VGGFace-nnabla
About
This repository is implementation of VGG-Face CNN with Sony's Neural Network Libraries, NNabla.
Requirement
- Python2
- NNabla 0.9.1rc3
- OpenCV 2.4.11
Usage
$ wget https://github.com/opencv/opencv/blob/master/data/haarcascades/haarcascade_frontalface_alt.xml
$ wget https://github.com/hogefugabar/VGGFace-nnabla/releases/download/v0.0.1-alpha/vggface_weights.pkl
Feature extraction
from vggface import VGGFace
import nnabla as nn
import numpy as np
improt cv2
batch_size = 50
"""convolution feature"""
x = nn.Variable([batch_size, 3, 224, 224])
model = VGGFace(x, include_top=False, pooling='avg')
feature = model.output
""""""
"""first FC layer feature"""
x = nn.Variable([batch_size, 3, 224, 224])
model = VGGFace(x, include_top=True)
feature = model.layers['fc1']
""""""
image = cv2.imread("./image.jpg")
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (224, 224))[np.newaxis].transpose(0, 3, 1, 2)/255.
x.d = image
feature.forward()
print(feature.d.copy()[0]) # this is feature of ./image.jpgFinetuning
import nnabla as nn
import nnabla.functions as F
import nnabla.parametric_functions as PF
import nnabla.solvers as S
from vggface import VGGFace
n_hidden = 4096
n_classes = 10
batch_size = 50
x = nn.Variable([batch_size, 3, 224, 224])
t = nn.Variable([batch_size, 1])
model = VGGFace(x, include_top=False)
h = model.output
h = F.relu(PF.affine(h, n_hidden, name='fc1'))
h = F.relu(PF.affine(h, n_hidden, name='fc2'))
y = PF.affine(h, n_classes)
loss = F.mean(F.softmax_cross_entropy(y, t))
"""insert your training loop"""Prediction
from vggface import VGGFace
import numpy as np
import cv2
image = cv2.imread("./image.jpg")
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (224, 224))[np.newaxis].transpose(0, 3, 1, 2)/255.
x = nn.Variable([1, 3, 244, 244])
model = VGGFace(x)
y = model.output
x.d = image
y.forward()
print(y.d[0].argmax())