【机器学习课程-华盛顿大学】：1 案例研究 1.6 深度学习（2）深度学习特征作为输入训练逻辑回归分类

1、导入库和数据

import graphlab
# Limit number of worker processes. This preserves system memory, which prevents hosted notebooks from crashing.
graphlab.set_runtime_config('GRAPHLAB_DEFAULT_NUM_PYLAMBDA_WORKERS', 4)

image_train = graphlab.SFrame('image_train_data/')
image_test = graphlab.SFrame('image_test_data/')

graphlab.canvas.set_target('browser')
image_train['image'].show()

 

2、直接把图像像素点值作为输入，训练逻辑回归分类器

raw_pixel_model = graphlab.logistic_classifier.create(image_train,target='label',
                                              features=['image_array'])

前3张图像全部分类错误：

raw_pixel_model.evaluate(image_test)

分类准确率只有46%

 

3、采用深度学习后的特征作为输入，训练逻辑回归分类器

deep_learning_model = graphlab.load_model('http://s3.amazonaws.com/GraphLab-Datasets/deeplearning/imagenet_model_iter45')
image_train['deep_features'] = deep_learning_model.extract_features(image_train)
deep_features_model = graphlab.logistic_classifier.create(image_train,
                                                         features=['deep_features'],
                                                         target='label')

 

测试：

（1）前3张图像分类正确

image_test[0:3]['image'].show()
deep_features_model.predict(image_test[0:3])

 

（2）整体分类正确率提高到78%

deep_features_model.evaluate(image_test)