——Tesla M40 + singularity + keras2 + jupyter notebook

• 硬件：Nvidia Tesla M40 顯存24GB
• 軟件環(huán)境：CentOS 7 + singularity + jupyter notebook
• 深度學(xué)習(xí)框架：keras2 ( 底層 tensorflow-gpu 1.2.0 )
• 語言環(huán)境：python 3.5

本案例也可以用 Ubuntu + keras2 或 Ubuntu + docker + keras2 實(shí)現(xiàn)索赏，當(dāng)然最好使用 GPU 跑， CPU 太慢。因此對(duì)環(huán)境的要求就是只要能跑起來 keras 調(diào)用 GPU 就 OK 了茂卦。

遷移學(xué)習(xí)的定義：在 ImageNet 已經(jīng)得到一個(gè)預(yù)訓(xùn)練好的 ConvNet 網(wǎng)絡(luò)痘绎，刪除網(wǎng)絡(luò)的最后一個(gè)全連接層弧腥，然后將 ConvNet 網(wǎng)絡(luò)的剩余部分作為新數(shù)據(jù)集的特征提取層本辐。一旦你提取了所有圖像的特征，就可以開始訓(xùn)練新數(shù)據(jù)集分類器赖淤。
微調(diào)：更換并重新訓(xùn)練 ConvNet 的網(wǎng)絡(luò)層蜀漆，還可以通過反向傳播算法對(duì)預(yù)訓(xùn)練網(wǎng)絡(luò)的權(quán)重進(jìn)行微調(diào)。

在 jupyter notebook 命令行下咱旱，首先引入

%matplotlib inline
import os
import sys
import glob
# import argparse  #這個(gè)模塊是命令行參數(shù)傳入确丢，在nb中不需要
import matplotlib.pyplot as plt

可以使用 jupyter notebook绷耍，也可以直接在編輯器(sublime text , Atom , pycharm 等)中編輯然后在終端用命令行運(yùn)行。
本文用 notebook蠕嫁，如果想在終端來運(yùn)行，就調(diào)用 argparse 模塊毯盈，運(yùn)行時(shí)給主函數(shù)傳入?yún)?shù)剃毒。

引入一些必要的模塊

from keras import __version__
from keras.applications.inception_v3 import InceptionV3, preprocess_input
from keras.models import Model
from keras.layers import Dense, GlobalAveragePooling2D
from keras.preprocessing.image import ImageDataGenerator
from keras.optimizers import SGD

引入 Inception V3 的 模型，第一次使用時(shí)搂赋，首先下載赘阀，大約88MB，會(huì)保存在 ~/.keras/models 下脑奠，以后再用就不用下載基公。

定義一些全局變量，這些全局變量是可以通過 argparse 模塊從命令行獲取宋欺，傳遞給主函數(shù)轰豆。在notebook中，調(diào)用主函數(shù)的時(shí)候直接傳遞給主函數(shù)齿诞。注意 keras2 中已經(jīng)將 nb_epochs 修改為 epochs 了酸休。

定義全連接層數(shù)為 FC_SIZE 變量（遷移學(xué)習(xí)需要傳遞的參數(shù)），定義凍結(jié)層數(shù)為 NB_IV3_LAYERS_TO_FREEZE 變量（微調(diào)需要傳遞的參數(shù)）祷杈。

IM_WIDTH, IM_HEIGHT = 299, 299    #修正 InceptionV3 的尺寸參數(shù)
EPOCHS = 10
BAT_SIZE = 40
FC_SIZE = 1024
NB_IV3_LAYERS_TO_FREEZE = 172

# 定義一個(gè)方法——獲取訓(xùn)練集和驗(yàn)證集中的樣本數(shù)量斑司，即nb_train_samples,nb_val_samples

def get_nb_files(directory):
    """Get number of files by searching directory recursively"""
    if not os.path.exists(directory):
        return 0
    cnt = 0
    for r, dirs, files in os.walk(directory):
        for dr in dirs:
            cnt += len(glob.glob(os.path.join(r, dr + "/*")))       # glob模塊是用來查找匹配文件的，后面接匹配規(guī)則但汞。
    return cnt

定義遷移學(xué)習(xí)函數(shù)宿刮，凍結(jié)所有的 base_model 層，不訓(xùn)練私蕾。

# 定義遷移學(xué)習(xí)的函數(shù)僵缺，不需要訓(xùn)練的部分。

def setup_to_transfer_learn(model, base_model):
    """Freeze all layers and compile the model"""
    for layer in base_model.layers:
        layer.trainable = False
    model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])

定義增加最后一個(gè)全連接層的函數(shù)踩叭，1024層

# 定義增加最后一個(gè)全連接層的函數(shù)

def add_new_last_layer(base_model, nb_classes):
    """Add last layer to the convnet

    Args:
        base_model: keras model excluding top
        nb_classes: # of classes

    Returns:
        new keras model with last layer
    """
    x = base_model.output
    x = GlobalAveragePooling2D()(x)
    x = Dense(FC_SIZE, activation='relu')(x) #new FC layer, random init
    predictions = Dense(nb_classes, activation='softmax')(x) #new softmax layer
    model = Model(inputs=base_model.input, outputs=predictions)
    return model

定義微調(diào)函數(shù)谤饭，凍結(jié)172層之前的層

# 定義微調(diào)函數(shù)

def setup_to_finetune(model):
    """Freeze the bottom NB_IV3_LAYERS and retrain the remaining top layers.

        note: NB_IV3_LAYERS corresponds to the top 2 inception blocks in the inceptionv3 arch

    Args:
        model: keras model
    """
    for layer in model.layers[:NB_IV3_LAYERS_TO_FREEZE]:
        layer.trainable = False
    for layer in model.layers[NB_IV3_LAYERS_TO_FREEZE:]:
        layer.trainable = True
    model.compile(optimizer=SGD(lr=0.0001, momentum=0.9), loss='categorical_crossentropy', metrics=['accuracy'])

訓(xùn)練結(jié)束后畫 acc-loss 圖，查看訓(xùn)練效果懊纳。

def plot_training(history):
    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.')
    plt.plot(epochs, val_acc, 'r')
    plt.title('Training and validation accuracy')

    plt.figure()
    plt.plot(epochs, loss, 'r.')
    plt.plot(epochs, val_loss, 'r-')
    plt.title('Training and validation loss')
    plt.show()

主函數(shù)揉抵，傳入一些參數(shù)。將圖片處理的代碼也放在主函數(shù)中嗤疯。

def train(train_dir, val_dir, epochs=EPOCHS, batch_size=BAT_SIZE, output_model_file="inceptionv3_25000.model"):
    """Use transfer learning and fine-tuning to train a network on a new dataset"""
    nb_train_samples = get_nb_files(train_dir)
    nb_classes = len(glob.glob(train_dir + "/*"))
    nb_val_samples = get_nb_files(val_dir)
    epochs = int(epochs)
    batch_size = int(batch_size)

    # data prep
    train_datagen =  ImageDataGenerator(
        preprocessing_function=preprocess_input,
        rotation_range=30,
        width_shift_range=0.2,
        height_shift_range=0.2,
        shear_range=0.2,
        zoom_range=0.2,
        horizontal_flip=True
    )
    test_datagen = ImageDataGenerator(
        preprocessing_function=preprocess_input,
        rotation_range=30,
        width_shift_range=0.2,
        height_shift_range=0.2,
        shear_range=0.2,
        zoom_range=0.2,
        horizontal_flip=True
    )

    train_generator = train_datagen.flow_from_directory(
        train_dir,
        target_size=(IM_WIDTH, IM_HEIGHT),
        batch_size=batch_size,
    )

    validation_generator = test_datagen.flow_from_directory(
        val_dir,
        target_size=(IM_WIDTH, IM_HEIGHT),
        batch_size=batch_size,
    )

    # 準(zhǔn)備跑起來冤今，首先給 base_model 和 model 賦值，遷移學(xué)習(xí)和微調(diào)都是使用 InceptionV3 的 notop 模型（看 inception_v3.py 源碼茂缚，此模型是打開了最后一個(gè)全連接層）戏罢，利用 add_new_last_layer 函數(shù)增加最后一個(gè)全連接層屋谭。

    base_model = InceptionV3(weights='imagenet', include_top=False) #include_top=False excludes final FC layer
    model = add_new_last_layer(base_model, nb_classes)

    print "開始遷移學(xué)習(xí):\n"

    # transfer learning
    setup_to_transfer_learn(model, base_model)

    history_tl = model.fit_generator(
        train_generator,
        steps_per_epoch=nb_train_samples // batch_size,
        epochs=epochs,
        validation_data=validation_generator,
        validation_steps=nb_val_samples // batch_size,
        class_weight='auto')
    
    print "開始微調(diào):\n"

    # fine-tuning
    setup_to_finetune(model)

    history_ft = model.fit_generator(
        train_generator,
        steps_per_epoch=nb_train_samples // batch_size,
        epochs=epochs,
        validation_data=validation_generator,
        validation_steps=nb_val_samples // batch_size,
        class_weight='auto')

    model.save(output_model_file)

    plot_training(history_ft)

# train(train_dir, val_dir, epochs=EPOCHS, batch_size=BAT_SIZE, output_model_file="inceptionv3_nbs.model")

train("./data/train", "./data/validation")

Found 20000 images belonging to 2 classes.Found 5000 images belonging to 2 classes.Epoch 1/10
500/500 [==============================] - 364s - loss: 0.8835 - acc: 0.8754 - val_loss: 0.0983 - val_acc: 0.9596
Epoch 2/10
500/500 [==============================] - 350s - loss: 0.1555 - acc: 0.9423 - val_loss: 0.1182 - val_acc: 0.9570
Epoch 3/10
500/500 [==============================] - 348s - loss: 0.1257 - acc: 0.9497 - val_loss: 0.0827 - val_acc: 0.9686
Epoch 4/10
500/500 [==============================] - 349s - loss: 0.1244 - acc: 0.9531 - val_loss: 0.0774 - val_acc: 0.9656
Epoch 5/10
500/500 [==============================] - 348s - loss: 0.1112 - acc: 0.9589 - val_loss: 0.1371 - val_acc: 0.9506
Epoch 6/10
500/500 [==============================] - 347s - loss: 0.1082 - acc: 0.9590 - val_loss: 0.0708 - val_acc: 0.9732
Epoch 7/10
500/500 [==============================] - 350s - loss: 0.1078 - acc: 0.9601 - val_loss: 0.0730 - val_acc: 0.9712
Epoch 8/10
500/500 [==============================] - 351s - loss: 0.1055 - acc: 0.9617 - val_loss: 0.1071 - val_acc: 0.9650
Epoch 9/10
500/500 [==============================] - 351s - loss: 0.1028 - acc: 0.9638 - val_loss: 0.1173 - val_acc: 0.9580
Epoch 10/10
500/500 [==============================] - 353s - loss: 0.1036 - acc: 0.9611 - val_loss: 0.0654 - val_acc: 0.9748
Epoch 1/10
500/500 [==============================] - 363s - loss: 0.0712 - acc: 0.9741 - val_loss: 0.0720 - val_acc: 0.9770
Epoch 2/10
500/500 [==============================] - 357s - loss: 0.0587 - acc: 0.9779 - val_loss: 0.0566 - val_acc: 0.9756
Epoch 3/10
500/500 [==============================] - 360s - loss: 0.0555 - acc: 0.9781 - val_loss: 0.0561 - val_acc: 0.9798
Epoch 4/10
500/500 [==============================] - 360s - loss: 0.0518 - acc: 0.9795 - val_loss: 0.0580 - val_acc: 0.9796
Epoch 5/10
500/500 [==============================] - 360s - loss: 0.0458 - acc: 0.9815 - val_loss: 0.0509 - val_acc: 0.9826
Epoch 6/10
500/500 [==============================] - 361s - loss: 0.0458 - acc: 0.9827 - val_loss: 0.0491 - val_acc: 0.9792
Epoch 7/10
500/500 [==============================] - 363s - loss: 0.0457 - acc: 0.9810 - val_loss: 0.0538 - val_acc: 0.9816
Epoch 8/10
500/500 [==============================] - 362s - loss: 0.0490 - acc: 0.9809 - val_loss: 0.0489 - val_acc: 0.9820
Epoch 9/10
500/500 [==============================] - 363s - loss: 0.0388 - acc: 0.9847 - val_loss: 0.0530 - val_acc: 0.9830
Epoch 10/10
500/500 [==============================] - 365s - loss: 0.0390 - acc: 0.9849 - val_loss: 0.0419 - val_acc: 0.9842