VGG-Face model for keras

keras VGG-Face Model.md

VGG-Face model for Keras

This is the Keras model of VGG-Face.

It has been obtained through the following method:

• vgg-face-keras:directly convert the vgg-face matconvnet model to keras model
• vgg-face-keras-fc:first convert vgg-face caffe model to mxnet model,and then convert it to keras model

Details about the network architecture can be found in the following paper:

Deep Face Recognition
O. M. Parkhi, A. Vedaldi, A. Zisserman
British Machine Vision Conference, 2015

Please cite the paper if you use the models.

Contents:

model and usage demo: see vgg-face-keras.py or vgg-face-keras-fc.py

The only difference between them is the last few layers(see the code and you'll understand),but they produce the same result.

weights: 

• Google Drive
• Baidu Yun(password:12i0)
• fc-version Google Drive
• fc-version Baidu Yun(password:d318)

Notice:

Please use this model in theano mode.

vgg-face-keras-fc.py

from keras.models import Model
from keras.layers import Input, Convolution2D, ZeroPadding2D, MaxPooling2D, Flatten, Dense, Dropout
from PIL import Image
import numpy as np

def vgg_face(weights_path=None):
    img = Input(shape=(3, 224, 224))

    pad1_1 = ZeroPadding2D(padding=(1, 1))(img)
    conv1_1 = Convolution2D(64, 3, 3, activation='relu', name='conv1_1')(pad1_1)
    pad1_2 = ZeroPadding2D(padding=(1, 1))(conv1_1)
    conv1_2 = Convolution2D(64, 3, 3, activation='relu', name='conv1_2')(pad1_2)
    pool1 = MaxPooling2D((2, 2), strides=(2, 2))(conv1_2)

    pad2_1 = ZeroPadding2D((1, 1))(pool1)
    conv2_1 = Convolution2D(128, 3, 3, activation='relu', name='conv2_1')(pad2_1)
    pad2_2 = ZeroPadding2D((1, 1))(conv2_1)
    conv2_2 = Convolution2D(128, 3, 3, activation='relu', name='conv2_2')(pad2_2)
    pool2 = MaxPooling2D((2, 2), strides=(2, 2))(conv2_2)

    pad3_1 = ZeroPadding2D((1, 1))(pool2)
    conv3_1 = Convolution2D(256, 3, 3, activation='relu', name='conv3_1')(pad3_1)
    pad3_2 = ZeroPadding2D((1, 1))(conv3_1)
    conv3_2 = Convolution2D(256, 3, 3, activation='relu', name='conv3_2')(pad3_2)
    pad3_3 = ZeroPadding2D((1, 1))(conv3_2)
    conv3_3 = Convolution2D(256, 3, 3, activation='relu', name='conv3_3')(pad3_3)
    pool3 = MaxPooling2D((2, 2), strides=(2, 2))(conv3_3)

    pad4_1 = ZeroPadding2D((1, 1))(pool3)
    conv4_1 = Convolution2D(512, 3, 3, activation='relu', name='conv4_1')(pad4_1)
    pad4_2 = ZeroPadding2D((1, 1))(conv4_1)
    conv4_2 = Convolution2D(512, 3, 3, activation='relu', name='conv4_2')(pad4_2)
    pad4_3 = ZeroPadding2D((1, 1))(conv4_2)
    conv4_3 = Convolution2D(512, 3, 3, activation='relu', name='conv4_3')(pad4_3)
    pool4 = MaxPooling2D((2, 2), strides=(2, 2))(conv4_3)

    pad5_1 = ZeroPadding2D((1, 1))(pool4)
    conv5_1 = Convolution2D(512, 3, 3, activation='relu', name='conv5_1')(pad5_1)
    pad5_2 = ZeroPadding2D((1, 1))(conv5_1)
    conv5_2 = Convolution2D(512, 3, 3, activation='relu', name='conv5_2')(pad5_2)
    pad5_3 = ZeroPadding2D((1, 1))(conv5_2)
    conv5_3 = Convolution2D(512, 3, 3, activation='relu', name='conv5_3')(pad5_3)
    pool5 = MaxPooling2D((2, 2), strides=(2, 2))(conv5_3)

    flat = Flatten()(pool5)
    fc6 = Dense(4096, activation='relu', name='fc6')(flat)
    fc6_drop = Dropout(0.5)(fc6)
    fc7 = Dense(4096, activation='relu', name='fc7')(fc6_drop)
    fc7_drop = Dropout(0.5)(fc7)
    out = Dense(2622, activation='softmax', name='fc8')(fc7_drop)

    model = Model(input=img, output=out)

    if weights_path:
        model.load_weights(weights_path)

    return model

if __name__ == "__main__":
    im = Image.open('A.J._Buckley.jpg')
    im = im.resize((224,224))
    im = np.array(im).astype(np.float32)
    #    im[:,:,0] -= 129.1863
    #    im[:,:,1] -= 104.7624
    #    im[:,:,2] -= 93.5940
    im = im.transpose((2,0,1))
    im = np.expand_dims(im, axis=0)

    # Test pretrained model
    model = vgg_face('vgg-face-keras-fc.h5')
    out = model.predict(im)
    print(out[0][0])
    

vgg-face-keras.py

from keras.models import Model
from keras.layers import Input, Convolution2D, ZeroPadding2D, MaxPooling2D, Flatten, Dropout, Activation
from PIL import Image
import numpy as np

def vgg_face(weights_path=None):
    img = Input(shape=(3, 224, 224))

    pad1_1 = ZeroPadding2D(padding=(1, 1))(img)
    conv1_1 = Convolution2D(64, 3, 3, activation='relu', name='conv1_1')(pad1_1)
    pad1_2 = ZeroPadding2D(padding=(1, 1))(conv1_1)
    conv1_2 = Convolution2D(64, 3, 3, activation='relu', name='conv1_2')(pad1_2)
    pool1 = MaxPooling2D((2, 2), strides=(2, 2))(conv1_2)

    pad2_1 = ZeroPadding2D((1, 1))(pool1)
    conv2_1 = Convolution2D(128, 3, 3, activation='relu', name='conv2_1')(pad2_1)
    pad2_2 = ZeroPadding2D((1, 1))(conv2_1)
    conv2_2 = Convolution2D(128, 3, 3, activation='relu', name='conv2_2')(pad2_2)
    pool2 = MaxPooling2D((2, 2), strides=(2, 2))(conv2_2)

    pad3_1 = ZeroPadding2D((1, 1))(pool2)
    conv3_1 = Convolution2D(256, 3, 3, activation='relu', name='conv3_1')(pad3_1)
    pad3_2 = ZeroPadding2D((1, 1))(conv3_1)
    conv3_2 = Convolution2D(256, 3, 3, activation='relu', name='conv3_2')(pad3_2)
    pad3_3 = ZeroPadding2D((1, 1))(conv3_2)
    conv3_3 = Convolution2D(256, 3, 3, activation='relu', name='conv3_3')(pad3_3)
    pool3 = MaxPooling2D((2, 2), strides=(2, 2))(conv3_3)

    pad4_1 = ZeroPadding2D((1, 1))(pool3)
    conv4_1 = Convolution2D(512, 3, 3, activation='relu', name='conv4_1')(pad4_1)
    pad4_2 = ZeroPadding2D((1, 1))(conv4_1)
    conv4_2 = Convolution2D(512, 3, 3, activation='relu', name='conv4_2')(pad4_2)
    pad4_3 = ZeroPadding2D((1, 1))(conv4_2)
    conv4_3 = Convolution2D(512, 3, 3, activation='relu', name='conv4_3')(pad4_3)
    pool4 = MaxPooling2D((2, 2), strides=(2, 2))(conv4_3)

    pad5_1 = ZeroPadding2D((1, 1))(pool4)
    conv5_1 = Convolution2D(512, 3, 3, activation='relu', name='conv5_1')(pad5_1)
    pad5_2 = ZeroPadding2D((1, 1))(conv5_1)
    conv5_2 = Convolution2D(512, 3, 3, activation='relu', name='conv5_2')(pad5_2)
    pad5_3 = ZeroPadding2D((1, 1))(conv5_2)
    conv5_3 = Convolution2D(512, 3, 3, activation='relu', name='conv5_3')(pad5_3)
    pool5 = MaxPooling2D((2, 2), strides=(2, 2))(conv5_3)

    fc6 = Convolution2D(4096, 7, 7, activation='relu', name='fc6')(pool5)
    fc6_drop = Dropout(0.5)(fc6)
    fc7 = Convolution2D(4096, 1, 1, activation='relu', name='fc7')(fc6_drop)
    fc7_drop = Dropout(0.5)(fc7)
    fc8 = Convolution2D(2622, 1, 1, name='fc8')(fc7_drop)
    flat = Flatten()(fc8)
    out = Activation('softmax')(flat)

    model = Model(input=img, output=out)

    if weights_path:
        model.load_weights(weights_path)

    return model

if __name__ == "__main__":
    im = Image.open('A.J._Buckley.jpg')
    im = im.resize((224,224))
    im = np.array(im).astype(np.float32)
#    im[:,:,0] -= 129.1863
#    im[:,:,1] -= 104.7624
#    im[:,:,2] -= 93.5940
    im = im.transpose((2,0,1))
    im = np.expand_dims(im, axis=0)

    # Test pretrained model
    model = vgg_face('vgg-face-keras.h5')
    out = model.predict(im)
    print(out[0][0])
    

what is the dataset used for this model?

Thanks, is anybody know the pytorch version?

Question: The mean subtraction done in

im[:,:,0] -= 129.1863

im[:,:,1] -= 104.7624

im[:,:,2] -= 93.5940

1: Are these the channel means obtained by the people who trained the VGG face?
2: If so, why are they commented?

I got this error when I ran the code:

VGG_Tensorflow_V2.py:15: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(64, (3, 3), activation="relu")`
  model.add(Convolution2D(64, 3, 3, activation='relu'))
VGG_Tensorflow_V2.py:17: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(64, (3, 3), activation="relu")`
  model.add(Convolution2D(64, 3, 3, activation='relu'))
VGG_Tensorflow_V2.py:21: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(128, (3, 3), activation="relu")`
  model.add(Convolution2D(128, 3, 3, activation='relu'))
VGG_Tensorflow_V2.py:23: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(128, (3, 3), activation="relu")`
  model.add(Convolution2D(128, 3, 3, activation='relu'))
Traceback (most recent call last):
  File "/home/mshafaei/.local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1576, in _create_c_op
    c_op = c_api.TF_FinishOperation(op_desc)
tensorflow.python.framework.errors_impl.InvalidArgumentError: Negative dimension size caused by subtracting 2 from 1 for 'max_pooling2d_2/MaxPool' (op: 'MaxPool') with input shapes: [?,1,112,128].

I also got this error when I ran vgg-face-keras.py:
tensorflow.python.framework.errors_impl.InvalidArgumentError: Negative dimension size caused by subtracting 2 from 1 for 'max_pooling2d_2/MaxPool' (op: 'MaxPool') with input shapes: [?,1,112,128].

I also got this error when I ran vgg-face-keras.py:
tensorflow.python.framework.errors_impl.InvalidArgumentError: Negative dimension size caused by subtracting 2 from 1 for 'max_pooling2d_2/MaxPool' (op: 'MaxPool') with input shapes: [?,1,112,128].

You could add channels_first as follow like this

from keras import backend as K
K.set_image_data_format('channels_first')

I got this error when I ran the code:

VGG_Tensorflow_V2.py:15: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(64, (3, 3), activation="relu")`
  model.add(Convolution2D(64, 3, 3, activation='relu'))
VGG_Tensorflow_V2.py:17: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(64, (3, 3), activation="relu")`
  model.add(Convolution2D(64, 3, 3, activation='relu'))
VGG_Tensorflow_V2.py:21: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(128, (3, 3), activation="relu")`
  model.add(Convolution2D(128, 3, 3, activation='relu'))
VGG_Tensorflow_V2.py:23: UserWarning: Update your `Conv2D` call to the Keras 2 API: `Conv2D(128, (3, 3), activation="relu")`
  model.add(Convolution2D(128, 3, 3, activation='relu'))
Traceback (most recent call last):
  File "/home/mshafaei/.local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1576, in _create_c_op
    c_op = c_api.TF_FinishOperation(op_desc)
tensorflow.python.framework.errors_impl.InvalidArgumentError: Negative dimension size caused by subtracting 2 from 1 for 'max_pooling2d_2/MaxPool' (op: 'MaxPool') with input shapes: [?,1,112,128].

You could add channels_first as follow like this

from keras import backend as K
K.set_image_data_format('channels_first')

Here is a test picture,the probability of the picture belonging to the first class should be 0.99953598.

Sorry but I ran the vgg-face-keras.py program using theano backend and the maximum probability is only 0.207927 (class 36, not class 1). Did I make any mistake?

Thanks!

Question: The mean subtraction done in
im[:,:,0] -= 129.1863
im[:,:,1] -= 104.7624
im[:,:,2] -= 93.5940
1: Are these the channel means obtained by the people who trained the VGG face?
2: If so, why are they commented?

these are my questions too... any one can answer how these means are computed? is it means we should also subtract these means from train images?

Here is a test picture,the probability of the picture belonging to the first class should be 0.99953598.

Sorry but I ran the vgg-face-keras.py program using theano backend and the maximum probability is only 0.207927 (class 36, not class 1). Did I make any mistake?

Thanks!

I run it with tf backend and max probability for that test image is 0.20376973. Does anybody know why we are not getting expected result?

Regards,
Aleksandar

@EncodeTS,
I run it with tf backend but I am getting system error :" Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2".Pls help me out.

@rhlshah
don't worry about this problem. any program you run will work as usual. the reason this happening is that you are using a CPU that supports AVX2, which is a set of instructions that your CPU can do. it will accelerate some vector operations. you can install another binary of TensorFlow but i don't recommend that since it's a hard thing to do.

Try this https://github.com/wayne841213/VGGFace/blob/master/vgg-face-keras.py
My result is 0.99910444

I am getting this error while using "vgg-face-keras.h5"
Negative dimension size caused by subtracting 2 from 1 for '{{node max_pooling2d_2/MaxPool}} = MaxPoolT=DT_FLOAT, data_format="NCHW", explicit_paddings=[], ksize=[1, 1, 2, 2], padding="VALID", strides=[1, 1, 2, 2]' with input shapes: [?,256,1,1].