在预训练的 CNN 中提取特征的有效方法

Question

有没有更有效的方法从数据集中提取特征然后如下：

def extract_features(directory, sample_count):
    features = np.zeros(shape=(sample_count, 6, 6, 512))
    labels = np.zeros(shape=(sample_count, 6))

    generator = 
    ImageDataGenerator(rescale=1./255).flow_from_directory(directory, 
    target_size=(Image_Size, Image_Size), batch_size = batch_size, 
    class_mode='categorical')

    i = 0

    print('Entering for loop...');

    for inputs_batch, labels_batch in generator:
        features_batch = conv_base.predict(inputs_batch)
        features[i * 20 : (i + 1) * 20] = features_batch
        labels[i * 20 : (i + 1) * 20] = labels_batch
        i += 1
        print(i);

        if (i * 20) >= sample_count:
            break

    return features, labels

由于我的数据集的大小，这个过程需要相当长的时间，我想知道是否有更好的方法来做到这一点？

提前致谢:)

完整代码：

from keras import layers
from keras import models
from keras import losses
from keras import optimizers
from keras.preprocessing.image import ImageDataGenerator
from keras.applications.vgg16 import VGG16

import matplotlib.pyplot as plt
import numpy as np

Train_DIR = '/Users/eoind/food/train'
Test_DIR = '/Users/eoind/food/test'
Validation_DIR = '/Users/eoind/food/validation'

Image_Size = 200 # Size of input images to be scaled to 
Train_Samples = 6000
Validation_Samples = 3000
Test_Samples = 3000

num_epochs = 30
batch_size = 20
steps_per_epoch = Train_Samples/batch_size

conv_base = VGG16(weights='imagenet', include_top=False, input_shape= 
(Image_Size, Image_Size, 3))

conv_base.summary()

print('Conv_Base Summary');

def extract_features(directory, sample_count):
    features = np.zeros(shape=(sample_count, 6, 6, 512))
    labels = np.zeros(shape=(sample_count, 6))

    generator = 
    ImageDataGenerator(rescale=1./255).flow_from_directory(directory, 
    target_size=(Image_Size, Image_Size), batch_size = batch_size, 
    class_mode='categorical')

    i = 0

    print('Entering for loop...');

    for inputs_batch, labels_batch in generator:
        features_batch = conv_base.predict(inputs_batch)
        features[i * 20 : (i + 1) * 20] = features_batch
        labels[i * 20 : (i + 1) * 20] = labels_batch
        i += 1
        print(i);
    
        if (i * 20) >= sample_count:
            break
    
    return features, labels

train_features, train_labels = extract_features(Train_DIR, Train_Samples)
validation_features, validation_labels = extract_features(Validation_DIR, 
Validation_Samples)
test_features, test_labels = extract_features(Test_DIR, Test_Samples)

print('Extracting Features');

train_features = np.reshape(train_features, (Train_Samples, 6 * 6 * 512))
validation_features = np.reshape(validation_features, (Validation_Samples, 6 * 
6 * 512))
test_features = np.reshape(test_features, (Test_Samples, 6 * 6 * 512))

print('Reshaping Features');

model = models.Sequential()
model.add(layers.Dense(256, activation='relu', input_dim=6 * 6 * 512))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(1, activation='sigmoid'))

model.summary()

print('Model Summary');

model.compile(optimizer=optimizers.RMSprop(lr=1e-4),
              loss=losses.categorical_crossentropy,
              metrics=['acc'])

print('Compiling Model');

hist = model.fit(train_features, train_labels,
                 steps_per_epoch = steps_per_epoch,
                 epochs = num_epochs,
                 batch_size = batch_size,
                 verbose = 1,
                 validation_data = (validation_features, validation_labels))

print('Fitting Model');

train_loss=hist.history['loss']
val_loss=hist.history['val_loss']
train_acc=hist.history['acc']
val_acc=hist.history['val_acc']
xc=range(num_epochs)

fig1=plt.figure(1,figsize=(7,5))
plt.plot(xc,train_loss)
plt.plot(xc,val_loss)
plt.xlabel('Number of Epochs')
plt.ylabel('Loss')
plt.title('Training Loss Vs. Validation Loss')
plt.grid(True)
plt.legend(['Training', 'Validation'])
plt.style.use(['classic'])
fig1.savefig('loss.png')

fig2=plt.figure(2,figsize=(7,5))
plt.plot(xc,train_acc)
plt.plot(xc,val_acc)
plt.xlabel('Number of Epochs')
plt.ylabel('Accuracy')
plt.title('Training Accuracy Vs. Validation Accuracy')
plt.grid(True)
plt.legend(['Training', 'Validation'], loc='upper left')
plt.style.use(['classic'])
fig2.savefig('acc.png')

model.save('food_pretrained.h5') # Save model

iPython 控制台输出

Layer (type)                 Output Shape              Param #   
=================================================================
input_19 (InputLayer)        (None, 200, 200, 3)       0         
_________________________________________________________________
block1_conv1 (Conv2D)        (None, 200, 200, 64)      1792      
_________________________________________________________________
block1_conv2 (Conv2D)        (None, 200, 200, 64)      36928     
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, 100, 100, 64)      0         
_________________________________________________________________
block2_conv1 (Conv2D)        (None, 100, 100, 128)     73856     
_________________________________________________________________
block2_conv2 (Conv2D)        (None, 100, 100, 128)     147584    
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, 50, 50, 128)       0         
_________________________________________________________________
block3_conv1 (Conv2D)        (None, 50, 50, 256)       295168    
_________________________________________________________________
block3_conv2 (Conv2D)        (None, 50, 50, 256)       590080    
_________________________________________________________________
block3_conv3 (Conv2D)        (None, 50, 50, 256)       590080    
_________________________________________________________________
block3_pool (MaxPooling2D)   (None, 25, 25, 256)       0         
_________________________________________________________________
block4_conv1 (Conv2D)        (None, 25, 25, 512)       1180160   
_________________________________________________________________
block4_conv2 (Conv2D)        (None, 25, 25, 512)       2359808   
_________________________________________________________________
block4_conv3 (Conv2D)        (None, 25, 25, 512)       2359808   
_________________________________________________________________
block4_pool (MaxPooling2D)   (None, 12, 12, 512)       0         
_________________________________________________________________
block5_conv1 (Conv2D)        (None, 12, 12, 512)       2359808   
_________________________________________________________________
block5_conv2 (Conv2D)        (None, 12, 12, 512)       2359808   
_________________________________________________________________
block5_conv3 (Conv2D)        (None, 12, 12, 512)       2359808   
_________________________________________________________________
block5_pool (MaxPooling2D)   (None, 6, 6, 512)         0         
=================================================================
Total params: 14,714,688
Trainable params: 14,714,688
Non-trainable params: 0
_________________________________________________________________
Conv_Base Summary
Found 6000 images belonging to 6 classes.
Entering for loop...
1
2
3
4
5
6
7
8
9
10
11
12...

Answer 1

我要说不，我认为没有更有效的方法。必须进行计算，尤其是在 CPU 上它们很慢。 您可以做的最好的事情是通过转换数据集、保存数组并在每次训练时加载它们来避免重新计算。

在我崩溃并买了一个 GPU 之前，我一直在经历同样的事情……从那以后我的生活压力就小多了。我强烈建议您进行投资，即使您买不起其他任何东西，也可以选择 1050。您可能需要弄清楚如何处理有限的 GPU 内存，但这会让您的事情变得更加顺利。