'] S_conv_3_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_3_1[0][0]'] S_conv_4_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_4_1[0][0]'] S_conv_5_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_5_1[0][0]'] S_conv_6_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_6_1[0][0]'] S_conv_7_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_7_1[0][0]'] S_conv_8_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_8_1[0][0]'] S_conv_9_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_9_1[0][0]'] S_conv_10_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_10_1[0][0]'] S_conv_11_2 (Conv2D) (None, 15, 17, 64) 18496 ['S_pool_11_1[0][0]'] Y_relu_1 (ReLU) (None, 1000, 64) 0 ['Y_norm_1[0][0]'] S_norm_0_2 (BatchNormaliza (None, 15, 17, 64) 256 ['S_conv_0_2[0][0]'] tion) S_norm_1_2 (BatchNormaliza (None, 15, 17, 64) 256 ['S_conv_1_2[0][0]'] tion) S_norm_2_2 (BatchNormaliza (None, 15, 17, 64) 256 ['S_conv_2_2[0][0]'] tion) S_norm_3_2 (BatchNormaliza (None, 15, 17, 64) 256 ['S_conv_3_2[0][0]'] tion) S_norm_4_2 (BatchNormaliza (None, 15, 17, 64) 256 ['S_conv_4_2[0][0]'] tion) S_norm_5_2 (BatchNormaliza (None, 15, 17, 64) 256 ['S_conv_5_2[0][0]'] tion) S_norm_6_2 (BatchNormaliza (None, 15, 17, 64) 256 ['S_conv_6_2[0][0]'] tion) S_norm_7_2 (BatchNormaliza (None, 15, 17, 64) 256 ['S_conv_7_2[0][0]'] tion) S_norm_8_2 (BatchNormaliza (None, 15, 17, 64) 256 ['S_conv_8_2[0][0]'] tion) S_norm_9_2 (BatchNormaliza (None, 15, 17, 64) 256 ['S_conv_9_2[0][0]'] tion) S_norm_10_2 (BatchNormaliz (None, 15, 17, 64) 256 ['S_conv_10_2[0][0]'] ation) S_norm_11_2 (BatchNormaliz (None, 15, 17, 64) 256 ['S_conv_11_2[0][0]'] ation) Y_pool_1 (MaxPooling1D) (None, 500, 64) 0 ['Y_relu_1[0][0]'] S_relu_0_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_0_2[0][0]'] S_relu_1_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_1_2[0][0]'] S_relu_2_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_2_2[0][0]'] S_relu_3_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_3_2[0][0]'] S_relu_4_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_4_2[0][0]'] S_relu_5_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_5_2[0][0]'] S_relu_6_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_6_2[0][0]'] S_relu_7_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_7_2[0][0]'] S_relu_8_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_8_2[0][0]'] S_relu_9_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_9_2[0][0]'] S_relu_10_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_10_2[0][0]'] S_relu_11_2 (ReLU) (None, 15, 17, 64) 0 ['S_norm_11_2[0][0]'] Y_conv_2 (Conv1D) (None, 500, 128) 24704 ['Y_pool_1[0][0]'] S_pool_0_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_0_2[0][0]'] S_pool_1_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_1_2[0][0]'] S_pool_2_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_2_2[0][0]'] S_pool_3_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_3_2[0][0]'] S_pool_4_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_4_2[0][0]'] S_pool_5_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_5_2[0][0]'] S_pool_6_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_6_2[0][0]'] S_pool_7_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_7_2[0][0]'] S_pool_8_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_8_2[0][0]'] S_pool_9_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_9_2[0][0]'] S_pool_10_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_10_2[0][0]'] S_pool_11_2 (MaxPooling2D) (None, 7, 8, 64) 0 ['S_relu_11_2[0][0]'] Y_norm_2 (BatchNormalizati (None, 500, 128) 512 ['Y_conv_2[0][0]'] on) S_conv_0_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_0_2[0][0]'] S_conv_1_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_1_2[0][0]'] S_conv_2_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_2_2[0][0]'] S_conv_3_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_3_2[0][0]'] S_conv_4_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_4_2[0][0]'] S_conv_5_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_5_2[0][0]'] S_conv_6_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_6_2[0][0]'] S_conv_7_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_7_2[0][0]'] S_conv_8_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_8_2[0][0]'] S_conv_9_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_9_2[0][0]'] S_conv_10_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_10_2[0][0]'] S_conv_11_3 (Conv2D) (None, 7, 8, 128) 73856 ['S_pool_11_2[0][0]'] Y_relu_2 (ReLU) (None, 500, 128) 0 ['Y_norm_2[0][0]'] S_norm_0_3 (BatchNormaliza (None, 7, 8, 128) 512 ['S_conv_0_3[0][0]'] tion) S_norm_1_3 (BatchNormaliza (None, 7, 8, 128) 512 ['S_conv_1_3[0][0]'] tion) S_norm_2_3 (BatchNormaliza (None, 7, 8, 128) 512 ['S_conv_2_3[0][0]'] tion) S_norm_3_3 (BatchNormaliza (None, 7, 8, 128) 512 ['S_conv_3_3[0][0]'] tion) S_norm_4_3 (BatchNormaliza (None, 7, 8, 128) 512 ['S_conv_4_3[0][0]'] tion) S_norm_5_3 (BatchNormaliza (None, 7, 8, 128) 512 ['S_conv_5_3[0][0]'] tion) S_norm_6_3 (BatchNormaliza (None, 7, 8, 128) 512 ['S_conv_6_3[0][0]'] tion) S_norm_7_3 (BatchNormaliza (None, 7, 8, 128) 512 ['S_conv_7_3[0][0]'] tion) S_norm_8_3 (BatchNormaliza (None, 7, 8, 128) 512 ['S_conv_8_3[0][0]'] tion) S_norm_9_3 (BatchNormaliza (None, 7, 8, 128) 512 ['S_conv_9_3[0][0]'] tion) S_norm_10_3 (BatchNormaliz (None, 7, 8, 128) 512 ['S_conv_10_3[0][0]'] ation) S_norm_11_3 (BatchNormaliz (None, 7, 8, 128) 512 ['S_conv_11_3[0][0]'] ation) X_inputs (InputLayer) [(None, 7)] 0 [] Y_pool_2 (MaxPooling1D) (None, 250, 128) 0 ['Y_relu_2[0][0]'] S_relu_0_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_0_3[0][0]'] S_relu_1_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_1_3[0][0]'] S_relu_2_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_2_3[0][0]'] S_relu_3_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_3_3[0][0]'] S_relu_4_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_4_3[0][0]'] S_relu_5_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_5_3[0][0]'] S_relu_6_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_6_3[0][0]'] S_relu_7_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_7_3[0][0]'] S_relu_8_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_8_3[0][0]'] S_relu_9_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_9_3[0][0]'] S_relu_10_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_10_3[0][0]'] S_relu_11_3 (ReLU) (None, 7, 8, 128) 0 ['S_norm_11_3[0][0]'] X_norm (Normalization) (None, 7) 15 ['X_inputs[0][0]'] Y_conv_3 (Conv1D) (None, 250, 256) 98560 ['Y_pool_2[0][0]'] S_pool_0_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_0_3[0][0]'] S_pool_1_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_1_3[0][0]'] S_pool_2_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_2_3[0][0]'] S_pool_3_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_3_3[0][0]'] S_pool_4_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_4_3[0][0]'] S_pool_5_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_5_3[0][0]'] S_pool_6_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_6_3[0][0]'] S_pool_7_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_7_3[0][0]'] S_pool_8_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_8_3[0][0]'] S_pool_9_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_9_3[0][0]'] S_pool_10_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_10_3[0][0]'] S_pool_11_3 (MaxPooling2D) (None, 3, 4, 128) 0 ['S_relu_11_3[0][0]'] X_dense_1 (Dense) (None, 32) 256 ['X_norm[0][0]'] Y_norm_3 (BatchNormalizati (None, 250, 256) 1024 ['Y_conv_3[0][0]'] on) S_aver_0 (GlobalAveragePoo (None, 128) 0 ['S_pool_0_3[0][0]'] ling2D) S_aver_1 (GlobalAveragePoo (None, 128) 0 ['S_pool_1_3[0][0]'] ling2D) S_aver_2 (GlobalAveragePoo (None, 128) 0 ['S_pool_2_3[0][0]'] ling2D) S_aver_3 (GlobalAveragePoo (None, 128) 0 ['S_pool_3_3[0][0]'] ling2D) S_aver_4 (GlobalAveragePoo (None, 128) 0 ['S_pool_4_3[0][0]'] ling2D) S_aver_5 (GlobalAveragePoo (None, 128) 0 ['S_pool_5_3[0][0]'] ling2D) S_aver_6 (GlobalAveragePoo (None, 128) 0 ['S_pool_6_3[0][0]'] ling2D) S_aver_7 (GlobalAveragePoo (None, 128) 0 ['S_pool_7_3[0][0]'] ling2D) S_aver_8 (GlobalAveragePoo (None, 128) 0 ['S_pool_8_3[0][0]'] ling2D) S_aver_9 (GlobalAveragePoo (None, 128) 0 ['S_pool_9_3[0][0]'] ling2D) S_aver_10 (GlobalAveragePo (None, 128) 0 ['S_pool_10_3[0][0]'] oling2D) S_aver_11 (GlobalAveragePo (None, 128) 0 ['S_pool_11_3[0][0]'] oling2D) X_drop_1 (Dropout) (None, 32) 0 ['X_dense_1[0][0]'] Y_relu_3 (ReLU) (None, 250, 256) 0 ['Y_norm_3[0][0]'] tf.expand_dims_12 (TFOpLam (None, 1, 128) 0 ['S_aver_0[0][0]'] bda) tf.expand_dims_13 (TFOpLam (None, 1, 128) 0 ['S_aver_1[0][0]'] bda) tf.expand_dims_14 (TFOpLam (None, 1, 128) 0 ['S_aver_2[0][0]'] bda) tf.expand_dims_15 (TFOpLam (None, 1, 128) 0 ['S_aver_3[0][0]'] bda) tf.expand_dims_16 (TFOpLam (None, 1, 128) 0 ['S_aver_4[0][0]'] bda) tf.expand_dims_17 (TFOpLam (None, 1, 128) 0 ['S_aver_5[0][0]'] bda) tf.expand_dims_18 (TFOpLam (None, 1, 128) 0 ['S_aver_6[0][0]'] bda) tf.expand_dims_19 (TFOpLam (None, 1, 128) 0 ['S_aver_7[0][0]'] bda) tf.expand_dims_20 (TFOpLam (None, 1, 128) 0 ['S_aver_8[0][0]'] bda) tf.expand_dims_21 (TFOpLam (None, 1, 128) 0 ['S_aver_9[0][0]'] bda) tf.expand_dims_22 (TFOpLam (None, 1, 128) 0 ['S_aver_10[0][0]'] bda) tf.expand_dims_23 (TFOpLam (None, 1, 128) 0 ['S_aver_11[0][0]'] bda) X_dense_2 (Dense) (None, 32) 1056 ['X_drop_1[0][0]'] Y_aver (GlobalAveragePooli (None, 256) 0 ['Y_relu_3[0][0]'] ng1D) S_concat (Concatenate) (None, 12, 128) 0 ['tf.expand_dims_12[0][0]', 'tf.expand_dims_13[0][0]', 'tf.expand_dims_14[0][0]', 'tf.expand_dims_15[0][0]', 'tf.expand_dims_16[0][0]', 'tf.expand_dims_17[0][0]', 'tf.expand_dims_18[0][0]', 'tf.expand_dims_19[0][0]', 'tf.expand_dims_20[0][0]', 'tf.expand_dims_21[0][0]', 'tf.expand_dims_22[0][0]', 'tf.expand_dims_23[0][0]'] X_drop_2 (Dropout) (None, 32) 0 ['X_dense_2[0][0]'] Y_drop (Dropout) (None, 256) 0 ['Y_aver[0][0]'] S_flatten (Flatten) (None, 1536) 0 ['S_concat[0][0]'] Z_concat (Concatenate) (None, 1824) 0 ['X_drop_2[0][0]', 'Y_drop[0][0]', 'S_flatten[0][0]'] Z_dense_1 (Dense) (None, 128) 233600 ['Z_concat[0][0]'] Z_dense_2 (Dense) (None, 128) 16512 ['Z_dense_1[0][0]'] Z_drop_1 (Dropout) (None, 128) 0 ['Z_dense_2[0][0]'] Z_outputs (Dense) (None, 5) 645 ['Z_drop_1[0][0]'] ================================================================================================== Total params: 1505424 (5.74 MB) Trainable params: 1499109 (5.72 MB) Non-trainable params: 6315 (24.68 KB) __________________________________________________________________________________________________
In [48]:
MODEL_CHECKPOINT = '/kaggle/working/model/model05.ckpt' callbacks_list = [ keras.callbacks.EarlyStopping(monitor='val_binary_accuracy', patience=10), keras.callbacks.ModelCheckpoint(filepath=MODEL_CHECKPOINT, monitor='val_binary_accuracy', save_best_only=True) ] history = model05.fit([X_train, Y_train, S_train], Z_train, epochs=100, batch_size=32, callbacks=callbacks_list, validation_data=([X_valid, Y_valid, S_valid], Z_valid)) model05 = keras.models.load_model(MODEL_CHECKPOINT) model05.save('/kaggle/working/model/model05.keras') Epoch 1/100 546/546 [==============================] - 73s 78ms/step - loss: 0.4367 - binary_accuracy: 0.8027 - precision: 0.6567 - recall: 0.4733 - val_loss: 0.3634 - val_binary_accuracy: 0.8370 - val_precision: 0.7266 - val_recall: 0.5812 Epoch 2/100 546/546 [==============================] - 41s 75ms/step - loss: 0.3563 - binary_accuracy: 0.8468 - precision: 0.7428 - recall: 0.6099 - val_loss: 0.3341 - val_binary_accuracy: 0.8513 - val_precision: 0.7411 - val_recall: 0.6432 Epoch 3/100 546/546 [==============================] - 40s 74ms/step - loss: 0.3281 - binary_accuracy: 0.8630 - precision: 0.7700 - recall: 0.6594 - val_loss: 0.3133 - val_binary_accuracy: 0.8662 - val_precision: 0.7847 - val_recall: 0.6568 Epoch 4/100 546/546 [==============================] - 20s 37ms/step - loss: 0.3108 - binary_accuracy: 0.8724 - precision: 0.7913 - recall: 0.6782 - val_loss: 0.3220 - val_binary_accuracy: 0.8632 - val_precision: 0.7745 - val_recall: 0.6557 Epoch 5/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2937 - binary_accuracy: 0.8803 - precision: 0.8030 - recall: 0.7025 - val_loss: 0.3798 - val_binary_accuracy: 0.8473 - val_precision: 0.7033 - val_recall: 0.6968 Epoch 6/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2888 - binary_accuracy: 0.8830 - precision: 0.8116 - recall: 0.7042 - val_loss: 0.3121 - val_binary_accuracy: 0.8658 - val_precision: 0.7907 - val_recall: 0.6457 Epoch 7/100 546/546 [==============================] - 41s 76ms/step - loss: 0.2806 - binary_accuracy: 0.8865 - precision: 0.8162 - recall: 0.7160 - val_loss: 0.2973 - val_binary_accuracy: 0.8792 - val_precision: 0.8164 - val_recall: 0.6803 Epoch 8/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2742 - binary_accuracy: 0.8895 - precision: 0.8209 - recall: 0.7246 - val_loss: 0.3024 - val_binary_accuracy: 0.8788 - val_precision: 0.7843 - val_recall: 0.7253 Epoch 9/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2694 - binary_accuracy: 0.8917 - precision: 0.8264 - recall: 0.7278 - val_loss: 0.3001 - val_binary_accuracy: 0.8779 - val_precision: 0.7831 - val_recall: 0.7224 Epoch 10/100 546/546 [==============================] - 20s 36ms/step - loss: 0.2631 - binary_accuracy: 0.8948 - precision: 0.8342 - recall: 0.7329 - val_loss: 0.3032 - val_binary_accuracy: 0.8782 - val_precision: 0.7867 - val_recall: 0.7185 Epoch 11/100 546/546 [==============================] - 20s 38ms/step - loss: 0.2601 - binary_accuracy: 0.8970 - precision: 0.8354 - recall: 0.7421 - val_loss: 0.2918 - val_binary_accuracy: 0.8787 - val_precision: 0.8005 - val_recall: 0.7000 Epoch 12/100 546/546 [==============================] - 21s 38ms/step - loss: 0.2547 - binary_accuracy: 0.8981 - precision: 0.8377 - recall: 0.7446 - val_loss: 0.3232 - val_binary_accuracy: 0.8709 - val_precision: 0.8395 - val_recall: 0.6118 Epoch 13/100 546/546 [==============================] - 40s 74ms/step - loss: 0.2512 - binary_accuracy: 0.8993 - precision: 0.8397 - recall: 0.7475 - val_loss: 0.2859 - val_binary_accuracy: 0.8839 - val_precision: 0.8128 - val_recall: 0.7092 Epoch 14/100 546/546 [==============================] - 41s 76ms/step - loss: 0.2454 - binary_accuracy: 0.9016 - precision: 0.8437 - recall: 0.7533 - val_loss: 0.2838 - val_binary_accuracy: 0.8839 - val_precision: 0.8075 - val_recall: 0.7167 Epoch 15/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2376 - binary_accuracy: 0.9052 - precision: 0.8501 - recall: 0.7623 - val_loss: 0.2959 - val_binary_accuracy: 0.8769 - val_precision: 0.7828 - val_recall: 0.7174 Epoch 16/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2359 - binary_accuracy: 0.9067 - precision: 0.8521 - recall: 0.7668 - val_loss: 0.2951 - val_binary_accuracy: 0.8798 - val_precision: 0.8153 - val_recall: 0.6850 Epoch 17/100 546/546 [==============================] - 40s 74ms/step - loss: 0.2270 - binary_accuracy: 0.9094 - precision: 0.8576 - recall: 0.7726 - val_loss: 0.2973 - val_binary_accuracy: 0.8844 - val_precision: 0.7964 - val_recall: 0.7356 Epoch 18/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2248 - binary_accuracy: 0.9102 - precision: 0.8584 - recall: 0.7756 - val_loss: 0.3116 - val_binary_accuracy: 0.8782 - val_precision: 0.7734 - val_recall: 0.7403 Epoch 19/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2181 - binary_accuracy: 0.9126 - precision: 0.8630 - recall: 0.7809 - val_loss: 0.3419 - val_binary_accuracy: 0.8779 - val_precision: 0.7921 - val_recall: 0.7082 Epoch 20/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2097 - binary_accuracy: 0.9167 - precision: 0.8688 - recall: 0.7927 - val_loss: 0.3222 - val_binary_accuracy: 0.8805 - val_precision: 0.8009 - val_recall: 0.7089 Epoch 21/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2054 - binary_accuracy: 0.9184 - precision: 0.8726 - recall: 0.7960 - val_loss: 0.3747 - val_binary_accuracy: 0.8642 - val_precision: 0.7620 - val_recall: 0.6818 Epoch 22/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2005 - binary_accuracy: 0.9207 - precision: 0.8780 - recall: 0.8002 - val_loss: 0.4820 - val_binary_accuracy: 0.8082 - val_precision: 0.6178 - val_recall: 0.6547 Epoch 23/100 546/546 [==============================] - 20s 37ms/step - loss: 0.2109 - binary_accuracy: 0.9171 - precision: 0.8695 - recall: 0.7939 - val_loss: 0.4633 - val_binary_accuracy: 0.8668 - val_precision: 0.7798 - val_recall: 0.6671 Epoch 24/100 546/546 [==============================] - 20s 37ms/step - loss: 0.1875 - binary_accuracy: 0.9260 - precision: 0.8849 - recall: 0.8158 - val_loss: 0.5360 - val_binary_accuracy: 0.8647 - val_precision: 0.7722 - val_recall: 0.6675 Epoch 25/100 546/546 [==============================] - 20s 37ms/step - loss: 0.1782 - binary_accuracy: 0.9305 - precision: 0.8928 - recall: 0.8265 - val_loss: 0.4001 - val_binary_accuracy: 0.8487 - val_precision: 0.7212 - val_recall: 0.6654 Epoch 26/100 546/546 [==============================] - 20s 37ms/step - loss: 0.1826 - binary_accuracy: 0.9285 - precision: 0.8915 - recall: 0.8191 - val_loss: 0.3431 - val_binary_accuracy: 0.8698 - val_precision: 0.7674 - val_recall: 0.7039 Epoch 27/100 546/546 [==============================] - 20s 37ms/step - loss: 0.1625 - binary_accuracy: 0.9355 - precision: 0.8986 - recall: 0.8419 - val_loss: 0.3582 - val_binary_accuracy: 0.8689 - val_precision: 0.7750 - val_recall: 0.6868
Takto se tedy trénovalo:
In [49]:
sns.relplot(data=pd.DataFrame(history.history), kind='line', height=4, aspect=4)
plt.show()
Nakonec můžu model vyhodnotit proti testovacím sadám dat pro X, Y a S:
In [50]:
Z_pred = np.around(model05.predict([X_test, Y_test, S_test], verbose=0))
metrics = pd.concat([metrics.drop(model05.name, level=0, errors='ignore'), model_quality_reporting(Z_test, Z_pred, model_name=model05.name)])
precision recall f1-score support
NORM 0.82 0.91 0.86 964
MI 0.93 0.63 0.75 553
STTC 0.78 0.73 0.75 523
CD 0.71 0.72 0.72 498
HYP 0.71 0.49 0.58 263
micro avg 0.80 0.75 0.77 2801
macro avg 0.79 0.69 0.73 2801
weighted avg 0.81 0.75 0.77 2801
samples avg 0.78 0.75 0.75 2801
Vyhodnocení všech modelů
Tak si nejdříve ukážeme, jak dopadly metriky u všech zkoušených modelů:
In [51]:
metrics.index.set_names(['model', 'group'], inplace=True)
metrics.reset_index(inplace=True)
Metrika Precision
In [52]:
fig, ax = plt.subplots(figsize=(8, 3)) ax = sns.barplot(metrics, x='group', y='precision', hue='model', ax=ax) sns.move_legend(ax, "upper left", bbox_to_anchor=(1, 1)) plt.show()
Metrika Recall
In [53]:
fig, ax = plt.subplots(figsize=(8, 3)) ax = sns.barplot(metrics, x='group', y='recall', hue='model', ax=ax) sns.move_legend(ax, "upper left", bbox_to_anchor=(1, 1)) plt.show()
Metrika F1-score
In [54]:
fig, ax = plt.subplots(figsize=(8, 3)) ax = sns.barplot(metrics, x='group', y='f1-score', hue='model', ax=ax) sns.move_legend(ax, "upper left", bbox_to_anchor=(1, 1)) plt.show()
Závěrečné hodnocení
O celkovém výsledku si udělejte prosím názor sami.
Z mého pohledu vypadá nejlépe kombinovaný model EKG křivky společně s metadaty, a to především proto, že má nejlepší výsledky u těch nejméně zastoupených klasifikačních skupin.
Přidejte svůj názor jako první
Žádné názory
