def getImagesWithID(path):
imagePaths, labels, labels_dict = facenet.get_image_paths_and_labels(
dataset)
#print(imagePaths) 路徑
#print(labels) 數量
#print(labels_dict) ids
faces = []
Ids = []
for imagePath in imagePaths:
# 開啟圖像轉換成numpy
faceImg = Image.open(imagePath).convert('L') # 轉灰階
# PIL圖像轉換為numpy
faceNp = np.array(faceImg, 'uint8')
# 從圖片名稱中獲取使用者ID
ID = int(os.path.split(imagePath)[-1].split('.')[1])
# iamge
faces.append(faceNp)
# ids
Ids.append(ID)
#print ID
cv2.imshow("training", faceNp)
cv2.waitKey(10)
return np.array(Ids), np.array(faces) # 回傳ids faces
def getImagesWithID(DATA_PATH):
#-------------載入相關模型--------------#
dataset = facenet.get_dataset(DATA_PATH)
imagePaths, labels, labels_dict = facenet.get_image_paths_and_labels(
dataset)
#print(imagePaths) 路徑
#print(labels) 數量
#print(labels_dict) ids
faces = []
Ids = []
for imagePath in imagePaths:
# 開啟圖像轉換成numpy
faceImg = Image.open(imagePath).convert('L') # 轉灰階
# PIL圖像轉換為numpy
faceNp = np.array(faceImg, 'uint8')
# 從圖片名稱中獲取使用者ID
ID = int(os.path.split(imagePath)[-1].split('.')[1])
# iamge
faces.append(faceNp)
# ids
Ids.append(ID)
#print ID
return np.array(Ids), np.array(faces) # 回傳ids faces
def load_MTCNN(DATA_PATH):
with tf.Graph().as_default():
with tf.Session() as sess:
# DATA_PATH = os.path.join(BASE_DIR, "data") #data目錄
IMG_OUT_PATH = os.path.join(DATA_PATH, "dataset") #image目錄
# FACENET_DATA_PATH = os.path.join(DATA_PATH, "facenet","20180402-114759","20180402-114759.pb") #dacenet路徑
datadir = IMG_OUT_PATH # 經過偵測、對齊 & 裁剪後的人臉圖像目錄
dataset = facenet.get_dataset(
datadir) # 取得人臉類別(ImageClass)的列表與圖像路徑
paths, labels, labels_dict = facenet.get_image_paths_and_labels(
dataset) #取得每個人臉的圖像路徑跟ID標籤
#print (paths) #test
#print (labels) #test
#print (labels_dict) #test
print('Origin: Number of classes: %d' % len(labels_dict)) #人臉種類
print('Origin: Number of images: %d' % len(paths)) #人臉總數
#------------載入Facenet模型------------#
modeldir = '/content/gdrive/My Drive/Colab Notebooks/facenet/20180402-114759/20180402-114759.pb'
facenet.load_model(modeldir)
images_placeholder = tf.get_default_graph().get_tensor_by_name(
"input:0")
embeddings = tf.get_default_graph().get_tensor_by_name(
"embeddings:0")
phase_train_placeholder = tf.get_default_graph(
).get_tensor_by_name("phase_train:0")
embedding_size = embeddings.get_shape()[1]
#------------計算人臉特徵向量------------#
batch_size = 5 # 一次輸入的樣本數量
image_size = 140 # 要做為Facenet的圖像輸入的大小
times_pohto = 10.0 # 每張照片看的次數
nrof_images = len(paths) # 總共要處理的人臉圖像
# 計算總共要跑的批次數
nrof_batches_per_epoch = int(
math.ceil(times_pohto * nrof_images / batch_size))
# 構建一個變數來保存"人臉特徵向量"
emb_array = np.zeros(
(nrof_images, embedding_size)) # <-- Face Embedding
for i in tqdm(range(nrof_batches_per_epoch)): # 實際訓練 facenet
start_index = i * batch_size
end_index = min((i + 1) * batch_size, nrof_images)
paths_batch = paths[start_index:end_index]
images = facenet.load_data(paths_batch, False, False,
image_size)
feed_dict = {
images_placeholder: images,
phase_train_placeholder: False
}
emb_array[start_index:end_index, :] = sess.run(
embeddings, feed_dict=feed_dict)
return emb_array, labels, labels_dict
def trainer_photo(request):
DATA_PATH = os.path.join(BASE_DIR, "data") #data目錄
FACENET_DATA_PATH = os.path.join(DATA_PATH, "facenet", "20180402-114759",
"20180402-114759.pb") #dacenet路徑
SVM_DATA_PATH = os.path.join(DATA_PATH, "serializer",
"lfw_svm_classifier.pkl") #svm路徑
IMG_OUT_PATH = os.path.join(DATA_PATH, "dataset") #裁切後人臉目錄
#----------載入MTCNN-----------#
with tf.Graph().as_default():
with tf.Session() as sess:
datadir = IMG_OUT_PATH # 經過偵測、對齊 & 裁剪後的人臉圖像目錄
dataset = facenet.get_dataset(
datadir) # 取得人臉類別(ImageClass)的列表與圖像路徑
paths, labels, labels_dict = facenet.get_image_paths_and_labels(
dataset) #取得每個人臉的圖像路徑跟ID標籤
#print (paths) #test
#print (labels) #test
#print (labels_dict) #test
print('Origin: Number of classes: %d' % len(labels_dict)) #人臉種類
print('Origin: Number of images: %d' % len(paths)) #人臉總數
#------------載入Facenet模型------------#
modeldir = FACENET_DATA_PATH
facenet.load_model(modeldir)
images_placeholder = tf.get_default_graph().get_tensor_by_name(
"input:0")
embeddings = tf.get_default_graph().get_tensor_by_name(
"embeddings:0")
phase_train_placeholder = tf.get_default_graph(
).get_tensor_by_name("phase_train:0")
embedding_size = embeddings.get_shape()[1]
#------------計算人臉特徵向量------------#
batch_size = 5 # 一次輸入的樣本數量
image_size = 140 # 要做為Facenet的圖像輸入的大小
times_pohto = 10.0 # 每張照片看的次數
nrof_images = len(paths) # 總共要處理的人臉圖像
# 計算總共要跑的批次數
nrof_batches_per_epoch = int(
math.ceil(times_pohto * nrof_images / batch_size))
# 構建一個變數來保存"人臉特徵向量"
emb_array = np.zeros(
(nrof_images, embedding_size)) # <-- Face Embedding
for i in tqdm(range(nrof_batches_per_epoch)): # 實際訓練 facenet
start_index = i * batch_size
end_index = min((i + 1) * batch_size, nrof_images)
paths_batch = paths[start_index:end_index]
images = facenet.load_data(paths_batch, False, False,
image_size)
feed_dict = {
images_placeholder: images,
phase_train_placeholder: False
}
emb_array[start_index:end_index, :] = sess.run(
embeddings, feed_dict=feed_dict)
#--------------保存facenet.pkl-------------#
# 人臉特徵
emb_features_file = open(
os.path.join(DATA_PATH + '/recognizer', 'lfw_emb_features.pkl'), 'wb')
pickle.dump(emb_array, emb_features_file)
emb_features_file.close()
# 矩陣
emb_lables_file = open(
os.path.join(DATA_PATH + '/recognizer', 'lfw_emb_labels.pkl'), 'wb')
pickle.dump(labels, emb_lables_file)
emb_lables_file.close()
# user_ids
emb_lables_dict_file = open(
os.path.join(DATA_PATH + '/recognizer', 'lfw_emb_labels_dict.pkl'),
'wb')
pickle.dump(labels_dict, emb_lables_dict_file)
emb_lables_dict_file.close()
#------------------載入pkl------------------#
# 人臉特徵
with open(os.path.join(DATA_PATH + '/recognizer', 'lfw_emb_features.pkl'),
'rb') as emb_features_file:
emb_features = pickle.load(emb_features_file)
# 每一張的人臉標籤
with open(os.path.join(DATA_PATH + '/recognizer', 'lfw_emb_labels.pkl'),
'rb') as emb_lables_file:
emb_labels = pickle.load(emb_lables_file)
# 總共的人臉標籤種類
with open(
os.path.join(DATA_PATH + '/recognizer', 'lfw_emb_labels_dict.pkl'),
'rb') as emb_lables_dict_file:
emb_labels_dict = pickle.load(emb_lables_dict_file)
#-------------------測試--------------------#
print("人臉特徵數量: {}, shape: {}, type: {}".format(len(emb_features),
emb_features.shape,
type(emb_features)))
print("人臉標籤數量: {}, type: {}".format(len(emb_labels), type(emb_labels)))
print("人臉標籤種類: {}, type: {}", len(emb_labels_dict), type(emb_labels_dict))
#-------------------準備相關變數-----------------#
# 訓練/測試變數
X_train = []
y_train = []
X_test = []
y_test = []
# 保存己經有處理過的人臉label
processed = set()
# 分割訓練資料集與驗證資料集
for (emb_feature, emb_label) in zip(emb_features, emb_labels):
if emb_label in processed:
X_train.append(emb_feature)
y_train.append(emb_label)
else:
X_test.append(emb_feature)
y_test.append(emb_label)
processed.add(emb_label)
# 結果
print('X_train: {}, y_train: {}'.format(len(X_train), len(y_train)))
print('X_test: {}, y_test: {}'.format(len(X_test), len(y_test)))
#----------------訓練人臉分類器(SVM Classifier)-----------------#
#使用scikit-learn的SVM分類器來進行訓練。
#使用linearSvc來訓練
# 訓練分類器
print('Training classifier')
linearsvc_classifier = LinearSVC(C=1, multi_class='crammer_singer')
# 進行訓練
linearsvc_classifier.fit(X_train, y_train)
classifier_filename = SVM_DATA_PATH
class_names = []
for key in sorted(emb_labels_dict.keys()):
class_names.append(emb_labels_dict[key].replace('_', ' '))
with open(classifier_filename, 'wb') as outfile:
pickle.dump((linearsvc_classifier, class_names), outfile)
print('Saved classifier model to file "%s"' % classifier_filename)
return redirect('/')