TFLearn實現ConvNet對CIFAR-10分類

一、數據集

CIFAR-10 and CIFAR-100 datasets http://www.cs.toronto.edu/~kriz/cifar.html

下載CIFAR-10數據集，解壓放在.py同一目錄，如\cifar-10-batches-py\cifar-10-batches-py

或下載該資源解壓即可：https://download.csdn.net/download/lly1122334/10966600

二、代碼

# coding=utf-8
from __future__ import division, print_function, absolute_import
import tflearn
from tflearn.data_utils import shuffle, to_categorical
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.estimator import regression
from tflearn.data_preprocessing import ImagePreprocessing
from tflearn.data_augmentation import ImageAugmentation

from tflearn.datasets import cifar10

(X, Y), (X_test, Y_test) = cifar10.load_data()
X, Y = shuffle(X, Y)
Y = to_categorical(Y, 10)
Y_test = to_categorical(Y_test, 10)

# Real-time data preprocessing
img_prep = ImagePreprocessing()
img_prep.add_featurewise_zero_center()
img_prep.add_featurewise_stdnorm()

img_aug = ImageAugmentation()
img_aug.add_random_flip_leftright()
img_aug.add_random_rotation(max_angle=25.)

network = input_data(shape=[None, 32, 32, 3],
                     data_preprocessing=img_prep,
                     data_augmentation=img_aug)
network = conv_2d(network, 32, 3, activation='relu')
network = max_pool_2d(network, 2)
network = conv_2d(network, 64, 3, activation='relu')
network = conv_2d(network, 64, 3, activation='relu')
network = max_pool_2d(network, 2)
network = fully_connected(network, 512, activation='relu')
network = dropout(network, 0.5)
network = fully_connected(network, 10, activation='softmax')
network = regression(network, optimizer='adam',
                     loss='categorical_crossentropy',
                     learning_rate=0.001)

model = tflearn.DNN(network, tensorboard_verbose=0)
model.fit(X, Y, n_epoch=50, shuffle=True, validation_set=(X_test, Y_test),
          show_metric=True, batch_size=96, run_id='cifar10_cnn')

三、運行結果

50輪訓練後分類準確率達88.43%，訓練時間約爲139.226s×50÷60s≈116min