from keras.preprocessing.image import ImageDataGenerator
import tensorflow as tf
import numpy as n
# gene model define
class gene(tf.keras.Model):
def __init__(self):
super().__init__()
self.conv1 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[5,5],
input_shape=(150,150,1))
self.leak1 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.conv2 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[5,5])
self.leak2 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.conv3 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[4,4])
self.leak3 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.conv4 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[4,4])
self.leak4 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.conv5 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[4,4])
self.leak5 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.conv6 = tf.keras.layers.Conv2D(filters=64,
kernel_size=[3,3])
self.leak6 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.conv7 = tf.keras.layers.Conv2D(filters=64,
kernel_size=[3,3])
self.leak7 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.conv8 = tf.keras.layers.Conv2D(filters=64,
kernel_size=[3,3])
self.leak8 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.deconv1 = tf.keras.layers.Conv2DTranspose(64,
kernel_size=[3,3])
self.deleak1 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.add1 = tf.keras.layers.Add()
self.deconv2 = tf.keras.layers.Conv2DTranspose(64,
kernel_size=[3,3])
self.deleak2 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.add2 = tf.keras.layers.Add()
self.deconv3 = tf.keras.layers.Conv2DTranspose(128,
kernel_size=[3,3])
self.deleak3 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.add3 = tf.keras.layers.Add()
self.deconv4 = tf.keras.layers.Conv2DTranspose(128,
kernel_size=[4,4])
self.deleak4 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.add4 = tf.keras.layers.Add()
self.deconv5 = tf.keras.layers.Conv2DTranspose(128,
kernel_size=[4,4])
self.deleak5 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.add5 = tf.keras.layers.Add()
self.deconv6 = tf.keras.layers.Conv2DTranspose(128,
kernel_size=[4,4])
self.deleak6 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.add6 = tf.keras.layers.Add()
self.deconv7 = tf.keras.layers.Conv2DTranspose(128,
kernel_size=[5,5])
self.deleak7 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.add7= tf.keras.layers.Add()
self.deconv8 = tf.keras.layers.Conv2DTranspose(128,
kernel_size=[5,5])
self.deleak8 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.dense = tf.keras.layers.Dense(1,
activation=tf.nn.relu)
self.add = tf.keras.layers.Add()
def call(self,inputs):
x1 = self.conv1(inputs)
x1 = self.leak1(x1)
x2 = self.conv2(x1)
x2 = self.leak2(x2)
x3 = self.conv3(x2)
x3 = self.leak3(x3)
x4 = self.conv4(x3)
x4 = self.leak4(x4)
x5 = self.conv5(x4)
x5 = self.leak5(x5)
x6 = self.conv6(x5)
x6 = self.leak6(x6)
x7 = self.conv7(x6)
x7 = self.leak7(x7)
x8 = self.conv8(x7)
x8 = self.leak8(x8)
y1 = self.deconv1(x8)
y1 = self.deleak1(y1)
y1 = self.add1([x7,y1])
y2 = self.deconv2(y1)
y2 = self.deleak2(y2)
y2 = self.add2([x6,y2])
y3 = self.deconv3(y2)
y3 = self.deleak3(y3)
y3 = self.add3([x5,y3])
y4 = self.deconv4(y3)
y4 = self.deleak4(y4)
y4 = self.add4([x4,y4])
y5 = self.deconv5(y4)
y5 = self.deleak5(y5)
y5 = self.add5([x3,y5])
y6 = self.deconv6(y5)
y6 = self.deleak6(y6)
y6 = self.add6([x2,y6])
y7 = self.deconv7(y6)
y7 = self.deleak7(y7)
y7 = self.add7([x1,y7])
y8 = self.deconv8(y7)
y8 = self.deleak8(y8)
x = self.dense(y8)
output = self.add([x,inputs])
return output
class dis(tf.keras.Model):
def __init__(self):
super().__init__()
self.con1 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[5,5],
input_shape=(150,150,1))
self.leak1 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.con2 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[5,5])
self.leak2 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.con3 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[5,5])
self.leak3 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.con4 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[5,5])
self.leak4 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.con5 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[5,5])
self.leak5 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.con6 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[5,5])
self.leak6 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.con7 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[5,5])
self.leak7 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.con8 = tf.keras.layers.Conv2D(filters=128,
kernel_size=[5,5])
self.leak8 = tf.keras.layers.LeakyReLU(alpha =0.4)
self.flatten = tf.keras.layers.Flatten()
self.dense1 = tf.keras.layers.Dense(1,
activation=tf.nn.softmax)
def call(self,inputs):
x = self.con1(inputs)
x = self.leak1(x)
x = self.con2(x)
x = self.leak2(x)
x = self.con3(x)
x = self.leak3(x)
x = self.con4(x)
x = self.leak4(x)
x = self.con5(x)
x = self.leak5(x)
x = self.con6(x)
x = self.leak6(x)
x = self.con7(x)
x = self.leak7(x)
x = self.con8(x)
x = self.leak8(x)
x = self.flatten(x)
x = tf.nn.dropout(x, 0.5)
output = self.dense1(x)
return output
gene1 = gene()
dis1 = dis()
这是模型定义
#imagegenerator
num=0
learning_rate = 0.001
optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)
train_datagen = ImageDataGenerator(rescale=1./255)
train_gene = train_datagen.flow_from_directory(
'data/',
color_mode='grayscale',
target_size=(150, 150),
batch_size=10,
class_mode=None)
图像生成器
from undersample import undersample
for num in range(1000):
for x in train_gene:
x_u = []
for i in range(x.shape[0]):
x_u.append(undersample(x[i,:,:,:],0.7))
x_u = n.array(x_u)
x_r = gene1(x_u)
D_train = n.vstack((x,x_r))
D_label = n.vstack((n.zeros((x.shape[0],1)),n.ones((x.shape[0],1))))
with tf.GradientTape() as tape:
D = dis1(D_train)
loss = tf.keras.losses.binary_crossentropy(y_true=D_label,y_pred=D)
loss = tf.reduce_mean(loss)
print("batch %d: D_loss %f" % (num, loss.numpy()))
grads = tape.gradient(loss , dis1.variables)
optimizer.apply_gradients(grads_and_vars=zip(grads , dis1.variables))
break
for x in train_gene:
x_u = []
for i in range(x.shape[0]):
x_u.append(undersample(x[i,:,:,:],0.7))
x_u = n.array(x_u)
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(gene1.variables)
D = dis1(gene1(x_u))
loss = tf.keras.losses.binary_crossentropy(y_true=n.zeros((x.shape[0],1)),y_pred=D)
loss = tf.reduce_mean(loss)
print('batch %d: G_loss %f' % (num, loss.numpy()))
grads = tape.gradient(loss , gene1.variables)
optimizer.apply_gradients(grads_and_vars=zip(grads , gene1.variables))
break
列车损耗
我想训练一个 gan 模型，该模型将欠采样 MRI 图像重建为正常的 MRI 图像。但是当我使用上面的代码训练模型时，我得到了 D_loss/G_loss，即 7.666619/15.333239，并且永远不会改变。这是代码问题器官模型火车不是这样工作的吗？
你想要的是输入发生器噪声并要求生成MRI图像。但是，据我所知，您正在输入MRI图像，然后要求它检查MRI图像。
您需要发电机中的噪声，当它出来时，您需要使用原始图像mean_squared_error，然后，您还需要使用原始标签和通过鉴别器获得的标签进行binary_crossentropy。
经过一些纪元，鉴于您的模型是合理的，您将看到良好的图像。
当您开始时，鉴别器将具有很高的准确性，并且随着时间的减少而减少。这就是你想要的。
https://towardsdatascience.com/gan-by-example-using-keras-on-tensorflow-backend-1a6d515a60d0