ResUnet实现（tensorflow2）
tensorflow
一个面向所有人的开源机器学习框架
项目地址：https://gitcode.com/gh_mirrors/te/tensorflow
免费下载资源
·
我们将在本篇博客实现Resunet。
首先上Unet图
然后结合Unet和Resnet，就是一个新的网络。
def bn_act(x, act=True):
'batch normalization layer with an optinal activation layer'
x = tf.keras.layers.BatchNormalization()(x)
if act == True:
x = tf.keras.layers.Activation('relu')(x)
return x
以下方框代表数据，圆形代表方法，虚线代表判断
bn_act方法图
def conv_block(x, filters, kernel_size=3, padding='same', strides=1):
'convolutional layer which always uses the batch normalization layer'
conv = bn_act(x)
conv = Conv2D(filters, kernel_size, padding=padding, strides=strides)(conv)
return conv
conv_block方法图
stem方法图
def stem(x, filters, kernel_size=3, padding='same', strides=1):
conv = Conv2D(filters, kernel_size, padding=padding, strides=strides)(x)
conv = conv_block(conv, filters, kernel_size, padding, strides)
shortcut = Conv2D(filters, kernel_size=1, padding=padding, strides=strides)(x)
shortcut = bn_act(shortcut, act=False)
output = Add()([conv, shortcut])
return output
tensorflow
一个面向所有人的开源机器学习框架
项目地址：https://gitcode.com/gh_mirrors/te/tensorflow
residual_block方法图
def residual_block(x, filters, kernel_size=3, padding='same', strides=1):
res = conv_block(x, filters, k_size, padding, strides)
res = conv_block(res, filters, k_size, padding, 1)
shortcut = Conv2D(filters, kernel_size, padding=padding, strides=strides)(x)
shortcut = bn_act(shortcut, act=False)
output = Add()([shortcut, res])
return output
def upsample_concat_block(x, xskip):
u = UpSampling2D((2,2))(x)
c = Concatenate()([u, xskip])
return c
ResUNet方法图
def ResUNet(img_h, img_w):
f = [16, 32, 64, 128, 256]
inputs = Input((img_h, img_w, 1))
## Encoder
e0 = inputs
e1 = stem(e0, f[0])
e2 = residual_block(e1, f[1], strides=2)
e3 = residual_block(e2, f[2], strides=2)
e4 = residual_block(e3, f[3], strides=2)
e5 = residual_block(e4, f[4], strides=2)
## Bridge
b0 = conv_block(e5, f[4], strides=1)
b1 = conv_block(b0, f[4], strides=1)
## Decoder
u1 = upsample_concat_block(b1, e4)
d1 = residual_block(u1, f[4])
u2 = upsample_concat_block(d1, e3)
d2 = residual_block(u2, f[3])
u3 = upsample_concat_block(d2, e2)
d3 = residual_block(u3, f[2])
u4 = upsample_concat_block(d3, e1)
d4 = residual_block(u4, f[1])
outputs = tf.keras.layers.Conv2D(4, (1, 1), padding="same", activation="sigmoid")(d4)
model = tf.keras.models.Model(inputs, outputs)
return model
这个图就不画了。
可以看下这个模型的概括。
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 256, 800, 1) 0
__________________________________________________________________________________________________
conv2d_167 (Conv2D) (None, 256, 800, 16) 160 input_1[0][0]
__________________________________________________________________________________________________
batch_normalization_75 (BatchNo (None, 256, 800, 16) 64 conv2d_167[0][0]
__________________________________________________________________________________________________
activation_74 (Activation) (None, 256, 800, 16) 0 batch_normalization_75[0][0]
__________________________________________________________________________________________________
conv2d_169 (Conv2D) (None, 256, 800, 16) 32 input_1[0][0]
__________________________________________________________________________________________________
conv2d_168 (Conv2D) (None, 256, 800, 16) 2320 activation_74[0][0]
__________________________________________________________________________________________________
batch_normalization_76 (BatchNo (None, 256, 800, 16) 64 conv2d_169[0][0]
__________________________________________________________________________________________________
add (Add) (None, 256, 800, 16) 0 conv2d_168[0][0]
batch_normalization_76[0][0]
__________________________________________________________________________________________________
batch_normalization_77 (BatchNo (None, 256, 800, 16) 64 add[0][0]
__________________________________________________________________________________________________
activation_75 (Activation) (None, 256, 800, 16) 0 batch_normalization_77[0][0]
__________________________________________________________________________________________________
conv2d_170 (Conv2D) (None, 128, 400, 32) 4640 activation_75[0][0]
__________________________________________________________________________________________________
batch_normalization_78 (BatchNo (None, 128, 400, 32) 128 conv2d_170[0][0]
__________________________________________________________________________________________________
conv2d_172 (Conv2D) (None, 128, 400, 32) 4640 add[0][0]
__________________________________________________________________________________________________
activation_76 (Activation) (None, 128, 400, 32) 0 batch_normalization_78[0][0]
__________________________________________________________________________________________________
batch_normalization_79 (BatchNo (None, 128, 400, 32) 128 conv2d_172[0][0]
__________________________________________________________________________________________________
conv2d_171 (Conv2D) (None, 128, 400, 32) 9248 activation_76[0][0]
__________________________________________________________________________________________________
add_1 (Add) (None, 128, 400, 32) 0 batch_normalization_79[0][0]
conv2d_171[0][0]
__________________________________________________________________________________________________
batch_normalization_80 (BatchNo (None, 128, 400, 32) 128 add_1[0][0]
__________________________________________________________________________________________________
activation_77 (Activation) (None, 128, 400, 32) 0 batch_normalization_80[0][0]
__________________________________________________________________________________________________
conv2d_173 (Conv2D) (None, 64, 200, 64) 18496 activation_77[0][0]
__________________________________________________________________________________________________
batch_normalization_81 (BatchNo (None, 64, 200, 64) 256 conv2d_173[0][0]
__________________________________________________________________________________________________
conv2d_175 (Conv2D) (None, 64, 200, 64) 18496 add_1[0][0]
__________________________________________________________________________________________________
activation_78 (Activation) (None, 64, 200, 64) 0 batch_normalization_81[0][0]
__________________________________________________________________________________________________
batch_normalization_82 (BatchNo (None, 64, 200, 64) 256 conv2d_175[0][0]
__________________________________________________________________________________________________
conv2d_174 (Conv2D) (None, 64, 200, 64) 36928 activation_78[0][0]
__________________________________________________________________________________________________
add_2 (Add) (None, 64, 200, 64) 0 batch_normalization_82[0][0]
conv2d_174[0][0]
__________________________________________________________________________________________________
batch_normalization_83 (BatchNo (None, 64, 200, 64) 256 add_2[0][0]
__________________________________________________________________________________________________
activation_79 (Activation) (None, 64, 200, 64) 0 batch_normalization_83[0][0]
__________________________________________________________________________________________________
conv2d_176 (Conv2D) (None, 32, 100, 128) 73856 activation_79[0][0]
__________________________________________________________________________________________________
batch_normalization_84 (BatchNo (None, 32, 100, 128) 512 conv2d_176[0][0]
__________________________________________________________________________________________________
conv2d_178 (Conv2D) (None, 32, 100, 128) 73856 add_2[0][0]
__________________________________________________________________________________________________
activation_80 (Activation) (None, 32, 100, 128) 0 batch_normalization_84[0][0]
__________________________________________________________________________________________________
batch_normalization_85 (BatchNo (None, 32, 100, 128) 512 conv2d_178[0][0]
__________________________________________________________________________________________________
conv2d_177 (Conv2D) (None, 32, 100, 128) 147584 activation_80[0][0]
__________________________________________________________________________________________________
add_3 (Add) (None, 32, 100, 128) 0 batch_normalization_85[0][0]
conv2d_177[0][0]
__________________________________________________________________________________________________
batch_normalization_86 (BatchNo (None, 32, 100, 128) 512 add_3[0][0]
__________________________________________________________________________________________________
activation_81 (Activation) (None, 32, 100, 128) 0 batch_normalization_86[0][0]
__________________________________________________________________________________________________
conv2d_179 (Conv2D) (None, 16, 50, 256) 295168 activation_81[0][0]
__________________________________________________________________________________________________
batch_normalization_87 (BatchNo (None, 16, 50, 256) 1024 conv2d_179[0][0]
__________________________________________________________________________________________________
conv2d_181 (Conv2D) (None, 16, 50, 256) 295168 add_3[0][0]
__________________________________________________________________________________________________
activation_82 (Activation) (None, 16, 50, 256) 0 batch_normalization_87[0][0]
__________________________________________________________________________________________________
batch_normalization_88 (BatchNo (None, 16, 50, 256) 1024 conv2d_181[0][0]
__________________________________________________________________________________________________
conv2d_180 (Conv2D) (None, 16, 50, 256) 590080 activation_82[0][0]
__________________________________________________________________________________________________
add_4 (Add) (None, 16, 50, 256) 0 batch_normalization_88[0][0]
conv2d_180[0][0]
__________________________________________________________________________________________________
batch_normalization_89 (BatchNo (None, 16, 50, 256) 1024 add_4[0][0]
__________________________________________________________________________________________________
activation_83 (Activation) (None, 16, 50, 256) 0 batch_normalization_89[0][0]
__________________________________________________________________________________________________
conv2d_182 (Conv2D) (None, 16, 50, 256) 590080 activation_83[0][0]
__________________________________________________________________________________________________
batch_normalization_90 (BatchNo (None, 16, 50, 256) 1024 conv2d_182[0][0]
__________________________________________________________________________________________________
activation_84 (Activation) (None, 16, 50, 256) 0 batch_normalization_90[0][0]
__________________________________________________________________________________________________
conv2d_183 (Conv2D) (None, 16, 50, 256) 590080 activation_84[0][0]
__________________________________________________________________________________________________
up_sampling2d (UpSampling2D) (None, 32, 100, 256) 0 conv2d_183[0][0]
__________________________________________________________________________________________________
concatenate_36 (Concatenate) (None, 32, 100, 384) 0 up_sampling2d[0][0]
add_3[0][0]
__________________________________________________________________________________________________
batch_normalization_91 (BatchNo (None, 32, 100, 384) 1536 concatenate_36[0][0]
__________________________________________________________________________________________________
activation_85 (Activation) (None, 32, 100, 384) 0 batch_normalization_91[0][0]
__________________________________________________________________________________________________
conv2d_184 (Conv2D) (None, 32, 100, 256) 884992 activation_85[0][0]
__________________________________________________________________________________________________
batch_normalization_92 (BatchNo (None, 32, 100, 256) 1024 conv2d_184[0][0]
__________________________________________________________________________________________________
conv2d_186 (Conv2D) (None, 32, 100, 256) 884992 concatenate_36[0][0]
__________________________________________________________________________________________________
activation_86 (Activation) (None, 32, 100, 256) 0 batch_normalization_92[0][0]
__________________________________________________________________________________________________
batch_normalization_93 (BatchNo (None, 32, 100, 256) 1024 conv2d_186[0][0]
__________________________________________________________________________________________________
conv2d_185 (Conv2D) (None, 32, 100, 256) 590080 activation_86[0][0]
__________________________________________________________________________________________________
add_5 (Add) (None, 32, 100, 256) 0 batch_normalization_93[0][0]
conv2d_185[0][0]
__________________________________________________________________________________________________
up_sampling2d_1 (UpSampling2D) (None, 64, 200, 256) 0 add_5[0][0]
__________________________________________________________________________________________________
concatenate_37 (Concatenate) (None, 64, 200, 320) 0 up_sampling2d_1[0][0]
add_2[0][0]
__________________________________________________________________________________________________
batch_normalization_94 (BatchNo (None, 64, 200, 320) 1280 concatenate_37[0][0]
__________________________________________________________________________________________________
activation_87 (Activation) (None, 64, 200, 320) 0 batch_normalization_94[0][0]
__________________________________________________________________________________________________
conv2d_187 (Conv2D) (None, 64, 200, 128) 368768 activation_87[0][0]
__________________________________________________________________________________________________
batch_normalization_95 (BatchNo (None, 64, 200, 128) 512 conv2d_187[0][0]
__________________________________________________________________________________________________
conv2d_189 (Conv2D) (None, 64, 200, 128) 368768 concatenate_37[0][0]
__________________________________________________________________________________________________
activation_88 (Activation) (None, 64, 200, 128) 0 batch_normalization_95[0][0]
__________________________________________________________________________________________________
batch_normalization_96 (BatchNo (None, 64, 200, 128) 512 conv2d_189[0][0]
__________________________________________________________________________________________________
conv2d_188 (Conv2D) (None, 64, 200, 128) 147584 activation_88[0][0]
__________________________________________________________________________________________________
add_6 (Add) (None, 64, 200, 128) 0 batch_normalization_96[0][0]
conv2d_188[0][0]
__________________________________________________________________________________________________
up_sampling2d_2 (UpSampling2D) (None, 128, 400, 128 0 add_6[0][0]
__________________________________________________________________________________________________
concatenate_38 (Concatenate) (None, 128, 400, 160 0 up_sampling2d_2[0][0]
add_1[0][0]
__________________________________________________________________________________________________
batch_normalization_97 (BatchNo (None, 128, 400, 160 640 concatenate_38[0][0]
__________________________________________________________________________________________________
activation_89 (Activation) (None, 128, 400, 160 0 batch_normalization_97[0][0]
__________________________________________________________________________________________________
conv2d_190 (Conv2D) (None, 128, 400, 64) 92224 activation_89[0][0]
__________________________________________________________________________________________________
batch_normalization_98 (BatchNo (None, 128, 400, 64) 256 conv2d_190[0][0]
__________________________________________________________________________________________________
conv2d_192 (Conv2D) (None, 128, 400, 64) 92224 concatenate_38[0][0]
__________________________________________________________________________________________________
activation_90 (Activation) (None, 128, 400, 64) 0 batch_normalization_98[0][0]
__________________________________________________________________________________________________
batch_normalization_99 (BatchNo (None, 128, 400, 64) 256 conv2d_192[0][0]
__________________________________________________________________________________________________
conv2d_191 (Conv2D) (None, 128, 400, 64) 36928 activation_90[0][0]
__________________________________________________________________________________________________
add_7 (Add) (None, 128, 400, 64) 0 batch_normalization_99[0][0]
conv2d_191[0][0]
__________________________________________________________________________________________________
up_sampling2d_3 (UpSampling2D) (None, 256, 800, 64) 0 add_7[0][0]
__________________________________________________________________________________________________
concatenate_39 (Concatenate) (None, 256, 800, 80) 0 up_sampling2d_3[0][0]
add[0][0]
__________________________________________________________________________________________________
batch_normalization_100 (BatchN (None, 256, 800, 80) 320 concatenate_39[0][0]
__________________________________________________________________________________________________
activation_91 (Activation) (None, 256, 800, 80) 0 batch_normalization_100[0][0]
__________________________________________________________________________________________________
conv2d_193 (Conv2D) (None, 256, 800, 32) 23072 activation_91[0][0]
__________________________________________________________________________________________________
batch_normalization_101 (BatchN (None, 256, 800, 32) 128 conv2d_193[0][0]
__________________________________________________________________________________________________
conv2d_195 (Conv2D) (None, 256, 800, 32) 23072 concatenate_39[0][0]
__________________________________________________________________________________________________
activation_92 (Activation) (None, 256, 800, 32) 0 batch_normalization_101[0][0]
__________________________________________________________________________________________________
batch_normalization_102 (BatchN (None, 256, 800, 32) 128 conv2d_195[0][0]
__________________________________________________________________________________________________
conv2d_194 (Conv2D) (None, 256, 800, 32) 9248 activation_92[0][0]
__________________________________________________________________________________________________
add_8 (Add) (None, 256, 800, 32) 0 batch_normalization_102[0][0]
conv2d_194[0][0]
__________________________________________________________________________________________________
conv2d_196 (Conv2D) (None, 256, 800, 4) 132 add_8[0][0]
==================================================================================================
Total params: 6,287,508
Trainable params: 6,280,212
Non-trainable params: 7,296
__________________________________________________________________________________________________
推荐内容
阅读全文
AI总结
GitHub 加速计划 / te / tensorflow
28
4
下载
一个面向所有人的开源机器学习框架
最近提交(Master分支：2 个月前 )
4f64a3d5
Instead, check for this case in `ResolveUsers` and `ResolveOperand`, by querying whether the `fused_expression_root` is part of the `HloFusionAdaptor`.
This prevents us from stepping into nested fusions.
PiperOrigin-RevId: 724311958
2 个月前
aa7e952e
Fix a bug in handling negative strides, and add a test case that exposes it.
We can have negative strides that are not just -1, e.g. with a combining
reshape.
PiperOrigin-RevId: 724293790
2 个月前
旨在为数千万中国开发者提供一个无缝且高效的云端环境，以支持学习、使用和贡献开源项目。
更多推荐
4
0- 0
已为社区贡献1条内容
相关推荐
查看更多
tensorflow
28
一个面向所有人的开源机器学习框架
tensorflow
0
TensorFlow for R
TensorFlow
0
Project containig related material for my TensorFlow articles
所有评论(0)