问题描述
我使用 TensorFlow 2.2 来实现 Unet。 这是我正在构建的代码，它给了我 logits vs labels 的错误。已附上错误图片。
**imports**
import numpy as np
import os
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
**Double Conv**
class Down(layers.Layer):
def __init__(self,channels):
super(Down,self).__init__()
self.conv1 = layers.Conv2D(channels,(3,3),padding='valid',activation='relu')
self.conv2 = layers.Conv2D(channels,activation='relu')
def call(self,input_tensor,training=False):
x = self.conv1(input_tensor,training=training)
x = self.conv2(x,training=training)
return x
**crop tensor**
def crp_img(tensor,target):
x = tensor.get_shape().as_list()
t_h = x[1]
y = target.get_shape().as_list()
tg_h= y[1]
delta = t_h - tg_h
delta = delta // 2
#target = tf.image.resize_with_pad(target,t_h,t_w)
return tensor[:,delta:t_h-delta,:]
**Unet**
class U_net(keras.Model):
def __init__(self):
super(U_net,self).__init__()
self.down1 = Down(64)
self.down2 = Down(128)
self.down3 = Down(256)
self.down4 = Down(512)
self.down5 = Down(1024)
self.pool = layers.MaxPooling2D((2,2),2)
self.cv_T1= layers.Conv2DTranspose(512,(2,strides=(2,2))
self.cv_T2 = layers.Conv2DTranspose(256,2))
self.cv_T3 = layers.Conv2DTranspose(128,2))
self.cv_T4 = layers.Conv2DTranspose(64,2))
self.up1 = Down(512)
self.up2 = Down(256)
self.up3 = Down(128)
self.up4 = Down(64)
def call(self,training=False):
d1 = self.down1(input_tensor,training=training)#
x1 = self.pool(d1)
print("Down_1 ",x1.shape)
d2 = self.down2(x1,training=training)#
x2 = self.pool(d2)
print("Down_2 ",x2.shape)
d3 = self.down3(x2,training=training)#
x3 = self.pool(d3)
print("Down_3 ",x3.shape)
d4 = self.down4(x3,training=training)#
x4 = self.pool(d4)
print("Down_4 ",x4.shape)
x5 = self.down5(x4,training=training)
#mid = layers.Conv2D(1024,padding='same',activation='relu')(x4)
print("Down_5 ",x5.shape)
up_1 = self.cv_T1(x5)
y_1 = crp_img(d4,up_1)
x_11 = layers.concatenate([up_1,y_1])
#print(x_11.shape)
up = self.up1(x_11)
print(up.shape)
up_2 = self.cv_T2(up)
y_2 = crp_img(d3,up_2)
x_22 = layers.concatenate([up_2,y_2])
#rint(x_22.shape)
up_c = self.up2(x_22)
print(up_c.shape)
up_3 = self.cv_T3(up_c)
print(up_3.shape)
#up_3 = self.cv_T3(up_c)
y_3 = crp_img(d2,up_3)
#print(y_3.shape)
x_33 = layers.concatenate([up_3,y_3])
up_x = self.up3(x_33)
print(up_x.shape)
up_4 = self.cv_T4(up_x)
y_4 = crp_img(d1,up_4)
x_44 = layers.concatenate([up_4,y_4])
#print(x_44.shape)
up_z = self.up4(x_44)
# print(up_z.shape)
print(up_z.shape)
#print(conv_x.shape)
z = layers.Conv2D(1,(1,1),activation='sigmoid')(up_z)
print(z.shape)
return z
def model(self):
x = keras.Input(shape=(None,None,3))
return keras.Model(inputs=[x],outputs=self.call(x))
**Model Call and Training**
model = U_net()
model.compile(optimizer=keras.optimizers.Adam(lr=0.001),loss=keras.losses.BinaryCrossentropy(from_logits=False),metrics=["accuracy"])
model.fit(train_set,epochs=epochs,steps_per_epoch=len(train_images)//bs,verbose=1)
当我尝试训练上述模型时，会产生 logits 与标签的错误 ValueError：logits 和标签必须具有相同的形状 ((None,388,1) vs (None,572,1))
解决方法
暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！
如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。
小编邮箱:dio#foxmail.com (将#修改为@）