# Ken0001-ML_DenseNet
###### tags: `note`
## Single label
```python=
def densenet(input_shape, num_class, f=32):
# depth
r = 6, 12, 24, 16
```
f = channel數，論文設定32
### Dense Block
```python=
def dense_block(a, r):
for i in range(r):
x = BatchNormalization(epsilon=1.001e-5)(a)
x = ReLU()(x)
x = Conv2D(f*4, (1, 1), strides=1, padding='same')(x)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = ReLU()(x)
x = Conv2D(f, (3, 3), strides=1, padding='same')(x)
a = Concatenate()([a,x])
return a
```
### Transition Block
```python=
def transition(x):
print(K.int_shape(x)[-1]//2)
x = Conv2D(K.int_shape(x)[-1]//2, (1,1), strides=1, padding='same')(x)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = ReLU()(x)
x = AveragePooling2D(2, strides=2)(x)
return x
```
int_shape = 顯示該feature的大小
```python=
input = Input(input_shape)
# Conv2D(filters, kernel_size, strides=(1, 1), padding='valid', ...)
x = Conv2D(64, (7, 7), strides=2, padding='same')(input)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = ReLU()(x)
x = MaxPooling2D(3, strides=2, padding='same')(x)
for i in r:
d = dense_block(x, i)
x = transition(d)
x = GlobalMaxPooling2D()(d)
output = Dense(num_class)(x)
model = Model(input, output)
return model
```
---
## Multi label
```python=
def mldensenet(input_shape, num_class, f=32):
# depth
r = 6, 12, 24, 16
```
### Dense Block
```python=
def dense_block(a, r):
for i in range(r):
x = Conv2D(f*4, (1, 1), strides=1, padding='same')(a)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = ReLU()(x)
x = Conv2D(f, (3, 3), strides=1, padding='same')(x)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = ReLU()(x)
a = Concatenate()([a,x])
return a
```
### Transition Block
```python=
def transition(x):
print(K.int_shape(x)[-1]//2)
x = Conv2D(K.int_shape(x)[-1]//2, (1,1), strides=1, padding='same')(x)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = ReLU()(x)
x = AveragePooling2D(2, strides=2)(x)
return x
```
```python=
input = Input(input_shape)
# Conv2D(filters, kernel_size, strides=(1, 1), padding='valid', ...)
x = Conv2D(64, (7, 7), strides=2, padding='same')(input)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = ReLU()(x)
x = MaxPooling2D(3, strides=2, padding='same')(x)
for i in r:
d = dense_block(x, i)
x = transition(d)
```
### Branch
```python=
final = []
for j in range(num_class):
branch = Conv2D(512, (3,3), strides=2, padding='same')(d)
branch = Conv2D(163, (3,3), strides=2, padding='same')(branch)
branch = Flatten()(branch)
branch = Dense(1024)(branch)
output = Dense(1, activation='sigmoid')(branch)
final.append(output)
final = Concatenate()(final)
model = Model(input, final)
return model
```
借我放圖
### 新版完整程式碼
```python=3.7
"""
ML-DenseNet
"""
from tensorflow.keras.models import Model, Sequential
from tensorflow.keras.layers import Input, Dense, Conv2D, MaxPooling2D, AveragePooling2D, Flatten, GlobalAvgPool2D, BatchNormalization, Concatenate, ReLU
import tensorflow.keras.backend as K
def mldensenet(input_shape, n_classes, mltype=0, finalAct='softmax', f=32):
# Define repetition
if mltype == 0 or mltype == 2 or mltype == 3:
repetitions = 6,12,48,32
elif mltype == 1 or mltype == 4 or mltype == 5:
repetitions = 6,12,32
""" Dense block
1x1x128 conv + 3x3x32 conv
"""
def dense_block(tensor, r):
for i in range(r):
x = BatchNormalization(epsilon=1.001e-5)(tensor)
x = ReLU()(x)
x = Conv2D(f*4, (1, 1), strides=1, padding='same')(x)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = ReLU()(x)
x = Conv2D(f, (3, 3), strides=1, padding='same')(x)
tensor = Concatenate()([tensor,x])
return tensor
""" Transition_layer
1x1x(input/2) conv + 2x2 pooling
"""
def transition_layer(x):
x = Conv2D(K.int_shape(x)[-1]//2, (1,1), strides=1, padding='same')(x)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = ReLU()(x)
x = AveragePooling2D(2, strides=2)(x)
return x
""" Dense Block on branch
Same as default dense block
"""
def dense_block_branch(x, outputs):
depth = 16
for i in range(n_classes):
d = dense_block(x, depth)
branch = GlobalAvgPool2D()(d)
output = Dense(1, activation=finalAct)(branch)
outputs.append(output)
outputs = Concatenate()(outputs)
return outputs
""" Multilayer Perceptron on branch
3 fully connected layers
"""
def fc_branch(x, outputs):
x = GlobalAvgPool2D()(d)
for i in range(n_classes):
x = Dense(1024, activation='relu')(x)
x = Dense(512, activation='relu')(x)
x = Dense(256, activation='relu')(x)
output = Dense(1, activation=finalAct)(x)
outputs.append(output)
outputs = Concatenate()(outputs)
return outputs
""" Convolution Nerual Network on branch
6 conv layers + 3 fully connected layers
"""
def cnn_branch(x, outputs):
for i in range(n_classes):
v = Conv2D(256, (3, 3), activation='relu', padding='same')(x)
v = Conv2D(256, (3, 3), activation='relu', padding='same')(v)
v = Conv2D(256, (3, 3), activation='relu', padding='same')(v)
v = MaxPooling2D((2, 2), strides=(2, 2))(v)
v = Conv2D(512, (3, 3), activation='relu', padding='same')(v)
v = Conv2D(512, (3, 3), activation='relu', padding='same')(v)
v = Conv2D(512, (3, 3), activation='relu', padding='same')(v)
v = MaxPooling2D((2, 2), strides=(2, 2))(v)
v = Flatten()(v)
d = Dense(2048, activation='relu')(v)
d = Dense(2048, activation='relu')(d)
d = Dense(1024, activation='relu')(d)
output = Dense(1, activation=finalAct)(d)
outputs.append(output)
outputs = Concatenate()(outputs)
return outputs
input = Input(input_shape)
x = Conv2D(64, (7, 7), strides=2, padding='same')(input)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = ReLU()(x)
x = MaxPooling2D(3, strides=2, padding='same')(x)
for r in repetitions:
d = dense_block(x, r)
x = transition_layer(d)
outputs = []
# Original DenseNet
if mltype==0:
x = GlobalAvgPool2D()(d)
outputs = Dense(n_classes, activation=finalAct)(x)
# ML-DenseNet
elif mltype==1:
outputs = dense_block_branch(x, outputs)
elif mltype==2 or mltype == 4:
outputs = fc_branch(x, outputs)
elif mltype==3:
outputs = cnn_branch(d, outputs)
elif mltype==5:
outputs = cnn_branch(x, outputs)
model = Model(input, outputs)
return model
```
Sign in with Wallet
Connect another wallet