keras - keras 逐层复制 VGG16 预训练权重

Question

我想将一些 VGG16 层权重逐层复制到另一个具有相似层的小型网络，但我收到一条错误消息：

  File "/home/d/Desktop/s/copyweights.py", line 78, in <module>
    list(f["model_weights"].keys())
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "/usr/local/lib/python3.5/dist-packages/h5py/_hl/group.py", line 262, in __getitem__
    oid = h5o.open(self.id, self._e(name), lapl=self._lapl)
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "h5py/h5o.pyx", line 190, in h5py.h5o.open
KeyError: "Unable to open object (object 'model_weights' doesn't exist)"

该文件肯定在路径中，我再次下载该文件只是为了确保它没有损坏（并且当我使用 model.load_weights 通常不会导致错误 - 我也安装了 HDF5

这是代码：

from keras import applications
from keras.preprocessing.image import ImageDataGenerator
from keras import optimizers
from keras.models import Sequential, Model
from keras.layers import Dropout, Flatten, Dense, GlobalAveragePooling2D
from keras import backend as k
from keras.callbacks import ModelCheckpoint, LearningRateScheduler, TensorBoard, EarlyStopping
from keras import layers
from keras import models
from keras import optimizers
from keras.layers import Dropout
from keras.regularizers import l2
from keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau
import matplotlib.pyplot as plt
from keras.preprocessing.image import ImageDataGenerator
import os

epochs = 50
callbacks = []
#schedule = None
decay = 0.0

earlyStopping = EarlyStopping(monitor='val_loss', patience=10, verbose=0, mode='min')
mcp_save = ModelCheckpoint('.mdl_wts.hdf5', save_best_only=True, monitor='val_loss', mode='min')
reduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1, epsilon=1e-5, mode='min')

base_model = models.Sequential()
base_model.add(layers.Conv2D(64, (3, 3), activation='relu', name='block1_conv1',  input_shape=(256, 256, 3)))
base_model.add(layers.Conv2D(64, (3, 3), activation='relu', name='block1_conv2'))
base_model.add(layers.MaxPooling2D((2, 2)))
#model.add(Dropout(0.2))
base_model.add(layers.Conv2D(128, (3, 3), activation='relu', name='block2_conv1'))
base_model.add(layers.Conv2D(128, (3, 3), activation='relu', name='block2_conv2'))
base_model.add(layers.MaxPooling2D((2, 2), name='block2_pool'))
#model.add(Dropout(0.2))

base_model.summary()
"""
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 256, 256, 3)       0         
_________________________________________________________________
block1_conv1 (Conv2D)        (None, 256, 256, 64)      1792      
_________________________________________________________________
block1_conv2 (Conv2D)        (None, 256, 256, 64)      36928     
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, 128, 128, 64)      0         
_________________________________________________________________
block2_conv1 (Conv2D)        (None, 128, 128, 128)     73856     
_________________________________________________________________
block2_conv2 (Conv2D)        (None, 128, 128, 128)     147584    
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, 64, 64, 128)       0         
=================================================================
Total params: 260,160.0
Trainable params: 260,160.0
Non-trainable params: 0.0
"""

layer_dict = dict([(layer.name, layer) for layer in base_model.layers])
[layer.name for layer in base_model.layers]
"""
['input_1',
 'block1_conv1',
 'block1_conv2',
 'block1_pool',
 'block2_conv1',
 'block2_conv2',
 'block2_pool']
"""

import h5py
weights_path = '/home/d/Desktop/s/vgg16_weights_new.h5' # ('https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg19_weights_tf_dim_ordering_tf_kernels.h5)
f = h5py.File(weights_path)

list(f["model_weights"].keys())
"""
['block1_conv1',
 'block1_conv2',
 'block1_pool',
 'block2_conv1',
 'block2_conv2',
 'block2_pool',
 'block3_conv1',
 'block3_conv2',
 'block3_conv3',
 'block3_conv4',
 'block3_pool',
 'block4_conv1',
 'block4_conv2',
 'block4_conv3',
 'block4_conv4',
 'block4_pool',
 'block5_conv1',
 'block5_conv2',
 'block5_conv3',
 'block5_conv4',
 'block5_pool',
 'dense_1',
 'dense_2',
 'dense_3',
 'dropout_1',
 'global_average_pooling2d_1',
 'input_1']
"""

# list all the layer names which are in the model.
layer_names = [layer.name for layer in base_model.layers]


"""
# Here we are extracting model_weights for each and every layer from the .h5 file
>>> f["model_weights"]["block1_conv1"].attrs["weight_names"]
array([b'block1_conv1/kernel:0', b'block1_conv1/bias:0'], 
      dtype='|S21')
# we are assiging this array to weight_names below 
>>> f["model_weights"]["block1_conv1"]["block1_conv1/kernel:0]
<HDF5 dataset "kernel:0": shape (3, 3, 3, 64), type "<f4">
# The list comprehension (weights) stores these two weights and bias of both the layers 
>>>layer_names.index("block1_conv1")
1
>>> model.layers[1].set_weights(weights)
# This will set the weights for that particular layer.
With a for loop we can set_weights for the entire network.
"""
for i in layer_dict.keys():
    weight_names = f["model_weights"][i].attrs["weight_names"]
    weights = [f["model_weights"][i][j] for j in weight_names]
    index = layer_names.index(i)
    base_model.layers[index].set_weights(weights)

base_model.add(layers.Flatten())
base_model.add(layers.Dropout(0.5))  #Dropout for regularization
base_model.add(layers.Dense(256, activation='relu'))
base_model.add(layers.Dense(1, activation='sigmoid'))  #Sigmoid function at the end because we have just two classes


# compile the model with a SGD/momentum optimizer
# and a very slow learning rate.
base_model.compile(loss='binary_crossentropy',
                   optimizer=optimizers.Adam(lr=1e-4, decay=decay),
                   metrics=['accuracy'])



os.environ["CUDA_VISIBLE_DEVICES"]="0"

train_dir = '/home/d/Desktop/s/data/train'
eval_dir = '/home/d/Desktop/s/data/eval'
test_dir = '/home/d/Desktop/s/data/test'


# create a data generator
train_datagen = ImageDataGenerator(rescale=1./255,   #Scale the image between 0 and 1
                                    rotation_range=40,
                                    width_shift_range=0.2,
                                    height_shift_range=0.2,
                                    shear_range=0.2,
                                    zoom_range=0.2,
                                    horizontal_flip=True,)

val_datagen = ImageDataGenerator(rescale=1./255)  #We do not augment validation data. we only perform rescale

test_datagen = ImageDataGenerator(rescale=1./255)  #We do not augment validation data. we only perform rescale

# load and iterate training dataset
train_generator = train_datagen.flow_from_directory(train_dir,  target_size=(224,224),class_mode='binary', batch_size=16, shuffle='True', seed=42)
# load and iterate validation dataset
val_generator = val_datagen.flow_from_directory(eval_dir,  target_size=(224,224),class_mode='binary', batch_size=16, shuffle='True', seed=42)
# load and iterate test dataset
test_generator = test_datagen.flow_from_directory(test_dir,  target_size=(224,224), class_mode=None, batch_size=1, shuffle='False', seed=42)

#The training part
#We train for 64 epochs with about 100 steps per epoch
history = base_model.fit_generator(train_generator,
                              steps_per_epoch=train_generator.n // train_generator.batch_size,
                              epochs=epochs,
                              validation_data=val_generator,
                              validation_steps=val_generator.n // val_generator.batch_size,
                              callbacks=[earlyStopping, mcp_save, reduce_lr_loss])

#Save the model
#base_model.save_weights('/home/d/Desktop/s/base_model_weights.h5')
#base_model.save('/home/d/Desktop/s/base_model_keras.h5')

#lets plot the train and val curve
#get the details form the history object
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(1, len(acc) + 1)

#Train and validation accuracy
plt.plot(epochs, acc, 'b', label='Training accuracy')
plt.plot(epochs, val_acc, 'r', label='Validation accuracy')
plt.title('Training and Validation accurarcy')
plt.legend()

plt.figure()
#Train and validation loss
plt.plot(epochs, loss, 'b', label='Training loss')
plt.plot(epochs, val_loss, 'r', label='Validation loss')
plt.title('Training and Validation loss')
plt.legend()

plt.show()

标签： kerash5pypre-trained-modeltransfer-learningvgg-net

Answer 1

直接取VGG16模型：

from keras.applications import VGG16
vgg = VGG16(choose parameters)

for layer in vgg.layers:
    layer_weights_list = layer.get_weights()