Tímto článkem navazuji na předchozí: Rozpoznání zápalu plic z RTG snímků – díl první .

Model VGG16

Jedná se o jeden ze základních modelů pro klasifikaci obrázků založených na konvolučních vrstvách. Principem je postupné zmenšování plošných dimenzí a naopak rozšiřování velikosti vlastností. Názorné schéma modelu vypadá takto:

Jako podklad pro detailnější studium by mohl posloužit například tento článek (ale další dostupných zdrojů je pochopitelně více): Everything you need to know about VGG16

VGG16 trénovaný z čistého stolu

Nejdříve si vyzkouším napsat implementaci modelu sám a také jej vytrénovat s využitím vlastních dat.

Takto si tedy vytvořím model:

In [13]:

def create_model_VGG16(X_shape, classes=2, name="VGG16"):

    def mlp(x, hidden_units, activation='relu', dropout_rate=0.3, name=""):
        for i, units in enumerate(hidden_units):
            x = layers.Dense(units, activation=activation, name=f"{name}_{i}_dense")(x)
            x = layers.Dropout(dropout_rate, name=f"{name}_{i}_dropout")(x)
        return x

    def conv_block(x, filters, *, kernels=None, steps=None, dropout=0.2, name=""):
        for i in range(len(filters)):
            x = layers.Conv2D(filters[i], kernels[i] if kernels else (3, 3), strides=steps[i] if steps else (1, 1), padding='same', name=f'{name}_conv_{i}')(x)
            x = layers.BatchNormalization(name=f'{name}_norm_{i}')(x)
            x = layers.Activation('relu', name=f'{name}_relu_{i}')(x)
        x = layers.MaxPooling2D((2, 2), name=f'{name}_maxpool')(x)
        return x

    inputs = Input(X_shape[-3:], name='inputs')

    x = conv_block(inputs, (64, 64), name="block_1")
    x = conv_block(x, (128, 128), name="block_2")
    x = conv_block(x, (256, 256, 256), name="block_3")
    x = conv_block(x, (512, 512, 512), name="block_4")
    x = conv_block(x, (512, 512, 512), name="block_5")

    x = layers.Flatten(name="flatten")(x)

    x = mlp(x, (1024, 512), name="dense")
    outputs = layers.Dense(classes, activation='softmax', name='outputs')(x)

    return Model(inputs=inputs, outputs=outputs, name=name)

Vidíte, že je poskládán z pěti konvolučních bloků, ve kterých se postupně zmenšují prostorové dimenze (to má na svědomí vrstva MaxPooling2D) a naopak se zvětšují počty vlastností (to je dáno počtem filtrů).

Po konvolučních blocích následuje úplné opuštění prostorových dimenzí – vrstva Flatten.

Závěr je zajištěn plně propojenými vrstvami ve formě multi-layer perceptron (MLP) a závěrečné klasifikační vrstvy s aktivací softmax.

Před testováním opět příprava dat. Budu používat snímky v odstínech šedi:

In [14]:

x_train, x_valid, y_train, y_valid = train_test_split(*get_datasource(DATA_TRAIN, DATA_VALID), test_size=0.2)
x_test, y_test = get_datasource(DATA_TEST)

x_train = np.expand_dims(x_train, axis=-1)
x_valid = np.expand_dims(x_valid, axis=-1)
x_test  = np.expand_dims(x_test, axis=-1)

datagen = ImageDataGenerator(
        rotation_range = 30,
        zoom_range = 0.2,
        width_shift_range=0.1,
        height_shift_range=0.1,
        horizontal_flip = True,
        vertical_flip=False)

datagen.fit(x_train)

A takto vypadá vlastní vyhodnocení modelu:

In [15]:

evaluate_model(create_model_VGG16(x_train.shape, 2, name='VGG16'), forced_training=False)
=== MODEL EVALUATION =================================================

Model: "VGG16"
_________________________________________________________________
 Layer (type)                Output Shape              Param #
=================================================================
 inputs (InputLayer)         [(None, 224, 224, 1)]     0

 block_1_conv_0 (Conv2D)     (None, 224, 224, 64)      640
 block_1_norm_0 (BatchNorma  (None, 224, 224, 64)      256
 lization)
 block_1_relu_0 (Activation  (None, 224, 224, 64)      0
 )
 block_1_conv_1 (Conv2D)     (None, 224, 224, 64)      36928
 block_1_norm_1 (BatchNorma  (None, 224, 224, 64)      256
 lization)
 block_1_relu_1 (Activation  (None, 224, 224, 64)      0
 )
 block_1_maxpool (MaxPoolin  (None, 112, 112, 64)      0
 g2D)
 block_2_conv_0 (Conv2D)     (None, 112, 112, 128)     73856
 block_2_norm_0 (BatchNorma  (None, 112, 112, 128)     512
 lization)
 block_2_relu_0 (Activation  (None, 112, 112, 128)     0
 )
 block_2_conv_1 (Conv2D)     (None, 112, 112, 128)     147584
 block_2_norm_1 (BatchNorma  (None, 112, 112, 128)     512
 lization)
 block_2_relu_1 (Activation  (None, 112, 112, 128)     0
 )
 block_2_maxpool (MaxPoolin  (None, 56, 56, 128)       0
 g2D)
 block_3_conv_0 (Conv2D)     (None, 56, 56, 256)       295168
 block_3_norm_0 (BatchNorma  (None, 56, 56, 256)       1024
 lization)
 block_3_relu_0 (Activation  (None, 56, 56, 256)       0
 )
 block_3_conv_1 (Conv2D)     (None, 56, 56, 256)       590080
 block_3_norm_1 (BatchNorma  (None, 56, 56, 256)       1024
 lization)
 block_3_relu_1 (Activation  (None, 56, 56, 256)       0
 )
 block_3_conv_2 (Conv2D)     (None, 56, 56, 256)       590080
 block_3_norm_2 (BatchNorma  (None, 56, 56, 256)       1024
 lization)
 block_3_relu_2 (Activation  (None, 56, 56, 256)       0
 )
 block_3_maxpool (MaxPoolin  (None, 28, 28, 256)       0
 g2D)
 block_4_conv_0 (Conv2D)     (None, 28, 28, 512)       1180160
 block_4_norm_0 (BatchNorma  (None, 28, 28, 512)       2048
 lization)
 block_4_relu_0 (Activation  (None, 28, 28, 512)       0
 )
 block_4_conv_1 (Conv2D)     (None, 28, 28, 512)       2359808
 block_4_norm_1 (BatchNorma  (None, 28, 28, 512)       2048
 lization)
 block_4_relu_1 (Activation  (None, 28, 28, 512)       0
 )
 block_4_conv_2 (Conv2D)     (None, 28, 28, 512)       2359808
 block_4_norm_2 (BatchNorma  (None, 28, 28, 512)       2048
 lization)
 block_4_relu_2 (Activation  (None, 28, 28, 512)       0
 )
 block_4_maxpool (MaxPoolin  (None, 14, 14, 512)       0
 g2D)
 block_5_conv_0 (Conv2D)     (None, 14, 14, 512)       2359808
 block_5_norm_0 (BatchNorma  (None, 14, 14, 512)       2048
 lization)
 block_5_relu_0 (Activation  (None, 14, 14, 512)       0
 )
 block_5_conv_1 (Conv2D)     (None, 14, 14, 512)       2359808
 block_5_norm_1 (BatchNorma  (None, 14, 14, 512)       2048
 lization)
 block_5_relu_1 (Activation  (None, 14, 14, 512)       0
 )
 block_5_conv_2 (Conv2D)     (None, 14, 14, 512)       2359808
 block_5_norm_2 (BatchNorma  (None, 14, 14, 512)       2048
 lization)
 block_5_relu_2 (Activation  (None, 14, 14, 512)       0
 )
 block_5_maxpool (MaxPoolin  (None, 7, 7, 512)         0
 g2D)
 flatten (Flatten)           (None, 25088)             0
 dense_0_dense (Dense)       (None, 1024)              25691136
 dense_0_dropout (Dropout)   (None, 1024)              0
 dense_1_dense (Dense)       (None, 512)               524800
 dense_1_dropout (Dropout)   (None, 512)               0
 outputs (Dense)             (None, 2)                 1026
=================================================================
Total params: 40947394 (156.20 MB)
Trainable params: 40938946 (156.17 MB)
Non-trainable params: 8448 (33.00 KB)
_________________________________________________________________

--- Model training ---------------------------------------------------

Epoch 1/40
131/131 [==============================] - ETA: 0s - loss: 1.3506 - accuracy: 0.6944 - auc: 0.7577
Epoch 1: val_auc improved from -inf to 0.74785, saving model to /kaggle/working/model/VGG16.ckpt
131/131 [==============================] - 62s 369ms/step - loss: 1.3506 - accuracy: 0.6944 - auc: 0.7577 - val_loss: 43.0450 - val_accuracy: 0.7479 - val_auc: 0.7479
Epoch 2/40
131/131 [==============================] - ETA: 0s - loss: 0.6281 - accuracy: 0.7376 - auc: 0.8404
Epoch 2: val_auc did not improve from 0.74785
131/131 [==============================] - 36s 276ms/step - loss: 0.6281 - accuracy: 0.7376 - auc: 0.8404 - val_loss: 12.9182 - val_accuracy: 0.7479 - val_auc: 0.7479
Epoch 3/40
131/131 [==============================] - ETA: 0s - loss: 0.4900 - accuracy: 0.8256 - auc: 0.9112
Epoch 3: val_auc did not improve from 0.74785
131/131 [==============================] - 37s 279ms/step - loss: 0.4900 - accuracy: 0.8256 - auc: 0.9112 - val_loss: 7.0124 - val_accuracy: 0.7479 - val_auc: 0.7479
Epoch 4/40
131/131 [==============================] - ETA: 0s - loss: 0.3516 - accuracy: 0.8597 - auc: 0.9375
Epoch 4: val_auc did not improve from 0.74785
131/131 [==============================] - 37s 278ms/step - loss: 0.3516 - accuracy: 0.8597 - auc: 0.9375 - val_loss: 6.2788 - val_accuracy: 0.7479 - val_auc: 0.7479
Epoch 5/40
131/131 [==============================] - ETA: 0s - loss: 0.2710 - accuracy: 0.8822 - auc: 0.9561
Epoch 5: val_auc improved from 0.74785 to 0.75282, saving model to /kaggle/working/model/VGG16.ckpt
131/131 [==============================] - 44s 332ms/step - loss: 0.2710 - accuracy: 0.8822 - auc: 0.9561 - val_loss: 0.6298 - val_accuracy: 0.6619 - val_auc: 0.7528
Epoch 6/40
131/131 [==============================] - ETA: 0s - loss: 0.2601 - accuracy: 0.8941 - auc: 0.9623
Epoch 6: val_auc improved from 0.75282 to 0.96774, saving model to /kaggle/working/model/VGG16.ckpt
131/131 [==============================] - 44s 332ms/step - loss: 0.2601 - accuracy: 0.8941 - auc: 0.9623 - val_loss: 0.2389 - val_accuracy: 0.8968 - val_auc: 0.9677
Epoch 7/40
131/131 [==============================] - ETA: 0s - loss: 0.2816 - accuracy: 0.8877 - auc: 0.9578
Epoch 7: val_auc did not improve from 0.96774
131/131 [==============================] - 37s 278ms/step - loss: 0.2816 - accuracy: 0.8877 - auc: 0.9578 - val_loss: 0.7441 - val_accuracy: 0.7832 - val_auc: 0.8990
Epoch 8/40
131/131 [==============================] - ETA: 0s - loss: 0.2391 - accuracy: 0.9020 - auc: 0.9658
Epoch 8: val_auc did not improve from 0.96774
131/131 [==============================] - 36s 277ms/step - loss: 0.2391 - accuracy: 0.9020 - auc: 0.9658 - val_loss: 0.7869 - val_accuracy: 0.6285 - val_auc: 0.6936
Epoch 9/40
131/131 [==============================] - ETA: 0s - loss: 0.2318 - accuracy: 0.9075 - auc: 0.9690
Epoch 9: val_auc did not improve from 0.96774
131/131 [==============================] - 37s 278ms/step - loss: 0.2318 - accuracy: 0.9075 - auc: 0.9690 - val_loss: 0.7344 - val_accuracy: 0.5979 - val_auc: 0.6878
Epoch 10/40
131/131 [==============================] - ETA: 0s - loss: 0.2295 - accuracy: 0.9109 - auc: 0.9692
Epoch 10: val_auc did not improve from 0.96774
131/131 [==============================] - 36s 277ms/step - loss: 0.2295 - accuracy: 0.9109 - auc: 0.9692 - val_loss: 0.3142 - val_accuracy: 0.8739 - val_auc: 0.9428
Epoch 11/40
131/131 [==============================] - ETA: 0s - loss: 0.2212 - accuracy: 0.9202 - auc: 0.9721
Epoch 11: val_auc did not improve from 0.96774
131/131 [==============================] - 36s 276ms/step - loss: 0.2212 - accuracy: 0.9202 - auc: 0.9721 - val_loss: 0.8512 - val_accuracy: 0.6122 - val_auc: 0.6439
Epoch 12/40
131/131 [==============================] - ETA: 0s - loss: 0.2093 - accuracy: 0.9183 - auc: 0.9746
Epoch 12: val_auc did not improve from 0.96774
131/131 [==============================] - 37s 279ms/step - loss: 0.2093 - accuracy: 0.9183 - auc: 0.9746 - val_loss: 0.2938 - val_accuracy: 0.8892 - val_auc: 0.9558
Epoch 13/40
131/131 [==============================] - ETA: 0s - loss: 0.1989 - accuracy: 0.9302 - auc: 0.9765
Epoch 13: val_auc did not improve from 0.96774
131/131 [==============================] - 36s 278ms/step - loss: 0.1989 - accuracy: 0.9302 - auc: 0.9765 - val_loss: 0.2497 - val_accuracy: 0.9035 - val_auc: 0.9614
Epoch 14/40
131/131 [==============================] - ETA: 0s - loss: 0.1884 - accuracy: 0.9345 - auc: 0.9783
Epoch 14: val_auc did not improve from 0.96774
131/131 [==============================] - 37s 279ms/step - loss: 0.1884 - accuracy: 0.9345 - auc: 0.9783 - val_loss: 0.5664 - val_accuracy: 0.7650 - val_auc: 0.8407
Epoch 15/40
131/131 [==============================] - ETA: 0s - loss: 0.2049 - accuracy: 0.9188 - auc: 0.9750
Epoch 15: val_auc did not improve from 0.96774
131/131 [==============================] - 36s 278ms/step - loss: 0.2049 - accuracy: 0.9188 - auc: 0.9750 - val_loss: 0.8792 - val_accuracy: 0.7698 - val_auc: 0.8510
Epoch 16/40
131/131 [==============================] - ETA: 0s - loss: 0.1908 - accuracy: 0.9309 - auc: 0.9776
Epoch 16: val_auc did not improve from 0.96774
131/131 [==============================] - 37s 279ms/step - loss: 0.1908 - accuracy: 0.9309 - auc: 0.9776 - val_loss: 0.4967 - val_accuracy: 0.7584 - val_auc: 0.8660

--- Training history -------------------------------------------------

--- Test Predictions and Metrics -------------------------------------

              precision    recall  f1-score   support

      NORMAL       0.84      0.58      0.69       234
   PNEUMONIA       0.79      0.94      0.85       390

    accuracy                           0.80       624
   macro avg       0.82      0.76      0.77       624
weighted avg       0.81      0.80      0.79       624


=== MODEL EVALUATION FINISHED ========================================

VGG16 transfer learning

V předchozí kapitole jsem si model vytvořil sám a sám jsem jej také trénoval na vlastních datech. Druhou možností, a ta je také hodně oblíbená, je využití již dříve připravených a trénovaných modelů. Takových již připravených modelů je velké množství, například některé jako součást distribuce Keras. Vyzkouším si jeden takový použít.

Problémem je, že tyto modely byly trénovány pro jinou datovou sadu, než mám já, a také na jiné klasifikační třídy (prostě klasifikaci do třídy NORMAL a PNEUMONIA byste tam nenašli). To se řeší tak, že se použije jen část připraveného modelu, především ta konvoluční, a klasifikační vrstvy se tam dodají až následně. No a pak se trénují pouze ty doplněné vrstvy.

Takto nějak by to mohlo vypadat pro model VGG16:

In [16]:

from keras.applications import VGG16

def create_model_VGG16Trans(X_shape, classes=2, name="VGG16Trans"):

    def mlp(x, hidden_units, activation='relu', dropout_rate=0.3, name=""):
        for i, units in enumerate(hidden_units):
            x = layers.Dense(units, activation=activation, name=f"{name}_{i}_dense")(x)
            x = layers.Dropout(dropout_rate, name=f"{name}_{i}_dropout")(x)
        return x

    base_model = VGG16(include_top=False, input_shape=tuple(X_shape)[-3:])
    base_model.trainable = False

    inputs = Input(X_shape[-3:], name='inputs')

    x = base_model(inputs, training=False)
    x = layers.Flatten(name="flatten")(x)

    x = mlp(x, (1024, 512), name="dense")
    outputs = layers.Dense(classes, activation='softmax', name='outputs')(x)

    return Model(inputs=inputs, outputs=outputs, name=name)

Načetl jsem si model z distribuce Keras s vyloučením klasifikačních vrstev. Ty jsem následně doplnil sám ve formě MLP a poslední plně propojené vrstvy s aktivační funkcí softmax do dvou klasifikačních tříd.

Musím si ještě připravit data. V tomto případě musím obrázky načíst jako RGB snímky, protože původní model takto trénován byl:

In [17]:

x_train, x_valid, y_train, y_valid = train_test_split(*get_datasource(DATA_TRAIN, DATA_VALID, flag=cv2.IMREAD_COLOR), test_size=0.2)
x_test, y_test = get_datasource(DATA_TEST, flag=cv2.IMREAD_COLOR)

datagen = ImageDataGenerator(
        rotation_range = 30,
        zoom_range = 0.2,
        width_shift_range=0.1,
        height_shift_range=0.1,
        horizontal_flip = True,
        vertical_flip=False)

datagen.fit(x_train)

A nyní již vlastní vyhodnocení:

In [18]:

evaluate_model(create_model_VGG16Trans(x_train.shape, 2), forced_training=False)
Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/vgg16/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5
58889256/58889256 [==============================] - 0s 0us/step
=== MODEL EVALUATION =================================================

Model: "VGG16Trans"
_________________________________________________________________
 Layer (type)                Output Shape              Param #
=================================================================
 inputs (InputLayer)         [(None, 224, 224, 3)]     0
 vgg16 (Functional)          (None, 7, 7, 512)         14714688
 flatten (Flatten)           (None, 25088)             0
 dense_0_dense (Dense)       (None, 1024)              25691136
 dense_0_dropout (Dropout)   (None, 1024)              0
 dense_1_dense (Dense)       (None, 512)               524800
 dense_1_dropout (Dropout)   (None, 512)               0
 outputs (Dense)             (None, 2)                 1026
=================================================================
Total params: 40931650 (156.14 MB)
Trainable params: 26216962 (100.01 MB)
Non-trainable params: 14714688 (56.13 MB)
_________________________________________________________________

--- Model training ---------------------------------------------------

Epoch 1/40
131/131 [==============================] - ETA: 0s - loss: 0.7880 - accuracy: 0.8380 - auc: 0.9009
Epoch 1: val_auc improved from -inf to 0.97700, saving model to /kaggle/working/model/VGG16Trans.ckpt
131/131 [==============================] - 60s 427ms/step - loss: 0.7880 - accuracy: 0.8380 - auc: 0.9009 - val_loss: 0.1930 - val_accuracy: 0.9236 - val_auc: 0.9770
Epoch 2/40
131/131 [==============================] - ETA: 0s - loss: 0.2414 - accuracy: 0.9054 - auc: 0.9660
Epoch 2: val_auc improved from 0.97700 to 0.98849, saving model to /kaggle/working/model/VGG16Trans.ckpt
131/131 [==============================] - 54s 414ms/step - loss: 0.2414 - accuracy: 0.9054 - auc: 0.9660 - val_loss: 0.1473 - val_accuracy: 0.9484 - val_auc: 0.9885
Epoch 3/40
131/131 [==============================] - ETA: 0s - loss: 0.2004 - accuracy: 0.9197 - auc: 0.9756
Epoch 3: val_auc did not improve from 0.98849
131/131 [==============================] - 52s 397ms/step - loss: 0.2004 - accuracy: 0.9197 - auc: 0.9756 - val_loss: 0.1584 - val_accuracy: 0.9341 - val_auc: 0.9851
Epoch 4/40
131/131 [==============================] - ETA: 0s - loss: 0.1825 - accuracy: 0.9266 - auc: 0.9795
Epoch 4: val_auc improved from 0.98849 to 0.98866, saving model to /kaggle/working/model/VGG16Trans.ckpt
131/131 [==============================] - 55s 420ms/step - loss: 0.1825 - accuracy: 0.9266 - auc: 0.9795 - val_loss: 0.1331 - val_accuracy: 0.9503 - val_auc: 0.9887
Epoch 5/40
131/131 [==============================] - ETA: 0s - loss: 0.1848 - accuracy: 0.9271 - auc: 0.9789
Epoch 5: val_auc did not improve from 0.98866
131/131 [==============================] - 51s 392ms/step - loss: 0.1848 - accuracy: 0.9271 - auc: 0.9789 - val_loss: 0.1578 - val_accuracy: 0.9446 - val_auc: 0.9832
Epoch 6/40
131/131 [==============================] - ETA: 0s - loss: 0.1545 - accuracy: 0.9407 - auc: 0.9850
Epoch 6: val_auc improved from 0.98866 to 0.98935, saving model to /kaggle/working/model/VGG16Trans.ckpt
131/131 [==============================] - 55s 420ms/step - loss: 0.1545 - accuracy: 0.9407 - auc: 0.9850 - val_loss: 0.1329 - val_accuracy: 0.9436 - val_auc: 0.9893
Epoch 7/40
131/131 [==============================] - ETA: 0s - loss: 0.1591 - accuracy: 0.9395 - auc: 0.9842
Epoch 7: val_auc did not improve from 0.98935
131/131 [==============================] - 51s 389ms/step - loss: 0.1591 - accuracy: 0.9395 - auc: 0.9842 - val_loss: 0.1284 - val_accuracy: 0.9484 - val_auc: 0.9893
Epoch 8/40
131/131 [==============================] - ETA: 0s - loss: 0.1554 - accuracy: 0.9381 - auc: 0.9850
Epoch 8: val_auc improved from 0.98935 to 0.99141, saving model to /kaggle/working/model/VGG16Trans.ckpt
131/131 [==============================] - 55s 417ms/step - loss: 0.1554 - accuracy: 0.9381 - auc: 0.9850 - val_loss: 0.1194 - val_accuracy: 0.9513 - val_auc: 0.9914
Epoch 9/40
131/131 [==============================] - ETA: 0s - loss: 0.1449 - accuracy: 0.9417 - auc: 0.9869
Epoch 9: val_auc did not improve from 0.99141
131/131 [==============================] - 52s 395ms/step - loss: 0.1449 - accuracy: 0.9417 - auc: 0.9869 - val_loss: 0.1535 - val_accuracy: 0.9389 - val_auc: 0.9860
Epoch 10/40
131/131 [==============================] - ETA: 0s - loss: 0.1368 - accuracy: 0.9467 - auc: 0.9883
Epoch 10: val_auc did not improve from 0.99141
131/131 [==============================] - 51s 387ms/step - loss: 0.1368 - accuracy: 0.9467 - auc: 0.9883 - val_loss: 0.1240 - val_accuracy: 0.9561 - val_auc: 0.9909
Epoch 11/40
131/131 [==============================] - ETA: 0s - loss: 0.1611 - accuracy: 0.9352 - auc: 0.9839
Epoch 11: val_auc did not improve from 0.99141
131/131 [==============================] - 51s 390ms/step - loss: 0.1611 - accuracy: 0.9352 - auc: 0.9839 - val_loss: 0.1199 - val_accuracy: 0.9503 - val_auc: 0.9912
Epoch 12/40
131/131 [==============================] - ETA: 0s - loss: 0.1677 - accuracy: 0.9295 - auc: 0.9829
Epoch 12: val_auc did not improve from 0.99141
131/131 [==============================] - 51s 389ms/step - loss: 0.1677 - accuracy: 0.9295 - auc: 0.9829 - val_loss: 0.1267 - val_accuracy: 0.9522 - val_auc: 0.9894
Epoch 13/40
131/131 [==============================] - ETA: 0s - loss: 0.1464 - accuracy: 0.9405 - auc: 0.9867
Epoch 13: val_auc did not improve from 0.99141
131/131 [==============================] - 51s 391ms/step - loss: 0.1464 - accuracy: 0.9405 - auc: 0.9867 - val_loss: 0.1261 - val_accuracy: 0.9494 - val_auc: 0.9906
Epoch 14/40
131/131 [==============================] - ETA: 0s - loss: 0.1295 - accuracy: 0.9498 - auc: 0.9893
Epoch 14: val_auc improved from 0.99141 to 0.99262, saving model to /kaggle/working/model/VGG16Trans.ckpt
131/131 [==============================] - 54s 415ms/step - loss: 0.1295 - accuracy: 0.9498 - auc: 0.9893 - val_loss: 0.1198 - val_accuracy: 0.9589 - val_auc: 0.9926
Epoch 15/40
131/131 [==============================] - ETA: 0s - loss: 0.1374 - accuracy: 0.9503 - auc: 0.9881
Epoch 15: val_auc did not improve from 0.99262
131/131 [==============================] - 52s 395ms/step - loss: 0.1374 - accuracy: 0.9503 - auc: 0.9881 - val_loss: 0.1211 - val_accuracy: 0.9580 - val_auc: 0.9906
Epoch 16/40
131/131 [==============================] - ETA: 0s - loss: 0.1329 - accuracy: 0.9486 - auc: 0.9889
Epoch 16: val_auc did not improve from 0.99262
131/131 [==============================] - 51s 392ms/step - loss: 0.1329 - accuracy: 0.9486 - auc: 0.9889 - val_loss: 0.1131 - val_accuracy: 0.9618 - val_auc: 0.9907
Epoch 17/40
131/131 [==============================] - ETA: 0s - loss: 0.1353 - accuracy: 0.9503 - auc: 0.9880
Epoch 17: val_auc did not improve from 0.99262
131/131 [==============================] - 51s 393ms/step - loss: 0.1353 - accuracy: 0.9503 - auc: 0.9880 - val_loss: 0.1136 - val_accuracy: 0.9647 - val_auc: 0.9905
Epoch 18/40
131/131 [==============================] - ETA: 0s - loss: 0.1287 - accuracy: 0.9520 - auc: 0.9898
Epoch 18: val_auc did not improve from 0.99262
131/131 [==============================] - 51s 391ms/step - loss: 0.1287 - accuracy: 0.9520 - auc: 0.9898 - val_loss: 0.1262 - val_accuracy: 0.9589 - val_auc: 0.9911
Epoch 19/40
131/131 [==============================] - ETA: 0s - loss: 0.1313 - accuracy: 0.9453 - auc: 0.9895
Epoch 19: val_auc did not improve from 0.99262
131/131 [==============================] - 51s 390ms/step - loss: 0.1313 - accuracy: 0.9453 - auc: 0.9895 - val_loss: 0.1219 - val_accuracy: 0.9589 - val_auc: 0.9897
Epoch 20/40
131/131 [==============================] - ETA: 0s - loss: 0.1250 - accuracy: 0.9534 - auc: 0.9897
Epoch 20: val_auc did not improve from 0.99262
131/131 [==============================] - 51s 393ms/step - loss: 0.1250 - accuracy: 0.9534 - auc: 0.9897 - val_loss: 0.1068 - val_accuracy: 0.9637 - val_auc: 0.9922
Epoch 21/40
131/131 [==============================] - ETA: 0s - loss: 0.1168 - accuracy: 0.9536 - auc: 0.9913
Epoch 21: val_auc did not improve from 0.99262
131/131 [==============================] - 52s 396ms/step - loss: 0.1168 - accuracy: 0.9536 - auc: 0.9913 - val_loss: 0.1225 - val_accuracy: 0.9522 - val_auc: 0.9909
Epoch 22/40
131/131 [==============================] - ETA: 0s - loss: 0.1160 - accuracy: 0.9558 - auc: 0.9909
Epoch 22: val_auc did not improve from 0.99262
131/131 [==============================] - 51s 386ms/step - loss: 0.1160 - accuracy: 0.9558 - auc: 0.9909 - val_loss: 0.1367 - val_accuracy: 0.9532 - val_auc: 0.9884
Epoch 23/40
131/131 [==============================] - ETA: 0s - loss: 0.1098 - accuracy: 0.9572 - auc: 0.9925
Epoch 23: val_auc did not improve from 0.99262
131/131 [==============================] - 52s 398ms/step - loss: 0.1098 - accuracy: 0.9572 - auc: 0.9925 - val_loss: 0.1116 - val_accuracy: 0.9551 - val_auc: 0.9924
Epoch 24/40
131/131 [==============================] - ETA: 0s - loss: 0.1162 - accuracy: 0.9534 - auc: 0.9913
Epoch 24: val_auc did not improve from 0.99262
131/131 [==============================] - 51s 392ms/step - loss: 0.1162 - accuracy: 0.9534 - auc: 0.9913 - val_loss: 0.1244 - val_accuracy: 0.9580 - val_auc: 0.9894

--- Training history -------------------------------------------------

--- Test Predictions and Metrics -------------------------------------

              precision    recall  f1-score   support

      NORMAL       0.89      0.87      0.88       234
   PNEUMONIA       0.92      0.94      0.93       390

    accuracy                           0.91       624
   macro avg       0.91      0.91      0.91       624
weighted avg       0.91      0.91      0.91       624


=== MODEL EVALUATION FINISHED ========================================

V následujícím dílu budu pokračovat modelem ResNet50.