公司新闻
Tensorflow 2.* 网络训练(一) compile(optimizer, loss, metrics, loss_weights)
以下是使用ConvRNN2D层替换ConvLSTM2D层的修改后的代码:
```python
from tensorflow.keras.layers import Conv2D, Conv2DTranspose, ConvRNN2D, Input, BatchNormalization, LeakyReLU, Flatten, Dense, Reshape
from tensorflow.keras.models import Model
class Model():
def __init__(self):
self.img_seq_shape=(10,128,128,3)
self.img_shape=(128,128,3)
self.train_img=dataset
patch=int(128 / 2 ** 4)
self.disc_patch=(patch, patch, 1)
self.optimizer=tf.keras.optimizers.Adam(learning_rate=0.001)
self.build_generator=self.build_generator()
self.build_discriminator=self.build_discriminator()
self.build_discriminator.compile(loss='binary_crossentropy',
optimizer=self.optimizer,
metrics=['accuracy'])
self.build_generator.compile(loss='binary_crossentropy',
optimizer=self.optimizer)
img_seq_A=Input(shape=(10,128,128,3)) #输入图片
img_B=Input(shape=self.img_shape) #目标图片
fake_B=self.build_generator(img_seq_A) #生成的伪目标图片
self.build_discriminator.trainable=False
valid=self.build_discriminator([img_seq_A, fake_B])
self.combined=Model([img_seq_A, img_B], [valid, fake_B])
self.combined.compile(loss=['binary_crossentropy', 'mse'],
loss_weights=[1, 100],
optimizer=self.optimizer,
metrics=['accuracy'])
def build_generator(self):
def res_net(inputs, filters):
x=inputs
net=conv2d(x, filters // 2, (1, 1), 1)
net=conv2d(net, filters, (3, 3), 1)
net=net + x
# net=tf.keras.layers.LeakyReLU(0.2)(net)
return net
def conv2d(inputs, filters, kernel_size, strides):
x=Conv2D(filters, kernel_size, strides, 'same')(inputs)
x=BatchNormalization()(x)
x=LeakyReLU(alpha=0.2)(x)
return x
d0=Input(shape=(10, 128, 128, 3))
out=ConvRNN2D(filters=32, kernel_size=3, padding='same')(d0)
out=conv2d(out, 3, 1, 1)
return Model(inputs=d0, outputs=out)
def build_discriminator(self):
def d_layer(layer_input, filters, f_size=4, normalization=True):
"""Discriminator layer"""
d=Conv2D(filters, kernel_size=f_size, strides=2, padding='same')(layer_input)
if normalization:
d=BatchNormalization()(d)
d=LeakyReLU(alpha=0.2)(d)
return d
img_A=Input(shape=self.img_seq_shape)
img_B=Input(shape=self.img_shape)
combined_imgs=tf.keras.layers.concatenate([img_A, img_B])
d1=d_layer(combined_imgs, 64, normalization=False)
d2=d_layer(d1, 128)
d3=d_layer(d2, 256)
d4=d_layer(d3, 512)
validity=Conv2D(1, kernel_size=4, strides=1, padding='same')(d4)
return Model([img_A, img_B], validity)
```
在代码中,我们将ConvLSTM2D层替换为了ConvRNN2D层,并添加了新的模块`convrnn`。