def create_seg_model(self,
net,
n=21,
backbone='mobilenetv2',
load_weights=False,
multi_gpu=False):
'''
Net is:
1. original deeplab v3+
2. original deeplab v3+ and subpixel upsampling layer
'''
model = Deeplabv3(weights=None,
input_tensor=None,
infer=False,
input_shape=self.sz + (3, ),
classes=21,
backbone=backbone,
OS=8,
alpha=1)
base_model = Model(model.input, model.layers[-5].output)
for layer in base_model.layers:
layer.trainable = False
self.net = net
self.modelpath = 'weights/{}_{}.h5'.format(backbone, net)
if backbone == 'xception':
scale = 4
else:
scale = 8
if net == 'original':
x = Conv2D(n, (1, 1), padding='same',
name='conv_upsample')(base_model.output)
x = Lambda(lambda x: K.tf.image.resize_bilinear(
x, size=(self.sz[0], self.sz[1])))(x)
x = Reshape((self.sz[0] * self.sz[1], -1))(x)
x = Activation('softmax', name='pred_mask')(x)
model = Model(base_model.input, x, name='deeplabv3p')
elif net == 'subpixel':
x = Subpixel(n, 1, scale, padding='same')(base_model.output)
x = Reshape((self.sz[0] * self.sz[1], -1))(x)
x = Activation('softmax', name='pred_mask')(x)
model = Model(base_model.input, x, name='deeplabv3p_subpixel')
# Do ICNR
for layer in model.layers:
if type(layer) == Subpixel:
c, b = layer.get_weights()
w = icnr_weights(scale=scale, shape=c.shape)
layer.set_weights([w, b])
if load_weights:
model.load_weights('weights/{}_{}.h5'.format(backbone, net))
if multi_gpu:
from keras.utils import multi_gpu_model
model = multi_gpu_model(model, gpus=len(get_available_gpus()))
self.model = model
return model
def get_uncompiled_model(input_shape, num_classes, backbone, infer=False):
model = Deeplabv3(weights=None, input_tensor=None, infer=infer,
input_shape=input_shape, classes=num_classes,
backbone=backbone, OS=16, alpha=1)
base_model = Model(model.input, model.layers[-5].output)
#self.net = net
#modelpath = 'weights/{}_{}.h5'.format(backbone, net)
if backbone=='xception':
scale = 4
else:
scale = 8
#elif net == 'subpixel':
x = Subpixel(num_classes, 1, scale, padding='same')(base_model.output)
#x = Reshape((input_shape[0]*input_shape[1], -1)) (x)
#x = Reshape((input_shape[0]*input_shape[1], num_classes)) (x)
x = Activation('softmax', name = 'pred_mask')(x)
model = Model(base_model.input, x, name='deeplabv3p_subpixel')
# Do ICNR
for layer in model.layers:
if type(layer) == Subpixel:
c, b = layer.get_weights()
w = ICNR(scale=scale)(shape=c.shape)
#W = tf.convert_to_tensor(w, dtype=tf.float32)
layer.set_weights([w, b])
return model
def create_seg_model(self,
opt,
net,
load_weights=False,
multi_gpu=True,
to_compile=True):
self.net = net
n_classes = len(get_VOC2012_classes()) - 1
if net == 'enet':
self.modelpath = 'weights/enet.h5'
model, _ = build(nc=n_classes,
w=self.sz[1],
h=self.sz[0],
plot=False)
elif net == 'mobileunet':
self.modelpath = 'weights/mobileunet.h5'
model = MobileUNet(input_shape=(self.sz[1], self.sz[0], 3),
alpha=1,
alpha_up=1,
depth_multiplier=1,
n_classes=n_classes)
elif net == 'unet':
self.modelpath = 'weights/unet.h5'
model = create_unet(self.sz, n_classes)
elif net == 'segnet':
self.modelpath = 'weights/segnet.h5'
model = create_segnet(self.sz, n_classes)
elif net == 'tiramisu':
self.modelpath = 'weights/tiramisu.h5'
model = Tiramisu(self.sz + (3, ), n_classes)
else:
self.modelpath = 'weights/deeplabv3.h5'
model = Deeplabv3(weights=None,
input_tensor=None,
input_shape=self.sz + (3, ),
classes=21,
backbone='mobilenetv2',
OS=16,
alpha=1,
use_coordconv=self.coords)
if self.coords:
self.modelpath = 'weights/deeplabv3_coordconv.h5'
if multi_gpu:
from keras.utils import multi_gpu_model
model = multi_gpu_model(model, gpus=len(get_available_gpus()))
if to_compile:
model.compile(
optimizer=opt,
sample_weight_mode="temporal",
loss=softmax_sparse_crossentropy_ignoring_last_label,
#loss = wisense_loss(30),
metrics=[
Mean_IOU, background_sparse_accuracy,
sparse_accuracy_ignoring_last_label,
foreground_sparse_accuracy, 'acc'
])
if load_weights:
self.load_weights(model)
return model