將新的圖像通過訓練好的卷積神經網絡直到瓶頸層的過程,可以看成是對圖像進行特徵提取的過程。在訓練好的inceptionV3模型中,因爲將瓶頸層的輸出再通過一個單層的全連接層神經網絡可以很好地區分1000種類別的圖像,所以有理由認爲瓶頸層輸出的節點向量可以被作爲任何圖像的一個更加精簡且表達能力更強的特徵向量。
一般來說,在數據量足夠的情況下,遷移學習的效果不如完全重新訓練。
案例來源於 《TensorFlow實戰Google深度學習框架》。在源代碼基礎上,做了一些修改後,能夠保存完整的pb文件。
一、資料下載
1、谷歌提供的訓練好的Inception-v3模型 https://storage.googleapis.com/download.tensorflow.org/models/inception_dec_2015.zip
2、案例使用的數據集 http://download.tensorflow.org/example_images/flower_photos.tgz
二、代碼
1、訓練代碼
# -*- coding: utf-8 -*-
"""
卷積神經網絡 Inception-v3模型 遷移學習
"""
import glob
import os.path
import random
import numpy as np
import tensorflow as tf
from tensorflow.python.platform import gfile
from tensorflow.python.framework import graph_util
# inception-v3 模型瓶頸層的節點個數
BOTTLENECK_TENSOR_SIZE = 2048
# inception-v3 模型中代表瓶頸層結果的張量名稱
BOTTLENECK_TENSOR_NAME = 'pool_3/_reshape:0'
# 圖像輸入張量所對應的名稱
JPEG_DATA_TENSOR_NAME = 'DecodeJpeg/contents:0'
# 下載的谷歌訓練好的inception-v3模型文件目錄
MODEL_DIR = './inceptionV3'
# 下載的谷歌訓練好的inception-v3模型文件名
MODEL_FILE = 'tensorflow_inception_graph.pb'
# 保存訓練數據通過瓶頸層後提取的特徵向量
CACHE_DIR = './inceptionV3/tmp/bottleneck'
# 圖片數據的文件夾
INPUT_DATA = './inceptionV3/flower_photos'
# 驗證的數據百分比
VALIDATION_PERCENTAGE = 10
# 測試的數據百分比
TEST_PERCENTACE = 10
# 定義神經網路的設置
LEARNING_RATE = 0.01
STEPS = 1000
BATCH = 100
# 這個函數把數據集分成訓練,驗證,測試三部分
def create_image_lists(testing_percentage, validation_percentage):
"""
這個函數把數據集分成訓練,驗證,測試三部分
:param testing_percentage:測試的數據百分比 10
:param validation_percentage:驗證的數據百分比 10
:return:
"""
result = {}
# 獲取目錄下所有子目錄
sub_dirs = [x[0] for x in os.walk(INPUT_DATA)]
# ['/path/to/flower_data', '/path/to/flower_data\\daisy', '/path/to/flower_data\\dandelion',
# '/path/to/flower_data\\roses', '/path/to/flower_data\\sunflowers', '/path/to/flower_data\\tulips']
# 數組中的第一個目錄是當前目錄,這裏設置標記,不予處理
is_root_dir = True
for sub_dir in sub_dirs: # 遍歷目錄數組,每次處理一種
if is_root_dir:
is_root_dir = False
continue
# 獲取當前目錄下所有的有效圖片文件
extensions = ['jpg', 'jepg', 'JPG', 'JPEG']
file_list = []
dir_name = os.path.basename(sub_dir) # 返回路徑名路徑的基本名稱,如:daisy|dandelion|roses|sunflowers|tulips
for extension in extensions:
file_glob = os.path.join(INPUT_DATA, dir_name, '*.' + extension) # 將多個路徑組合後返回
file_list.extend(glob.glob(file_glob)) # glob.glob返回所有匹配的文件路徑列表,extend往列表中追加另一個列表
if not file_list: continue
# 通過目錄名獲取類別名稱
label_name = dir_name.lower() # 返回其小寫
# 初始化當前類別的訓練數據集、測試數據集、驗證數據集
training_images = []
testing_images = []
validation_images = []
for file_name in file_list: # 遍歷此類圖片的每張圖片的路徑
base_name = os.path.basename(file_name) # 路徑的基本名稱也就是圖片的名稱,如:102841525_bd6628ae3c.jpg
# 隨機講數據分到訓練數據集、測試集和驗證集
chance = np.random.randint(100)
if chance < validation_percentage:
validation_images.append(base_name)
elif chance < (testing_percentage + validation_percentage):
testing_images.append(base_name)
else:
training_images.append(base_name)
result[label_name] = {
'dir': dir_name,
'training': training_images,
'testing': testing_images,
'validation': validation_images
}
return result
# 這個函數通過類別名稱、所屬數據集和圖片編號獲取一張圖片的地址
def get_image_path(image_lists, image_dir, label_name, index, category):
"""
:param image_lists:所有圖片信息
:param image_dir:根目錄 ( 圖片特徵向量根目錄 CACHE_DIR | 圖片原始路徑根目錄 INPUT_DATA )
:param label_name:類別的名稱( daisy|dandelion|roses|sunflowers|tulips )
:param index:編號
:param category:所屬的數據集( training|testing|validation )
:return: 一張圖片的地址
"""
# 獲取給定類別的圖片集合
label_lists = image_lists[label_name]
# 獲取這種類別的圖片中,特定的數據集(base_name的一維數組)
category_list = label_lists[category]
mod_index = index % len(category_list) # 圖片的編號%此數據集中圖片數量
# 獲取圖片文件名
base_name = category_list[mod_index]
sub_dir = label_lists['dir']
# 拼接地址
full_path = os.path.join(image_dir, sub_dir, base_name)
return full_path
# 圖片的特徵向量的文件地址
def get_bottleneck_path(image_lists, label_name, index, category):
return get_image_path(image_lists, CACHE_DIR, label_name, index, category) + '.txt' # CACHE_DIR 特徵向量的根地址
# 計算特徵向量
def run_bottleneck_on_image(sess, image_data, image_data_tensor, bottleneck_tensor):
"""
:param sess:
:param image_data:圖片內容
:param image_data_tensor:
:param bottleneck_tensor:
:return:
"""
bottleneck_values = sess.run(bottleneck_tensor, {image_data_tensor: image_data})
bottleneck_values = np.squeeze(bottleneck_values)
return bottleneck_values
# 獲取一張圖片對應的特徵向量的路徑
def get_or_create_bottleneck(sess, image_lists, label_name, index, category, jpeg_data_tensor, bottleneck_tensor):
"""
:param sess:
:param image_lists:
:param label_name:類別名
:param index:圖片編號
:param category:
:param jpeg_data_tensor:
:param bottleneck_tensor:
:return:
"""
label_lists = image_lists[label_name]
sub_dir = label_lists['dir']
sub_dir_path = os.path.join(CACHE_DIR, sub_dir) # 到類別的文件夾
if not os.path.exists(sub_dir_path): os.makedirs(sub_dir_path)
bottleneck_path = get_bottleneck_path(image_lists, label_name, index, category) # 獲取圖片特徵向量的路徑
if not os.path.exists(bottleneck_path): # 如果不存在
# 獲取圖片原始路徑
image_path = get_image_path(image_lists, INPUT_DATA, label_name, index, category)
# 獲取圖片內容
image_data = gfile.FastGFile(image_path, 'rb').read()
# 計算圖片特徵向量
bottleneck_values = run_bottleneck_on_image(sess, image_data, jpeg_data_tensor, bottleneck_tensor)
# 將特徵向量存儲到文件
bottleneck_string = ','.join(str(x) for x in bottleneck_values)
with open(bottleneck_path, 'w') as bottleneck_file:
bottleneck_file.write(bottleneck_string)
else:
# 讀取保存的特徵向量文件
with open(bottleneck_path, 'r') as bottleneck_file:
bottleneck_string = bottleneck_file.read()
# 字符串轉float數組
bottleneck_values = [float(x) for x in bottleneck_string.split(',')]
return bottleneck_values
# 隨機獲取一個batch的圖片作爲訓練數據(特徵向量,類別)
def get_random_cached_bottlenecks(sess, n_classes, image_lists, how_many, category, jpeg_data_tensor,
bottleneck_tensor):
"""
:param sess:
:param n_classes: 類別數量
:param image_lists:
:param how_many: 一個batch的數量
:param category: 所屬的數據集
:param jpeg_data_tensor:
:param bottleneck_tensor:
:return: 特徵向量列表,類別列表
"""
bottlenecks = []
ground_truths = []
for _ in range(how_many):
# 隨機一個類別和圖片編號加入當前的訓練數據
label_index = random.randrange(n_classes)
label_name = list(image_lists.keys())[label_index] # 隨機圖片的類別名
image_index = random.randrange(65536) # 隨機圖片的編號
bottleneck = get_or_create_bottleneck(sess, image_lists, label_name, image_index, category, jpeg_data_tensor,
bottleneck_tensor) # 計算此圖片的特徵向量
ground_truth = np.zeros(n_classes, dtype=np.float32)
ground_truth[label_index] = 1.0
bottlenecks.append(bottleneck)
ground_truths.append(ground_truth)
return bottlenecks, ground_truths
# 獲取全部的測試數據
def get_test_bottlenecks(sess, image_lists, n_classes, jpeg_data_tensor, bottleneck_tensor):
bottlenecks = []
ground_truths = []
label_name_list = list(image_lists.keys()) # ['dandelion', 'daisy', 'sunflowers', 'roses', 'tulips']
for label_index, label_name in enumerate(label_name_list): # 枚舉每個類別,如:0 sunflowers
category = 'testing'
for index, unused_base_name in enumerate(image_lists[label_name][category]): # 枚舉此類別中的測試數據集中的每張圖片
'''''
print(index, unused_base_name)
0 10386503264_e05387e1f7_m.jpg
1 1419608016_707b887337_n.jpg
2 14244410747_22691ece4a_n.jpg
...
105 9467543719_c4800becbb_m.jpg
106 9595857626_979c45e5bf_n.jpg
107 9922116524_ab4a2533fe_n.jpg
'''
bottleneck = get_or_create_bottleneck(
sess, image_lists, label_name, index, category, jpeg_data_tensor, bottleneck_tensor)
ground_truth = np.zeros(n_classes, dtype=np.float32)
ground_truth[label_index] = 1.0
bottlenecks.append(bottleneck)
ground_truths.append(ground_truth)
return bottlenecks, ground_truths
def create_inception_graph():
with tf.Graph().as_default() as graph:
model_filename = os.path.join(
MODEL_DIR, MODEL_FILE)
with gfile.FastGFile(model_filename, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
bottleneck_tensor, jpeg_data_tensor = tf.import_graph_def(graph_def, name='', return_elements=[
BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME])
return graph, bottleneck_tensor, jpeg_data_tensor
def add_final_training_ops(class_count, bottleneck_tensor):
# 輸入
bottleneck_input = tf.placeholder_with_default(bottleneck_tensor, [None, BOTTLENECK_TENSOR_SIZE], name='BottleneckInputPlaceholder')
ground_truth_input = tf.placeholder(tf.float32, [None, class_count], name='GroundTruthInput')
# 全連接層
with tf.name_scope('output'):
weights = tf.Variable(tf.truncated_normal([BOTTLENECK_TENSOR_SIZE, class_count], stddev=0.001))
biases = tf.Variable(tf.zeros([class_count]))
logits = tf.matmul(bottleneck_input, weights) + biases
final_tensor = tf.nn.softmax(logits, name='prob')
# 損失
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=ground_truth_input)
cross_entropy_mean = tf.reduce_mean(cross_entropy)
train_step = tf.train.GradientDescentOptimizer(LEARNING_RATE).minimize(cross_entropy_mean)
# 正確率
with tf.name_scope('evaluation'):
correct_prediction = tf.equal(tf.argmax(final_tensor, 1), tf.argmax(ground_truth_input, 1))
evaluation_step = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
return (train_step,evaluation_step, cross_entropy_mean, bottleneck_input, ground_truth_input)
def train():
image_lists = create_image_lists(TEST_PERCENTACE, VALIDATION_PERCENTAGE)
n_classes = len(image_lists.keys())
print('n_classes:',n_classes)
graph, bottleneck_tensor, jpeg_data_tensor=create_inception_graph()
print(bottleneck_tensor.graph is tf.get_default_graph())
with tf.Session(graph=graph) as sess:
train_step,evaluation_step,cross_entropy_mean,bottleneck_input,ground_truth_input=add_final_training_ops(n_classes,bottleneck_tensor)
# 初始化參數
init = tf.global_variables_initializer()
sess.run(init)
for i in range(STEPS):
# 每次獲取一個batch的訓練數據
train_bottlenecks, train_ground_truth = get_random_cached_bottlenecks(sess, n_classes, image_lists, BATCH,
'training', jpeg_data_tensor,
bottleneck_tensor)
# 訓練
sess.run(train_step,
feed_dict={bottleneck_input: train_bottlenecks, ground_truth_input: train_ground_truth})
# 驗證
if i % 100 == 0 or i + 1 == STEPS:
validation_bottlenecks, validation_ground_truth = get_random_cached_bottlenecks(sess, n_classes,
image_lists, BATCH,
'validation',
jpeg_data_tensor,
bottleneck_tensor)
validation_accuracy = sess.run(evaluation_step, feed_dict={bottleneck_input: validation_bottlenecks,
ground_truth_input: validation_ground_truth})
print('Step %d: Validation accuracy on random sampled %d examples = %.1f%%' % (
i, BATCH, validation_accuracy * 100))
# 測試
test_bottlenecks, test_ground_truth = get_test_bottlenecks(sess, image_lists, n_classes, jpeg_data_tensor,
bottleneck_tensor)
test_accuracy = sess.run(evaluation_step,
feed_dict={bottleneck_input: test_bottlenecks, ground_truth_input: test_ground_truth})
print('Final test accuracy = %.1f%%' % (test_accuracy * 100))
constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, ["output/prob"])
with tf.gfile.FastGFile("./pbtxt/nn.pb", mode='wb') as f:
f.write(constant_graph.SerializeToString())
2、預測代碼
import cv2
def predict():
strings = ['daisy', 'dandelion', 'roses', 'sunflowers', 'tulips']
def id_to_string(node_id):
return strings[node_id]
with tf.gfile.FastGFile('./pbtxt/nn.pb', 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
tf.import_graph_def(graph_def, name='')
with tf.Session() as sess:
softmax_tensor = sess.graph.get_tensor_by_name('output/prob:0')
# 遍歷目錄
for root, dirs, files in os.walk('./inceptionV3/predict_images/'):
for file in files:
# 載入圖片
image_data = tf.gfile.FastGFile(os.path.join(root, file), 'rb').read()
predictions = sess.run(softmax_tensor, {'DecodeJpeg/contents:0': image_data}) # 圖片格式是jpg格式
predictions = np.squeeze(predictions) # 把結果轉爲1維數據
# 打印圖片路徑及名稱
image_path = os.path.join(root, file)
print(image_path)
# 排序
top_k = predictions.argsort()[::-1]
print(top_k)
for node_id in top_k:
# 獲取分類名稱
human_string = id_to_string(node_id)
# 獲取該分類的置信度
score = predictions[node_id]
print('%s (score = %.5f)' % (human_string, score))
print()
img = cv2.imread(image_path)
cv2.imshow('image', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
三、參考博文
1、http://blog.csdn.net/tz_zs/article/details/77728391?ABstrategy=codes_snippets_optimize_v3
2、https://blog.csdn.net/l18930738887/article/details/72812689