deepface/deepface/DeepFace.py
Sefik Ilkin Serengil 31de8d7461 fire removed
2022-05-08 20:17:52 +01:00

828 lines
26 KiB
Python

import warnings
warnings.filterwarnings("ignore")
import os
#os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import time
from os import path
import numpy as np
import pandas as pd
from tqdm import tqdm
import pickle
from deepface.basemodels import VGGFace, OpenFace, Facenet, Facenet512, FbDeepFace, DeepID, DlibWrapper, ArcFace, Boosting
from deepface.extendedmodels import Age, Gender, Race, Emotion
from deepface.commons import functions, realtime, distance as dst
import tensorflow as tf
tf_version = int(tf.__version__.split(".")[0])
if tf_version == 2:
import logging
tf.get_logger().setLevel(logging.ERROR)
def build_model(model_name):
"""
This function builds a deepface model
Parameters:
model_name (string): face recognition or facial attribute model
VGG-Face, Facenet, OpenFace, DeepFace, DeepID for face recognition
Age, Gender, Emotion, Race for facial attributes
Returns:
built deepface model
"""
global model_obj #singleton design pattern
models = {
'VGG-Face': VGGFace.loadModel,
'OpenFace': OpenFace.loadModel,
'Facenet': Facenet.loadModel,
'Facenet512': Facenet512.loadModel,
'DeepFace': FbDeepFace.loadModel,
'DeepID': DeepID.loadModel,
'Dlib': DlibWrapper.loadModel,
'ArcFace': ArcFace.loadModel,
'Emotion': Emotion.loadModel,
'Age': Age.loadModel,
'Gender': Gender.loadModel,
'Race': Race.loadModel
}
if not "model_obj" in globals():
model_obj = {}
if not model_name in model_obj.keys():
model = models.get(model_name)
if model:
model = model()
model_obj[model_name] = model
#print(model_name," built")
else:
raise ValueError('Invalid model_name passed - {}'.format(model_name))
return model_obj[model_name]
def verify(img1_path, img2_path = '', model_name = 'VGG-Face', distance_metric = 'cosine', model = None, enforce_detection = True, detector_backend = 'opencv', align = True, prog_bar = True, normalization = 'base'):
"""
This function verifies an image pair is same person or different persons.
Parameters:
img1_path, img2_path: exact image path, numpy array (BGR) or based64 encoded images could be passed. If you are going to call verify function for a list of image pairs, then you should pass an array instead of calling the function in for loops.
e.g. img1_path = [
['img1.jpg', 'img2.jpg'],
['img2.jpg', 'img3.jpg']
]
model_name (string): VGG-Face, Facenet, OpenFace, DeepFace, DeepID, Dlib, ArcFace or Ensemble
distance_metric (string): cosine, euclidean, euclidean_l2
model: Built deepface model. A face recognition model is built every call of verify function. You can pass pre-built face recognition model optionally if you will call verify function several times.
model = DeepFace.build_model('VGG-Face')
enforce_detection (boolean): If any face could not be detected in an image, then verify function will return exception. Set this to False not to have this exception. This might be convenient for low resolution images.
detector_backend (string): set face detector backend as retinaface, mtcnn, opencv, ssd or dlib
prog_bar (boolean): enable/disable a progress bar
Returns:
Verify function returns a dictionary. If img1_path is a list of image pairs, then the function will return list of dictionary.
{
"verified": True
, "distance": 0.2563
, "max_threshold_to_verify": 0.40
, "model": "VGG-Face"
, "similarity_metric": "cosine"
}
"""
tic = time.time()
img_list, bulkProcess = functions.initialize_input(img1_path, img2_path)
resp_objects = []
#--------------------------------
if model_name == 'Ensemble':
model_names = ["VGG-Face", "Facenet", "OpenFace", "DeepFace"]
metrics = ["cosine", "euclidean", "euclidean_l2"]
else:
model_names = []; metrics = []
model_names.append(model_name)
metrics.append(distance_metric)
#--------------------------------
if model == None:
if model_name == 'Ensemble':
models = Boosting.loadModel()
else:
model = build_model(model_name)
models = {}
models[model_name] = model
else:
if model_name == 'Ensemble':
Boosting.validate_model(model)
models = model.copy()
else:
models = {}
models[model_name] = model
#------------------------------
disable_option = (False if len(img_list) > 1 else True) or not prog_bar
pbar = tqdm(range(0,len(img_list)), desc='Verification', disable = disable_option)
for index in pbar:
instance = img_list[index]
if type(instance) == list and len(instance) >= 2:
img1_path = instance[0]; img2_path = instance[1]
ensemble_features = []
for i in model_names:
custom_model = models[i]
#img_path, model_name = 'VGG-Face', model = None, enforce_detection = True, detector_backend = 'mtcnn'
img1_representation = represent(img_path = img1_path
, model_name = model_name, model = custom_model
, enforce_detection = enforce_detection, detector_backend = detector_backend
, align = align
, normalization = normalization
)
img2_representation = represent(img_path = img2_path
, model_name = model_name, model = custom_model
, enforce_detection = enforce_detection, detector_backend = detector_backend
, align = align
, normalization = normalization
)
#----------------------
#find distances between embeddings
for j in metrics:
if j == 'cosine':
distance = dst.findCosineDistance(img1_representation, img2_representation)
elif j == 'euclidean':
distance = dst.findEuclideanDistance(img1_representation, img2_representation)
elif j == 'euclidean_l2':
distance = dst.findEuclideanDistance(dst.l2_normalize(img1_representation), dst.l2_normalize(img2_representation))
else:
raise ValueError("Invalid distance_metric passed - ", distance_metric)
distance = np.float64(distance) #causes trobule for euclideans in api calls if this is not set (issue #175)
#----------------------
#decision
if model_name != 'Ensemble':
threshold = dst.findThreshold(i, j)
if distance <= threshold:
identified = True
else:
identified = False
resp_obj = {
"verified": identified
, "distance": distance
, "threshold": threshold
, "model": model_name
, "detector_backend": detector_backend
, "similarity_metric": distance_metric
}
if bulkProcess == True:
resp_objects.append(resp_obj)
else:
return resp_obj
else: #Ensemble
#this returns same with OpenFace - euclidean_l2
if i == 'OpenFace' and j == 'euclidean':
continue
else:
ensemble_features.append(distance)
#----------------------
if model_name == 'Ensemble':
boosted_tree = Boosting.build_gbm()
prediction = boosted_tree.predict(np.expand_dims(np.array(ensemble_features), axis=0))[0]
verified = np.argmax(prediction) == 1
score = prediction[np.argmax(prediction)]
resp_obj = {
"verified": verified
, "score": score
, "distance": ensemble_features
, "model": ["VGG-Face", "Facenet", "OpenFace", "DeepFace"]
, "similarity_metric": ["cosine", "euclidean", "euclidean_l2"]
}
if bulkProcess == True:
resp_objects.append(resp_obj)
else:
return resp_obj
#----------------------
else:
raise ValueError("Invalid arguments passed to verify function: ", instance)
#-------------------------
toc = time.time()
if bulkProcess == True:
resp_obj = {}
for i in range(0, len(resp_objects)):
resp_item = resp_objects[i]
resp_obj["pair_%d" % (i+1)] = resp_item
return resp_obj
def analyze(img_path, actions = ('emotion', 'age', 'gender', 'race') , models = None, enforce_detection = True, detector_backend = 'opencv', prog_bar = True):
"""
This function analyzes facial attributes including age, gender, emotion and race
Parameters:
img_path: exact image path, numpy array (BGR) or base64 encoded image could be passed. If you are going to analyze lots of images, then set this to list. e.g. img_path = ['img1.jpg', 'img2.jpg']
actions (tuple): The default is ('age', 'gender', 'emotion', 'race'). You can drop some of those attributes.
models: (Optional[dict]) facial attribute analysis models are built in every call of analyze function. You can pass pre-built models to speed the function up.
models = {}
models['age'] = DeepFace.build_model('Age')
models['gender'] = DeepFace.build_model('Gender')
models['emotion'] = DeepFace.build_model('Emotion')
models['race'] = DeepFace.build_model('Race')
enforce_detection (boolean): The function throws exception if a face could not be detected. Set this to True if you don't want to get exception. This might be convenient for low resolution images.
detector_backend (string): set face detector backend as retinaface, mtcnn, opencv, ssd or dlib.
prog_bar (boolean): enable/disable a progress bar
Returns:
The function returns a dictionary. If img_path is a list, then it will return list of dictionary.
{
"region": {'x': 230, 'y': 120, 'w': 36, 'h': 45},
"age": 28.66,
"gender": "woman",
"dominant_emotion": "neutral",
"emotion": {
'sad': 37.65260875225067,
'angry': 0.15512987738475204,
'surprise': 0.0022171278033056296,
'fear': 1.2489334680140018,
'happy': 4.609785228967667,
'disgust': 9.698561953541684e-07,
'neutral': 56.33133053779602
}
"dominant_race": "white",
"race": {
'indian': 0.5480832420289516,
'asian': 0.7830780930817127,
'latino hispanic': 2.0677512511610985,
'black': 0.06337375962175429,
'middle eastern': 3.088453598320484,
'white': 93.44925880432129
}
}
"""
actions = list(actions)
if not models:
models = {}
img_paths, bulkProcess = functions.initialize_input(img_path)
#---------------------------------
built_models = list(models.keys())
#---------------------------------
#pre-trained models passed but it doesn't exist in actions
if len(built_models) > 0:
if 'emotion' in built_models and 'emotion' not in actions:
actions.append('emotion')
if 'age' in built_models and 'age' not in actions:
actions.append('age')
if 'gender' in built_models and 'gender' not in actions:
actions.append('gender')
if 'race' in built_models and 'race' not in actions:
actions.append('race')
#---------------------------------
if 'emotion' in actions and 'emotion' not in built_models:
models['emotion'] = build_model('Emotion')
if 'age' in actions and 'age' not in built_models:
models['age'] = build_model('Age')
if 'gender' in actions and 'gender' not in built_models:
models['gender'] = build_model('Gender')
if 'race' in actions and 'race' not in built_models:
models['race'] = build_model('Race')
#---------------------------------
resp_objects = []
disable_option = (False if len(img_paths) > 1 else True) or not prog_bar
global_pbar = tqdm(range(0,len(img_paths)), desc='Analyzing', disable = disable_option)
for j in global_pbar:
img_path = img_paths[j]
resp_obj = {}
disable_option = (False if len(actions) > 1 else True) or not prog_bar
pbar = tqdm(range(0, len(actions)), desc='Finding actions', disable = disable_option)
img_224 = None # Set to prevent re-detection
region = [] # x, y, w, h of the detected face region
region_labels = ['x', 'y', 'w', 'h']
is_region_set = False
#facial attribute analysis
for index in pbar:
action = actions[index]
pbar.set_description("Action: %s" % (action))
if action == 'emotion':
emotion_labels = ['angry', 'disgust', 'fear', 'happy', 'sad', 'surprise', 'neutral']
img, region = functions.preprocess_face(img = img_path, target_size = (48, 48), grayscale = True, enforce_detection = enforce_detection, detector_backend = detector_backend, return_region = True)
emotion_predictions = models['emotion'].predict(img)[0,:]
sum_of_predictions = emotion_predictions.sum()
resp_obj["emotion"] = {}
for i in range(0, len(emotion_labels)):
emotion_label = emotion_labels[i]
emotion_prediction = 100 * emotion_predictions[i] / sum_of_predictions
resp_obj["emotion"][emotion_label] = emotion_prediction
resp_obj["dominant_emotion"] = emotion_labels[np.argmax(emotion_predictions)]
elif action == 'age':
if img_224 is None:
img_224, region = functions.preprocess_face(img = img_path, target_size = (224, 224), grayscale = False, enforce_detection = enforce_detection, detector_backend = detector_backend, return_region = True)
age_predictions = models['age'].predict(img_224)[0,:]
apparent_age = Age.findApparentAge(age_predictions)
resp_obj["age"] = int(apparent_age) #int cast is for the exception - object of type 'float32' is not JSON serializable
elif action == 'gender':
if img_224 is None:
img_224, region = functions.preprocess_face(img = img_path, target_size = (224, 224), grayscale = False, enforce_detection = enforce_detection, detector_backend = detector_backend, return_region = True)
gender_prediction = models['gender'].predict(img_224)[0,:]
if np.argmax(gender_prediction) == 0:
gender = "Woman"
elif np.argmax(gender_prediction) == 1:
gender = "Man"
resp_obj["gender"] = gender
elif action == 'race':
if img_224 is None:
img_224, region = functions.preprocess_face(img = img_path, target_size = (224, 224), grayscale = False, enforce_detection = enforce_detection, detector_backend = detector_backend, return_region = True) #just emotion model expects grayscale images
race_predictions = models['race'].predict(img_224)[0,:]
race_labels = ['asian', 'indian', 'black', 'white', 'middle eastern', 'latino hispanic']
sum_of_predictions = race_predictions.sum()
resp_obj["race"] = {}
for i in range(0, len(race_labels)):
race_label = race_labels[i]
race_prediction = 100 * race_predictions[i] / sum_of_predictions
resp_obj["race"][race_label] = race_prediction
resp_obj["dominant_race"] = race_labels[np.argmax(race_predictions)]
#-----------------------------
if is_region_set != True:
resp_obj["region"] = {}
is_region_set = True
for i, parameter in enumerate(region_labels):
resp_obj["region"][parameter] = int(region[i]) #int cast is for the exception - object of type 'float32' is not JSON serializable
#---------------------------------
if bulkProcess == True:
resp_objects.append(resp_obj)
else:
return resp_obj
if bulkProcess == True:
resp_obj = {}
for i in range(0, len(resp_objects)):
resp_item = resp_objects[i]
resp_obj["instance_%d" % (i+1)] = resp_item
return resp_obj
def find(img_path, db_path, model_name ='VGG-Face', distance_metric = 'cosine', model = None, enforce_detection = True, detector_backend = 'opencv', align = True, prog_bar = True, normalization = 'base', silent=False):
"""
This function applies verification several times and find an identity in a database
Parameters:
img_path: exact image path, numpy array (BGR) or based64 encoded image. If you are going to find several identities, then you should pass img_path as array instead of calling find function in a for loop. e.g. img_path = ["img1.jpg", "img2.jpg"]
db_path (string): You should store some .jpg files in a folder and pass the exact folder path to this.
model_name (string): VGG-Face, Facenet, OpenFace, DeepFace, DeepID, Dlib or Ensemble
distance_metric (string): cosine, euclidean, euclidean_l2
model: built deepface model. A face recognition models are built in every call of find function. You can pass pre-built models to speed the function up.
model = DeepFace.build_model('VGG-Face')
enforce_detection (boolean): The function throws exception if a face could not be detected. Set this to True if you don't want to get exception. This might be convenient for low resolution images.
detector_backend (string): set face detector backend as retinaface, mtcnn, opencv, ssd or dlib
prog_bar (boolean): enable/disable a progress bar
Returns:
This function returns pandas data frame. If a list of images is passed to img_path, then it will return list of pandas data frame.
"""
tic = time.time()
img_paths, bulkProcess = functions.initialize_input(img_path)
#-------------------------------
if os.path.isdir(db_path) == True:
if model == None:
if model_name == 'Ensemble':
if not silent: print("Ensemble learning enabled")
models = Boosting.loadModel()
else: #model is not ensemble
model = build_model(model_name)
models = {}
models[model_name] = model
else: #model != None
if not silent: print("Already built model is passed")
if model_name == 'Ensemble':
Boosting.validate_model(model)
models = model.copy()
else:
models = {}
models[model_name] = model
#---------------------------------------
if model_name == 'Ensemble':
model_names = ['VGG-Face', 'Facenet', 'OpenFace', 'DeepFace']
metric_names = ['cosine', 'euclidean', 'euclidean_l2']
elif model_name != 'Ensemble':
model_names = []; metric_names = []
model_names.append(model_name)
metric_names.append(distance_metric)
#---------------------------------------
file_name = "representations_%s.pkl" % (model_name)
file_name = file_name.replace("-", "_").lower()
if path.exists(db_path+"/"+file_name):
if not silent: print("WARNING: Representations for images in ",db_path," folder were previously stored in ", file_name, ". If you added new instances after this file creation, then please delete this file and call find function again. It will create it again.")
f = open(db_path+'/'+file_name, 'rb')
representations = pickle.load(f)
if not silent: print("There are ", len(representations)," representations found in ",file_name)
else: #create representation.pkl from scratch
employees = []
for r, d, f in os.walk(db_path): # r=root, d=directories, f = files
for file in f:
if ('.jpg' in file.lower()) or ('.png' in file.lower()):
exact_path = r + "/" + file
employees.append(exact_path)
if len(employees) == 0:
raise ValueError("There is no image in ", db_path," folder! Validate .jpg or .png files exist in this path.")
#------------------------
#find representations for db images
representations = []
pbar = tqdm(range(0,len(employees)), desc='Finding representations', disable = prog_bar)
#for employee in employees:
for index in pbar:
employee = employees[index]
instance = []
instance.append(employee)
for j in model_names:
custom_model = models[j]
representation = represent(img_path = employee
, model_name = model_name, model = custom_model
, enforce_detection = enforce_detection, detector_backend = detector_backend
, align = align
, normalization = normalization
)
instance.append(representation)
#-------------------------------
representations.append(instance)
f = open(db_path+'/'+file_name, "wb")
pickle.dump(representations, f)
f.close()
if not silent: print("Representations stored in ",db_path,"/",file_name," file. Please delete this file when you add new identities in your database.")
#----------------------------
#now, we got representations for facial database
if model_name != 'Ensemble':
df = pd.DataFrame(representations, columns = ["identity", "%s_representation" % (model_name)])
else: #ensemble learning
columns = ['identity']
[columns.append('%s_representation' % i) for i in model_names]
df = pd.DataFrame(representations, columns = columns)
df_base = df.copy() #df will be filtered in each img. we will restore it for the next item.
resp_obj = []
global_pbar = tqdm(range(0, len(img_paths)), desc='Analyzing', disable = prog_bar)
for j in global_pbar:
img_path = img_paths[j]
#find representation for passed image
for j in model_names:
custom_model = models[j]
target_representation = represent(img_path = img_path
, model_name = model_name, model = custom_model
, enforce_detection = enforce_detection, detector_backend = detector_backend
, align = align
, normalization = normalization
)
for k in metric_names:
distances = []
for index, instance in df.iterrows():
source_representation = instance["%s_representation" % (j)]
if k == 'cosine':
distance = dst.findCosineDistance(source_representation, target_representation)
elif k == 'euclidean':
distance = dst.findEuclideanDistance(source_representation, target_representation)
elif k == 'euclidean_l2':
distance = dst.findEuclideanDistance(dst.l2_normalize(source_representation), dst.l2_normalize(target_representation))
distances.append(distance)
#---------------------------
if model_name == 'Ensemble' and j == 'OpenFace' and k == 'euclidean':
continue
else:
df["%s_%s" % (j, k)] = distances
if model_name != 'Ensemble':
threshold = dst.findThreshold(j, k)
df = df.drop(columns = ["%s_representation" % (j)])
df = df[df["%s_%s" % (j, k)] <= threshold]
df = df.sort_values(by = ["%s_%s" % (j, k)], ascending=True).reset_index(drop=True)
resp_obj.append(df)
df = df_base.copy() #restore df for the next iteration
#----------------------------------
if model_name == 'Ensemble':
feature_names = []
for j in model_names:
for k in metric_names:
if model_name == 'Ensemble' and j == 'OpenFace' and k == 'euclidean':
continue
else:
feature = '%s_%s' % (j, k)
feature_names.append(feature)
#print(df.head())
x = df[feature_names].values
#--------------------------------------
boosted_tree = Boosting.build_gbm()
y = boosted_tree.predict(x)
verified_labels = []; scores = []
for i in y:
verified = np.argmax(i) == 1
score = i[np.argmax(i)]
verified_labels.append(verified)
scores.append(score)
df['verified'] = verified_labels
df['score'] = scores
df = df[df.verified == True]
#df = df[df.score > 0.99] #confidence score
df = df.sort_values(by = ["score"], ascending=False).reset_index(drop=True)
df = df[['identity', 'verified', 'score']]
resp_obj.append(df)
df = df_base.copy() #restore df for the next iteration
#----------------------------------
toc = time.time()
if not silent: print("find function lasts ",toc-tic," seconds")
if len(resp_obj) == 1:
return resp_obj[0]
return resp_obj
else:
raise ValueError("Passed db_path does not exist!")
return None
def represent(img_path, model_name = 'VGG-Face', model = None, enforce_detection = True, detector_backend = 'opencv', align = True, normalization = 'base'):
"""
This function represents facial images as vectors.
Parameters:
img_path: exact image path, numpy array (BGR) or based64 encoded images could be passed.
model_name (string): VGG-Face, Facenet, OpenFace, DeepFace, DeepID, Dlib, ArcFace.
model: Built deepface model. A face recognition model is built every call of verify function. You can pass pre-built face recognition model optionally if you will call verify function several times. Consider to pass model if you are going to call represent function in a for loop.
model = DeepFace.build_model('VGG-Face')
enforce_detection (boolean): If any face could not be detected in an image, then verify function will return exception. Set this to False not to have this exception. This might be convenient for low resolution images.
detector_backend (string): set face detector backend as retinaface, mtcnn, opencv, ssd or dlib
normalization (string): normalize the input image before feeding to model
Returns:
Represent function returns a multidimensional vector. The number of dimensions is changing based on the reference model. E.g. FaceNet returns 128 dimensional vector; VGG-Face returns 2622 dimensional vector.
"""
if model is None:
model = build_model(model_name)
#---------------------------------
#decide input shape
input_shape_x, input_shape_y = functions.find_input_shape(model)
#detect and align
img = functions.preprocess_face(img = img_path
, target_size=(input_shape_y, input_shape_x)
, enforce_detection = enforce_detection
, detector_backend = detector_backend
, align = align)
#---------------------------------
#custom normalization
img = functions.normalize_input(img = img, normalization = normalization)
#---------------------------------
#represent
embedding = model.predict(img)[0].tolist()
return embedding
def stream(db_path = '', model_name ='VGG-Face', detector_backend = 'opencv', distance_metric = 'cosine', enable_face_analysis = True, source = 0, time_threshold = 5, frame_threshold = 5):
"""
This function applies real time face recognition and facial attribute analysis
Parameters:
db_path (string): facial database path. You should store some .jpg files in this folder.
model_name (string): VGG-Face, Facenet, OpenFace, DeepFace, DeepID, Dlib or Ensemble
detector_backend (string): opencv, ssd, mtcnn, dlib, retinaface
distance_metric (string): cosine, euclidean, euclidean_l2
enable_facial_analysis (boolean): Set this to False to just run face recognition
source: Set this to 0 for access web cam. Otherwise, pass exact video path.
time_threshold (int): how many second analyzed image will be displayed
frame_threshold (int): how many frames required to focus on face
"""
if time_threshold < 1:
raise ValueError("time_threshold must be greater than the value 1 but you passed "+str(time_threshold))
if frame_threshold < 1:
raise ValueError("frame_threshold must be greater than the value 1 but you passed "+str(frame_threshold))
realtime.analysis(db_path, model_name, detector_backend, distance_metric, enable_face_analysis
, source = source, time_threshold = time_threshold, frame_threshold = frame_threshold)
def detectFace(img_path, target_size = (224, 224), detector_backend = 'opencv', enforce_detection = True, align = True):
"""
This function applies pre-processing stages of a face recognition pipeline including detection and alignment
Parameters:
img_path: exact image path, numpy array (BGR) or base64 encoded image
detector_backend (string): face detection backends are retinaface, mtcnn, opencv, ssd or dlib
Returns:
deteced and aligned face in numpy format
"""
img = functions.preprocess_face(img = img_path, target_size = target_size, detector_backend = detector_backend
, enforce_detection = enforce_detection, align = align)[0] #preprocess_face returns (1, 224, 224, 3)
return img[:, :, ::-1] #bgr to rgb
#---------------------------
#main
def __main__():
pass
functions.initialize_folder()