Usuario:
Modelo vgg16
Descripción 
El modelo vgg16 se especializa en análisis de imágenes y esta disponible desde Keras. Se carga a tensorflow (tf) mediante keras.applications:
import numpy as np import tensorflow as tf from tensorflow import keras from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Activation, Dense, Flatten, BatchNormalization, Conv2D, MaxPool2D # cargar modelo vgg16 vgg16_model = tf.keras.applications.vgg16.VGG16()
ID:(13759, 0)
Formar modelo
Descripción
Antes de proceder a entrenar el modelo se deben bloquear las capas que no deben ser entrenadas ya que ya lo fueron en la estructuración del modelo.
# definir modelo
vgg_model = Sequential()
# Bloqueo de 16 capas para que no sean entrenadas
for layer in vgg16_model.layers[:-1]:
vgg_model.add(layer)
ID:(13760, 0)
Mostrar modelo
Descripción
Se puede mostrar con summary el resumen del modelo:
# mostrar el resumen del modelo vgg_model.summary()
Hay que hacer notar que
- los comandos MaxPooling van reduciendo el tamaño de las imágenes desde los 224 a 112, 56, 28, 14 y 7
- las convoluciones (Conv2D) para ello van aumentando las dimensiones de la paleta de colores de 64 a 128, 256 y 512
Model: 'sequential_1'
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
block1_conv1 (Conv2D) (None, 224, 224, 64) 1792
_________________________________________________________________
block1_conv2 (Conv2D) (None, 224, 224, 64) 36928
_________________________________________________________________
block1_pool (MaxPooling2D) (None, 112, 112, 64) 0
_________________________________________________________________
block2_conv1 (Conv2D) (None, 112, 112, 128) 73856
_________________________________________________________________
block2_conv2 (Conv2D) (None, 112, 112, 128) 147584
_________________________________________________________________
block2_pool (MaxPooling2D) (None, 56, 56, 128) 0
_________________________________________________________________
block3_conv1 (Conv2D) (None, 56, 56, 256) 295168
_________________________________________________________________
block3_conv2 (Conv2D) (None, 56, 56, 256) 590080
_________________________________________________________________
block3_conv3 (Conv2D) (None, 56, 56, 256) 590080
_________________________________________________________________
block3_pool (MaxPooling2D) (None, 28, 28, 256) 0
_________________________________________________________________
block4_conv1 (Conv2D) (None, 28, 28, 512) 1180160
_________________________________________________________________
block4_conv2 (Conv2D) (None, 28, 28, 512) 2359808
_________________________________________________________________
block4_conv3 (Conv2D) (None, 28, 28, 512) 2359808
_________________________________________________________________
block4_pool (MaxPooling2D) (None, 14, 14, 512) 0
_________________________________________________________________
block5_conv1 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block5_conv2 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block5_conv3 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block5_pool (MaxPooling2D) (None, 7, 7, 512) 0
_________________________________________________________________
flatten (Flatten) (None, 25088) 0
_________________________________________________________________
fc1 (Dense) (None, 4096) 102764544
_________________________________________________________________
fc2 (Dense) (None, 4096) 16781312
=================================================================
Total params: 134,260,544
Trainable params: 134,260,544
Non-trainable params: 0
_________________________________________________________________
ID:(13761, 0)
Bloquear capas ya entrenadas
Descripción
Para evitar que en el proceso de aprendizaje de vuelva a entrenar las capas del modelo original se procede al bloqueo del parámetro trainable:
# bloqueo de capas
for layer in vgg_model.layers:
layer.trainable = False
Para aprovechar la información ya obtenida en el aprendizaje en modelo original se debe bloquear el aprendizaje de dichas capa.
ID:(13762, 0)
Agregar capa final
Descripción
El modelo se adapta agregando una capa final del tamaño de las categorías que se quieren predecir:
# agregar capa final (densa) vgg_model.add(Dense(units=len(classes),activation='softmax'))
El units debe ser igualado al numero de clases que se pretenden pronosticar.
ID:(13763, 0)
Mostrar modelo modificado
Descripción
Para mostrar como se modifico el modelo se puede nuevamente usar la función summary:
# mostrar modelo vgg vgg_model.summary()
En el listado de capas se observa al final la capa densa con las clases que se desean pronosticar:
Model: 'sequential_1'
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
block1_conv1 (Conv2D) (None, 224, 224, 64) 1792
_________________________________________________________________
block1_conv2 (Conv2D) (None, 224, 224, 64) 36928
_________________________________________________________________
block1_pool (MaxPooling2D) (None, 112, 112, 64) 0
_________________________________________________________________
block2_conv1 (Conv2D) (None, 112, 112, 128) 73856
_________________________________________________________________
block2_conv2 (Conv2D) (None, 112, 112, 128) 147584
_________________________________________________________________
block2_pool (MaxPooling2D) (None, 56, 56, 128) 0
_________________________________________________________________
block3_conv1 (Conv2D) (None, 56, 56, 256) 295168
_________________________________________________________________
block3_conv2 (Conv2D) (None, 56, 56, 256) 590080
_________________________________________________________________
block3_conv3 (Conv2D) (None, 56, 56, 256) 590080
_________________________________________________________________
block3_pool (MaxPooling2D) (None, 28, 28, 256) 0
_________________________________________________________________
block4_conv1 (Conv2D) (None, 28, 28, 512) 1180160
_________________________________________________________________
block4_conv2 (Conv2D) (None, 28, 28, 512) 2359808
_________________________________________________________________
block4_conv3 (Conv2D) (None, 28, 28, 512) 2359808
_________________________________________________________________
block4_pool (MaxPooling2D) (None, 14, 14, 512) 0
_________________________________________________________________
block5_conv1 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block5_conv2 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block5_conv3 (Conv2D) (None, 14, 14, 512) 2359808
_________________________________________________________________
block5_pool (MaxPooling2D) (None, 7, 7, 512) 0
_________________________________________________________________
flatten (Flatten) (None, 25088) 0
_________________________________________________________________
fc1 (Dense) (None, 4096) 102764544
_________________________________________________________________
fc2 (Dense) (None, 4096) 16781312
_________________________________________________________________
dense_1 (Dense) (None, 290) 1188130
=================================================================
Total params: 135,448,674
Trainable params: 1,188,130
Non-trainable params: 134,260,544
_________________________________________________________________
ID:(13764, 0)
Compilar modelo
Descripción
Para armar el modelo se debe proceder a compilarlo:
# Compilar el modelo vgg_model.compile(optimizer=Adam(learning_rate=0.0001), loss='categorical_crossentropy',metrics=['accuracy'])
ID:(13765, 0)
Proceso de aprendizaje
Descripción
# Proceso de aprendizaje con el modelo vgg16 modificado vgg_model.fit(x=train_batches, validation_data=validate_batches, epochs=5,verbose=2)
Epoch 1/5
373/373 - 336s - loss: 2.4201 - accuracy: 0.2790 - val_loss: 2.2614 - val_accuracy: 0.3023
Epoch 2/5
373/373 - 340s - loss: 1.7795 - accuracy: 0.4362 - val_loss: 2.0955 - val_accuracy: 0.3639
Epoch 3/5
373/373 - 339s - loss: 1.5116 - accuracy: 0.5241 - val_loss: 2.0217 - val_accuracy: 0.3868
Epoch 4/5
373/373 - 339s - loss: 1.3313 - accuracy: 0.5754 - val_loss: 1.9925 - val_accuracy: 0.3926
Epoch 5/5
373/373 - 341s - loss: 1.1973 - accuracy: 0.6296 - val_loss: 1.9473 - val_accuracy: 0.4169
ID:(13766, 0)