¿Cómo implementar un clasificador simple de imágenes?
Clasificando Imágenes
Hoy en día se piensa que las redes neuronales son la cúspide de la complejidad en el Machine Learning. Sí, es verdad, todos los días salen investigaciones nuevas que demuestran la capacidad que estos sistemas tienen, pero lo que uno a veces no sabe es que nuestros computadores tienen capacidad suficiente para resolver algunos de estos problemas.
La idea de este tutorial es implementar una red convolucional simple que permita clasificar imágenes utilizando Keras, en este caso, se tienen imágenes de bicicletas, autos y aviones. Para ello se utilizarán las siguientes imágenes, las cuales se pueden descargar de aquí:
Rara vez todas las imágenes tendrán las mismas características. En este caso las imágenes son de distintos tamaños, colores, y además están en distintas direcciones. Es por eso que el preprocesamiento debe considerar escalar todas las imágenes a un tamaño común que permita suficiente capacidad de clasificación pero también que permita caber en memoria.
from tensorflow.keras.preprocessing.image import load_img, img_to_array, array_to_img
from tensorflow.image import resize
import glob
import numpy as np
import matplotlib.pyplot as plt
np.random.seed(1)
Datos de Entrenamiento
c_lista = glob.glob('fotos/c[1-5].jpg')
b_lista = glob.glob('fotos/b[1-5].jpg')
p_lista = glob.glob('fotos/p[1-5].jpg')
lista_train = c_lista + b_lista + p_lista
En este caso se utilizarán 15 imágenes para entrenamiento, es decir, las imágenes del 1 al 5 en cada clase.
Una vez se tiene claro el set de entrenamiento, es entonces necesario generar un preprocesamiento. En este caso, debido a que la característica principal para poder determinar cada clase es la forma, es que se utilizarán sólo imágenes en blanco y negro y se reescalarán a 100 $\times$ 100 pixeles.
Para ello se utilizan utility functions de tensorflow:
def import_pic(path, size = (100,100)):
return img_to_array(load_img(path,color_mode='grayscale', target_size = (100,100)))
fotos = np.array([import_pic(p) for p in lista_train])
Las imágenes importadas se ven así:
plt.figure(figsize=(10, 10))
for i in range(15):
ax = plt.subplot(3, 5, i + 1)
plt.imshow(fotos[i])
#plt.title(int(labels[i]))
plt.axis("off")
Luego, se define el vector objetivo, en este caso se trata de un vector multiclase, en el que la clase 1 corresponde a autos, la clase 2 será bicicletas y la clase 3 serán aviones.
y_train = np.repeat([[1,0,0],[0,1,0],[0,0,1]], [5,5,5], axis = 0)
Datos de Testing
Es importante que todos los procesos a los que se sometió a los datos de entrenamiento sea replicado de la misma manera para los datos de testing.
lista_test = glob.glob('fotos/c6.jpg')+glob.glob('fotos/b6.jpg')+glob.glob('fotos/p6.jpg')
fotos_test = np.array([import_pic(p) for p in lista_test])
y_test = np.array([[1,0,0],[0,1,0],[0,0,1]])
plt.figure(figsize=(10, 10))
for i in range(3):
ax = plt.subplot(3, 1, i + 1)
plt.imshow(fotos_test[i])
#plt.title(int(labels[i]))
plt.axis("off")
Implementación del Modelo en Tensorflow
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Conv2D, MaxPooling2D, Flatten, Dropout
En este caso se creará una red convolucional similar
a la arquitectura VGG-16.
model = Sequential(name = 'Convolutional_Network')
model.add(Conv2D(name = 'Conv_1', filters = 32, kernel_size=(3,3), activation='relu', input_shape = (100,100,1)))
model.add(Conv2D(name = 'Conv_2', filters=32, kernel_size=(3,3), activation='relu'))
model.add(MaxPooling2D(name = 'Pool1', pool_size = (2,2)))
model.add(Dropout(name = 'Drop_1', rate = 0.25))
model.add(Conv2D(name = 'Conv_3', filters=64, kernel_size=(3,3), activation='relu'))
model.add(Conv2D(name = 'Conv_4', filters = 64, kernel_size = (3,3), activation = 'relu'))
model.add(MaxPooling2D(name = 'Pool2', pool_size = (2,2)))
model.add(Dropout(name = 'Drop_2', rate = 0.25))
model.add(Flatten(name = 'Flat'))
model.add(Dense(name = 'Dense_1', units = 256, activation='relu'))
model.add(Dropout(name = 'Drop_3', rate = 0.25))
model.add(Dense(name = 'Output', units = 3, activation='softmax'))
model.compile(loss = 'categorical_crossentropy', optimizer = 'sgd', metrics = 'accuracy')
model.summary()
Model: "Convolutional_Network"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
Conv_1 (Conv2D) (None, 98, 98, 32) 320
_________________________________________________________________
Conv_2 (Conv2D) (None, 96, 96, 32) 9248
_________________________________________________________________
Pool1 (MaxPooling2D) (None, 48, 48, 32) 0
_________________________________________________________________
Drop_1 (Dropout) (None, 48, 48, 32) 0
_________________________________________________________________
Conv_3 (Conv2D) (None, 46, 46, 64) 18496
_________________________________________________________________
Conv_4 (Conv2D) (None, 44, 44, 64) 36928
_________________________________________________________________
Pool2 (MaxPooling2D) (None, 22, 22, 64) 0
_________________________________________________________________
Drop_2 (Dropout) (None, 22, 22, 64) 0
_________________________________________________________________
Flat (Flatten) (None, 30976) 0
_________________________________________________________________
Dense_1 (Dense) (None, 256) 7930112
_________________________________________________________________
Drop_3 (Dropout) (None, 256) 0
_________________________________________________________________
Output (Dense) (None, 3) 771
=================================================================
Total params: 7,995,875
Trainable params: 7,995,875
Non-trainable params: 0
_________________________________________________________________
En este caso la red cuenta con dos secciones convolucionales. Cada sección posee 32 y 64 filtros respectivamente y cuenta con dos capas convolucionales, un pooling y un dropout. Luego se pasa por una traansformación Flatten, para aplanar el outout y permitir el traspaso a una capa densa que permitirá la clasificación de cada imágen. En total la red requiere calcular cerca de 8 millones de parámetros.
Es importante recalcar que debido a que se trata de un problema multiclase es que la loss function a utilizar es la categorical crossentropy.
Finalmente se entrena el modelo, escalando el valor de cada pixel por 255, y se entrena por 200 epochs luego de ser mezclado.
%%time
model.fit(fotos/255,y_train, epochs = 200, shuffle = True)
Epoch 1/200
1/1 [==============================] - 0s 2ms/step - loss: 1.1311 - accuracy: 0.2000
Epoch 2/200
1/1 [==============================] - 0s 5ms/step - loss: 1.1029 - accuracy: 0.2667
Epoch 3/200
1/1 [==============================] - 0s 5ms/step - loss: 1.0923 - accuracy: 0.3333
Epoch 4/200
1/1 [==============================] - 0s 5ms/step - loss: 1.0671 - accuracy: 0.5333
Epoch 5/200
1/1 [==============================] - 0s 6ms/step - loss: 1.0561 - accuracy: 0.4000
Epoch 6/200
1/1 [==============================] - 0s 7ms/step - loss: 1.0438 - accuracy: 0.4667
Epoch 7/200
1/1 [==============================] - 0s 6ms/step - loss: 1.1003 - accuracy: 0.2667
Epoch 8/200
1/1 [==============================] - 0s 6ms/step - loss: 1.0388 - accuracy: 0.5333
Epoch 9/200
1/1 [==============================] - 0s 6ms/step - loss: 1.0468 - accuracy: 0.3333
Epoch 10/200
1/1 [==============================] - 0s 6ms/step - loss: 1.0816 - accuracy: 0.3333
Epoch 11/200
1/1 [==============================] - 0s 6ms/step - loss: 1.0738 - accuracy: 0.4000
Epoch 12/200
1/1 [==============================] - 0s 5ms/step - loss: 1.0323 - accuracy: 0.4667
Epoch 13/200
1/1 [==============================] - 0s 5ms/step - loss: 1.0177 - accuracy: 0.4667
Epoch 14/200
1/1 [==============================] - 0s 5ms/step - loss: 0.9835 - accuracy: 0.5333
Epoch 15/200
1/1 [==============================] - 0s 5ms/step - loss: 1.0549 - accuracy: 0.4000
Epoch 16/200
1/1 [==============================] - 0s 5ms/step - loss: 1.0462 - accuracy: 0.4667
Epoch 17/200
1/1 [==============================] - 0s 6ms/step - loss: 0.9802 - accuracy: 0.8000
Epoch 18/200
1/1 [==============================] - 0s 5ms/step - loss: 0.9999 - accuracy: 0.4667
Epoch 19/200
1/1 [==============================] - 0s 5ms/step - loss: 0.9641 - accuracy: 0.5333
Epoch 20/200
1/1 [==============================] - 0s 5ms/step - loss: 0.9304 - accuracy: 0.6000
Epoch 21/200
1/1 [==============================] - 0s 5ms/step - loss: 0.9563 - accuracy: 0.5333
Epoch 22/200
1/1 [==============================] - 0s 5ms/step - loss: 0.8624 - accuracy: 0.6667
Epoch 23/200
1/1 [==============================] - 0s 5ms/step - loss: 0.9080 - accuracy: 0.6000
Epoch 24/200
1/1 [==============================] - 0s 5ms/step - loss: 0.8809 - accuracy: 0.5333
Epoch 25/200
1/1 [==============================] - 0s 6ms/step - loss: 0.8548 - accuracy: 0.6667
Epoch 26/200
1/1 [==============================] - 0s 7ms/step - loss: 0.8035 - accuracy: 0.6667
Epoch 27/200
1/1 [==============================] - 0s 5ms/step - loss: 0.7307 - accuracy: 0.8000
Epoch 28/200
1/1 [==============================] - 0s 6ms/step - loss: 0.7353 - accuracy: 0.7333
Epoch 29/200
1/1 [==============================] - 0s 6ms/step - loss: 0.7121 - accuracy: 0.6667
Epoch 30/200
1/1 [==============================] - 0s 5ms/step - loss: 0.7371 - accuracy: 0.7333
Epoch 31/200
1/1 [==============================] - 0s 6ms/step - loss: 0.6615 - accuracy: 0.6667
Epoch 32/200
1/1 [==============================] - 0s 6ms/step - loss: 0.6892 - accuracy: 0.6667
Epoch 33/200
1/1 [==============================] - 0s 6ms/step - loss: 0.9558 - accuracy: 0.6000
Epoch 34/200
1/1 [==============================] - 0s 6ms/step - loss: 1.1499 - accuracy: 0.6000
Epoch 35/200
1/1 [==============================] - 0s 5ms/step - loss: 0.8030 - accuracy: 0.5333
Epoch 36/200
1/1 [==============================] - 0s 6ms/step - loss: 0.6960 - accuracy: 0.6667
Epoch 37/200
1/1 [==============================] - 0s 5ms/step - loss: 0.6028 - accuracy: 0.8667
Epoch 38/200
1/1 [==============================] - 0s 5ms/step - loss: 0.6218 - accuracy: 0.6667
Epoch 39/200
1/1 [==============================] - 0s 6ms/step - loss: 0.5688 - accuracy: 0.8667
Epoch 40/200
1/1 [==============================] - 0s 6ms/step - loss: 0.5789 - accuracy: 0.6667
Epoch 41/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4852 - accuracy: 0.8000
Epoch 42/200
1/1 [==============================] - 0s 6ms/step - loss: 0.6028 - accuracy: 0.6000
Epoch 43/200
1/1 [==============================] - 0s 6ms/step - loss: 0.8221 - accuracy: 0.6000
Epoch 44/200
1/1 [==============================] - 0s 5ms/step - loss: 1.0544 - accuracy: 0.6000
Epoch 45/200
1/1 [==============================] - 0s 5ms/step - loss: 0.7402 - accuracy: 0.6667
Epoch 46/200
1/1 [==============================] - 0s 5ms/step - loss: 0.6138 - accuracy: 0.6667
Epoch 47/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4751 - accuracy: 0.8000
Epoch 48/200
1/1 [==============================] - 0s 5ms/step - loss: 0.5481 - accuracy: 0.8000
Epoch 49/200
1/1 [==============================] - 0s 5ms/step - loss: 0.4765 - accuracy: 0.7333
Epoch 50/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4316 - accuracy: 0.8667
Epoch 51/200
1/1 [==============================] - 0s 6ms/step - loss: 0.3946 - accuracy: 0.8000
Epoch 52/200
1/1 [==============================] - 0s 6ms/step - loss: 0.5281 - accuracy: 0.6667
Epoch 53/200
1/1 [==============================] - 0s 5ms/step - loss: 0.6698 - accuracy: 0.6667
Epoch 54/200
1/1 [==============================] - 0s 5ms/step - loss: 0.8859 - accuracy: 0.6000
Epoch 55/200
1/1 [==============================] - 0s 6ms/step - loss: 0.5392 - accuracy: 0.6667
Epoch 56/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4646 - accuracy: 0.8667
Epoch 57/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4750 - accuracy: 0.8667
Epoch 58/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4122 - accuracy: 0.9333
Epoch 59/200
1/1 [==============================] - 0s 5ms/step - loss: 0.3788 - accuracy: 0.9333
Epoch 60/200
1/1 [==============================] - 0s 5ms/step - loss: 0.4162 - accuracy: 0.8000
Epoch 61/200
1/1 [==============================] - 0s 5ms/step - loss: 0.4748 - accuracy: 0.8000
Epoch 62/200
1/1 [==============================] - 0s 6ms/step - loss: 0.5568 - accuracy: 0.6667
Epoch 63/200
1/1 [==============================] - 0s 5ms/step - loss: 0.4309 - accuracy: 0.8000
Epoch 64/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4492 - accuracy: 0.7333
Epoch 65/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4759 - accuracy: 0.6667
Epoch 66/200
1/1 [==============================] - 0s 5ms/step - loss: 0.5108 - accuracy: 0.6667
Epoch 67/200
1/1 [==============================] - 0s 6ms/step - loss: 0.5261 - accuracy: 0.6667
Epoch 68/200
1/1 [==============================] - 0s 5ms/step - loss: 0.5230 - accuracy: 0.7333
Epoch 69/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4543 - accuracy: 0.8000
Epoch 70/200
1/1 [==============================] - 0s 5ms/step - loss: 0.3860 - accuracy: 0.7333
Epoch 71/200
1/1 [==============================] - 0s 5ms/step - loss: 0.3412 - accuracy: 0.8000
Epoch 72/200
1/1 [==============================] - 0s 5ms/step - loss: 0.3883 - accuracy: 0.8000
Epoch 73/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4140 - accuracy: 0.8000
Epoch 74/200
1/1 [==============================] - 0s 5ms/step - loss: 0.6120 - accuracy: 0.7333
Epoch 75/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4134 - accuracy: 0.8000
Epoch 76/200
1/1 [==============================] - 0s 6ms/step - loss: 0.3200 - accuracy: 0.9333
Epoch 77/200
1/1 [==============================] - 0s 5ms/step - loss: 0.3044 - accuracy: 0.9333
Epoch 78/200
1/1 [==============================] - 0s 5ms/step - loss: 0.2896 - accuracy: 0.8667
Epoch 79/200
1/1 [==============================] - 0s 5ms/step - loss: 0.2509 - accuracy: 1.0000
Epoch 80/200
1/1 [==============================] - 0s 6ms/step - loss: 0.3629 - accuracy: 0.8000
Epoch 81/200
1/1 [==============================] - 0s 5ms/step - loss: 0.4773 - accuracy: 0.8000
Epoch 82/200
1/1 [==============================] - 0s 5ms/step - loss: 0.6619 - accuracy: 0.7333
Epoch 83/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4416 - accuracy: 0.7333
Epoch 84/200
1/1 [==============================] - 0s 6ms/step - loss: 0.3173 - accuracy: 0.8667
Epoch 85/200
1/1 [==============================] - 0s 5ms/step - loss: 0.2936 - accuracy: 0.8667
Epoch 86/200
1/1 [==============================] - 0s 6ms/step - loss: 0.2473 - accuracy: 1.0000
Epoch 87/200
1/1 [==============================] - 0s 5ms/step - loss: 0.2481 - accuracy: 1.0000
Epoch 88/200
1/1 [==============================] - 0s 6ms/step - loss: 0.3151 - accuracy: 0.8000
Epoch 89/200
1/1 [==============================] - 0s 6ms/step - loss: 0.3888 - accuracy: 0.8000
Epoch 90/200
1/1 [==============================] - 0s 5ms/step - loss: 0.4922 - accuracy: 0.7333
Epoch 91/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4382 - accuracy: 0.7333
Epoch 92/200
1/1 [==============================] - 0s 6ms/step - loss: 0.3617 - accuracy: 0.7333
Epoch 93/200
1/1 [==============================] - 0s 5ms/step - loss: 0.2617 - accuracy: 0.9333
Epoch 94/200
1/1 [==============================] - 0s 6ms/step - loss: 0.2286 - accuracy: 0.9333
Epoch 95/200
1/1 [==============================] - 0s 6ms/step - loss: 0.2101 - accuracy: 1.0000
Epoch 96/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1811 - accuracy: 0.9333
Epoch 97/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1959 - accuracy: 1.0000
Epoch 98/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1525 - accuracy: 1.0000
Epoch 99/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1802 - accuracy: 0.9333
Epoch 100/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1608 - accuracy: 1.0000
Epoch 101/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1246 - accuracy: 1.0000
Epoch 102/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1201 - accuracy: 1.0000
Epoch 103/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0910 - accuracy: 1.0000
Epoch 104/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0957 - accuracy: 0.9333
Epoch 105/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1012 - accuracy: 1.0000
Epoch 106/200
1/1 [==============================] - 0s 6ms/step - loss: 0.3639 - accuracy: 0.8000
Epoch 107/200
1/1 [==============================] - 0s 6ms/step - loss: 1.1047 - accuracy: 0.7333
Epoch 108/200
1/1 [==============================] - 0s 6ms/step - loss: 0.5726 - accuracy: 0.6667
Epoch 109/200
1/1 [==============================] - 0s 5ms/step - loss: 0.3250 - accuracy: 0.8000
Epoch 110/200
1/1 [==============================] - 0s 6ms/step - loss: 0.2097 - accuracy: 1.0000
Epoch 111/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1716 - accuracy: 1.0000
Epoch 112/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1640 - accuracy: 1.0000
Epoch 113/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1720 - accuracy: 1.0000
Epoch 114/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0978 - accuracy: 1.0000
Epoch 115/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1654 - accuracy: 0.9333
Epoch 116/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1490 - accuracy: 0.9333
Epoch 117/200
1/1 [==============================] - 0s 6ms/step - loss: 0.2189 - accuracy: 0.8667
Epoch 118/200
1/1 [==============================] - 0s 5ms/step - loss: 0.4318 - accuracy: 0.7333
Epoch 119/200
1/1 [==============================] - 0s 6ms/step - loss: 0.5190 - accuracy: 0.7333
Epoch 120/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1658 - accuracy: 1.0000
Epoch 121/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1210 - accuracy: 1.0000
Epoch 122/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1245 - accuracy: 1.0000
Epoch 123/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1225 - accuracy: 1.0000
Epoch 124/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1460 - accuracy: 1.0000
Epoch 125/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1592 - accuracy: 0.9333
Epoch 126/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0912 - accuracy: 1.0000
Epoch 127/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0838 - accuracy: 1.0000
Epoch 128/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0996 - accuracy: 1.0000
Epoch 129/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0948 - accuracy: 1.0000
Epoch 130/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0813 - accuracy: 1.0000
Epoch 131/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1437 - accuracy: 0.9333
Epoch 132/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0994 - accuracy: 1.0000
Epoch 133/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1402 - accuracy: 0.9333
Epoch 134/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1606 - accuracy: 0.8667
Epoch 135/200
1/1 [==============================] - 0s 5ms/step - loss: 0.2365 - accuracy: 0.9333
Epoch 136/200
1/1 [==============================] - 0s 6ms/step - loss: 0.2089 - accuracy: 0.8667
Epoch 137/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1223 - accuracy: 0.9333
Epoch 138/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0547 - accuracy: 1.0000
Epoch 139/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0975 - accuracy: 0.9333
Epoch 140/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1617 - accuracy: 0.9333
Epoch 141/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0851 - accuracy: 1.0000
Epoch 142/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0541 - accuracy: 1.0000
Epoch 143/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0496 - accuracy: 1.0000
Epoch 144/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0681 - accuracy: 1.0000
Epoch 145/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0416 - accuracy: 1.0000
Epoch 146/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0324 - accuracy: 1.0000
Epoch 147/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0288 - accuracy: 1.0000
Epoch 148/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0322 - accuracy: 1.0000
Epoch 149/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0443 - accuracy: 1.0000
Epoch 150/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0460 - accuracy: 1.0000
Epoch 151/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0512 - accuracy: 1.0000
Epoch 152/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0512 - accuracy: 1.0000
Epoch 153/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0796 - accuracy: 1.0000
Epoch 154/200
1/1 [==============================] - 0s 5ms/step - loss: 0.1010 - accuracy: 0.9333
Epoch 155/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0719 - accuracy: 1.0000
Epoch 156/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0385 - accuracy: 1.0000
Epoch 157/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0267 - accuracy: 1.0000
Epoch 158/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0248 - accuracy: 1.0000
Epoch 159/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0322 - accuracy: 1.0000
Epoch 160/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0256 - accuracy: 1.0000
Epoch 161/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0241 - accuracy: 1.0000
Epoch 162/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0313 - accuracy: 1.0000
Epoch 163/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0368 - accuracy: 1.0000
Epoch 164/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0247 - accuracy: 1.0000
Epoch 165/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0239 - accuracy: 1.0000
Epoch 166/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0214 - accuracy: 1.0000
Epoch 167/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0275 - accuracy: 1.0000
Epoch 168/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0116 - accuracy: 1.0000
Epoch 169/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0380 - accuracy: 1.0000
Epoch 170/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0183 - accuracy: 1.0000
Epoch 171/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0164 - accuracy: 1.0000
Epoch 172/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0217 - accuracy: 1.0000
Epoch 173/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0081 - accuracy: 1.0000
Epoch 174/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0366 - accuracy: 1.0000
Epoch 175/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0361 - accuracy: 1.0000
Epoch 176/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0168 - accuracy: 1.0000
Epoch 177/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0282 - accuracy: 1.0000
Epoch 178/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0175 - accuracy: 1.0000
Epoch 179/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0064 - accuracy: 1.0000
Epoch 180/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0117 - accuracy: 1.0000
Epoch 181/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0202 - accuracy: 1.0000
Epoch 182/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0252 - accuracy: 1.0000
Epoch 183/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0097 - accuracy: 1.0000
Epoch 184/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0596 - accuracy: 0.9333
Epoch 185/200
1/1 [==============================] - 0s 5ms/step - loss: 0.2113 - accuracy: 0.9333
Epoch 186/200
1/1 [==============================] - 0s 6ms/step - loss: 1.6986 - accuracy: 0.6667
Epoch 187/200
1/1 [==============================] - 0s 6ms/step - loss: 1.2900 - accuracy: 0.6667
Epoch 188/200
1/1 [==============================] - 0s 5ms/step - loss: 0.7224 - accuracy: 0.6667
Epoch 189/200
1/1 [==============================] - 0s 6ms/step - loss: 0.4613 - accuracy: 0.7333
Epoch 190/200
1/1 [==============================] - 0s 6ms/step - loss: 0.3255 - accuracy: 0.8000
Epoch 191/200
1/1 [==============================] - 0s 6ms/step - loss: 0.2301 - accuracy: 0.9333
Epoch 192/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1540 - accuracy: 1.0000
Epoch 193/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1042 - accuracy: 1.0000
Epoch 194/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0850 - accuracy: 1.0000
Epoch 195/200
1/1 [==============================] - 0s 6ms/step - loss: 0.1039 - accuracy: 1.0000
Epoch 196/200
1/1 [==============================] - 0s 5ms/step - loss: 0.0898 - accuracy: 1.0000
Epoch 197/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0558 - accuracy: 1.0000
Epoch 198/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0598 - accuracy: 1.0000
Epoch 199/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0588 - accuracy: 1.0000
Epoch 200/200
1/1 [==============================] - 0s 6ms/step - loss: 0.0430 - accuracy: 1.0000
CPU times: user 3.21 s, sys: 414 ms, total: 3.63 s
Wall time: 3.51 s
En mi caso esto tomó 4 segundos, ya que tengo una RTX 2070 que paraleliza esto muy bien. Correr esto en CPU podría tomar un poquito más.
Inferencia
Luego del proceso de Entrenamiento, se predecirá los resultados utilizando el set de imágenes para testear:
model.predict(fotos_test)
array([[1., 0., 0.],
[0., 1., 0.],
[0., 0., 1.]], dtype=float32)
Es posible ver que la red aprendió rápidamente la tarea encomendada y predijo exitosamente todos los casos de testeo dados, primero un auto, segundo una bicicleta y tercero un avión.
Espero les haya gustado.