以圖搜圖之模型篇： 基於 InceptionV3 的模型 finetune

在以圖搜圖的過程中，需要以來模型提取特征，通過特征之間的歐式距離來找到相似的圖形。

本次我們主要講訴以圖搜圖模型創建的方法。

圖片預處理方法，看這里： https://keras.io/zh/preprocessing/image/

本文主要參考了這位大神的文章， 傳送門在此： InceptionV3進行fine-tuning

 

訓練模型代碼如下：

# 基本流程
# 
import os
import sys
import glob
import argparse
import matplotlib.pyplot as plt

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

# 一、定義函數
IM_WIDTH, IM_HEIGHT = 299, 299   # inceptionV3 指定圖片尺寸
FC_SIZE = 1024                   # 全連接層的數量

# 二、數據處理
# 圖片歸類放在不同文件夾下
train_dir = 'E:/Project/Image/data/finetune/train'  # 訓練集數據
val_dir = 'E:/Project/Image/data/finetune/test' # 驗證集數據
nb_epoch = 1
batch_size = 15
nb_classes = len(glob.glob(train_dir + "/*"))  # 分類數


# 圖片增強
# ImageDataGenerator 會自動根據路徑下的文件夾創建標簽，所以在代碼中只看到輸入的 x, 看不到 y
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
)

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

validation_generator = train_datagen.flow_from_directory(
    val_dir, target_size=(IM_WIDTH, IM_HEIGHT),batch_size=batch_size, class_mode='categorical'
)

# 三、使用 bottleneck finetune
# 去掉 模型最外層的全連接層，添加上自己的 全連接層
# 添加新層函數
def add_new_last_layer(base_model, nb_classes):
    x = base_model.output
    x = GlobalAveragePooling2D()(x) # 下采樣
    x = Dense(FC_SIZE, activation='relu')(x)
    predict_bottle_feat = Dense(nb_classes, activation='softmax')(x)
    model = Model(input=base_model.input, output=predict_bottle_feat)
    return model

# 凍結 base_model 所有層
def setup_to_transfer_learn(model, base_model):
    for layer in base_model.layers:
        layer.trainable = False
    model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])

# 定義網絡框架
base_model = InceptionV3(weights='imagenet', include_top=False)
model = add_new_last_layer(base_model, nb_classes)
setup_to_transfer_learn(model, base_model)

# 訓練
# 模式一訓練
steps = 20 # 可以自由定義，越大結果越精准，但過大容易過擬合
history_tl = model.fit_generator(
  train_generator,
  epochs=nb_epoch,
  steps_per_epoch=steps,
  validation_data=validation_generator,
  validation_steps=steps,
  class_weight='auto')

# 保存模型
model.save("my_inceptionV3.h5")

 

使用模型提取指定層的特征： 

from keras.preprocessing import image
from keras_applications.inception_v3 import preprocess_input
from keras.models import Model, load_model
import numpy as np

target_size = (229, 229) #fixed size for InceptionV3 architecture
base_model = load_model(filepath="my_inceptionV3.h5")

# 需要提取那一層的特征，此處就寫入指定層的名稱
model = Model(input=base_model.input, output=base_model.get_layer('block4_pool').output)

img_path = "C:/Users/Administrator/Pictures/搜圖/horse.jpg"
img = image.load_img(img_path, target_size=target_size)
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)

block4_pool_features = model.predict(x)

 

使用模型進行預測：

from keras.preprocessing import image
from  keras.models import load_model
import numpy as np
import json
from keras_applications.imagenet_utils import decode_predictions


def predict(model, img, target_size):
  """Run model prediction on image
  Args:
    model: keras model
    img: PIL format image
    target_size: (w,h) tuple
  Returns:
    list of predicted labels and their probabilities
  """
  if img.size != target_size:
    img = img.resize(target_size)

  x = image.img_to_array(img)
  x = np.expand_dims(x, axis=0)
  x = preprocess_input(x)
  preds = model.predict(x)   # 此處獲取的為
  return preds[0]            # 返回 numpy array [classes, ]

def decode_predict(probalities_list):
  with open("img_classes.json", 'r') as load_f:
    load_dict = json.load(load_f)
  index = probalities_list.index(max(probalities_list))
  target_class = load_dict[str(index)]
  return target_class

target_size = (229, 229) #fixed size for InceptionV3 architecture
model = load_model(filepath="my_inceptionV3.h5")
img = image.load_img("C:/Users/Administrator/Pictures/搜圖/horse.jpg")

res_numpy = predict(model, img, target_size=target_size)
res_list = res_numpy.tolist()
target_class = decode_predict(res_list)
print(target_class)