數(shù)據(jù)科學家Prakash Jay介紹了遷移學習的原理，基于Keras實現(xiàn)遷移學習，以及遷移學習的常見情形。

Inception-V3

什么是遷移學習？

機器學習中的遷移學習問題，關注如何保存解決一個問題時獲得的知識，并將其應用于另一個相關的不同問題。

為什么遷移學習？

在實踐中，很少有人從頭訓練一個卷積網(wǎng)絡，因為很難獲取足夠的數(shù)據(jù)集。使用預訓練的網(wǎng)絡有助于解決大多數(shù)手頭的問題。

訓練深度網(wǎng)絡代價高昂。即使使用數(shù)百臺配備了昂貴的GPU的機器，訓練最復雜的模型也需要好多周。

決定深度學習的拓撲/特色/訓練方法/超參數(shù)是沒有多少理論指導的黑魔法。

我的經(jīng)驗

不要試圖成為英雄。

—— Andrej Karapathy

我面對的大多數(shù)計算機視覺問題沒有非常大的數(shù)據(jù)集（5000-40000圖像）。即使使用極端的數(shù)據(jù)增強策略，也很難達到像樣的精確度。而在少量數(shù)據(jù)集上訓練數(shù)百萬參數(shù)的網(wǎng)絡通常會導致過擬合。所以遷移學習是我的救星。

遷移學習為何有效？

讓我們看下深度學習網(wǎng)絡學習了什么，靠前的層嘗試檢測邊緣，中間層嘗試檢測形狀，而靠后的層嘗試檢測高層數(shù)據(jù)特征。這些訓練好的網(wǎng)絡通常有助于解決其他計算機視覺問題。

下面，讓我們看下如何使用Keras實現(xiàn)遷移學習，以及遷移學習的常見情形。

基于Keras的簡單實現(xiàn)

from keras import applications

from keras.preprocessing.image importImageDataGenerator

from keras import optimizers

from keras.models importSequential, Model

from keras.layers importDropout, Flatten, Dense, GlobalAveragePooling2D

from keras import backend as k

from keras.callbacks importModelCheckpoint, LearningRateScheduler, TensorBoard, EarlyStopping

img_width, img_height = 256, 256

train_data_dir = "data/train"

validation_data_dir = "data/val"

nb_train_samples = 4125

nb_validation_samples = 466

batch_size = 16

epochs = 50

model = applications.VGG19(weights = "imagenet", include_top=False, input_shape = (img_width, img_height, 3))

"""

層 (類型) 輸出形狀 參數(shù)數(shù)量

=================================================================

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

_________________________________________________________________

block3_conv1 (Conv2D) (None, 64, 64, 256) 295168

_________________________________________________________________

block3_conv2 (Conv2D) (None, 64, 64, 256) 590080

_________________________________________________________________

block3_conv3 (Conv2D) (None, 64, 64, 256) 590080

_________________________________________________________________

block3_conv4 (Conv2D) (None, 64, 64, 256) 590080

_________________________________________________________________

block3_pool (MaxPooling2D) (None, 32, 32, 256) 0

_________________________________________________________________

block4_conv1 (Conv2D) (None, 32, 32, 512) 1180160

_________________________________________________________________

block4_conv2 (Conv2D) (None, 32, 32, 512) 2359808

_________________________________________________________________

block4_conv3 (Conv2D) (None, 32, 32, 512) 2359808

_________________________________________________________________

block4_conv4 (Conv2D) (None, 32, 32, 512) 2359808

_________________________________________________________________

block4_pool (MaxPooling2D) (None, 16, 16, 512) 0

_________________________________________________________________

block5_conv1 (Conv2D) (None, 16, 16, 512) 2359808

_________________________________________________________________

block5_conv2 (Conv2D) (None, 16, 16, 512) 2359808

_________________________________________________________________

block5_conv3 (Conv2D) (None, 16, 16, 512) 2359808

_________________________________________________________________

block5_conv4 (Conv2D) (None, 16, 16, 512) 2359808

_________________________________________________________________

block5_pool (MaxPooling2D) (None, 8, 8, 512) 0

=================================================================

總參數(shù)： 20,024,384.0

可訓練參數(shù)： 20,024,384.0

不可訓練參數(shù)： 0.0

"""

# 凍結不打算訓練的層。這里我凍結了前5層。

for layer in model.layers[:5]:

layer.trainable = False

# 增加定制層

x = model.output

x = Flatten()(x)

x = Dense(1024, activation="relu")(x)

x = Dropout(0.5)(x)

x = Dense(1024, activation="relu")(x)

predictions = Dense(16, activation="softmax")(x)

# 創(chuàng)建最終模型

model_final = Model(input = model.input, output = predictions)

# 編譯最終模型

model_final.compile(loss = "categorical_crossentropy", optimizer = optimizers.SGD(lr=0.0001, momentum=0.9), metrics=["accuracy"])

# 數(shù)據(jù)增強

train_datagen = ImageDataGenerator(

rescale = 1./255,

horizontal_flip = True,

fill_mode = "nearest",

zoom_range = 0.3,

width_shift_range = 0.3,

height_shift_range=0.3,

rotation_range=30)

test_datagen = ImageDataGenerator(

rescale = 1./255,

horizontal_flip = True,

fill_mode = "nearest",

zoom_range = 0.3,

width_shift_range = 0.3,

height_shift_range=0.3,

rotation_range=30)

train_generator = train_datagen.flow_from_directory(

train_data_dir,

target_size = (img_height, img_width),

batch_size = batch_size,

class_mode = "categorical")

validation_generator = test_datagen.flow_from_directory(

validation_data_dir,

target_size = (img_height, img_width),

class_mode = "categorical")

# 保存模型

checkpoint = ModelCheckpoint("vgg16_1.h5", monitor='val_acc', verbose=1, save_best_only=True, save_weights_only=False, mode='auto', period=1)

early = EarlyStopping(monitor='val_acc', min_delta=0, patience=10, verbose=1, mode='auto')

# 訓練模型

model_final.fit_generator(

train_generator,

samples_per_epoch = nb_train_samples,

epochs = epochs,

validation_data = validation_generator,

nb_val_samples = nb_validation_samples,

callbacks = [checkpoint, early])

遷移學習的常見情形

別忘了，靠前的層中的卷積特征更通用，靠后的層中的卷積特征更針對原本的數(shù)據(jù)集。遷移學習有4種主要場景：

1. 新數(shù)據(jù)集較小，和原數(shù)據(jù)集相似

如果我們嘗試訓練整個網(wǎng)絡，容易導致過擬合。由于新數(shù)據(jù)和原數(shù)據(jù)相似，因此我們期望卷積網(wǎng)絡中的高層特征和新數(shù)據(jù)集相關。因此，建議凍結所有卷積層，只訓練分類器（比如，線性分類器）：

for layer in model.layers:

layer.trainable = False

2. 新數(shù)據(jù)集較大，和原數(shù)據(jù)集相似

由于我們有更多數(shù)據(jù)，我們更有自信，如果嘗試對整個網(wǎng)絡進行精細調(diào)整，不會導致過擬合。

for layer in model.layers:

layer.trainable = True

其實默認值就是True，上面的代碼明確指定所有層可訓練，是為了更清楚地強調(diào)這一點。

由于開始的幾層檢測邊緣，你也可以選擇凍結這些層。比如，以下代碼凍結VGG19的前5層：

for layer in model.layers[:5]:

layer.trainable = False

3. 新數(shù)據(jù)集很小，但和原數(shù)據(jù)很不一樣

由于數(shù)據(jù)集很小，我們大概想要從靠前的層提取特征，然后在此之上訓練一個分類器：（假定你對h5py有所了解）

from keras import applications

from keras.preprocessing.image importImageDataGenerator

from keras import optimizers

from keras.models importSequential, Model

from keras.layers importDropout, Flatten, Dense, GlobalAveragePooling2D

from keras import backend as k

from keras.callbacks importModelCheckpoint, LearningRateScheduler, TensorBoard, EarlyStopping

img_width, img_height = 256, 256

### 創(chuàng)建網(wǎng)絡

img_input = Input(shape=(256, 256, 3))

x = Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv1')(img_input)

x = Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv2')(x)

x = MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x)

# 塊2

x = Conv2D(128, (3, 3), activation='relu', padding='same', name='block2_conv1')(x)

x = Conv2D(128, (3, 3), activation='relu', padding='same', name='block2_conv2')(x)

x = MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x)

model = Model(input = img_input, output = x)

model.summary()

"""

_________________________________________________________________

層 (類型) 輸出形狀 參數(shù)數(shù)量

=================================================================

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

=================================================================

總參數(shù)：260,160.0

可訓練參數(shù)：260,160.0

不可訓練參數(shù)：0.0

"""

layer_dict = dict([(layer.name, layer) for layer in model.layers])

[layer.name for layer in model.layers]

"""

['input_1',

'block1_conv1',

'block1_conv2',

'block1_pool',

'block2_conv1',

'block2_conv2',

'block2_pool']

"""

import h5py

weights_path = 'vgg19_weights.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']

"""

# 列出模型中的所有層的名稱

layer_names = [layer.name for layer in model.layers]

"""

# 提取`.h5`文件中每層的模型權重

>>> f["model_weights"]["block1_conv1"].attrs["weight_names"]

array([b'block1_conv1/kernel:0', b'block1_conv1/bias:0'],

dtype='|S21')

# 將這一數(shù)組分配給weight_names

>>> f["model_weights"]["block1_conv1"]["block1_conv1/kernel:0]

# 列表推導（weights）儲存層的權重和偏置

>>>layer_names.index("block1_conv1")

1

>>> model.layers[1].set_weights(weights)

# 為特定層設置權重。

使用for循環(huán)我們可以為整個網(wǎng)絡設置權重。

"""

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)

model.layers[index].set_weights(weights)

import cv2

import numpy as np

import pandas as pd

from tqdm import tqdm

import itertools

import glob

features = []

for i in tqdm(files_location):

im = cv2.imread(i)

im = cv2.resize(cv2.cvtColor(im, cv2.COLOR_BGR2RGB), (256, 256)).astype(np.float32) / 255.0

im = np.expand_dims(im, axis =0)

outcome = model_final.predict(im)

features.append(outcome)

## 收集這些特征，創(chuàng)建一個dataframe，在其上訓練一個分類器

以上代碼提取block2_pool特征。通常而言，由于這層有64 x 64 x 128特征，在其上訓練一個分類器可能于事無補。我們可以加上一些全連接層，然后在其基礎上訓練神經(jīng)網(wǎng)絡。

增加少量全連接層和一個輸出層。

為靠前的層設置權重，然后凍結。

訓練網(wǎng)絡。

4. 新數(shù)據(jù)集很大，和原數(shù)據(jù)很不一樣

由于你有一個很大的數(shù)據(jù)集，你可以設計你自己的網(wǎng)絡，或者使用現(xiàn)有的網(wǎng)絡。

你可以基于隨機初始化權重或預訓練網(wǎng)絡權重初始化訓練網(wǎng)絡。一般選擇后者。

你可以使用不同的網(wǎng)絡，或者基于現(xiàn)有網(wǎng)絡做些改動。