Tensorflow：更新不可训练模型层的权重

Question

我有一个使用 Keras 创建的经过训练的模型。在这个模型上，我想通过冻结除最后一个卷积层之外的所有层来应用迁移学习。但是，当我在冻结图层后拟合模型时，我注意到一些冻结的图层具有不同的权重。我怎样才能避免这种情况？

我尝试使用model.trainable = False 冻结整个模型，但这也没有成功。

我正在使用 python 3.5.0、tensorflow 1.0.1 和 Keras 2.0.3

---

示例脚本

import os
import timeit
import datetime
import numpy as np
from keras.layers.core import Activation, Reshape, Permute
from keras.layers.convolutional import Convolution2D, MaxPooling2D, UpSampling2D, ZeroPadding2D
from keras.layers.normalization import BatchNormalization
from keras.optimizers import Adam
from keras import models
from keras import backend as K
K.set_image_dim_ordering('th')

def conv_model(input_shape, data_shape, kern_size, filt_size, pad_size,\
                               maxpool_size, n_classes, compile_model=True):
    """
    Create a small conv neural network
    input_shape: input shape of the images
    data_shape: 1d shape of the data
    kern_size: Kernel size used in all convolutional2d layers
    filt_size: Filter size of the first and last convolutional2d layer
    pad_size: size of padding
    maxpool_size: Pool size of all maxpooling2d and upsampling2d layers
    n_classes: number of output classes
    compile_model: True if the model should be compiled

    output: Keras deep learning model
    """
    #keep track of compilation time
    start_time = timeit.default_timer()
    model = models.Sequential()
    # Add a noise layer to get a denoising autoencoder. This helps avoid overfitting
    model.add(ZeroPadding2D(padding=(pad_size, pad_size), input_shape=input_shape))

    #Encoding layers
    model.add(Convolution2D(filt_size, kern_size, kern_size, border_mode='valid'))
    model.add(BatchNormalization())
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(maxpool_size, maxpool_size)))
    model.add(UpSampling2D(size=(maxpool_size, maxpool_size)))
    model.add(ZeroPadding2D(padding=(pad_size, pad_size)))
    model.add(Convolution2D(filt_size, kern_size, kern_size, border_mode='valid'))
    model.add(BatchNormalization())
    model.add(Convolution2D(n_classes, 1, 1, border_mode='valid'))
    model.add(Reshape((n_classes, data_shape), input_shape=(n_classes,)+input_shape[1:]))
    model.add(Permute((2, 1)))
    model.add(Activation('softmax'))

    if compile_model:
        model.compile(loss="categorical_crossentropy", optimizer='adam', metrics=["accuracy"])
    print('Model compiled in {0} seconds'.format(datetime.timedelta(seconds=round(\
          timeit.default_timer() - start_time))))
    return model

if __name__ == '__main__':
    #Create some random training data
    train_data = np.random.randint(0, 10, 3*512*512*20, dtype='uint8').reshape(-1, 3, 512, 512)
    train_labels = np.random.randint(0, 1, 7*512*512*20, dtype='uint8').reshape(-1, 512*512, 7)
    #Get dims of the data
    data_dims = train_data.shape[2:]
    data_shape = np.prod(data_dims)
    #Create initial model
    initial_model = conv_model((train_data.shape[1], train_data.shape[2], train_data.shape[3]),\
                               data_shape, 3, 4, 1, 2, train_labels.shape[-1])
    #Train initial model on first part of the training data
    initial_model.fit(train_data[0:10], train_labels[0:10], verbose=2)
    #Store initial weights
    initial_weights = initial_model.get_weights()

    #Create transfer learning model
    transf_model = conv_model((train_data.shape[1], train_data.shape[2], train_data.shape[3]),\
                              data_shape, 3, 4, 1, 2, train_labels.shape[-1], False)
    #Set transfer model weights
    transf_model.set_weights(initial_weights)
    #Set all layers trainable to False (except final conv layer)
    for layer in transf_model.layers:
        layer.trainable = False
    transf_model.layers[9].trainable = True
    print(transf_model.layers[9])
    #Compile model
    transf_model.compile(loss="categorical_crossentropy", optimizer=Adam(lr=1e-4),\
                         metrics=["accuracy"])
    #Train model on second part of the data
    transf_model.fit(train_data[10:20], train_labels[10:20], verbose=2)
    #Store transfer model weights
    transf_weights = transf_model.get_weights()

    #Check where the weights have changed
    for i in range(len(initial_weights)):
        update_w = np.sum(initial_weights[i] != transf_weights[i])
        if update_w != 0:
            print(str(update_w)+' updated weights for layer '+str(transf_model.layers[i]))

Answer 1

编译模型后，您会丢失之前的权重，因为它们被重新采样。你需要先转移它们，将权重设置为不可训练，然后编译：

#Compile model
transf_model.set_weights(initial_weights)

#Set all layers trainable to False (except final conv layer)
for layer in transf_model.layers:
    layer.trainable = False

transf_model.layers[9].trainable = True

transf_model.compile(loss="categorical_crossentropy", optimizer=Adam(lr=1e-4),\
                     metrics=["accuracy"])

否则 - 重采样时权重会发生变化。

编辑：

模型应该在更改后编译 - 因为在编译期间keras 正在将所有可训练/不可训练的权重设置在一个不会进一步更改的列表中。

Answer 2

您应该将 Keras 升级到 Keras v2.1.3

这个问题刚刚解决，冻结 BatchNormalization 层的最后一个功能现在在最近的版本中可用：

BatchNormalization 中的 trainable 属性现在禁用批量统计信息的更新（即，如果 trainable == False，该层现在将在推理模式下 100% 运行）。

错误原因：

在之前的版本中，BatchNormalization 层的方差和均值参数无法设置untrainable，并且不起作用，虽然你坐了layer.trainable = False。

现在，它可以工作了！