我正在Keras中使用VGG16(解码器部分)训练U-Net。该模型训练得很好并且正在学习-我看到验证集的成绩有所提高。
但是,当我尝试在图像上调用predict
时,我收到的矩阵的所有值都相同。
下面是模型:
class Gray2VGGInput(Layer):
"""Custom conversion layer"""
def build(self, x):
self.image_mean = K.variable(value=np.array([103.939, 116.779, 123.68]).reshape([1,1,1,3]).astype('float32'),
dtype='float32',
name='imageNet_mean' )
self.built = True
return
def call(self, x):
rgb_x = K.concatenate([x,x,x], axis=-1 )
norm_x = rgb_x - self.image_mean
return norm_x
def compute_output_shape(self, input_shape):
return input_shape[:3] + (3,)
def UNET1_VGG16(img_rows=864, img_cols=1232):
'''
UNET with pretrained layers from VGG16
'''
def upsampleLayer(in_layer, concat_layer, input_size):
'''
Upsampling (=Decoder) layer building block
Parameters
----------
in_layer: input layer
concat_layer: layer with which to concatenate
input_size: input size fot convolution
'''
upsample = Conv2DTranspose(input_size, (2, 2), strides=(2, 2), padding='same')(in_layer)
upsample = concatenate([upsample, concat_layer])
conv = Conv2D(input_size, (1, 1), activation='relu', kernel_initializer='he_normal', padding='same')(upsample)
conv = BatchNormalization()(conv)
conv = Dropout(0.2)(conv)
conv = Conv2D(input_size, (1, 1), activation='relu', kernel_initializer='he_normal', padding='same')(conv)
conv = BatchNormalization()(conv)
return conv
#--------
#INPUT
#--------
#batch, height, width, channels
inputs_1 = Input((img_rows, img_cols, 1))
#-----------------------
#INPUT CONVERTER & VGG16
#-----------------------
inputs_3 = Gray2VGGInput(name='gray_to_rgb')(inputs_1) #shape=(img_rows, img_cols, 3)
base_VGG16 = VGG16(include_top=False, weights='imagenet', input_tensor=inputs_3)
#--------
#DECODER
#--------
c1 = base_VGG16.get_layer("block1_conv2").output #(None, 864, 1232, 64)
c2 = base_VGG16.get_layer("block2_conv2").output #(None, 432, 616, 128)
c3 = base_VGG16.get_layer("block3_conv2").output #(None, 216, 308, 256)
c4 = base_VGG16.get_layer("block4_conv2").output #(None, 108, 154, 512)
#--------
#BOTTLENECK
#--------
c5 = base_VGG16.get_layer("block5_conv2").output #(None, 54, 77, 512)
#--------
#ENCODER
#--------
c6 = upsampleLayer(in_layer=c5, concat_layer=c4, input_size=512)
c7 = upsampleLayer(in_layer=c6, concat_layer=c3, input_size=256)
c8 = upsampleLayer(in_layer=c7, concat_layer=c2, input_size=128)
c9 = upsampleLayer(in_layer=c8, concat_layer=c1, input_size=64)
#--------
#DENSE OUTPUT
#--------
outputs = Conv2D(1, (1, 1), activation='sigmoid')(c9)
model = Model(inputs=inputs_1, outputs=outputs)
#Freeze layers
for layer in model.layers[:16]:
layer.trainable = False
print(model.summary())
model.compile(optimizer='adam',
loss=fr.diceCoefLoss,
metrics=[fr.diceCoef])
return model
然后,我加载模型并调用predict
:
model = un.UNET1_VGG16()
pth_to_model = PTH_OUTPUT + 'weights__L_01.h5'
model.load_weights(pth_to_model)
preds = model.predict(X_image_test, verbose=1)
但是,结果如下:
[[0.4567569 0.4567569 0.4567569 ... 0.4567569 0.4567569 0.4567569]
[0.4567569 0.4567569 0.4567569 ... 0.4567569 0.4567569 0.4567569]
[0.4567569 0.4567569 0.4567569 ... 0.4567569 0.4567569 0.4567569]
...
[0.4567569 0.4567569 0.4567569 ... 0.4567569 0.4567569 0.4567569]
[0.4567569 0.4567569 0.4567569 ... 0.4567569 0.4567569 0.4567569]
[0.4567569 0.4567569 0.4567569 ... 0.4567569 0.4567569 0.4567569]]
我在没有VGG16的其他型号上使用相同的步骤,并且一切正常。因此,我认为与VGG16相关的某些问题是错误的。也许是输入层,我正在将其转换为“伪” RGB图像?
答案 0 :(得分:2)
问题在于VGG
冻结的图层。如果您的数据集与imagenet完全不同,也许您应该端到端训练整个模型。另外,很明显,如果冻结BatchNormalization
层,它们的行为可能会很奇怪。供参考,请参阅此discussion。
答案 1 :(得分:1)
如果您在某些特定的约束条件下训练网络(例如,减去平均值),则在测试网络时(由于通过测试,您还会通过网络进行前向传递),因此您需要减去平均值(就像在训练中)。
这可以解决您的问题。