Face recognition is a computer vision task of identifying and verifying a person based on a photograph of their face.
FaceNet is a face recognition system developed in 2015 by researchers at Google that achieved then state-of-the-art results on a range of face recognition benchmark datasets. The FaceNet system can be used broadly thanks to multiple third-party open source implementations of the model and the availability of pre-trained models.
The FaceNet system can be used to extract high-quality features from faces, called face embeddings, that can then be used to train a face identification system.
In this tutorial, you will discover how to develop a face detection system using FaceNet and an SVM classifier to identify people from photographs.
After completing this tutorial, you will know:
About the FaceNet face recognition system developed by Google and open source implementations and pre-trained models.
How to prepare a face detection dataset including first extracting faces via a face detection system and then extracting face features via face embeddings.
How to fit, evaluate, and demonstrate an SVM model to predict identities from faces embeddings.
Kick-start your project with my new book Deep Learning for Computer Vision, including step-by-step tutorials and the Python source code files for all examples.
Let’s get started.
Update Nov/2019: Updated for TensorFlow v2.0 and MTCNN v0.1.0.
How to Develop a Face Recognition System Using FaceNet in Keras and an SVM Classifier Photo by Peter Valverde, some rights reserved.
Tutorial Overview
This tutorial is divided into five parts; they are:
Face Recognition
FaceNet Model
How to Load a FaceNet Model in Keras
How to Detect Faces for Face Recognition
How to Develop a Face Classification System
Face Recognition
Face recognition is the general task of identifying and verifying people from photographs of their face.
The 2011 book on face recognition titled “Handbook of Face Recognition” describes two main modes for face recognition, as:
Face Verification. A one-to-one mapping of a given face against a known identity (e.g. is this the person?).
Face Identification. A one-to-many mapping for a given face against a database of known faces (e.g. who is this person?).
A face recognition system is expected to identify faces present in images and videos automatically. It can operate in either or both of two modes: (1) face verification (or authentication), and (2) face identification (or recognition).
It is a system that, given a picture of a face, will extract high-quality features from the face and predict a 128 element vector representation these features, called a face embedding.
FaceNet, that directly learns a mapping from face images to a compact Euclidean space where distances directly correspond to a measure of face similarity.
The model is a deep convolutional neural network trained via a triplet loss function that encourages vectors for the same identity to become more similar (smaller distance), whereas vectors for different identities are expected to become less similar (larger distance). The focus on training a model to create embeddings directly (rather than extracting them from an intermediate layer of a model) was an important innovation in this work.
Our method uses a deep convolutional network trained to directly optimize the embedding itself, rather than an intermediate bottleneck layer as in previous deep learning approaches.
These face embeddings were then used as the basis for training classifier systems on standard face recognition benchmark datasets, achieving then-state-of-the-art results.
Our system cuts the error rate in comparison to the best published result by 30% …
The paper also explores other uses of the embeddings, such as clustering to group like-faces based on their extracted features.
It is a robust and effective face recognition system, and the general nature of the extracted face embeddings lends the approach to a range of applications.
How to Load a FaceNet Model in Keras
There are a number of projects that provide tools to train FaceNet-based models and make use of pre-trained models.
Perhaps the most prominent is called OpenFace that provides FaceNet models built and trained using the PyTorch deep learning framework. There is a port of OpenFace to Keras, called Keras OpenFace, but at the time of writing, the models appear to require Python 2, which is quite limiting.
Another prominent project is called FaceNet by David Sandberg that provides FaceNet models built and trained using TensorFlow. The project looks mature, although at the time of writing does not provide a library-based installation nor clean API. Usefully, David’s project provides a number of high-performing pre-trained FaceNet models and there are a number of projects that port or convert these models for use in Keras.
A notable example is Keras FaceNet by Hiroki Taniai. His project provides a script for converting the Inception ResNet v1 model from TensorFlow to Keras. He also provides a pre-trained Keras model ready for use.
We will use the pre-trained Keras FaceNet model provided by Hiroki Taniai in this tutorial. It was trained on MS-Celeb-1M dataset and expects input images to be color, to have their pixel values whitened (standardized across all three channels), and to have a square shape of 160×160 pixels.
Download the model file and place it in your current working directory with the filename ‘facenet_keras.h5‘.
We can load the model directly in Keras using the load_model() function; for example:
1
2
3
4
5
6
7
# example of loading the keras facenet model
from keras.models import load_model
# load the model
model=load_model('facenet_keras.h5')
# summarize input and output shape
print(model.inputs)
print(model.outputs)
Running the example loads the model and prints the shape of the input and output tensors.
We can see that the model indeed expects square color images as input with the shape 160×160, and will output a face embedding as a 128 element vector.
Now that we have a FaceNet model, we can explore using it.
How to Detect Faces for Face Recognition
Before we can perform face recognition, we need to detect faces.
Face detection is the process of automatically locating faces in a photograph and localizing them by drawing a bounding box around their extent.
In this tutorial, we will also use the Multi-Task Cascaded Convolutional Neural Network, or MTCNN, for face detection, e.g. finding and extracting faces from photos. This is a state-of-the-art deep learning model for face detection, described in the 2016 paper titled “Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks.”
We can confirm that the library was installed correctly by importing the library and printing the version; for example:
1
2
3
4
# confirm mtcnn was installed correctly
import mtcnn
# print version
print(mtcnn.__version__)
Running the example prints the current version of the library.
1
0.1.0
We can use the mtcnn library to create a face detector and extract faces for our use with the FaceNet face detector models in subsequent sections.
The first step is to load an image as a NumPy array, which we can achieve using the PIL library and the open() function. We will also convert the image to RGB, just in case the image has an alpha channel or is black and white.
1
2
3
4
5
6
# load image from file
image=Image.open(filename)
# convert to RGB, if needed
image=image.convert('RGB')
# convert to array
pixels=asarray(image)
Next, we can create an MTCNN face detector class and use it to detect all faces in the loaded photograph.
1
2
3
4
# create the detector, using default weights
detector=MTCNN()
# detect faces in the image
results=detector.detect_faces(pixels)
The result is a list of bounding boxes, where each bounding box defines a lower-left-corner of the bounding box, as well as the width and height.
If we assume there is only one face in the photo for our experiments, we can determine the pixel coordinates of the bounding box as follows. Sometimes the library will return a negative pixel index, and I think this is a bug. We can fix this by taking the absolute value of the coordinates.
1
2
3
4
5
# extract the bounding box from the first face
x1,y1,width,height=results[0]['box']
# bug fix
x1,y1=abs(x1),abs(y1)
x2,y2=x1+width,y1+height
We can use these coordinates to extract the face.
1
2
# extract the face
face=pixels[y1:y2,x1:x2]
We can then use the PIL library to resize this small image of the face to the required size; specifically, the model expects square input faces with the shape 160×160.
1
2
3
4
# resize pixels to the model size
image=Image.fromarray(face)
image=image.resize((160,160))
face_array=asarray(image)
Tying all of this together, the function extract_face() will load a photograph from the loaded filename and return the extracted face. It assumes that the photo contains one face and will return the first face detected.
We can use this function to extract faces as needed in the next section that can be provided as input to the FaceNet model.
How to Develop a Face Classification System
In this section, we will develop a face detection system to predict the identity of a given face.
The model will be trained and tested using the ‘5 Celebrity Faces Dataset‘ that contains many photographs of five different celebrities.
We will use an MTCNN model for face detection, the FaceNet model will be used to create a face embedding for each detected face, then we will develop a Linear Support Vector Machine (SVM) classifier model to predict the identity of a given face.
5 Celebrity Faces Dataset
The 5 Celebrity Faces Dataset is a small dataset that contains photographs of celebrities.
The dataset was prepared and made available by Dan Becker and provided for free download on Kaggle. Note, a Kaggle account is required to download the dataset.
Download the dataset (this may require a Kaggle login), data.zip (2.5 megabytes), and unzip it in your local directory with the folder name ‘5-celebrity-faces-dataset‘.
You should now have a directory with the following structure (note, there are spelling mistakes in some directory names, and they were left as-is in this example):
1
2
3
4
5
6
7
8
9
10
11
12
13
5-celebrity-faces-dataset
├── train
│ ├── ben_afflek
│ ├── elton_john
│ ├── jerry_seinfeld
│ ├── madonna
│ └── mindy_kaling
└── val
├── ben_afflek
├── elton_john
├── jerry_seinfeld
├── madonna
└── mindy_kaling
We can see that there is a training dataset and a validation or test dataset.
Looking at some of the photos in the directories, we can see that the photos provide faces with a range of orientations, lighting, and in various sizes. Importantly, each photo contains one face of the person.
We will use this dataset as the basis for our classifier, trained on the ‘train‘ dataset only and classify faces in the ‘val‘ dataset. You can use this same structure to develop a classifier with your own photographs.
Detect Faces
The first step is to detect the face in each photograph and reduce the dataset to a series of faces only.
Let’s test out our face detector function defined in the previous section, specifically extract_face().
Looking in the ‘5-celebrity-faces-dataset/train/ben_afflek/‘ directory, we can see that there are 14 photographs of Ben Affleck in the training dataset. We can detect the face in each photograph, and create a plot with 14 faces, with two rows of seven images each.
The complete example is listed below.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
# demonstrate face detection on 5 Celebrity Faces Dataset
Running the example takes a moment and reports the progress of each loaded photograph along the way and the shape of the NumPy array containing the face pixel data.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
1 (160, 160, 3)
2 (160, 160, 3)
3 (160, 160, 3)
4 (160, 160, 3)
5 (160, 160, 3)
6 (160, 160, 3)
7 (160, 160, 3)
8 (160, 160, 3)
9 (160, 160, 3)
10 (160, 160, 3)
11 (160, 160, 3)
12 (160, 160, 3)
13 (160, 160, 3)
14 (160, 160, 3)
A figure is created containing the faces detected in the Ben Affleck directory.
We can see that each face was correctly detected and that we have a range of lighting, skin tones, and orientations in the detected faces.
Plot of 14 Faces of Ben Affleck Detected From the Training Dataset of the 5 Celebrity Faces Dataset
So far, so good.
Next, we can extend this example to step over each subdirectory for a given dataset (e.g. ‘train‘ or ‘val‘), extract the faces, and prepare a dataset with the name as the output label for each detected face.
The load_faces() function below will load all of the faces into a list for a given directory, e.g. ‘5-celebrity-faces-dataset/train/ben_afflek/‘.
1
2
3
4
5
6
7
8
9
10
11
12
# load images and extract faces for all images in a directory
def load_faces(directory):
faces=list()
# enumerate files
forfilename inlistdir(directory):
# path
path=directory+filename
# get face
face=extract_face(path)
# store
faces.append(face)
returnfaces
We can call the load_faces() function for each subdirectory in the ‘train‘ or ‘val‘ folders. Each face has one label, the name of the celebrity, which we can take from the directory name.
The load_dataset() function below takes a directory name such as ‘5-celebrity-faces-dataset/train/‘ and detects faces for each subdirectory (celebrity), assigning labels to each detected face.
It returns the X and y elements of the dataset as NumPy arrays.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
# load a dataset that contains one subdir for each class that in turn contains images
def load_dataset(directory):
X,y=list(),list()
# enumerate folders, on per class
forsubdir inlistdir(directory):
# path
path=directory+subdir+'/'
# skip any files that might be in the dir
ifnotisdir(path):
continue
# load all faces in the subdirectory
faces=load_faces(path)
# create labels
labels=[subdir for_inrange(len(faces))]
# summarize progress
print('>loaded %d examples for class: %s'%(len(faces),subdir))
# store
X.extend(faces)
y.extend(labels)
returnasarray(X),asarray(y)
We can then call this function for the ‘train’ and ‘val’ folders to load all of the data, then save the results in a single compressed NumPy array file via the savez_compressed() function.
First, all of the photos in the ‘train‘ dataset are loaded, then faces are extracted, resulting in 93 samples with square face input and a class label string as output. Then the ‘val‘ dataset is loaded, providing 25 samples that can be used as a test dataset.
Both datasets are then saved to a compressed NumPy array file called ‘5-celebrity-faces-dataset.npz‘ that is about three megabytes and is stored in the current working directory.
1
2
3
4
5
6
7
8
9
10
11
12
>loaded 14 examples for class: ben_afflek
>loaded 19 examples for class: madonna
>loaded 17 examples for class: elton_john
>loaded 22 examples for class: mindy_kaling
>loaded 21 examples for class: jerry_seinfeld
(93, 160, 160, 3) (93,)
>loaded 5 examples for class: ben_afflek
>loaded 5 examples for class: madonna
>loaded 5 examples for class: elton_john
>loaded 5 examples for class: mindy_kaling
>loaded 5 examples for class: jerry_seinfeld
(25, 160, 160, 3) (25,)
This dataset is ready to be provided to a face detection model.
Create Face Embeddings
The next step is to create a face embedding.
A face embedding is a vector that represents the features extracted from the face. This can then be compared with the vectors generated for other faces. For example, another vector that is close (by some measure) may be the same person, whereas another vector that is far (by some measure) may be a different person.
The classifier model that we want to develop will take a face embedding as input and predict the identity of the face. The FaceNet model will generate this embedding for a given image of a face.
The FaceNet model can be used as part of the classifier itself, or we can use the FaceNet model to pre-process a face to create a face embedding that can be stored and used as input to our classifier model. This latter approach is preferred as the FaceNet model is both large and slow to create a face embedding.
We can, therefore, pre-compute the face embeddings for all faces in the train and test (formally ‘val‘) sets in our 5 Celebrity Faces Dataset.
Next, we can load our FaceNet model ready for converting faces into face embeddings.
1
2
3
# load the facenet model
model=load_model('facenet_keras.h5')
print('Loaded Model')
We can then enumerate each face in the train and test datasets to predict an embedding.
To predict an embedding, first the pixel values of the image need to be suitably prepared to meet the expectations of the FaceNet model. This specific implementation of the FaceNet model expects that the pixel values are standardized.
1
2
3
4
5
# scale pixel values
face_pixels=face_pixels.astype('float32')
# standardize pixel values across channels (global)
mean,std=face_pixels.mean(),face_pixels.std()
face_pixels=(face_pixels-mean)/std
In order to make a prediction for one example in Keras, we must expand the dimensions so that the face array is one sample.
1
2
# transform face into one sample
samples=expand_dims(face_pixels,axis=0)
We can then use the model to make a prediction and extract the embedding vector.
1
2
3
4
# make prediction to get embedding
yhat=model.predict(samples)
# get embedding
embedding=yhat[0]
The get_embedding() function defined below implements these behaviors and will return a face embedding given a single image of a face and the loaded FaceNet model.
1
2
3
4
5
6
7
8
9
10
11
12
# get the face embedding for one face
def get_embedding(model,face_pixels):
# scale pixel values
face_pixels=face_pixels.astype('float32')
# standardize pixel values across channels (global)
mean,std=face_pixels.mean(),face_pixels.std()
face_pixels=(face_pixels-mean)/std
# transform face into one sample
samples=expand_dims(face_pixels,axis=0)
# make prediction to get embedding
yhat=model.predict(samples)
returnyhat[0]
Tying all of this together, the complete example of converting each face into a face embedding in the train and test datasets is listed below.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
# calculate a face embedding for each face in the dataset using facenet
from numpy import load
from numpy import expand_dims
from numpy import asarray
from numpy import savez_compressed
from keras.models import load_model
# get the face embedding for one face
def get_embedding(model,face_pixels):
# scale pixel values
face_pixels=face_pixels.astype('float32')
# standardize pixel values across channels (global)
Running the example reports progress along the way.
We can see that the face dataset was loaded correctly and so was the model. The train dataset was then transformed into 93 face embeddings, each comprised of a 128 element vector. The 25 examples in the test dataset were also suitably converted to face embeddings.
The resulting datasets were then saved to a compressed NumPy array that is about 50 kilobytes with the name ‘5-celebrity-faces-embeddings.npz‘ in the current working directory.
Next, the data requires some minor preparation prior to modeling.
First, it is a good practice to normalize the face embedding vectors. It is a good practice because the vectors are often compared to each other using a distance metric.
In this context, vector normalization means scaling the values until the length or magnitude of the vectors is 1 or unit length. This can be achieved using the Normalizer class in scikit-learn. It might even be more convenient to perform this step when the face embeddings are created in the previous step.
1
2
3
4
# normalize input vectors
in_encoder=Normalizer(norm='l2')
trainX=in_encoder.transform(trainX)
testX=in_encoder.transform(testX)
Next, the string target variables for each celebrity name need to be converted to integers.
It is common to use a Linear Support Vector Machine (SVM) when working with normalized face embedding inputs. This is because the method is very effective at separating the face embedding vectors. We can fit a linear SVM to the training data using the SVC class in scikit-learn and setting the ‘kernel‘ attribute to ‘linear‘. We may also want probabilities later when making predictions, which can be configured by setting ‘probability‘ to ‘True‘.
1
2
3
# fit model
model=SVC(kernel='linear')
model.fit(trainX,trainy)
Next, we can evaluate the model.
This can be achieved by using the fit model to make a prediction for each example in the train and test datasets and then calculating the classification accuracy.
Running the example first confirms that the number of samples in the train and test datasets is as we expect
Next, the model is evaluated on the train and test dataset, showing perfect classification accuracy. This is not surprising given the size of the dataset and the power of the face detection and face recognition models used.
1
2
Dataset: train=93, test=25
Accuracy: train=100.000, test=100.000
We can make it more interesting by plotting the original face and the prediction.
First, we need to load the face dataset, specifically the faces in the test dataset. We could also load the original photos to make it even more interesting.
1
2
3
# load faces
data=load('5-celebrity-faces-dataset.npz')
testX_faces=data['arr_2']
The rest of the example is the same up until we fit the model.
First, we need to select a random example from the test set, then get the embedding, face pixels, expected class prediction, and the corresponding name for the class.
1
2
3
4
5
6
# test model on a random example from the test dataset
A different random example from the test dataset will be selected each time the code is run.
Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.
In this case, a photo of Jerry Seinfeld is selected and correctly predicted.
1
2
Predicted: jerry_seinfeld (88.476)
Expected: jerry_seinfeld
A plot of the chosen face is also created, showing the predicted name and probability in the image title.
Detected Face of Jerry Seinfeld, Correctly Identified by the SVM Classifier
Further Reading
This section provides more resources on the topic if you are looking to go deeper.
In this tutorial, you discovered how to develop a face detection system using FaceNet and an SVM classifier to identify people from photographs.
Specifically, you learned:
About the FaceNet face recognition system developed by Google and open source implementations and pre-trained models.
How to prepare a face detection dataset including first extracting faces via a face detection system and then extracting face features via face embeddings.
How to fit, evaluate, and demonstrate an SVM model to predict identities from faces embeddings.
Do you have any questions?
Ask your questions in the comments below and I will do my best to answer.
It provides self-study tutorials on topics like: classification, object detection (yolo and rcnn), face recognition (vggface and facenet), data preparation and much more...
Finally Bring Deep Learning to your Vision Projects
Was looking at whether Transfer learning, Siamese network and triplet loss approaches are applicable to animal face(eg a sheep, goat etc) recognition particularly mobileNet(or otherwise) when your crystal clear blog came up.
Kindly shed more light on its applicability and any other auxiliary hints.
Sorry Sir! but it’s not helping me out because I’m creating my own datasets of 15 people so that’s why facing that error can you help me about what changes should I change in your code ??? and how do i train my own new facenet model ????
What do you suggest when we have tens of thousands of classes.
A Facenet model itself as a classifier or a specific classifier model is to be trained. In terms of scalability and performance which is the preferred method.
Referring to:
“The FaceNet model can be used as part of the classifier itself, or we can use the FaceNet model to pre-process a face to create a face embedding that can be stored and used as input to our classifier model. This latter approach is preferred as the FaceNet model is both large and slow to create a face embedding.”
Hi jason,
As per my understanding The triplet loss is used so that the model can also learn the dissimilarity between the classes rather than only learning similarity between same class.
But here we are not training our model on the actual image dataset on which we need our classification to be done. Rather we are using SVM for that.
So, how can we make use of triplet loss in this method of face recognition?
Hi Jason. nice tutor!
But i wonder that, if i want to identify who is stranger. should i make a folder for ‘stranger’ contains alot of stranger faces???(exclude your 5 Celebrity Faces Dataset ??)
Thanks for your tutorial.
But when i flollow you, i have a warning :
embeddings.npz is not UTF-8 encoded
UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.
Do you know how to fix it.
Thank you so much.
But when I have a new image to recognize, do I need to put it to the validate folder and rerun the code ?
And how can we use this to recognite face in a video ?
Hello Jason,
Thanks for your wonderful tutorial, I’d like to know what is the best solution to apply recognition part if I have a very small dataset -only one face per each identity- in this case, I think SVM wouldn’t help.
Hello Jason,
Thank you for this amazing tutorial, I used python 3 to run this code. I would like to know why am i getting this error (No module named ‘mtcnn’) and how can I correct it?
I want to use transfer learning for masked face recognition. But i didn’t found any better masked face recognition dataset. I need a masked face dataset with proper labeling of each candidate. So is there any better masked face dataset available? where can i find this dataset?
But I have another question, because sometime when i run the code, all worked perfectly and when I run the code again, sometime i have this error warning in load_model part although the face recognition still work
UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.
warnings.warn(‘No training configuration found in save file: ‘
Hi Jason， while I was executing the code “load_model(‘facenet_keras.h5’)”, the exception “tuple index out of range” is thrown, can you tell me why? thanks in advance.
Hi Jason, again wonderful article and tutorial you provided to us. I wonder how I can customize dataset for my needs such as my friends dataset and perform training on it?
Thanks Jason, really helpful as always but I got a weird “invalid argument” error. But I fixed it by changing ‘f’ to ‘F’ in facenet_keras.h5 because I notice it couldn’t recognize character ‘f’. Maybe because it’s trained on Ubuntu but I run your code on Windows 10. I don’t know!
Thank you! Dear Jason, could you please tell me how I can get access to other properties of model. I mean I don’t need model.predict. I need other properties. Is there a way to list all of them such as different convs or avgpool. I tried __dict__ and dir() but they don’t give what I want. For example, how did you know model has a property called “.predict”? Where can I find all of them? Thank you!
Which one would you suggest? If I have to develop Face Verification system then there are few approaches (listing two approaches from your article):
Approach 1: Detect face using MTCNN, train VGGFACE2 on the collected dataset which helps to predict the probability of a given face belonging to a particular class
Approach 2: Detect face using MTCNN, get face embedding vector using facenet keras model and then apply SVM or Neural Network to predict classes
Which approach would you recommend? Can you please explain?
Thanks for the tutorial.
Unit length normalization isn’t for SVM. For SVM you typically use range scaling – MinMaxScaler, or standardization – StandardScaler. The goal is to make different features uniform. Actually, it’s a surprise that unit length normalization produced 100% accuracy in your case. That’s probably due to small data. It does not work for SVM in general and didn’t work for me.
Great article. You have explained all the necessary steps to implement a face recognition system. I am working on a similar problem but in a bigger scale. I am in a belief that a classification based face identification is not a scalable solution. Please give me your opinion.
If I want to recognise a thousand faces in real time manner then, what type of changes do I need to make to your implementation.
I believe it would be really helpful if you create an article about large scale face recognition.
Once i have trained the model on 5 class (each class having 50 images). Now i use the model to detect images that it has not seen, it correctly guesses that the person in the image is class A for example with an accuracy ( prediction ) 65%. Is it possible to now add such image back to training and expect to get better results ?
Thanks for this tutorial. Its really helpful. I wanted to know why you used train and val dataset. I mean are these two used for training purpose. What is the use of val here.?
In the face classification, I am not able to understand where are you selecting the random photo to test against your dataset. How can I add my jpg photo to test again the dataset. Can you explain please. Thanks
Got it. Thanks
What I have seen is that in train dataset I put my pictures more than 30 images and in val dataset I put 1 image of mine for testing. So it was recognizing me fine. But when put some other person pic in val dataset, it was still recognizing it as me
I have trained on 30 classes with
45 images in Train folder and
15 images in Test folder (val)
after this upon testing with a new image which belongs to a class
im getting good results:
Image A – class A (99.996 %) which is correct
Image X – class A (99.996 %) it belongs to an unkown class to the model but still it says that it belongs to class A with extremely high confidence.
Any thoughts on why this occurs ??
Thanks for your reply.
Could you please explain or guide to towards the direction of
“””You might need to train the model on “you” vs “not-you”, or people vs unknown.”””
So when we train the model, do i put a unkown folder?
like train folder :
class A (30 images)
class B (30 images)
unkown ???.
Sorry if this doesnt make sense, its a bit hard to understand what you mean by train the model on “you” vs “not-you”.
Yes, if your goal is to fit a model that can detect you, and when you show it photos of other people, it knows it’s not you, then you need a dataset of lots of photos of you and photos of other people (unknown group) so the model can learn the difference.
Highly appreciated for your assistance.
What if we have more than 2 classes.
Training dataset has 5 classes.
During testing if we feed unseen face of one of the above classes, it may predict the face as one of the class.
But what if we feed faces of unknown class? I would expect the model to predict unknown.
So how we can accomplish this scenario?
AttributeError: ‘JpegImageFile’ object has no attribute ‘getexif’ or
AttributeError: module ‘PIL.Image’ has no attribute ‘Exif’
This error occurs when i use the Image.open command to import the image.
few examples for the images i am getting an error are as follows:
httpssmediacacheakpinimgcomxfecfecaefaadfebejpg.jpg (traing data for elton john)
httpwwwjohnpauljonesarenacomeventimagesEltonCalendarVjpg.jpg (training data fro elton john)
I tried searching for this issue online but was not able to find any helpful solution. Do you have any idea how i may solve this issue?
Dear Jason,
How can we create the face embeddings and fit the classification model without re-training the whole training dataset? I would be grateful if you can share me the code.
Why is the problem of retraining the whole training dataset? Usually the problem is training the model from scratch, which takes a long time to converge. But if you are based on a network trained for one particular purpose and retrain it for a different use, it would not take too long.
@Adrian
When the dataset is big (e.g. in millions of classes) and it is regularly increases, how it is possible to retrain the whole dataset every time? For example my initial dataset contains 2000 classes/subjects and every day it increases by 300 subjects. How can I proceed without retraining the existing model?
Thank you
Usually the first half of the network need not be retrained if your output classes are changed. But if you need to change the output layer from size 2000 to 2300, you need to retrain it anyway. You might think of some other ways make it smarter, e.g., always set one of your output as “not matched” and try to train the network to tell if the face is not any of the known one. Then you can always create a new network for the increased subjects.
Hey Jason,
First of all thank you so much for putting out the effort and organizing this tutorial! You’re truly awesome! 🙂
So I extracted facial embeddings of 3 people(6 high-quality high-resolution 5MP+ images per person as dataset )and trained them using SVM and used my built-in DELL WEBCAM(need I mention it generates a mirror image , ie my left hand appears right on the screen; also it’s a 0.3 MP 640×480 resolution feed) to identify faces.
So my problem is that the probabilities are always different for the same trained face by sometimes a difference as great as 20% under the same lighting conditions! It’s mostly around 71% but sometimes dwindles to 51% for a trained face. For a stranger face it varies between 40% and 68% hence because of this variation, I can’t set a single probability value as a threshold and it’s really cumbersome.
Can these differences be because of the difference in webcam quality and the dataset quality, that the algorithm has a tough time identifying the faces and varies the probability all the time, given the embeddings generated by the dataset are of much higher quality than those of the feed and also that the feed generates a mirror image of the faces in the dataset?
I’ll do just that. Any idea how to downsample? A friend tried with same dataset but with Logitech C310 HD webcam and got a consistent probability score .It’s unlikely it’s the light level in my case as it shows variations in probability at the exact same light conditions.
Okay. One quick thing.
Does both the dataset and webCam feed have to be of the same quality? Cuz I trained my face, Emma Watson and Daniel Radcliffe’s faces (their images size around 5 kb) and my image quality around 70 kb and there’s still some variation in probability
Generally yes, but you can achieve the desired result by using data augmentation either during training, during testing, or both to get the model used to data of varying quality – or to downsample all images to the same quality prior to training/using the model.
In addition to previous suggestions, you can also limit what is analyzed. For example only take face detections of very high confidence. You can use the face detection data to reject highly oblique angles (like if the nose is further left than the left eye, as one of many examples).
Not shown in this tutorial – but is very easy, is to increase the quality of the input by growing the face detection box vertically or horizontally to make it square… do this before resizing to low res. This prevents stretching of faces.
You may also have a shortage of data for your known faces. If possible try to grow that dataset. Furthermore if your dataset is too sparse (you have only a few known faces) you may have trouble because of the SVM maths. It may be beneficial to litter your known faces with a moderately sized dataset from the web (they would need to be caught and handled in code – you could call them JaneDoe1, JaneDoe2, etc). By doing this the SVM should have more cases of ambiguity whereas with only a couple known faces it may have displayed inappropriate confidence.
sir, My problem is my model is not able to distinguish between a known and unknown person in the real world.
Do you have any idea about how to identify an unknown person in the real world?
I guess it’s not really “data augmentation” when 5 out of the 6 images for Daniel radcliffe are 6KB, the last 67…and my face image quality are on average 120 KB, whereas for Emma Watson 2 out of the 3 images are 7 kb and the last 70. (The images generated by the webcam feed, the “test set” are 70 kb.). I guess both the dataset and feed should be same baseline image quality right?
By creating a data set of 500 people of 50 images each and train the model, can I expect good accuracy regarding detection?
Can I try deploying the same model on a Raspberry Pi with a Pi Camera?
Can you suggest any idea about adding a new person’s face to the model once it is deployed?
Hi Jason
i run the code and add two of my own photos in both train and val dataset, SVM predict show correct class_index for my own photos in val dataset, but the SVM predict_proba show probability as below: class_index is 2(0.01181915, 0.01283217), it is the smallest value.
[0.14074676 0.21515797 0.01181915 0.10075247 0.15657367 0.37494997]
[0.1056123 0.20128499 0.01283217 0.1100254 0.23492927 0.33531586]
i see documents saying that SVM predict_proba show meanlingless result on small dataset, is it caused by that? How can i detect one-face class probability?
second question: can you show more code on how to train unknown people class?
I went through your tutorial and I got 100 train and test efficiency. Till now everything is clear.
But the problem arises when I apply the developed model (through your tutorial) to live frames through a webcam (my train ~700 images/class and test images ~300 images/class are captured using webcam).
The model does more misclassification when I apply the trained model to a frame from a webcam.
I am not sure how to normalize the face embedding vector in this case?
Thanks for the reply. It means I should normalize the input image rather than the embedding. If the input image is normalized then I don’t need to normalize the embedding.
I can’t figure out how to introduce the new unsorted pictures into the code. I tried making an npz file using the new pictures in one folder and loading them into the classifier(# load faces
data = load(‘5-celebrity-faces-dataset.npz’), but the classification result was pretty bad so im assuming what i did is not correct.
In following the your tutorial my directory for the pictures looked like this:
├── train
│ ├── A : pictures of person A
│ ├── B : pictures of person B
│ └── C : pictures of person C
│
│
└── val
├── A : pictures of person A
├── B : pictures of person B
└── C : pictures of person C
and got a successful result, with let’s say
“ABC_dataset.npz” & “ABC_embeddings.npz”
So, I’m trying to one step further, and added a folder to the directory
├── train
| ├── A : pictures of person A
| ├── B : pictures of person B
| └── C : pictures of person C
|
|
├── val
| ├── A : pictures of person A
| ├── B : pictures of person B
| └── C : pictures of person C
|
└── test : pictures of persons A, B, C
and the newly added test folder contains pictures of all A, B, C.
In an attempt to introduce the data from the “test” folder,
I extracted arrays of the faces from the pictures of the “test” folder saved the
extracted arrays into an npz file, let’s say “ABC_test_dataset.npz”
And loaded “ABC_test_dataset.npz” into the last part of the tutorial
# load faces
data = load(‘”ABC_test_dataset.npz”‘)
testX_faces = data[‘arr_2’]
# load face embeddings
data = load(‘ABC_embeddings.npz’)
trainX, trainy, testX, testy = data[‘arr_0’], data[‘arr_1’], data[‘arr_2’], data[‘arr_3’]
and so on.
When I tried this, the result I got was pretty bad so I’m assuming what I did is
a wrong way of introducing new dataset into the code.
What if the cause of poor performance is the specific pictures in test? What if you swap around some of the pictures from test with those in train and see if that lifts skill?
What if you confirm the pipeline and add pictures from train into test and confirm they are predicted correctly – they should be?
Thanks for the answer!
I will play with the picture data sets a bit more and tell you how it goes.
Last questions,
1. Is it ok to use gray-scale pictures for the training?
2. What should be the ratio between the number of pictures in the train and val
folders?
For example, train : 1000 pics & val : 100? or train : 1000 & val : 50 is fine?
Hey Jason!
I am trying to develop an attendance collector from video footage.
My problem arises during the classification part,it constantly interchanges the output label names.
Say X and Y are two people then X is always identified as Y and vice versa.
The code is error free and there is no error in the input labels.
How can correct this.Will this be solved if use something apart from SVM. If so what?
Or should i do some fine tuning as specified in one of your earlier answers?
Please guide me.
– First, extract only the face from an image using MTCNN. Please make sure to scale pixels.
– Then find the face embedding using the Facenet model.
– Train binary/multiclass classification model using face embedding vectors obtained in the above step.
– So now you have face embedding vectors for train images and compare them with unknow encoding (a frame from webcam) using Consine or another metric.
– If score < threshold then face is recognized and in order to get a label, predict the class using binary/multiclass classification model trained in step 3.
Hey Jason!
I wanted to try doing image super resolution using GANs to improve the result of my recognition process,but i’m not finding any suitable website or blog to learn from.
Could you please guide.
I prefer using Python for the same.
Thank You!
This model doesn’t able to differentiate between a live video of human and a picture of a human. Bcoz this model looks for the only face, it doesn’t look for the Liveness in the face.
This is another research area in my view. Perhaps for the quick try, you can look whether eyes are blinking or a person is speaking to detect the Liveness. There are other techniques too.
Hi jason, I am only looking at algorithm to tell if an image has face or not, face can be far, at any angle . Preesntly neiter open cv, nor any model works well, Do you know any transfer learning method to only detect if image has face or not with lot of other things in background, Some images has only body but no face, in those case it should say as no face.
Perhaps try detecting and extracting the face first, then pass to a pre-trained model to get a face embedding. Perhaps try a suite of models to see which is the best for your dataset – also fit different model types on the embedding to perform the recognition.
I am trying to classify 800 celebrity faces.
1-using MTCNN to extract faces
2- using facenet to prepare the embedding
3- using KNN to classify the faces.
Problem is-Though we get neglible ‘not detection’ but we get lots of wrong detection.
Is there any solution to reduce the wrong detection.
Other combination of pipeline used-
1-MTCNN+Facenet+SVM
2-One shot learning
3-Also tried regularisation,data augmentation.
(used 100 faces per face)
Hey!
In the Keras FaceNet Pre-Trained Model (88 megabytes) you have mentioned, how should it be downloaded, it has 2 files models and weights and each has “.h5” file. Could you please tell me which has to be downloaded.
I have tried both, but the code is not getting executed, it is stopping after load_model()
There are 2 models. The one that gives you a face embedding and one that classifies embeddings as people. The first model does not need to be retrained. The second model is only trained once and is then used to make predictions for people that it knows about (e.g. during training).
Hello
When I make a test for one image of one person, in every time give different result of probability. Is that right? Or should it give me the same result?
This was great thanks. I am wanting to know how to change the randomisation section to score the entire test dataset. Not clear as it appears it was built to score only on a single, randomly selected index from that array. Any guidance on this?
Hi Jason it’s a good tutorial, but when I try to run this tutorial in tensorflow 2.0 the mtcnn is not compatible with tensorflow v2.0. do you have any idea for this issue.
thank you
Sorry for the delay.
I want to make a faceprint system. So, when people come to register, I take one picture for each person and make some transformation in the picture. Will it correctly predict the person?
Hi Jason I used your code on my own set.In all there were about 20 people each having approximately 30 images.
For testing I decided to see how the code performs in real time. If a random person walks in front of a web cam then will it be able to distinguish between UNKNOWN and some other person in the dataset.
Although I got pretty good results,(each of the person in my set is being correctly identified with a probability of 99%),the unknown people were also being assigned a score of 99%.
From my understanding,the unknown category shouldn’t recieve such a high score.Do you know what is going wrong here.
Training and cross Val error are coming out to be 1 and 0.99 .So I am not sure.Do you think I should reduce the dataset size.
I have used the same code to detect the faces from the my own dataset. I am unable to store the dataset in npz format. My dataset has images around Train-750 pics and val – 400.
The python code is getting killed after certain time . Let me know the workaround
System Specification:
RAM – 16 GB
Swap – 2 GB
OS – Ubuntu 18.04
Please disregard my comment – if your code fails when importing load_model from keras.models then it’s maybe an installation or setup problem. In my case I thought TF2 had keras but nothing was working right until I installed keras after installing TF2
Thanks so much for this Jason Brownlee… I got everything working because of this walkthrough.
I also converted the h5 to tflite which is just one commandline command. Additionally you use the interpreter instead of calling “predict” directly on the model… which involves about 8 lines of additional code, found on Tensorflow’s website. By doing this the embedding is obtained 10 times faster (maybe more). I’m on CPU only btw.
Hi Jason, been following you for the past 3 years and you are very inspiring. Just wanted to say, I tried this code in Colab and eventually hit errors. Are there any alterations when using Google Colab? Thank you again.
Quick question for you, Are you using the same code for the predication or for training classification using face embeddings on nvidia driver supported GPU?
Can you let me know the procedure ?
Since i read in the scikit library documentation, it is published that SciKit lib will not support GPU training.
Hi Jason,
I am getting this warning whenever I load facenet_keras.h5
UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.
warnings.warn(‘No training configuration found in save file: ‘
Hi again,
please guide me how to handle the unknown faces in the input.
The model is working fine when it gets the known faces,
But when an unknown face comes, the model is classifying them to the classes on which it was trained.
Example- model is trained to detect me but when I showed it some other face it is again saying that’s you.
I have been getting not so properly cropped images from the MTCNN detect_faces method. And, I am doing it for a short term project and can’t really get into details inside MTCNN. Just wanted to know why my face recognition is not as per published in this blog. I have used the same datasets and I have followed all instructions thorough.
I have a database with only one photo per person. Does the model have good results in this scenario as well?
I don’t see how to make a classifier from just one sample per class. Is there how or should I calculate the distance between the vectors directly (without a classifier)?
I have implemented this FR system but I’m having a problem understanding why we are using the SVM classifier. I don’t have specific questions because I think my problem is in my formative understanding of FaceNet. Kindly tell me if the following statements are true:
1) The FaceNet model, once loaded and we pass the detected face “predicts” an embedding of the face.
2) We then compare the embedding with the stored embeddings of faces in our dataset using the SVM’s predict function and determine who is the likely owner of the face.
3)Is there a way that we can implement FaceNet without having to use a classifier in the last stage?
The neural net model gives features from an image and the model, e.g. svm, classifies them as different people.
Yes, but it would be fragile – requiring the entire facenet model to be re-trained for any change to your dataset. Or perhaps just the output/classifier part of the model as we would use transfer learning.
Hey Jason,
The tutorial is very specific and easy to understand, a very big thank you for that.
But I wanted to use an image in my local directory instead of a random selection.
According to the comments above I understand that I had to read https://machinelearningmastery.com/how-to-load-and-manipulate-images-for-deep-learning-in-python-with-pil-pillow/ and I was supposed to pass it to “selections” variable.
But I got really confused because selections has just 1D value and I guess I’m supposed to convert my image to a numpy array and then pass it on to selections right?
So what should i do after converting it to a numpy array.
Yeah so I tried everything but I can’t seem to be able to insert an image from local into “selections” could you please help me out LOL really need it for my project
I am having same issue and was wondering how Rujul solved it. I am using Python 3.8.3 and using load_image from keras (version 2.4.3) and not tf.keras. Any help would be greatly appreciated.
Thank you Jason!!! i was able to do that as well, one thing i have noticed, sometimes it confuses faces with someone else, lets i have a trained set for ‘Jimmy Fallon’ so sometimes in a video it detects a random person and says its Jimmy Fallon and that too with more than 90% probability is that normal???
I tried implementing the model but the accuracy obtained on the 5-Celebrity-dataset is nearly 75% and similarly I created my on custom dataset with around 15 images each training class and 5-8 images in test class , the test accuracy obtained was just 61% and I tried increasing the size of the training dataset to nearly 30 images in each dataset but the test accuracy did not increase it remained the same. What should I do to improve the accuracy ?
1- In the paper, it said that facenet layers equal to 140M, but the model that you use it equal to 22,808,144. what is the difference?
2- How can I make a good embedding function that fit my model? Because In the embedding function in facenet and vggface2 there is a difference and I do not understand it
Hi …
When I want to calculate the accuracy of the facenet model, it should the test-data contains known and unknown labels? Because I tried to put labels that not in the dataset, It gave me an error
I tried to put unknown faces and labels, but when I encode the labels, It gives me an error that the dataset labels do not contain the label that I want to test it.
the error in this line:
testy = out_encoder.transform(testy)
You only know whether a prediction is true when training the model and when evaluating the model. After that – when you are using your model in practice, you don’t know if individual predictions are correct or not. All you know is the expected performance of the model on average.
I change by below code, that will read by photo EXIF Orientation tag. Hope it helps.
original code: image=Image.open(filename)
photo alignment by exif: image = ImageOps.exif_transpose(Image.open(filename))
please I need too much an algorithm or tutorial to compare a photo of face with another and have the percentage of similarity after the possibility of comparing a single photo with thousands of photo of faces to have the percentage of similarity
Hii Jason,
I have implemented this face net model to train 5 person’s dataset and the accuracy is also too good but it takes very long time for recognition. Can you give any advice to reduce the computation time for recognition like running the code on gpu or any other classification algo that takes less time.
Firstly, thanks for such amazing tutorials. Your website has been really helpful. So my question goes like this – how can I input multiple face arrays to vggface2 or FaceNet? In both the examples you have posted, you have used the function extract_face which outputs a single face.
Each face is one sample, you can pass in multiple samples to a model as an array and get many predictions in return. Each face must be prepared first before being passed to the model as input.
Helo , your explanation helps me much ! ty for this tuto , well i’m a beginner and i find it usefull and clearly , but i have one question plz if u don’t mind !
i’ve tried to make some modifications on ur code in the final to make it display all the test data instead of just one random picture but i couldn’t , so plz can u write for me the instructions that makes me test model on all the test dataset instead of one exemple ?
This is really a good example and I manage to run it and predict from it. But I have really one concern here, It’s good that we’ve predicted on the test data but what if I want to predict it from any external image? or my video cam image? it seems to me too cumbersome to process the raw/cam images to that level (embedding and so on). Could really have been great if you’ve taken the examples on some external images rather than same test data.
I used MTCNN to detected faces and it worked fine. But now, when I run the same code on the same photos, it doesn’t detect faces anymore. Just returns an empty list as a result.
Thanks for the article! This is one of the very few codes on the topic that is explained step for step. It’s a great guide. I wanted to test the model with a single or multiple new images which are not a part of the original dataset. How can I do that? My efforts have not yielded results. I would appreciate if you could tell me the process or the functions that need to be called/modified for the same.
Another thing is, this seems to work for a limited dataset. I can’t figure out why but when a larger dataset is taken, the shape of the Numpy array changes when loading dataset (function: load_dataset) and the consequent functions since the changed structure of trainX, trainy, etc. is not compatible with the following functions.
Hi Jason and sorry for my trivial question but I have a problem when I load the model:
model = load_model(‘facenet_keras.h5’)
this give me the following error:
bad marshal data (unknown type code)
I suppose that the problem is generated because the facenet_keras.h5 was generated with the 1.x version of tensorflow (while I am usintg the 2.2). Infact to correctly import the method I wrote:
from tensorflow.keras.models import load_model
but when I load the model I have the error mentioned above, please have you some suggestion?
I believe the problem is because you are trying to load the model using tf.keras instead of the standalone Keras library, not the version of tensorflow.
Just to add. Rest all versions being same, for me, it worked in python 3.6, but didn’t work on Python 3.8. Did not test on Python 3.7. So it is python version issue and not that of libraries.
Great tutorial.
Was looking at whether Transfer learning, Siamese network and triplet loss approaches are applicable to animal face(eg a sheep, goat etc) recognition particularly mobileNet(or otherwise) when your crystal clear blog came up.
Kindly shed more light on its applicability and any other auxiliary hints.
I don’t see why not.
How do I solve this error ???
ValueError: Input 0 of layer Conv2d_1a_3x3 is incompatible with the layer: : expected min_ndim=4, found ndim=2. Full shape received: [None, 128]
Sorry to hear that you’re having trouble, some of these tips may help:
https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me
Sorry Sir! but it’s not helping me out because I’m creating my own datasets of 15 people so that’s why facing that error can you help me about what changes should I change in your code ??? and how do i train my own new facenet model ????
Perhaps start with the above tutorial that works, and then adapt it for your own dataset.
Start by ensuring the existing tutorial works on your environment.
Then ensure you have loaded your dataset correctly.
Finally, adapt the example to use your loaded data.
Hi Jason,
This is fantastic, thanks for sharing.
What do you suggest when we have tens of thousands of classes.
A Facenet model itself as a classifier or a specific classifier model is to be trained. In terms of scalability and performance which is the preferred method.
Referring to:
“The FaceNet model can be used as part of the classifier itself, or we can use the FaceNet model to pre-process a face to create a face embedding that can be stored and used as input to our classifier model. This latter approach is preferred as the FaceNet model is both large and slow to create a face embedding.”
Good question, the facenet embedding approach is a good starting point, but perhaps check the literature for more scalable approaches.
Hi jason,
As per my understanding The triplet loss is used so that the model can also learn the dissimilarity between the classes rather than only learning similarity between same class.
But here we are not training our model on the actual image dataset on which we need our classification to be done. Rather we are using SVM for that.
So, how can we make use of triplet loss in this method of face recognition?
See this post:
https://machinelearningmastery.com/one-shot-learning-with-siamese-networks-contrastive-and-triplet-loss-for-face-recognition/
Hi Jason,
I want to try this on Cat and Dog dataset. What do you thing the pre-trained networks face embeddings will work in this case?
No, I don’t think it will work for animals.
Interesting idea though.
Thanks for your reply.
What do you think how much effort will it take to train facenet from scratch?
And certainly how much data?
Not off hand, sorry.
Thanks for the response Jason.
Hi Jason,
Can MTCNN detect faces of cats and dogs from image?
I don’t see why not, but the model would have to be trained on dogs and cats.
Hi Jason,
Thanks for a very nice tutor. But i cant set up mtcnn in my python2? Is there a way to install mtcnn for python2?
I don’t know, I have only used it with Python 3.
Hi Jason. nice tutor!
But i wonder that, if i want to identify who is stranger. should i make a folder for ‘stranger’ contains alot of stranger faces???(exclude your 5 Celebrity Faces Dataset ??)
Good question.
No, if a face does not match any known faces, it would be “unknown”. E.g. has a low probability for all known faces.
Thank you very much. Keep writting tutorials to help thoudsands of thoudsands people like me a round the world learning ML,DL. <3
Thanks, I’m grateful for your support!
yeah! I figured out that MTCNN only require Python3.
Follow this bellow: https://pypi.org/project/mtcnn/
Nice.
Thanks for your tutorial.
But when i flollow you, i have a warning :
embeddings.npz is not UTF-8 encoded
UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.
Do you know how to fix it.
You can safely ignore these warnings.
Thank you so much.
But when I have a new image to recognize, do I need to put it to the validate folder and rerun the code ?
And how can we use this to recognite face in a video ?
You can use the model directly, e.g. in memory.
Perhaps you can process each frame of video as an image with the model?
Hello Jason,
Thanks for your wonderful tutorial, I’d like to know what is the best solution to apply recognition part if I have a very small dataset -only one face per each identity- in this case, I think SVM wouldn’t help.
I think a model that uses a face embedding as an input would be a great starting point.
Hello Jason,
Thank you for this amazing tutorial, I used python 3 to run this code. I would like to know why am i getting this error (No module named ‘mtcnn’) and how can I correct it?
The error suggests you must install the mtcnn library:
I want to use transfer learning for masked face recognition. But i didn’t found any better masked face recognition dataset. I need a masked face dataset with proper labeling of each candidate. So is there any better masked face dataset available? where can i find this dataset?
Perhaps you can take an existing face dataset and mask the faces using opencv or similar?
Thanks for response. Can you refer any work or blog like your for doing mask face using opencv or similar?
Sorry, I do not have a tutorial on this topic, perhaps in the future.
Hello Jason, great tutorial.
Im beginner in python.
I try to understand your code, and little bit confusing when you choice random example from dataset when doing classification
in line 28. selection = choice([i for i in range(testX.shape[0])]),
its choose random vector value in testX.shape[0] from embeddings.npz right?
so how if we want using spesific image from val folder?, Can you refer any work or blog to doing this
Thanks.
Yes, you can see how to load an image here:
https://machinelearningmastery.com/how-to-load-and-manipulate-images-for-deep-learning-in-python-with-pil-pillow/
Kindly share the code of using external images
Thank you so much for the respons,
well I tried and it worked.
But I have another question, because sometime when i run the code, all worked perfectly and when I run the code again, sometime i have this error warning in load_model part although the face recognition still work
UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.
warnings.warn(‘No training configuration found in save file: ‘
why did this happen?
Thanks.
Well done!
You can safely ignore that warning message.
Hey @AI – how did you do the “image loading” part instead of randomly choosing an array from validation? could you please share the code? Thanks!
Hi Jason， while I was executing the code “load_model(‘facenet_keras.h5’)”, the exception “tuple index out of range” is thrown, can you tell me why? thanks in advance.
Sorry, I have not seen that error. I have some suggestions here that may help:
https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me
Hi Jason, again wonderful article and tutorial you provided to us. I wonder how I can customize dataset for my needs such as my friends dataset and perform training on it?
Thanks Jason, really helpful as always but I got a weird “invalid argument” error. But I fixed it by changing ‘f’ to ‘F’ in facenet_keras.h5 because I notice it couldn’t recognize character ‘f’. Maybe because it’s trained on Ubuntu but I run your code on Windows 10. I don’t know!
Nice work!
Thank you! Dear Jason, could you please tell me how I can get access to other properties of model. I mean I don’t need model.predict. I need other properties. Is there a way to list all of them such as different convs or avgpool. I tried __dict__ and dir() but they don’t give what I want. For example, how did you know model has a property called “.predict”? Where can I find all of them? Thank you!
You can access all the weights via get_weights()
Jason Can you Please post a tutorial on how to convert David sandberg tensorflow model in keras using Hiroki Tanai script to convert it into keras
Thanks for the suggestion.
How can I convert this script to tensorflow lite format in order to be used in an android applicaton?? Pleaseeee helpp !!
Sorry, I don’t have experience with that transform.
This will work for many hundreds of people?
I don’t see why not.
Hello Jason,
Thanks for sharing the interesting article!
I have read your two articles on Face Verification: 1) this one and 2) https://machinelearningmastery.com/how-to-perform-face-recognition-with-vggface2-convolutional-neural-network-in-keras/
Which one would you suggest? If I have to develop Face Verification system then there are few approaches (listing two approaches from your article):
Approach 1: Detect face using MTCNN, train VGGFACE2 on the collected dataset which helps to predict the probability of a given face belonging to a particular class
Approach 2: Detect face using MTCNN, get face embedding vector using facenet keras model and then apply SVM or Neural Network to predict classes
Which approach would you recommend? Can you please explain?
Thanks for sharing views.
Perhaps prototype a few approaches and see what works well for your specific project and requirements?
It means, I can try both approaches and have a look at efficiency, and select an approach with the best accuracy.
Thank you!
You’re welcome.
Hi,
I am looking for Speech recognition tutorial on Deep Learning using Keras.
I have gone through your this URL: https://machinelearningmastery.com/category/deep-learning/ but I couldn’t find any tutorial.
Could you please point to the tutorial link (if you have)?
Thank you!
Sorry, I don’t have tutorials on that topic, I hope to cover it in the future.
Thanks for the tutorial.
Unit length normalization isn’t for SVM. For SVM you typically use range scaling – MinMaxScaler, or standardization – StandardScaler. The goal is to make different features uniform. Actually, it’s a surprise that unit length normalization produced 100% accuracy in your case. That’s probably due to small data. It does not work for SVM in general and didn’t work for me.
Thanks for your note.
I followed best practices when using face embeddings from the literature.
Hi Jason,
Great article. You have explained all the necessary steps to implement a face recognition system. I am working on a similar problem but in a bigger scale. I am in a belief that a classification based face identification is not a scalable solution. Please give me your opinion.
If I want to recognise a thousand faces in real time manner then, what type of changes do I need to make to your implementation.
I believe it would be really helpful if you create an article about large scale face recognition.
Good question, perhaps an efficient data structure like a kdtree for the pre-computed embeddings?
Can we extract eyes part out of the extracted face using mtcnn detector?? Any help..
I don’t see why not.
It will find them with a point, you can draw a circle around that point and roughly extract the eyes.
You might have to scale the faces/images to the same pixel size prior to modeling and extraction.
Let me know how you go.
bonsoir monsieur pouvez vous me dire cette application peut fonctionner avec une interface tkinter qui affiche le nom et prénom et la photo reconnu
Sorry, I don’t know about “tkinter”.
Hi Jason,
1 . Here we are using 5 faces, what if we have thousands of faces, how to get the identity or index of those faces.
2. If we have fingerprints or voice which pretrained model would be most suitable.
I don’t see why not.
You may need a different model for fingerprints/voice.
If I have thousands of faces, SVM takes a lot of time. What do I do to get a quick result?
Perhaps try as simpler linear model?
Perhaps try running on a large EC2 instance?
Perhaps try running the code in parallel?
…
Once i have trained the model on 5 class (each class having 50 images). Now i use the model to detect images that it has not seen, it correctly guesses that the person in the image is class A for example with an accuracy ( prediction ) 65%. Is it possible to now add such image back to training and expect to get better results ?
Yes.
Awesome post, thanks for sharing.
Thanks!
Hi Jason
Thanks for this tutorial. Its really helpful. I wanted to know why you used train and val dataset. I mean are these two used for training purpose. What is the use of val here.?
In the face classification, I am not able to understand where are you selecting the random photo to test against your dataset. How can I add my jpg photo to test again the dataset. Can you explain please. Thanks
Here, the validation set is a hold out or test set.
We fit on the train set, then confirm the model has skill by evaluating it on the test set (val).
You could add a new directory to train and val sets, to train and evaluate the model for new people.
Got it. Thanks
What I have seen is that in train dataset I put my pictures more than 30 images and in val dataset I put 1 image of mine for testing. So it was recognizing me fine. But when put some other person pic in val dataset, it was still recognizing it as me
Any idea how can this be solved
Yes, even i was wondering this.
I have trained on 30 classes with
45 images in Train folder and
15 images in Test folder (val)
after this upon testing with a new image which belongs to a class
im getting good results:
Image A – class A (99.996 %) which is correct
Image X – class A (99.996 %) it belongs to an unkown class to the model but still it says that it belongs to class A with extremely high confidence.
Any thoughts on why this occurs ??
You must define the problem and model the way you intend to use it.
If you want the model to classify unknown people as unknown, you must give examples during training.
You might need to train the model on “you” vs “not-you”, or people vs unknown.
Thanks for your reply.
Could you please explain or guide to towards the direction of
“””You might need to train the model on “you” vs “not-you”, or people vs unknown.”””
So when we train the model, do i put a unkown folder?
like train folder :
class A (30 images)
class B (30 images)
unkown ???.
Sorry if this doesnt make sense, its a bit hard to understand what you mean by train the model on “you” vs “not-you”.
Help would be appriciated.
Thanks
Yes, if your goal is to fit a model that can detect you, and when you show it photos of other people, it knows it’s not you, then you need a dataset of lots of photos of you and photos of other people (unknown group) so the model can learn the difference.
Highly appreciated for your assistance.
What if we have more than 2 classes.
Training dataset has 5 classes.
During testing if we feed unseen face of one of the above classes, it may predict the face as one of the class.
But what if we feed faces of unknown class? I would expect the model to predict unknown.
So how we can accomplish this scenario?
You might need to add an “unknown” class during training, or add some if-statements to interpret the predicted probbilities.
Hi Jason,
Thanks for the code. it is very helpful.
For a few images i am getting a error as follows
AttributeError: ‘JpegImageFile’ object has no attribute ‘getexif’ or
AttributeError: module ‘PIL.Image’ has no attribute ‘Exif’
This error occurs when i use the Image.open command to import the image.
few examples for the images i am getting an error are as follows:
httpssmediacacheakpinimgcomxfecfecaefaadfebejpg.jpg (traing data for elton john)
httpwwwjohnpauljonesarenacomeventimagesEltonCalendarVjpg.jpg (training data fro elton john)
I tried searching for this issue online but was not able to find any helpful solution. Do you have any idea how i may solve this issue?
Thanks
I’m sorry to hear that, I have not seen those issues before.
I have some suggestions here that might help:
https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me
is it possible in the real-time after training?
I don’t see why not.
please write a programme about how it works in real-world after training?
Do you mean making a prediction after it is trained?
If so, the tutorial above shows that.
Otherwise, can you please elaborate what you mean exactly?
Hi Jason,
Your content is so much helpful.
What about a system where new faces can be registered? Would I have to retrain the model, containing the new classes (new faces)?
Thanks.
No, just create the face embeddings and fit the classification model.
Dear Jason,
How can we create the face embeddings and fit the classification model without re-training the whole training dataset? I would be grateful if you can share me the code.
Why is the problem of retraining the whole training dataset? Usually the problem is training the model from scratch, which takes a long time to converge. But if you are based on a network trained for one particular purpose and retrain it for a different use, it would not take too long.
@Adrian
When the dataset is big (e.g. in millions of classes) and it is regularly increases, how it is possible to retrain the whole dataset every time? For example my initial dataset contains 2000 classes/subjects and every day it increases by 300 subjects. How can I proceed without retraining the existing model?
Thank you
Usually the first half of the network need not be retrained if your output classes are changed. But if you need to change the output layer from size 2000 to 2300, you need to retrain it anyway. You might think of some other ways make it smarter, e.g., always set one of your output as “not matched” and try to train the network to tell if the face is not any of the known one. Then you can always create a new network for the increased subjects.
Hey Jason,
First of all thank you so much for putting out the effort and organizing this tutorial! You’re truly awesome! 🙂
So I extracted facial embeddings of 3 people(6 high-quality high-resolution 5MP+ images per person as dataset )and trained them using SVM and used my built-in DELL WEBCAM(need I mention it generates a mirror image , ie my left hand appears right on the screen; also it’s a 0.3 MP 640×480 resolution feed) to identify faces.
So my problem is that the probabilities are always different for the same trained face by sometimes a difference as great as 20% under the same lighting conditions! It’s mostly around 71% but sometimes dwindles to 51% for a trained face. For a stranger face it varies between 40% and 68% hence because of this variation, I can’t set a single probability value as a threshold and it’s really cumbersome.
Can these differences be because of the difference in webcam quality and the dataset quality, that the algorithm has a tough time identifying the faces and varies the probability all the time, given the embeddings generated by the dataset are of much higher quality than those of the feed and also that the feed generates a mirror image of the faces in the dataset?
Hope this isn’t too much trouble 🙂
forgot to mention, the variations in probability happen whenever I run the program on different occasions
Could also be light level, etc.
Try data prep that counters these issues?
Yes, it is likely the image quality.
Perhaps downsample the quality of all images to the same baseline level before fitting/using models?
I’ll do just that. Any idea how to downsample? A friend tried with same dataset but with Logitech C310 HD webcam and got a consistent probability score .It’s unlikely it’s the light level in my case as it shows variations in probability at the exact same light conditions.
Yes, I show how to resize images here:
https://machinelearningmastery.com/how-to-load-and-manipulate-images-for-deep-learning-in-python-with-pil-pillow/
Thank you for your prompt replies!
Also, can mirroring the feed be the cause as I’ve mentioned my webcam does that?
Probably not related.
Okay. One quick thing.
Does both the dataset and webCam feed have to be of the same quality? Cuz I trained my face, Emma Watson and Daniel Radcliffe’s faces (their images size around 5 kb) and my image quality around 70 kb and there’s still some variation in probability
Generally yes, but you can achieve the desired result by using data augmentation either during training, during testing, or both to get the model used to data of varying quality – or to downsample all images to the same quality prior to training/using the model.
In addition to previous suggestions, you can also limit what is analyzed. For example only take face detections of very high confidence. You can use the face detection data to reject highly oblique angles (like if the nose is further left than the left eye, as one of many examples).
Not shown in this tutorial – but is very easy, is to increase the quality of the input by growing the face detection box vertically or horizontally to make it square… do this before resizing to low res. This prevents stretching of faces.
You may also have a shortage of data for your known faces. If possible try to grow that dataset. Furthermore if your dataset is too sparse (you have only a few known faces) you may have trouble because of the SVM maths. It may be beneficial to litter your known faces with a moderately sized dataset from the web (they would need to be caught and handled in code – you could call them JaneDoe1, JaneDoe2, etc). By doing this the SVM should have more cases of ambiguity whereas with only a couple known faces it may have displayed inappropriate confidence.
Excellent suggestions!
sir, My problem is my model is not able to distinguish between a known and unknown person in the real world.
Do you have any idea about how to identify an unknown person in the real world?
You must train the model on known and unknown people for it to learn this distinction.
I guess it’s not really “data augmentation” when 5 out of the 6 images for Daniel radcliffe are 6KB, the last 67…and my face image quality are on average 120 KB, whereas for Emma Watson 2 out of the 3 images are 7 kb and the last 70. (The images generated by the webcam feed, the “test set” are 70 kb.). I guess both the dataset and feed should be same baseline image quality right?
Yes.
Thanks for the great tutorial,
By creating a data set of 500 people of 50 images each and train the model, can I expect good accuracy regarding detection?
Can I try deploying the same model on a Raspberry Pi with a Pi Camera?
Can you suggest any idea about adding a new person’s face to the model once it is deployed?
Perhaps try it and see?
Yes, compute face embeddings for new people as needed and re-fit the model.
Hi!
My question is that you train the network every time you want to recognise a face set.
How can we train once and run it multiple times ,say like a new face set every day.
Is it possible to implement this in the example you have coded?
No, the network is trained once. We use face embeddings from the network as inputs to a classifier that has to be updated when new people are added.
Hi Jason
i run the code and add two of my own photos in both train and val dataset, SVM predict show correct class_index for my own photos in val dataset, but the SVM predict_proba show probability as below: class_index is 2(0.01181915, 0.01283217), it is the smallest value.
[0.14074676 0.21515797 0.01181915 0.10075247 0.15657367 0.37494997]
[0.1056123 0.20128499 0.01283217 0.1100254 0.23492927 0.33531586]
i see documents saying that SVM predict_proba show meanlingless result on small dataset, is it caused by that? How can i detect one-face class probability?
second question: can you show more code on how to train unknown people class?
Yes, SVMs cannot predict good probabilities, you must calibrate the probabilities before they can be useful:
https://machinelearningmastery.com/calibrated-classification-model-in-scikit-learn/
You can add new “people” directly by computing their embeddings and using them in your model during training/testing.
Hai – I believe the code on this webpage is not accounting for the fact that docs say predict and predict_proba do not match. See this for example
https://github.com/scikit-learn/scikit-learn/issues/13211
Hello Jason,
Thank you again for sharing the nice blog.
I went through your tutorial and I got 100 train and test efficiency. Till now everything is clear.
But the problem arises when I apply the developed model (through your tutorial) to live frames through a webcam (my train ~700 images/class and test images ~300 images/class are captured using webcam).
The model does more misclassification when I apply the trained model to a frame from a webcam.
I am not sure how to normalize the face embedding vector in this case?
Could you please guide me?
Thanking you,
Saurabh
Perhaps start with simple image preparation from here:
https://machinelearningmastery.com/how-to-normalize-center-and-standardize-images-with-the-imagedatagenerator-in-keras/
Also perhaps try alternate models for learning how to differentiate the embeddings?
how to improve probability?
The tutorials here suggest how to improve deep learning model performance:
https://machinelearningmastery.com/start-here/#better
Hello Jason,
Thanks for the reply. It means I should normalize the input image rather than the embedding. If the input image is normalized then I don’t need to normalize the embedding.
Please feel free to correct me!
Thanking you,
Saurabh
Perhaps try scaling pixels prior to using the embedding.
Thank you Jason, it’s working now. Thanks for the kind help! Looking forward to grow in the Deep Learning era under your guidance.
Happy to hear that.
@saurabh How did you do it for webcam can you please explain me the procedure or share the code…….
Hello Jason, Thank you for your tutorials, is there a method i can use to implement on the raspberry pi 3 ?
kind regards
jada
Sorry, I don’t know about “raspberry pi 3”.
hi, thanks for the tutorial! it was really helpful!!
I have followed the tutorial and got successful result with my own data set of pictures.
Let’s say I used person A, B, C to trained the model
Now I’m trying to introduce a unsorted pictures of the above 3 people(A, B, C) in one folder and sort them
based on the code from your tutorial.
However, I can’t figure out how to introduce the new and unsorted pictures into the above code.
please help?
Thank you in advance!
Sounds straightforward, what is the specific problem you’re having?
I can’t figure out how to introduce the new unsorted pictures into the code. I tried making an npz file using the new pictures in one folder and loading them into the classifier(# load faces
data = load(‘5-celebrity-faces-dataset.npz’), but the classification result was pretty bad so im assuming what i did is not correct.
Perhaps start with the code in the tutorial and slowly adapt it for your specific dataset?
I think my explanation wasn’t clear enough…
In following the your tutorial my directory for the pictures looked like this:
├── train
│ ├── A : pictures of person A
│ ├── B : pictures of person B
│ └── C : pictures of person C
│
│
└── val
├── A : pictures of person A
├── B : pictures of person B
└── C : pictures of person C
and got a successful result, with let’s say
“ABC_dataset.npz” & “ABC_embeddings.npz”
So, I’m trying to one step further, and added a folder to the directory
├── train
| ├── A : pictures of person A
| ├── B : pictures of person B
| └── C : pictures of person C
|
|
├── val
| ├── A : pictures of person A
| ├── B : pictures of person B
| └── C : pictures of person C
|
└── test : pictures of persons A, B, C
and the newly added test folder contains pictures of all A, B, C.
In an attempt to introduce the data from the “test” folder,
I extracted arrays of the faces from the pictures of the “test” folder saved the
extracted arrays into an npz file, let’s say “ABC_test_dataset.npz”
And loaded “ABC_test_dataset.npz” into the last part of the tutorial
# load faces
data = load(‘”ABC_test_dataset.npz”‘)
testX_faces = data[‘arr_2’]
# load face embeddings
data = load(‘ABC_embeddings.npz’)
trainX, trainy, testX, testy = data[‘arr_0’], data[‘arr_1’], data[‘arr_2’], data[‘arr_3’]
and so on.
When I tried this, the result I got was pretty bad so I’m assuming what I did is
a wrong way of introducing new dataset into the code.
Sorry for the VERY LONG question.
Thank you!
Thanks for the elaboration!
I think you’re on the right track, well done!
What if the cause of poor performance is the specific pictures in test? What if you swap around some of the pictures from test with those in train and see if that lifts skill?
What if you confirm the pipeline and add pictures from train into test and confirm they are predicted correctly – they should be?
Let me know how you go.
Thanks for the answer!
I will play with the picture data sets a bit more and tell you how it goes.
Last questions,
1. Is it ok to use gray-scale pictures for the training?
2. What should be the ratio between the number of pictures in the train and val
folders?
For example, train : 1000 pics & val : 100? or train : 1000 & val : 50 is fine?
Thanks!
Sure. Try it and see.
Great question. I like 50/50 splits if there are enough data. I don’t want to be fooled by an optimistic result.
Perhaps run tests to see how sensitive the model/system is to training data size?
Hi, I’m planning on doing something similar. Please post your progress, while I work on it as well
It would be very helpful
Thanks
Hey Jason!
I am trying to develop an attendance collector from video footage.
My problem arises during the classification part,it constantly interchanges the output label names.
Say X and Y are two people then X is always identified as Y and vice versa.
The code is error free and there is no error in the input labels.
How can correct this.Will this be solved if use something apart from SVM. If so what?
Or should i do some fine tuning as specified in one of your earlier answers?
Please guide me.
Awaiting your reply.
Perhaps some of the labels for these people were swapped in the training data?
No that’s not the case.Its correct
You may have to step through your system carefully in order to debug it. I don’t think its something obvious, sorry.
Hello RK,
I am facing the same problem. I think there is a problem with the last model i.e. binary classification.
You should try comparing unknown face embedding with know face embedding.
I think this will help you!
Kindly share the output with me as I am facing a similar problem. I will also update you if I will make progress.
It might be related to the label encoder. The mapping of names to labels must be consistent across code examples.
You can force the consistency with arguments to the encoder, or simply save the encoder object via pickle.
Does that help?
Hello RK,
Finally, I got the solution.
– First, extract only the face from an image using MTCNN. Please make sure to scale pixels.
– Then find the face embedding using the Facenet model.
– Train binary/multiclass classification model using face embedding vectors obtained in the above step.
– So now you have face embedding vectors for train images and compare them with unknow encoding (a frame from webcam) using Consine or another metric.
– If score < threshold then face is recognized and in order to get a label, predict the class using binary/multiclass classification model trained in step 3.
@Jason, please feel free to correct me!
Thanking you,
Saurabh
Look great!
Jason has already written the blog. You can find it here: https://machinelearningmastery.com/how-to-perform-face-recognition-with-vggface2-convolutional-neural-network-in-keras/
Thanks!
Thanks a lot Saurabh!I will try implementing what you have proposed.Thanks a lot for helping.And i will also go through the blog.
Thank you once again Jason
Thanking you,
RK
Hey Jason!
I wanted to try doing image super resolution using GANs to improve the result of my recognition process,but i’m not finding any suitable website or blog to learn from.
Could you please guide.
I prefer using Python for the same.
Thank You!
Thanks for the suggestion, I hope to cover the topic in the future.
Can this model differentiate between a live video of human and an picture of a human shown in front of the camera
I don’t know, perhaps try it out?
If that distinction is important to your application, I believe you could fit a model to detect “non live” examples.
Hello Virinchi,
This model doesn’t able to differentiate between a live video of human and a picture of a human. Bcoz this model looks for the only face, it doesn’t look for the Liveness in the face.
This is another research area in my view. Perhaps for the quick try, you can look whether eyes are blinking or a person is speaking to detect the Liveness. There are other techniques too.
Will this Work for sketch matching with real picture of that person?
Perhaps try it and see?
can you recommend any other method which will be good for matching sketch with picture of that person?
Not off the cuff, perhaps start with some pre-trained models in order to create embedding and see what happens.
ok Thanks Alot Sir
Hi jason, I am only looking at algorithm to tell if an image has face or not, face can be far, at any angle . Preesntly neiter open cv, nor any model works well, Do you know any transfer learning method to only detect if image has face or not with lot of other things in background, Some images has only body but no face, in those case it should say as no face.
Perhaps try detecting and extracting the face first, then pass to a pre-trained model to get a face embedding. Perhaps try a suite of models to see which is the best for your dataset – also fit different model types on the embedding to perform the recognition.
I am trying to classify 800 celebrity faces.
1-using MTCNN to extract faces
2- using facenet to prepare the embedding
3- using KNN to classify the faces.
Problem is-Though we get neglible ‘not detection’ but we get lots of wrong detection.
Is there any solution to reduce the wrong detection.
Other combination of pipeline used-
1-MTCNN+Facenet+SVM
2-One shot learning
3-Also tried regularisation,data augmentation.
(used 100 faces per face)
Sounds like a great project!
Perhaps test different models?
Perhaps test different data preparation?
Perhaps review the errors and see if there is commonality?
Hi Parul Mishra,
Can you share the code for the second mentioned approach? It will be very helpful.
Implement ion of one shot learning .
Thanks
thanks for this amazing work
can i use this trained model in android application ?
You’re welcome!
Perhaps try it and see.
Hey!
In the Keras FaceNet Pre-Trained Model (88 megabytes) you have mentioned, how should it be downloaded, it has 2 files models and weights and each has “.h5” file. Could you please tell me which has to be downloaded.
I have tried both, but the code is not getting executed, it is stopping after load_model()
Download the model, not the standalone weights.
hey!
how can i detect all faces in the pic ? this one takes only the first face
You can adapt the example to operate on all faces. I don’t have the capacity to make this change for you, sorry.
thanks for the amazing work sir
Is it normal to get F1 score =1 ? and i expanded the dataset to be like 250 pics pf 15 persons
and im getting this :
accuracy: train=100.000, test=100.000
precision recall f1-score support
0 1.00 1.00 1.00 6
1 1.00 1.00 1.00 6
2 1.00 1.00 1.00 6
3 1.00 1.00 1.00 6
4 1.00 1.00 1.00 5
5 1.00 1.00 1.00 5
6 1.00 1.00 1.00 5
7 1.00 1.00 1.00 5
8 1.00 1.00 1.00 5
accuracy 1.00 49
macro avg 1.00 1.00 1.00 49
weighted avg 1.00 1.00 1.00 49
is it okay to be 100% correct ?
also what other algorithms can i use to make comparison with SVC ?
Yes, if the prediction problem is trivial.
Hey!
Can I use this for real=time applications using raspberry pi?
Perhaps. I don’t know about raspberry pi.
thanks sir
What other algorithms that i can try beside the SVM ?
Here are some suggestions:
https://machinelearningmastery.com/hyperparameters-for-classification-machine-learning-algorithms/
You can try ANN but SVM is better
Hello. Can any one tell me why this line is returning a None value?
results = detector.detect_faces(pixels)
Thank you
Sorry to hear that, this may help:
https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me
hello , thanks for the code, it is very easy to read and go through.
but I am struggling to understand what you mean by
“No, the network is trained once. We use face embeddings from the network as inputs to a classifier that has to be updated when new people are added.”
so does that mean I do not need to put new person folder on the train folder? and just put the folder in the val folder and run code?
or do have have to rerun the code from start to end with the new people in val and train folder?
You’re welcome.
Not sure I follow.
There are 2 models. The one that gives you a face embedding and one that classifies embeddings as people. The first model does not need to be retrained. The second model is only trained once and is then used to make predictions for people that it knows about (e.g. during training).
Does that help?
sorry i don’t quite understand
for every new person I add,
do i put the images in Train and Val folder?
so you mean as i have run the code throughout,
i only need to run the svm part and the following code to guess the person?
many thanks for your response
If you have new people, you must train the SVM model on these people.
Is that posible to make embedding without prediction? Because I need to separate the train and test dataset from the picture that I want to predict
No.
I don’t see the problem?
Hi Jason. I’m new in Keras and I copied your code for see what will be. But there is an error. Please help me. Thank you.
Traceback (most recent call last):
File “C:\Users\train\untitled0.py”, line 60, in
model_scores = get_model_scores(faces)
File “C:\Users\train\untitled0.py”, line 55, in get_model_scores
return model.predict(samples)
File “C:\Users\Anaconda3\lib\site-packages\keras\engine\training.py”, line 1149, in predict
x, _, _ = self._standardize_user_data(x)
File “C:\Users\Anaconda3\lib\site-packages\keras\engine\training.py”, line 751, in _standardize_user_data
exception_prefix=’input’)
File “C:\Users\Anaconda3\lib\site-packages\keras\engine\training_utils.py”, line 128, in standardize_input_data
‘with shape ‘ + str(data_shape))
ValueError: Error when checking input: expected input_7 to have 4 dimensions, but got array with shape (2, 1, 224, 224, 3)
Sorry to hear that, this might help:
https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me
Great explanation thanks,
Since it’s a biometric systems how can i find the ROC and FAR and FFR?
Thanks!
Perhaps you can use the sklearn metrics:
https://scikit-learn.org/stable/modules/classes.html#sklearn-metrics-metrics
HI …
I have problem here results = detector.detect_faces(pixels) it give that the result is empty
It suggests it did not detect a face.
Hello
When I make a test for one image of one person, in every time give different result of probability. Is that right? Or should it give me the same result?
No, see this:
https://machinelearningmastery.com/faq/single-faq/why-do-i-get-different-results-each-time-i-run-the-code
Hello,
This was great thanks. I am wanting to know how to change the randomisation section to score the entire test dataset. Not clear as it appears it was built to score only on a single, randomly selected index from that array. Any guidance on this?
After the SVM model is fit, you can enumerate your dataset and perform the prediction.
Hi Jason it’s a good tutorial, but when I try to run this tutorial in tensorflow 2.0 the mtcnn is not compatible with tensorflow v2.0. do you have any idea for this issue.
thank you
Yes, the tutorial works with Python 3.6, TensorFlow 2.0, and MTCNN v0.1.0.
Perhaps confirm your versions?
Hello sir that’s great tutorial
I’ll use some of the code and expand the dataset and try different algorithms and ill use Kaggle is it okay ?
Sure, as long as you cite the source and link back to this blog post.
Is the algorithm good for this scenario? > If I take one image and make some transformation on it and put all images in the training dataset.
Thank u…
Not sure I follow. What are you trying to achieve exactly?
Sorry for the delay.
I want to make a faceprint system. So, when people come to register, I take one picture for each person and make some transformation in the picture. Will it correctly predict the person?
Perhaps test it on idealized versions of your task.
Thanks for providing the complete code.
You’re welcome!
How about Face matching, match two faces, How to approach it ?
You will need a distance measure. I don’t have an example sorry.
How to give my own image path from local storage to predict the face?
This will show you how to load an arbitrary image from file:
https://machinelearningmastery.com/how-to-load-and-manipulate-images-for-deep-learning-in-python-with-pil-pillow/
Hi Jason I used your code on my own set.In all there were about 20 people each having approximately 30 images.
For testing I decided to see how the code performs in real time. If a random person walks in front of a web cam then will it be able to distinguish between UNKNOWN and some other person in the dataset.
Although I got pretty good results,(each of the person in my set is being correctly identified with a probability of 99%),the unknown people were also being assigned a score of 99%.
From my understanding,the unknown category shouldn’t recieve such a high score.Do you know what is going wrong here.
Training and cross Val error are coming out to be 1 and 0.99 .So I am not sure.Do you think I should reduce the dataset size.
Perhaps experiment with various changes to the model and evaluate the impact.
Hey may i know how did you implement the code in real time?
So should I add more layers to the model.
To what end?
I have Python 3.6, TensorFlow 2.0, and MTCNN v0.1.0., but I get an error:
AttributeError: module ‘tensorflow’ has no attribute ‘get_default_graph’
I have found a reason – it is the version of keras. V2.2 caused the error, V2.3.1 works fine
Happy to hear that!
Ensure you are using Keras 2.3
Hi Jason,
I have used the same code to detect the faces from the my own dataset. I am unable to store the dataset in npz format. My dataset has images around Train-750 pics and val – 400.
The python code is getting killed after certain time . Let me know the workaround
System Specification:
RAM – 16 GB
Swap – 2 GB
OS – Ubuntu 18.04
Perhaps try less data?
Perhaps try a smaller model?
Perhaps try running on ec2?
I found the solution for it. Actually in the code we are loading the model MTCNN() every time. No need of doing it. Load the model only once.
In the code load the MTCNN() only once. Mention detector = MTCNN() outside of the function . It will solve the memory starvation problem.
Thanks
Happy to hear that.
I had to change your very first code from this
from keras.models import load_model
to this
from tensorflow.keras.models import load_model
I’m not sure but since you’re using keras in TF it seems your code might need to be changed according to the above.
Please disregard my comment – if your code fails when importing load_model from keras.models then it’s maybe an installation or setup problem. In my case I thought TF2 had keras but nothing was working right until I installed keras after installing TF2
You don’t have to change the code, you need to install the standalone Keras library:
https://machinelearningmastery.com/faq/single-faq/what-is-the-difference-between-keras-and-tf-keras
Thanks so much for this Jason Brownlee… I got everything working because of this walkthrough.
I also converted the h5 to tflite which is just one commandline command. Additionally you use the interpreter instead of calling “predict” directly on the model… which involves about 8 lines of additional code, found on Tensorflow’s website. By doing this the embedding is obtained 10 times faster (maybe more). I’m on CPU only btw.
Thank you!
Well done!
Great tip, thanks.
Hi Jason, been following you for the past 3 years and you are very inspiring. Just wanted to say, I tried this code in Colab and eventually hit errors. Are there any alterations when using Google Colab? Thank you again.
Thanks!
I don’t know about colab, sorry. Never used it.
Thanks Jason ! So what environment did you use for this tutorial. Regards
My workstation.
It also runs using this:
https://machinelearningmastery.com/develop-evaluate-large-deep-learning-models-keras-amazon-web-services/
Hi,
How can I improve the performance of the predict face process? It takes 4 seconds to predict(1.8 seconds to extract face)
Use a smaller model.
Use a faster computer.
Use smaller data.
– I am using GPU.
– I predict one image at a time.
– The dataset contains 11500 images (260 people)
Is this actually its performance or can I make it better?
@ Ammar,
Quick question for you, Are you using the same code for the predication or for training classification using face embeddings on nvidia driver supported GPU?
Can you let me know the procedure ?
Since i read in the scikit library documentation, it is published that SciKit lib will not support GPU training.
ref : https://scikit-learn.org/stable/faq.html#why-is-there-no-support-for-deep-or-reinforcement-learning-will-there-be-support-for-deep-or-reinforcement-learning-in-scikit-learn
Hi Jason,
I am getting this warning whenever I load facenet_keras.h5
UserWarning: No training configuration found in save file: the model was *not* compiled. Compile it manually.
warnings.warn(‘No training configuration found in save file: ‘
You can safely ignore it.
Thanks, It worked
Well done.
Hi again,
please guide me how to handle the unknown faces in the input.
The model is working fine when it gets the known faces,
But when an unknown face comes, the model is classifying them to the classes on which it was trained.
Example- model is trained to detect me but when I showed it some other face it is again saying that’s you.
Help me with this
Perhaps create a new class of “unknown” and add many different faces to it.
I have been getting not so properly cropped images from the MTCNN detect_faces method. And, I am doing it for a short term project and can’t really get into details inside MTCNN. Just wanted to know why my face recognition is not as per published in this blog. I have used the same datasets and I have followed all instructions thorough.
Perhaps try different images?
I tried those of Elton John and Ben Affleck. Any other suggestions?
Yes, the dataset has 1 million people in it:
https://www.microsoft.com/en-us/research/project/ms-celeb-1m-challenge-recognizing-one-million-celebrities-real-world/
“IndexError: tuple index out of range” i got this error on your facenet_keras.h5 how to solve it?
Sorry to hear that, perhaps this will help:
https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me
Hi Jason,
That’s great tutorial!!
I have a database with only one photo per person. Does the model have good results in this scenario as well?
I don’t see how to make a classifier from just one sample per class. Is there how or should I calculate the distance between the vectors directly (without a classifier)?
(Sorry about my english skills)
Thanks,
Kevin
Perhaps run some tests?
Hi Jason!
Could you please tell me how to incorporate a database into this? Which will store all the images of the people on whom I want to train the network?
I don’t have an example of loading images from a database, sorry.
When we are predicting the final images from the test set and using the pyplot.imshow(random_face_pixels) script
I am getting an error saying ‘Invalid shape for (128,0) for image data’
the shape of random_face_pixels list is (128,0)
I want it to show me the visual image of the prediction
Perhaps this will help:
https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me
I have implemented this FR system but I’m having a problem understanding why we are using the SVM classifier. I don’t have specific questions because I think my problem is in my formative understanding of FaceNet. Kindly tell me if the following statements are true:
1) The FaceNet model, once loaded and we pass the detected face “predicts” an embedding of the face.
2) We then compare the embedding with the stored embeddings of faces in our dataset using the SVM’s predict function and determine who is the likely owner of the face.
3)Is there a way that we can implement FaceNet without having to use a classifier in the last stage?
The neural net model gives features from an image and the model, e.g. svm, classifies them as different people.
Yes, but it would be fragile – requiring the entire facenet model to be re-trained for any change to your dataset. Or perhaps just the output/classifier part of the model as we would use transfer learning.
Thanks a lot Sir. I appreciate the clarification.
You’re welcome.
Hey Jason,
The tutorial is very specific and easy to understand, a very big thank you for that.
But I wanted to use an image in my local directory instead of a random selection.
According to the comments above I understand that I had to read https://machinelearningmastery.com/how-to-load-and-manipulate-images-for-deep-learning-in-python-with-pil-pillow/ and I was supposed to pass it to “selections” variable.
But I got really confused because selections has just 1D value and I guess I’m supposed to convert my image to a numpy array and then pass it on to selections right?
So what should i do after converting it to a numpy array.
Thanks.
The linked tutorial shows you how to load an image as a numpy array directly. Perhaps re-check it?
Hey so should I straight up insert the numpy array into the “selection” variable ?
Sure, try it.
Yeah so I tried everything but I can’t seem to be able to insert an image from local into “selections” could you please help me out LOL really need it for my project
Sorry, I don’t have the capacity to write code for you.
Perhaps you can hire an engineer on upwork.com?
Hi Jason,
when i try to load model it throws below error
code = marshal.loads(raw_code)
ValueError: bad marshal data (unknown type code)
ValueError: bad marshal data (unknown type code)
code = marshal.loads(raw_code)
ValueError: bad marshal data (unknown type code)
Sorry, I have not seen this error before, perhaps try searching on stackoverflow?
i go this down. it was an issue with my versions…i followed your previous answers.
One thing i have observed, does this work with multiple faces in an image?
I guess not because it detects only one faced images.
Well done.
Each face must be extracted first.
I am getting the same error – how did you solve this?
This may help:
https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me
How did you solve this error??
I am having same issue and was wondering how Rujul solved it. I am using Python 3.8.3 and using load_image from keras (version 2.4.3) and not tf.keras. Any help would be greatly appreciated.
What version of tensorflow are you use?
Sorry for the late reply…using Tensorflow 2.2
Same error message. Looking for solution. @Jason, could you list the version I should use for this tutorual? python, keras, tensorflow?
Works with most versions.
Try TF 2.2, Keras 2.3 or 2.4, Python 3.6.
Thank you Jason!!! i was able to do that as well, one thing i have noticed, sometimes it confuses faces with someone else, lets i have a trained set for ‘Jimmy Fallon’ so sometimes in a video it detects a random person and says its Jimmy Fallon and that too with more than 90% probability is that normal???
Yes, this can happen. No model is perfect.
Is it possible to recognize faces real time?
Yes, perhaps if this model operated on a subset of video frames per second.
I tried implementing the model but the accuracy obtained on the 5-Celebrity-dataset is nearly 75% and similarly I created my on custom dataset with around 15 images each training class and 5-8 images in test class , the test accuracy obtained was just 61% and I tried increasing the size of the training dataset to nearly 30 images in each dataset but the test accuracy did not increase it remained the same. What should I do to improve the accuracy ?
Sorry to hear that, try this:
https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me
Hi…
Thank you for this tutorial.
I have tow questions
1- In the paper, it said that facenet layers equal to 140M, but the model that you use it equal to 22,808,144. what is the difference?
2- How can I make a good embedding function that fit my model? Because In the embedding function in facenet and vggface2 there is a difference and I do not understand it
Don’t know, perhaps a difference in implementation.
Perhaps start with an embedding from the pretrained models and only move to your own model if you want to try and lift performance.
what I mean is the function that you built it in get_embedding in facenet and vggface2
I have another question
How I can get the value of thresholds of predict image.
Thank you for help
The predict() function returns a probability.
Hi …
When I want to calculate the accuracy of the facenet model, it should the test-data contains known and unknown labels? Because I tried to put labels that not in the dataset, It gave me an error
Yes, a test dataset should be images and labels not seen by the model during training.
I tried to put unknown faces and labels, but when I encode the labels, It gives me an error that the dataset labels do not contain the label that I want to test it.
the error in this line:
testy = out_encoder.transform(testy)
The labels must be known to the model during training.
So, How can I check for the true and false negative if it must be known labels?
I don’t follow.
You only know whether a prediction is true when training the model and when evaluating the model. After that – when you are using your model in practice, you don’t know if individual predictions are correct or not. All you know is the expected performance of the model on average.
Hello doctor
Thank you for this tutorial
Is the MTCNN using face alignment by default?
What do you mean “by default”? Default in what?
I mean is the MTCNN using face alignment when it detects faces? If not, are you using face alignment in your function or I should adjust it?
Thanks for your help
MTCNN is used for face detection, it does support face alignment.
I change by below code, that will read by photo EXIF Orientation tag. Hope it helps.
original code: image=Image.open(filename)
photo alignment by exif: image = ImageOps.exif_transpose(Image.open(filename))
please I need too much an algorithm or tutorial to compare a photo of face with another and have the percentage of similarity after the possibility of comparing a single photo with thousands of photo of faces to have the percentage of similarity
Perhaps you can use a distance measure between the embedding of the two faces.
Perhaps also check the literature for similar projects to get ideas.
Hii Jason,
I have implemented this face net model to train 5 person’s dataset and the accuracy is also too good but it takes very long time for recognition. Can you give any advice to reduce the computation time for recognition like running the code on gpu or any other classification algo that takes less time.
Thank you.
Yes, perhaps run on faster hardware or use smaller images or a smaller model.
Hi Jason,
Firstly, thanks for such amazing tutorials. Your website has been really helpful. So my question goes like this – how can I input multiple face arrays to vggface2 or FaceNet? In both the examples you have posted, you have used the function extract_face which outputs a single face.
You’re welcome.
Each face is one sample, you can pass in multiple samples to a model as an array and get many predictions in return. Each face must be prepared first before being passed to the model as input.
Helo , your explanation helps me much ! ty for this tuto , well i’m a beginner and i find it usefull and clearly , but i have one question plz if u don’t mind !
i’ve tried to make some modifications on ur code in the final to make it display all the test data instead of just one random picture but i couldn’t , so plz can u write for me the instructions that makes me test model on all the test dataset instead of one exemple ?
You’re welcome.
This is a common question that I answer here:
https://machinelearningmastery.com/faq/single-faq/can-you-change-the-code-in-the-tutorial-to-___
Hi,
This is really a good example and I manage to run it and predict from it. But I have really one concern here, It’s good that we’ve predicted on the test data but what if I want to predict it from any external image? or my video cam image? it seems to me too cumbersome to process the raw/cam images to that level (embedding and so on). Could really have been great if you’ve taken the examples on some external images rather than same test data.
Thanks.
We do use a separate train and test sets. Perhaps re-read the tutorial.
Hi.
I used MTCNN to detected faces and it worked fine. But now, when I run the same code on the same photos, it doesn’t detect faces anymore. Just returns an empty list as a result.
Maybe you know how to fix it or why it happened?
Thanks.
Sorry to hear that, perhaps you modified the code. Perhaps try to copy-paste the code directly from the tutorial again.
hi great work thank you so much
well in my case i want to take more than one image in last function of the code how can i do please ? thank you !
You can change the code to load more than one image. I cannot change the code for you. What is the problem exactly?
Thanks for the article! This is one of the very few codes on the topic that is explained step for step. It’s a great guide. I wanted to test the model with a single or multiple new images which are not a part of the original dataset. How can I do that? My efforts have not yielded results. I would appreciate if you could tell me the process or the functions that need to be called/modified for the same.
Thanks!
See the example at the end of the tutorial for testing the model on a single image.
Thanks! Got it.
Another thing is, this seems to work for a limited dataset. I can’t figure out why but when a larger dataset is taken, the shape of the Numpy array changes when loading dataset (function: load_dataset) and the consequent functions since the changed structure of trainX, trainy, etc. is not compatible with the following functions.
What do you suppose is the issue at play?
You will have more samples, that is all. Perhaps double check your image loading code and perhaps scale all images to the same input size?
I am very new to this system how to find out your current directory in IDLE and how to keep the file ‘facenet_keras.h5‘ please help e
I recommend working from the command line directly:
https://machinelearningmastery.com/faq/single-faq/how-do-i-run-a-script-from-the-command-line
Hi Jason and sorry for my trivial question but I have a problem when I load the model:
model = load_model(‘facenet_keras.h5’)
this give me the following error:
bad marshal data (unknown type code)
I suppose that the problem is generated because the facenet_keras.h5 was generated with the 1.x version of tensorflow (while I am usintg the 2.2). Infact to correctly import the method I wrote:
from tensorflow.keras.models import load_model
but when I load the model I have the error mentioned above, please have you some suggestion?
I believe the problem is because you are trying to load the model using tf.keras instead of the standalone Keras library, not the version of tensorflow.
Thank you for you prompt answer, I have tested under Windows with TF 2.2 and Python 3.7 and works properly, for the moment is enough for me
Nice work!
How did you solve it?
I have the same problem as Paolo, but for me it doesn’t work with any load_model(). I am currently using Python 3.8, TF 2.4.1 and Keras 2.4.3.
Any other suggestions?
Just to add. Rest all versions being same, for me, it worked in python 3.6, but didn’t work on Python 3.8. Did not test on Python 3.7. So it is python version issue and not that of libraries.
Thanks.
I believe Python 3.8 is not supported by most machine learning libraries.
What about python 3.9.5 version
I use python 3.6 for compatibility.
I don’t know about other versions, I believe newer versions are not supported by many libraries.