VGG16 迁移学习可变输出

Question

在使用 VGG16 进行迁移学习时观察到奇怪的行为。

model = VGG16(weights='imagenet',include_top=True)
model.layers.pop()
model.layers.pop()

for layer in model.layers:
    layer.trainable=False

new_layer = Dense(2,activation='softmax')
inp = model.input
out = new_layer(model.layers[-1].output)

model = Model(inp,out)

然而，当使用 model.predict(image) 时，输出在分类方面会有所不同，即有时它将图像分类为 Class 1，而下一次将同一图像分类为 Class 2.

Answer 1

那是因为你没有设置种子。试试这个

import numpy as np
seed_value = 0
np.random.seed(seed_value)

model = VGG16(weights='imagenet',include_top=True)
model.layers.pop()
model.layers.pop()

for layer in model.layers:
    layer.trainable=False

new_layer = Dense(2, activation='softmax',
                  kernel_initializer=keras.initializers.glorot_normal(seed=seed_value),
                  bias_initializer=keras.initializers.Zeros())
inp = model.input
out = new_layer(model.layers[-1].output)

model = Model(inp,out)

VGG16 迁移学习可变输出

VGG16 Transfer Learning varying output

python

keras

deep-learning

transfer-learning