def timer(model):
result = ''
for i, layer in enumerate(model.layers):
layer_name = layer.__class__.__name__
if layer_name == 'InputLayer' or layer_name == 'Flatten' \
or layer_name == 'MaxPooling3D' or layer_name == 'Activation' \
or layer_name == 'MaxPooling2D' or layer_name == 'ZeroPadding2D' \
or layer_name == 'Add' or layer_name == 'AveragePooling2D' \
or layer_name == 'GlobalAveragePooling2D' or layer_name == 'Lambda'\
or layer_name == 'Concatenate':
continue
# Because we split some of Convolution layers, which has multiple inputs
# cannot use normal method for getting int_shape.
try:
shape = layer.input_shape
except Exception:
shape = layer.get_input_shape_at(0)
assert not isinstance(shape[0], tuple), \
'{:s} has bad input shape {:s}'.format(layer_name, str(layer.input_shape))
input_shape = list(shape)[1:]
data = np.random.random_sample(input_shape)
input = InputLayer(input_shape)
config = layer.get_config()
if layer_name == 'Dense':
_layer = Dense.from_config(config)
elif layer_name == 'Conv3D':
_layer = Conv3D.from_config(config)
elif layer_name == 'Conv2D':
_layer = Conv2D.from_config(config)
elif layer_name == 'SeparableConv2D':
_layer = SeparableConv2D.from_config(config)
elif layer_name == 'DepthwiseConv2D':
_layer = DepthwiseConv2D.from_config(config)
else:
raise Exception(layer_name, ' layer name not included')
# Not the real model used for inference.
_model = Sequential()
_model.add(input)
_model.add(_layer)
# fake run because the model needs to be complied
_model.predict(np.array([data]))
start = time.time()
for _ in range(50):
_model.predict(np.array([data]))
result += '{:>20s}: {:.3f} s\n'.format(layer.name,
(time.time() - start) / 50)
return result
def FCN_8s_pretrained(classes, in_shape, l2_reg, nopad=False, test=False):
vgg16 = VGG16(include_top=False, weights="imagenet", input_shape=in_shape)
inputs = Input(shape=in_shape)
if nopad:
x = inputs
else:
x = ZeroPadding2D(padding=(100, 100))(inputs)
for layer in vgg16.layers[1:]:
if "conv" in layer.name:
W, b = layer.get_weights()
config = layer.get_config()
config["kernel_regularizer"] = l2(l2_reg)
config["kernel_initializer"] = Constant(W)
config["bias_initializer"] = Constant(b)
conv = Conv2D.from_config(config)
x = conv(x)
elif "pool" in layer.name:
x = MaxPooling2D()(x)
if layer.name == "block3_pool":
feat3 = x
elif layer.name == "block4_pool":
feat4 = x
x = Conv2D(filters=4096,
kernel_size=(7, 7),
padding="valid",
activation="relu",
kernel_regularizer=l2(l2_reg),
name="fc1")(x)
x = Dropout(0.5)(x)
x = Conv2D(filters=4096,
kernel_size=(1, 1),
padding="valid",
activation="relu",
kernel_regularizer=l2(l2_reg),
name="fc2")(x)
x = Dropout(0.5)(x)
score5 = Conv2D(filters=classes,
kernel_size=(1, 1),
kernel_regularizer=l2(l2_reg),
activation="relu")(x)
if nopad:
score5 = Conv2DTranspose(filters=classes,
kernel_size=(14, 14),
strides=(1, 1),
padding="valid",
activation="linear",
kernel_regularizer=l2(l2_reg),
kernel_initializer=Constant(
bilinear_upsample_weights(
"full", classes)
))(score5)
else:
score5 = Conv2DTranspose(filters=classes,
kernel_size=(4, 4),
strides=(2, 2),
padding="same",
activation="linear",
kernel_regularizer=l2(l2_reg),
kernel_initializer=Constant(
bilinear_upsample_weights(2, classes)
))(score5)
# pool3 のfeature mapを次元圧縮
score3 = Conv2D(filters=classes,
kernel_size=(1, 1),
kernel_regularizer=l2(l2_reg),
activation='relu')(feat3)
# pool4のfeature mapを次元圧縮
score4 = Conv2D(filters=classes,
kernel_size=(1, 1),
kernel_regularizer=l2(l2_reg),
activation="relu")(feat4)
# merge p4 and p5
score4 = CroppingLike2D(K.int_shape(score5))(score4)
score45 = Add()([score4, score5])
# p4+p5 を x2 upsampling
if not nopad:
score45 = ZeroPadding2D(padding=(1, 1))(score45)
score45 = Conv2DTranspose(filters=classes,
kernel_size=(4, 4),
strides=(2, 2),
padding="same",
activation="linear",
kernel_regularizer=l2(l2_reg),
kernel_initializer=Constant(
bilinear_upsample_weights(2, classes)
))(score45)
# p3とp45をmerge
score3 = CroppingLike2D(K.int_shape(score45))(score3)
score345 = Add()([score3, score45])
# p3+p4+p5を x8 upsampling
if not nopad:
score345 = ZeroPadding2D(padding=(1, 1))(score345)
x = Conv2DTranspose(filters=classes,
kernel_size=(16, 16),
strides=(8, 8),
padding="same",
activation="linear",
kernel_regularizer=l2(l2_reg),
kernel_initializer=Constant(
bilinear_upsample_weights(8, classes)
))(score345)
x = CroppingLike2D(K.int_shape(inputs))(x)
if test:
x = Activation("softmax")(x)
model = Model(inputs=inputs, outputs=x)
return model
