Deep Learning

Con Keras y TensorFlow

Creado por Roberto Saavedra, Alejandro Rayo y Sergio Rodriguez

IA, MACHINE LEARNING Y DEEP LEARNING

REDES NEURONALES

HANDS-ON DEEP LEARNING

¿DÓNDE ENCONTRAR UN COMPORTAMIENTO INTELIGENTE?


from tensorflow import keras

(imagenes_entrenamiento, etiquetas_entrenamiento), (imagenes_test, etiquetas_test) = keras.datasets.mnist.load_data()


					


red_neuronal = keras.models.Sequential([
    keras.layers.Flatten(input_shape=[28, 28]),
    keras.layers.Dense(100, activation="sigmoid"),
    keras.layers.Dense(10, activation="softmax")
])


					

red_neuronal.compile(loss='sparse_categorical_crossentropy',
              optimizer=keras.optimizers.SGD(lr=0.05),
              metrics=['accuracy'])

early_stopping = keras.callbacks.EarlyStopping()
red_neuronal.fit(imagenes_entrenamiento, etiquetas_entrenamiento,
epochs=10, validation_split=0.1, callbacks=[early_stopping])
					

						
Train on 54000 samples, validate on 6000 samples
Epoch 1/10
54000/54000 [==============================] - 4s 72us/sample
loss: 0.9310 - accuracy: 0.7128 - val_loss: 0.6660 - val_accuracy:
0.7867
Epoch 2/10
54000/54000 [==============================] - 3s 62us/sample
loss: 0.7030 - accuracy: 0.7799 - val_loss: 0.6760 - val_accuracy:
0.7887						

					

red_neuronal_2 = keras.models.Sequential([
    keras.layers.Flatten(input_shape=[28, 28]),
    keras.layers.BatchNormalization(),
    keras.layers.Dense(200, activation="selu",
	kernel_initializer="lecun_normal"),
    keras.layers.BatchNormalization(),
    keras.layers.Dense(100, activation="selu",
	kernel_initializer="lecun_normal"),
    keras.layers.BatchNormalization(),
    keras.layers.Dense(10, activation="softmax")
])

red_neuronal_2.compile(loss='sparse_categorical_crossentropy',
              optimizer=keras.optimizers.SGD(lr=0.05),
              metrics=['accuracy'])

red_neuronal_2.fit(imagenes_entrenamiento,
                   etiquetas_entrenamiento,epochs=10,
                   validation_split=0.1,
				   callbacks=[early_stopping])

REDES NEURONALES CONVOLUCIONALES


cnn = keras.models.Sequential([
        keras.layers.Conv2D(64, kernel_size=3, activation='relu', input_shape= (28,28, 1)),
        keras.layers.MaxPooling2D(),
        keras.layers.Flatten(),
        keras.layers.BatchNormalization(),
        keras.layers.Dense(100, activation="selu", kernel_initializer="lecun_normal"),
        keras.layers.BatchNormalization(),
        keras.layers.Dense(10, activation="softmax")
])

cnn.compile(optimizer=keras.optimizers.SGD(lr=0.05), loss='sparse_categorical_crossentropy',
metrics=['accuracy'])

imagenes_test = imagenes_test.reshape(10000, 28, 28, 1)
imagenes_entrenamiento = imagenes_entrenamiento.reshape(60000, 28, 28, 1)

cnn.fit(imagenes_entrenamiento, etiquetas_entrenamiento, validation_split=0.1,
epochs=30, callbacks=[early_stopping])						

					

prueba = 1 - (prueba/255)
prueba = tf.resize(prueba, [28, 28])

tf.predict_classes(prueba)
				

array([9, 8])

MODELOS PRE-ENTRENADOS

RESNET-50

+


from urllib.request import urlopen, Request
import matplotlib.pyplot as plt
import tensorflow as tf


headers = {'User-Agent': 'Mozilla/5.0 (Windows NT 6.1) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/41.0.2228.0 Safari/537.3'}
imagen_perro = plt.imread(urlopen(Request('https://cdn.pixabay.com/photo/2015/11/17/13/13/dogue-de-bordeaux-1047521_1280.jpg', headers=headers)), format='jpg')
castillo = plt.imread(urlopen(Request('https://www.audioguiaroma.com/imagenes/castillo-san-angelo.jpg', headers=headers)), format='jpg')

imagen_perro_crop = tf.image.resize_with_pad(imagen_perro, 224,224, antialias=True)
castillo_crop = tf.image.resize_with_pad(castillo, 224, 224,antialias=True)

imagenes = keras.applications.resnet50.preprocess_input(np.array([imagen_perro_crop, castillo_crop]))

neural_network_on_steroids = keras.applications.resnet50.ResNet50(weights="imagenet")
pred = neural_network_on_steroids.predict(imagenes)

Predicciones para perro Probabilidad Predicciones para castillo Probabilidad
vizsla 58.95% palace 45.57%
labrador_retriever 29.05% dome 23.81%
bull_mastif 1.68% castle 12.66%

3Blue1Brown: Deep Learning

Dot CSV: Aprendiendo Inteligencia Artificial

Standford Machine Learning taught by Andrew NG

Hands-on Machine Learning with Scikit-Learn, Keras, and TensorFlow

MIT Introduction to Deep Learning
😘