通过迁移学习，我正在尝试使用Google Colab在Keras中训练VGG16。以下是notebook中的代码：(注意:输出是作为注释编写的)
代码语言：javascript复制 from keras import Sequential
from keras.layers import Dense, Flatten
from keras.applications import vgg16
from keras.applications.vgg16 import preprocess_input as vgg_pi
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Model
base_model = vgg16.VGG16(include_top=False, weights='imagenet',
input_shape=(224, 224, 3))
for layer in base_model.layers:
layer.trainable = False
base_model.summary()
# Total params: 14,714,688
# Trainable params: 0
# Non-trainable params: 14,714,688
x = base_model.output
x = Flatten(name='flatten', input_shape=base_model.output_shape)(x)
x = Dense(10, activation='softmax', name='predictions')(x)
model = Model(inputs=base_model.input, outputs=x)
model.summary()
# Total params: 14,965,578
# Trainable params: 250,890
# Non-trainable params: 14,714,688
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
train_datagen = ImageDataGenerator(
rescale=1./255,
rotation_range=40,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.2,
fill_mode='nearest',
)
validation_datagen = ImageDataGenerator(
rescale=1./255,
)
train_generator = train_datagen.flow_from_directory(
'/content/drive/My Drive/Colab Notebooks/domat/solo-dataset/train/',
target_size=(224, 224),
batch_size=32,
class_mode='categorical',
)
validation_generator = validation_datagen.flow_from_directory(
'/content/drive/My Drive/Colab Notebooks/domat/solo-dataset/validation/',
target_size=(224, 224),
batch_size=32,
class_mode='categorical',
)
# Found 11614 images belonging to 10 classes.
# Found 2884 images belonging to 10 classes.
# check if GPU is running
import tensorflow as tf
device_name = tf.test.gpu_device_name()
if device_name != '/device:GPU:0':
raise SystemError('GPU device not found')
print('Found GPU at: {}'.format(device_name))
# Found GPU at: /device:GPU:0
t_steps = 11614 // 32
v_steps = 2884 // 32
history = model.fit_generator(train_generator,
epochs=500,
steps_per_epoch=t_steps,
validation_data=validation_generator,
validation_steps=v_steps,
)
# Epoch 1/500
# 8/362 [..............................] - ETA: 41:02 - loss: 2.9058 - acc: 0.2383所以，由于某些原因，一个时期需要大约40分钟，我真的不明白为什么它这么慢。
之前，我尝试了不同的参数(添加更多完全连接的层)，每个时期大约在3分钟内完成，尽管它显然是过度拟合的，因为有14mil的参数是免费的，并且数据集要小得多。
有谁知道如何处理这个问题吗？我已经尝试了上百万种方法，但都太慢了。我甚至无法返回到原始配置来查看我之前正在做的事情，以便每个时期在大约3分钟内完成。