def _generate_background(self, image, bboxes):
_, H, W = image.shape
bbox = generate_background_bbox((H, W), (224, 224), bboxes)
ymin, xmin, ymax, xmax = bbox
image = resnet.prepare(image[:, ymin:ymax, xmin:xmax])
label = len(roaddamage_label_names) + 1
return image, label
def feature_extract(img, model):
img = resnet.prepare(img, size=None)
img = np.asarray([img], dtype=np.float32)
img = Variable(cuda.to_gpu(img))
with chainer.using_config("train", False):
feature = cuda.to_cpu(model(img).data)
return feature
def preprocess(image, model_name):
if model_name == "mv2":
image /= 128.
image = transforms.resize(image, (224, 224))
elif model_name == "vgg16":
image = vgg.prepare(image, size=(224, 224))
elif model_name == "resnet50":
image = resnet.prepare(image, size=(224, 224))
else:
raise Exception("illegal model")
return image
def _generate_damage(self, image, bboxes, labels):
index = np.random.randint(len(labels))
label = labels[index]
ymin, xmin, ymax, xmax = bboxes[index]
damage = image[:, ymin:ymax, xmin:xmax]
background = np.zeros(image.shape)
background[:, ymin:ymax, xmin:xmax] = damage
image = resnet.prepare(background)
return image, label
def _layer_out(self, images, layer):
layers = ['conv1', 'bn1', 'res2', 'res3', 'res4', 'res5']
if layer not in layers:
raise ValueError('Layer {} does not exist.'.format(layer))
x = concat_examples([prepare(img) for img in images])
with chainer.function.no_backprop_mode(), chainer.using_config(
'train', False):
x = Variable(self.xp.asarray(x))
y = self(x, layers=[layer])[layer]
return y
def test_prepare(self):
x1 = numpy.random.uniform(0, 255, (320, 240, 3)).astype(numpy.uint8)
x2 = numpy.random.uniform(0, 255, (320, 240)).astype(numpy.uint8)
x3 = numpy.random.uniform(0, 255, (160, 120, 3)).astype(self.dtype)
x4 = numpy.random.uniform(0, 255, (1, 160, 120)).astype(self.dtype)
x5 = numpy.random.uniform(0, 255, (3, 160, 120)).astype(numpy.uint8)
y1 = resnet.prepare(x1)
assert y1.shape == (3, 224, 224)
assert y1.dtype == self.dtype
y2 = resnet.prepare(x2)
assert y2.shape == (3, 224, 224)
assert y2.dtype == self.dtype
y3 = resnet.prepare(x3, size=None)
assert y3.shape == (3, 160, 120)
assert y3.dtype == self.dtype
y4 = resnet.prepare(x4)
assert y4.shape == (3, 224, 224)
assert y4.dtype == self.dtype
y5 = resnet.prepare(x5, size=None)
assert y5.shape == (3, 160, 120)
assert y5.dtype == self.dtype
def test_prepare(self):
x1 = numpy.random.uniform(0, 255, (320, 240, 3)).astype(numpy.uint8)
x2 = numpy.random.uniform(0, 255, (320, 240)).astype(numpy.uint8)
x3 = numpy.random.uniform(0, 255, (160, 120, 3)).astype(numpy.float32)
x4 = numpy.random.uniform(0, 255, (1, 160, 120)).astype(numpy.float32)
x5 = numpy.random.uniform(0, 255, (3, 160, 120)).astype(numpy.uint8)
y1 = resnet.prepare(x1)
self.assertEqual(y1.shape, (3, 224, 224))
self.assertEqual(y1.dtype, numpy.float32)
y2 = resnet.prepare(x2)
self.assertEqual(y2.shape, (3, 224, 224))
self.assertEqual(y2.dtype, numpy.float32)
y3 = resnet.prepare(x3, size=None)
self.assertEqual(y3.shape, (3, 160, 120))
self.assertEqual(y3.dtype, numpy.float32)
y4 = resnet.prepare(x4)
self.assertEqual(y4.shape, (3, 224, 224))
self.assertEqual(y4.dtype, numpy.float32)
y5 = resnet.prepare(x5, size=None)
self.assertEqual(y5.shape, (3, 160, 120))
self.assertEqual(y5.dtype, numpy.float32)
def test_prepare(self):
x1 = numpy.random.uniform(0, 255, (320, 240, 3)).astype(numpy.uint8)
x2 = numpy.random.uniform(0, 255, (320, 240)).astype(numpy.uint8)
x3 = numpy.random.uniform(0, 255, (160, 120, 3)).astype(self.dtype)
x4 = numpy.random.uniform(0, 255, (1, 160, 120)).astype(self.dtype)
x5 = numpy.random.uniform(0, 255, (3, 160, 120)).astype(numpy.uint8)
y1 = resnet.prepare(x1)
assert y1.shape == (3, 224, 224)
assert y1.dtype == self.dtype
y2 = resnet.prepare(x2)
assert y2.shape == (3, 224, 224)
assert y2.dtype == self.dtype
y3 = resnet.prepare(x3, size=None)
assert y3.shape == (3, 160, 120)
assert y3.dtype == self.dtype
y4 = resnet.prepare(x4)
assert y4.shape == (3, 224, 224)
assert y4.dtype == self.dtype
y5 = resnet.prepare(x5, size=None)
assert y5.shape == (3, 160, 120)
assert y5.dtype == self.dtype
def prepare_images(self, images):
if self.xp != np:
device = images.data.device
images = F.copy(images, -1)
converted_images = [
resnet.prepare(image.data, size=None)
for image in F.separate(images, axis=0)
]
converted_images = F.stack(converted_images, axis=0)
if self.xp != np:
converted_images = F.copy(converted_images, device.id)
return converted_images
def img2resnet(img, xp=np, dtype=np.float32):
dst = resnet.prepare(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
ch, w, h = dst.shape
return xp.array(dst, dtype=dtype).reshape(-1, ch, w, h)
def imgs2resnet(imgs, xp=np):
dst = [
resnet.prepare(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)) for img in imgs
]
return xp.array(dst)
def extract_feature(image, model):
x = resnet.prepare(image, size=None)
x = chainer.Variable(xp.asarray([x], dtype=xp.float32))
with chainer.using_config('train', False):
feature = model(x)
return feature
def get_example(self, i):
image, label = self.base[i]
if self.crop:
image = self.random_crop(image)
image = prepare(image)
return image, label
def __call__(self, in_data):
img = in_data[0]
img = resnet.prepare(img, (self.size, self.size))
return (img, *in_data[1:])