最近在Coursera上学习迁移学习，使用的是TensorFlow.我们将采用迁移学习来使用Inception_v3来对Horses vs. Humans 进行二分类。代码如下
import
# Import all the necessary files!
import os
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras import Model
download Inception and load local weight
# Download the inception v3 weights
!wget --no-check-certificate \
https://storage.googleapis.com/mledu-datasets/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5 \
-O /tmp/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5
# Import the inception model
from tensorflow.keras.applications.inception_v3 import InceptionV3
# Create an instance of the inception model from the local pre-trained weights
local_weights_file = '/tmp/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5'
pre_trained_model = InceptionV3(input_shape=(150,150,3),
include_top=False,
weights=None)
pre_trained_model.load_weights(local_weights_file)
# Make all the layers in the pre-trained model non-trainable
for layer in pre_trained_model.layers:
layer.trainable=False
# Print the model summary
pre_trained_model.summary()
# Expected Output is extremely large, but should end with:
#batch_normalization_v1_281 (Bat (None, 3, 3, 192) 576 conv2d_281[0][0]
#__________________________________________________________________________________________________
#activation_273 (Activation) (None, 3, 3, 320) 0 batch_normalization_v1_273[0][0]
#__________________________________________________________________________________________________
#mixed9_1 (Concatenate) (None, 3, 3, 768) 0 activation_275[0][0]
# activation_276[0][0]
#__________________________________________________________________________________________________
#concatenate_5 (Concatenate) (None, 3, 3, 768) 0 activation_279[0][0]
# activation_280[0][0]
#__________________________________________________________________________________________________
#activation_281 (Activation) (None, 3, 3, 192) 0 batch_normalization_v1_281[0][0]
#__________________________________________________________________________________________________
#mixed10 (Concatenate) (None, 3, 3, 2048) 0 activation_273[0][0]
# mixed9_1[0][0]
# concatenate_5[0][0]
# activation_281[0][0]
#==================================================================================================
#Total params: 21,802,784
#Trainable params: 0
#Non-trainable params: 21,802,784
last_layer = pre_trained_model.get_layer('mixed7')
print('last layer output shape: ', last_layer.output_shape)
last_output = last_layer.output
# Expected Output:
# ('last layer output shape: ', (None, 7, 7, 768))
define myCallback
# Define a Callback class that stops training once accuracy reaches 99.9%
class myCallback(tf.keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs={}):
if(logs.get('acc')>0.999):
print("\nReached 99.9% accuracy so cancelling training!")
self.model.stop_training = True
NEW MODEL
from tensorflow.keras.optimizers import RMSprop
# Flatten the output layer to 1 dimension
x = layers.Flatten()(last_output)
# Add a fully connected layer with 1,024 hidden units and ReLU activation
x = layers.Dense(1024, activation='relu')(x)
# Add a dropout rate of 0.2
x = layers.Dropout(0.2)(x)
# Add a final sigmoid layer for classification
x = layers.Dense (1, activation='sigmoid')(x)
model = Model( pre_trained_model.input, x)
model.compile(optimizer = RMSprop(lr=0.0001),
loss = 'binary_crossentropy',
metrics = ['acc'])
model.summary()
# Expected output will be large. Last few lines should be:
# mixed7 (Concatenate) (None, 7, 7, 768) 0 activation_248[0][0]
# activation_251[0][0]
# activation_256[0][0]
# activation_257[0][0]
# __________________________________________________________________________________________________
# flatten_4 (Flatten) (None, 37632) 0 mixed7[0][0]
# __________________________________________________________________________________________________
# dense_8 (Dense) (None, 1024) 38536192 flatten_4[0][0]
# __________________________________________________________________________________________________
# dropout_4 (Dropout) (None, 1024) 0 dense_8[0][0]
# __________________________________________________________________________________________________
# dense_9 (Dense) (None, 1) 1025 dropout_4[0][0]
# ==================================================================================================
# Total params: 47,512,481
# Trainable params: 38,537,217
# Non-trainable params: 8,975,264
download dataset
# Get the Horse or Human dataset
!wget --no-check-certificate https://storage.googleapis.com/laurencemoroney-blog.appspot.com/horse-or-human.zip -O /tmp/horse-or-human.zip
# Get the Horse or Human Validation dataset
!wget --no-check-certificate https://storage.googleapis.com/laurencemoroney-blog.appspot.com/validation-horse-or-human.zip -O /tmp/validation-horse-or-human.zip
from tensorflow.keras.preprocessing.image import ImageDataGenerator
import os
import zipfile
local_zip = '//tmp/horse-or-human.zip'
zip_ref = zipfile.ZipFile(local_zip, 'r')
zip_ref.extractall('/tmp/training')
zip_ref.close()
local_zip = '//tmp/validation-horse-or-human.zip'
zip_ref = zipfile.ZipFile(local_zip, 'r')
zip_ref.extractall('/tmp/validation')
zip_ref.close()
train_horses_dir = os.path.join('/tmp/training','horses')
train_humans_dir = os.path.join('/tmp/training','humans')
validation_horses_dir = os.path.join('/tmp/validation','horses')
validation_humans_dir = os.path.join('/tmp/validation','humans')
train_horses_fnames = os.listdir(train_horses_dir)
train_humans_fnames = os.listdir(train_humans_dir)
validation_horses_fnames = os.listdir(validation_horses_dir)
validation_humans_fnames = os.listdir(validation_humans_dir)
print(len(train_horses_fnames))
print(len(train_humans_fnames))
print(len(validation_horses_fnames))
print(len(validation_humans_fnames))
# Expected Output:
# 500
# 527
# 128
# 128
# Define our example directories and files
train_dir = '/tmp/training'
validation_dir = '/tmp/validation'
# Add our data-augmentation parameters to ImageDataGenerator
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,
horizontal_flip = True)
# Note that the validation data should not be augmented!
test_datagen = ImageDataGenerator(rescale= 1./255.)
# Flow training images in batches of 20 using train_datagen generator
train_generator = train_datagen.flow_from_directory(train_dir,
batch_size=20,
class_mode = 'binary',
target_size = (150,150))
# Flow validation images in batches of 20 using test_datagen generator
validation_generator = test_datagen.flow_from_directory(validation_dir,
batch_size=20,
class_mode='binary',
target_size = (150,150))
# Expected Output:
# Found 1027 images belonging to 2 classes.
# Found 256 images belonging to 2 classes.
train
# Run this and see how many epochs it should take before the callback
# fires, and stops training at 99.9% accuracy
# (It should take less than 100 epochs)
callbacks = myCallback()
history = model.fit_generator(train_generator,
validation_data= validation_generator,
steps_per_epoch=100,
epochs=100,
validation_steps=50,
verbose=2,
callbacks=[callbacks])
plot
import matplotlib.pyplot as plt
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(len(acc))
plt.plot(epochs, acc, 'r', label='Training accuracy')
plt.plot(epochs, val_acc, 'b', label='Validation accuracy')
plt.title('Training and validation accuracy')
plt.legend(loc=0)
plt.figure()
plt.show()
在学习的过程中我发现TensorFlow官方提供的教程中，产生迁移学习的新模型的方式与Coursera课程中的不一样，官方采用了Sequential的方法，而Coursera课程采用了Model的方法。
model = tf.keras.Sequential([
base_model,
keras.layers.GlobalAveragePooling2D(),
keras.layers.Dense(1, activation='sigmoid')
])
Sequential的方法我们已经很熟悉了，但在这里Model方法似乎略胜一筹。因为Model方法与pre_trained_model.get_layer的配合使用，能让我们使用迁移来的网络中想要的那一层的输出。
但在官方教程中有一处微调的手法值得注意
All we need to do is unfreeze the base_model, and set the bottom layers be un-trainable. Then, recompile the model (necessary for these changes to take effect), and resume training.
base_model.trainable = True
# Let's take a look to see how many layers are in the base model
print("Number of layers in the base model: ", len(base_model.layers))
# Fine tune from this layer onwards
fine_tune_at = 100
# Freeze all the layers before the `fine_tune_at` layer
for layer in base_model.layers[:fine_tune_at]:
layer.trainable = False
