Deepface Implementation in Keras

[vigneshwer dhinakaran]

Hi All,

I am using keras distribution of locally connected layer (https://github.com/joelthchao/keras/blob/master/keras/layers/local.py) with WDB0400 database.

Deepface paper link : https://research.facebook.com/publications/deepface-closing-the-gap-to-human-level-performance-in-face-verification/

Model Architecture

        model = Sequential()

#Deepnet Archictecture

#c1

model.add(Convolution2D(nb_filter[0], nb_row[0], nb_col[0], border_mode=border_mode, input_shape=(no_of_channels,img_row,img_col),init=weight_init))

model.add(Activation(activation))

#m2

model.add(MaxPooling2D(pool_size = pool_size, strides = stride_size))

#c3

model.add(Convolution2D(nb_filter[1], nb_row[1], nb_col[1], border_mode=border_mode, init=weight_init))

model.add(Activation(activation))

#l4

model.add(LocallyConnected2D(nb_filter[2], nb_row[2], nb_col[2],border_mode=border_mode, init=weight_init))

model.add(Activation(activation))

# #l5

model.add(LocallyConnected2D(nb_filter[3], nb_row[3], nb_col[3],border_mode=border_mode, init=weight_init))

model.add(Activation(activation))

#L6         model.add(LocallyConnected2D(nb_filter[4], nb_row[4], nb_col[4],border_mode=border_mode, init=weight_init))

model.add(Activation(activation))

#Dropout

        model.add(Dropout(dropout_percent[0]))

model.add(Flatten())

#F7

model.add(Dense(output_fc[0],init=weight_init))

model.add(Activation(activation)) 

#normalization l2 

#F8

# model.add(Dense(output_fc[1],init=weight_init))

model.add(Dense(nb_classes, activation='softmax',init=weight_init))

#learning process

optimizer_sgd = SGD(lr=lr_rate, decay=1e-6, momentum=0.9, nesterov=True)

The probelm I am facing is I am not able to bring down the loss values or over fit the model. Is there any one who has implemented this paper in keras successfully.

Thanks

[nitish bhardwaj]

Could you please explain the parameters that you are using for this experiment?