conv_model = MODELS.Sequential()
loc_model = MODELS.Sequential()
model = MODELS.Sequential()
if conv1:
conv_model.add(
CONV.Convolution2D(conv1_filters,
conv1_filter_size,
conv1_filter_size,
subsample=(conv1_stride, conv1_stride),
border_mode='valid',
input_shape=(prev_frames, image_size, image_size)))
if pool1:
conv_model.add(CONV.MaxPooling2D(pool_size=(pool1_size, pool1_size)))
conv_model.add(CORE.Activation(conv1_act))
conv_model.add(CORE.Flatten())
conv_model.add(CORE.Dense(fc1_size))
conv_model.add(CORE.Activation(fc1_act))
loc_model.add(CORE.Dense(fc1_size, input_shape=(prev_frames * 4, )))
loc_model.add(CORE.Activation(fc1_act))
#model.add(CONV.Convolution2D(conv2_filters, conv2_filter_size, conv2_filter_size, border_mode='valid'))
#model.add(CONV.MaxPooling2D(pool_size=(pool2_size, pool2_size)))
#model.add(CORE.Activation(conv2_act))
model.add(CORE.Merge([conv_model, loc_model], mode='concat'))
model.add(CORE.Dense(4, init='zero'))
model.add(CORE.Activation(fc2_act))
print 'Building bouncing MNIST generator'
from data_handler import *
def test_activation(self):
layer = core.Activation('linear')
self._runner(layer)
def Unet(nClasses,
optimizer=None,
input_width=360,
input_height=480,
nChannels=1):
inputs = Input((nChannels, input_height, input_width))
conv1 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(inputs)
conv1 = Dropout(0.2)(conv1)
conv1 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(pool1)
conv2 = Dropout(0.2)(conv2)
conv2 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(pool2)
conv3 = Dropout(0.2)(conv3)
conv3 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(conv3)
up1 = merge([UpSampling2D(size=(2, 2))(conv3), conv2],
mode='concat',
concat_axis=1)
conv4 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(up1)
conv4 = Dropout(0.2)(conv4)
conv4 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(conv4)
up2 = merge([UpSampling2D(size=(2, 2))(conv4), conv1],
mode='concat',
concat_axis=1)
conv5 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(up2)
conv5 = Dropout(0.2)(conv5)
conv5 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv5)
conv6 = Convolution2D(nClasses,
1,
1,
activation='relu',
border_mode='same')(conv5)
conv6 = core.Reshape((nClasses, input_height * input_width))(conv6)
conv6 = core.Permute((2, 1))(conv6)
conv7 = core.Activation('softmax')(conv6)
model = Model(input=inputs, output=conv7)
if not optimizer is None:
model.compile(loss="categorical_crossentropy",
optimizer=optimizer,
metrics=['accuracy'])
return model
def test_activation():
layer = core.Activation('linear')
_runner(layer)
def get_unet(n_ch, patch_height, patch_width):
inputs = Input(shape=(n_ch, patch_height, patch_width))
#data_format：字符串，“channels_first”或“channels_last”之一，代表图像的通道维的位置。
#以128x128的RGB图像为例，“channels_first”应将数据组织为（3,128,128），而“channels_last”应将数据组织为（128,128,3）。该参数的默认值是~/.keras/keras.json中设置的值，若从未设置过，则为“channels_last”。
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(inputs)
conv1 = Dropout(0.2)(conv1)
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv1)
pool1 = MaxPooling2D((2, 2))(conv1)
#
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(pool1)
conv2 = Dropout(0.2)(conv2)
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv2)
pool2 = MaxPooling2D((2, 2))(conv2)
#
conv3 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(pool2)
conv3 = Dropout(0.2)(conv3)
conv3 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv3)
up1 = UpSampling2D(size=(2, 2))(conv3)
up1 = concatenate([conv2, up1], axis=1)
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up1)
conv4 = Dropout(0.2)(conv4)
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv4)
#
up2 = UpSampling2D(size=(2, 2))(conv4)
up2 = concatenate([conv1, up2], axis=1)
conv5 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up2)
conv5 = Dropout(0.2)(conv5)
conv5 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv5)
#
#1×1的卷积的作用
#大概有两个方面的作用：1. 实现跨通道的交互和信息整合2. 进行卷积核通道数的降维和升维。
conv6 = Conv2D(2, (1, 1),
activation='relu',
padding='same',
data_format='channels_first')(conv5)
conv6 = core.Reshape((2, patch_height * patch_width))(
conv6) #此时output的shape是(batchsize,2,patch_height*patch_width)
conv6 = core.Permute((2, 1))(
conv6
) #此时output的shape是(Npatch,patch_height*patch_width,2)即输出维度是(Npatch,2304,2)
############
conv7 = core.Activation('softmax')(conv6)
model = Model(inputs=inputs, outputs=conv7)
# sgd = SGD(lr=0.01, decay=1e-6, momentum=0.3, nesterov=False)
model.compile(optimizer=Adam(lr=0.001),
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
'''
def get_gnet(n_ch,patch_height,patch_width):
inputs = Input((n_ch, patch_height, patch_width))
conv1 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(inputs)
conv1 = Dropout(0.2)(conv1)
conv1 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(conv1)
up1 = UpSampling2D(size=(2, 2))(conv1)
#
conv2 = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(up1)
conv2 = Dropout(0.2)(conv2)
conv2 = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(conv2)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv2)
#
conv3 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(pool1)
conv3 = Dropout(0.2)(conv3)
conv3 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(conv3)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv3)
#
conv4 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(pool2)
conv4 = Dropout(0.2)(conv4)
conv4 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(conv4)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv4)
#
conv5 = Convolution2D(128, 3, 3, activation='relu', border_mode='same')(pool3)
conv5 = Dropout(0.2)(conv5)
conv5 = Convolution2D(128, 3, 3, activation='relu', border_mode='same')(conv5)
#
up2 = concatenate([UpSampling2D(size=(2, 2))(conv5), conv4], axis=3)
conv6 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(up2)
conv6 = Dropout(0.2)(conv6)
conv6 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(conv6)
#
up3 = concatenate([UpSampling2D(size=(2, 2))(conv6), conv3], axis=3)
conv7 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(up3)
conv7 = Dropout(0.2)(conv7)
conv7 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(conv7)
#
up4 = concatenate([UpSampling2D(size=(2, 2))(conv7), conv2], axis=3)
conv8 = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(up4)
conv8 = Dropout(0.2)(conv8)
conv8 = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(conv8)
#
pool4 = MaxPooling2D(pool_size=(2, 2))(conv8)
conv9 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(pool4)
conv9 = Dropout(0.2)(conv9)
conv9 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(conv9)
#
conv10 = Convolution2D(2, 1, 1, activation='relu', border_mode='same')(conv9)
conv10 = core.Reshape((2,patch_height*patch_width))(conv10)
conv10 = core.Permute((2,1))(conv10)
############
conv10 = core.Activation('softmax')(conv10)
model = Model(input=inputs, output=conv10)
# sgd = SGD(lr=0.01, decay=1e-6, momentum=0.3, nesterov=False)
model.compile(optimizer='sgd', loss='categorical_crossentropy',metrics=['accuracy'])
x, y = imageSegmentationGenerator(cfg.train_images, cfg.train_annotations, cfg.train_batch_size,
cfg.n_classes, cfg.input_height, cfg.input_width, cfg.output_height,
cfg.output_width)
model.fit(
x, y,
steps_per_epoch=int(cfg.train_data_number / cfg.train_batch_size),
max_queue_size=8, workers=4, validation_data=5, epochs=cfg.epochs
)
# return model
trainX = train[:, 1:].reshape(train.shape[0], 1, img_rows, img_cols)
trainX = trainX.astype(float)
trainX /= 255.0
trainY = kutils.to_categorical(train[:, 0])
nb_classes = trainY.shape[1]
cnn = models.Sequential()
cnn.add(conv.ZeroPadding2D(
(1, 1),
input_shape=(1, img_rows, img_cols),
))
cnn.add(conv.Convolution2D(filters[0], kernel, kernel))
cnn.add(core.Activation('relu'))
cnn.add(conv.MaxPooling2D(strides=(pool, pool)))
cnn.add(conv.ZeroPadding2D((1, 1)))
cnn.add(conv.Convolution2D(filters[1], kernel, kernel))
cnn.add(core.Activation('relu'))
cnn.add(conv.MaxPooling2D(strides=(pool, pool)))
cnn.add(conv.ZeroPadding2D((1, 1)))
cnn.add(core.Flatten())
cnn.add(core.Dropout(0.5))
cnn.add(core.Dense(128))
cnn.add(core.Activation('relu'))
cnn.add(core.Dense(nb_classes))
train_num = int(len(data) * p)
train_data = data.iloc[:train_num, :]
test_data = data.iloc[:(len(data) - train_num), :]
#
X_train = train_data.iloc[:, :-1]
y_train = train_data.iloc[:, -1]
X_test = test_data.iloc[:, :-1]
y_test = test_data.iloc[:, -1]
# print(test_data)
# 建立模型
model = models.Sequential()
# 添加输入层（3个点）到 隐藏层（10个点）的连接
model.add(core.Dense(input_dim=3, units=10))
# 隐藏层使用relu激活函数
model.add(core.Activation('relu'))
# 添加隐藏层（10个点）到输出层（1个点）的连接
model.add(core.Dense(input_dim=10, units=1))
# 输出层使用sigmoid激活函数
model.add(core.Activation('sigmoid'))
model.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['accuracy'])
# 训练模型
model.fit(X_train, y_train, epochs=100, batch_size=1)
model.save_weights('./net.model')
predict_result = model.predict_classes(X_train).reshape(len(X_train))
plt = cm_plot(y_train, predict_result)
plt.savefig('{0}.png'.format('./img/Figure_02_1'))
def get_unet(n_ch, patch_height, patch_width):
inputs = Input(shape=(n_ch, patch_height, patch_width))
# encoding path
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(inputs)
conv1 = Dropout(0.2)(conv1)
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv1) # 32*48*48
#
pool1 = MaxPooling2D((2, 2), data_format='channels_first')(conv1)
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(pool1)
conv2 = Dropout(0.2)(conv2)
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv2) # 64*24*24
#
pool2 = MaxPooling2D((2, 2), data_format='channels_first')(conv2)
conv3 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(pool2)
conv3 = Dropout(0.2)(conv3)
conv3 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name='conv3')(conv3) # 128*12*12
vgg = Model(inputs, conv3)
vgg.load_weights(VGG_PATH, by_name=True)
# decoding + concat path
up1 = UpSampling2D(size=(2, 2),
data_format='channels_first')(conv3) # 128*24*24
up1 = Conv2D(64, (3, 3), padding='same',
data_format='channels_first')(up1) # 64*24*24
up1 = Dropout(0.2)(up1)
up1 = Activation('relu')(up1)
x2 = Attention_block(up1, conv2, 32)
up1 = Concatenate(axis=1)([x2, up1])
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up1)
conv4 = Dropout(0.2)(conv4)
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv4)
#
up2 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv4)
up2 = Conv2D(32, (3, 3), padding='same', data_format='channels_first')(up2)
up2 = Dropout(0.2)(up2)
up2 = Activation('relu')(up2)
x1 = Attention_block(up2, conv1, 16)
up1 = Concatenate(axis=1)([x1, up2])
conv5 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up2)
conv5 = Dropout(0.2)(conv5)
conv5 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv5) # 32*48*48
#
conv6 = Conv2D(2, (1, 1),
activation='relu',
padding='same',
data_format='channels_first')(conv5) # 2*48*48
conv6 = core.Reshape((2, patch_height * patch_width))(conv6)
conv6 = core.Permute((2, 1))(conv6)
############
conv7 = core.Activation('softmax')(conv6)
model = Model(inputs=inputs, outputs=conv7)
#sgd = SGD(lr=0.001, decay=1e-6, momentum=0.3, nesterov=False)
#optimizer = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
#model.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics=['accuracy'])
model.compile(optimizer=Adam(lr=1e-3),
loss=losses.binary_crossentropy,
metrics=['accuracy'])
return model
data_train = data.loc[1994:2013].copy()
# 数据标准化
data_mean = data_train.mean()
data_std = data_train.std()
data_train = (data_train - data_mean) / data_std
# 取出数据作为训练数据
X_train = data_train[feature]
# 取出y值
y_train = data_train['y']
# 建立模型
model = models.Sequential()
# 添加输入层（6个点）到 隐藏层（12个点）的连接
model.add(core.Dense(input_dim=6, units=12))
# 隐藏层使用relu激活函数
model.add(core.Activation('relu'))
# 添加隐藏层（12个点）到输出层（1个点）的连接
model.add(core.Dense(input_dim=12, units=1))
model.compile(optimizer='adam',
loss='mean_squared_error',
metrics=['accuracy'])
# 训练模型
model.fit(X_train, y_train, epochs=10000, batch_size=32, verbose=1)
# 保存模型参数
model.save_weights('./temp/4-net.model')
# 预测 并还原结果
x = (data[feature] - data_mean[feature]) / data_std[feature]
data['y_pred'] = model.predict(x) * data_std['y'] + data_mean['y']
p = data[['y', 'y_pred']].plot(subplots=True, style=['b-o', 'r-*'])
plt.savefig('{0}.png'.format('./img/Figure_04'))
padding='same',
data_format='channels_first')(up2)
conv5 = Dropout(0.2)(conv5)
conv5 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv5)
conv6 = Conv2D(2, (1, 1),
activation='relu',
padding='same',
data_format='channels_first')(conv5)
conv6 = core.Reshape((2, patch_size**2))(conv6)
conv6 = core.Permute((2, 1))(conv6)
conv7 = core.Activation('softmax')(conv6)
model = Model(input=inputs, output=conv7)
model.compile(optimizer='sgd', loss=loss_name, metrics=['accuracy'])
# Custom callback --------------------------------------------------------------
output_dir = 'output_' + str(patch_size) + '-' + str(stride) + '-' + str(n_epochs) + '-' \
+ str(learning_rate) + '-' + loss_name
output_dir_1000 = 'output_' + str(patch_size) + '-' + str(stride) + '-' + str(n_epochs) + '-' \
+ str(learning_rate) + '-' + loss_name + '-times1000'
if not os.path.isdir(output_dir): os.mkdir(output_dir)
if not os.path.isdir(output_dir_1000): os.mkdir(output_dir_1000)
class SaveReconstructions(keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs={}):
def DUnet(n_ch, patch_height, patch_width):
inputs = Input(shape=(n_ch, patch_height, patch_width))
conv1 = CONV(inputs, 8)
dens1_48 = Conv2D(4, (1, 1),
strides=(1, 1),
padding='same',
data_format='channels_first')(conv1)
dens1_24 = Conv2D(4, (1, 1),
strides=(2, 2),
padding='same',
data_format='channels_first')(conv1)
dens1_12 = Conv2D(4, (1, 1),
strides=(4, 4),
padding='same',
data_format='channels_first')(conv1)
dens1_6 = Conv2D(4, (1, 1),
strides=(8, 8),
padding='same',
data_format='channels_first')(conv1)
pool1 = MaxPooling2D((2, 2))(conv1)
conv2 = CONV(pool1, 16)
dens2_48 = Conv2DTranspose(4, (1, 1),
strides=(2, 2),
padding='same',
data_format='channels_first')(conv2)
dens2_24 = Conv2D(4, (1, 1),
strides=(1, 1),
padding='same',
data_format='channels_first')(conv2)
dens2_12 = Conv2D(4, (1, 1),
strides=(2, 2),
padding='same',
data_format='channels_first')(conv2)
dens2_6 = Conv2D(4, (1, 1),
strides=(4, 4),
padding='same',
data_format='channels_first')(conv2)
pool2 = MaxPooling2D((2, 2))(conv2)
conc1 = concatenate([pool2, dens1_12], axis=1)
conv3 = CONV(conc1, 32)
dens3_48 = Conv2DTranspose(4, (1, 1),
strides=(4, 4),
padding='same',
data_format='channels_first')(conv3)
dens3_24 = Conv2DTranspose(4, (1, 1),
strides=(2, 2),
padding='same',
data_format='channels_first')(conv3)
dens3_12 = Conv2D(4, (1, 1),
strides=(1, 1),
padding='same',
data_format='channels_first')(conv3)
pool3 = MaxPooling2D((2, 2))(conv3)
conc2 = concatenate([pool3, dens1_6, dens2_6], axis=1)
conv4 = CONV(conc2, 64)
dens4_48 = Conv2DTranspose(4, (1, 1),
strides=(8, 8),
padding='same',
data_format='channels_first')(conv4)
dens4_24 = Conv2DTranspose(4, (1, 1),
strides=(4, 4),
padding='same',
data_format='channels_first')(conv4)
up1 = UpSampling2D(size=(2, 2))(conv4)
conc3 = concatenate([up1, dens1_12, dens2_12, dens3_12], axis=1)
conv5 = CONV(conc3, 32)
dens5_48 = Conv2DTranspose(4, (1, 1),
strides=(4, 4),
padding='same',
data_format='channels_first')(conv5)
up2 = UpSampling2D(size=(2, 2))(conv5)
conc4 = concatenate([up2, dens1_24, dens2_24, dens3_24, dens4_24], axis=1)
conv6 = CONV(conc4, 16)
up3 = UpSampling2D(size=(2, 2))(conv6)
conc5 = concatenate(
[up3, dens1_48, dens2_48, dens3_48, dens4_48, dens5_48], axis=1)
conv7 = CONV(conc5, 8)
conv8 = Conv2D(2, (1, 1),
activation='relu',
padding='same',
data_format='channels_first')(conv7)
conv8 = core.Reshape((2, patch_height * patch_width))(conv8)
conv8 = core.Permute((2, 1))(conv8)
conv8 = core.Activation('softmax')(conv8)
model = Model(inputs=inputs, outputs=conv8)
sgd = SGD(lr=0.01, decay=1e-6, momentum=0.3, nesterov=False)
model.compile(optimizer=sgd,
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def Unet(nClasses, input_length,optimizer=None, nChannels=1):
inputs = Input((input_length, nChannels))
# (None,1800,1)--->(None,1800,16)
conv1 = Conv1D(16, 32, activation='relu', padding='same', kernel_initializer='he_normal')(inputs)
# (None,1800,16)--->(None,1800,16)
conv1 = Conv1D(16, 32, activation='relu', padding='same', kernel_initializer='he_normal')(conv1)
# (None,1800,16)--->(None,900,16)
pool1 = MaxPooling1D(pool_size=2)(conv1)
# (None,900,16)--->(None,900,32)
conv2 = Conv1D(32, 32, activation='relu', padding='same', kernel_initializer='he_normal')(pool1)
# (None,900,32)--->(None,900,32)
conv2 = Dropout(0.2)(conv2)
# (None,900,32)--->(None,900,32)
conv2 = Conv1D(32, 32, activation='relu', padding='same', kernel_initializer='he_normal')(conv2)
# (None,900,32)--->(None,450,32)
pool2 = MaxPooling1D(pool_size=2)(conv2)
#(None, 450, 32)--->(None,450,64)
conv3 = Conv1D(64, 32, activation='relu', padding='same', kernel_initializer='he_normal')(pool2)
# (None, 450, 64)--->(None,450,64)
conv3 = Conv1D(64, 32, activation='relu', padding='same', kernel_initializer='he_normal')(conv3)
# (None, 450, 64)--->(None,225,64)
pool3 = MaxPooling1D(pool_size=2)(conv3)
#(None,225,64)--->(None,225,128)
conv4 = Conv1D(128, 32, activation='relu', padding='same', kernel_initializer='he_normal')(pool3)
# (None,225,128)--->(None,225,128)
conv4 = Dropout(0.5)(conv4)
# (None,225,128)--->(None,225,128)
conv4 = Conv1D(128, 32, activation='relu', padding='same', kernel_initializer='he_normal')(conv4)
#(None,225,128)--->(None,450,64)
up1 = Conv1D(64, 2, activation='relu', padding='same', kernel_initializer='he_normal')(UpSampling1D(size=2)(conv4))
#(None, 450, 64)--->(None,450,128)
merge1 = concatenate([up1, conv3], axis=-1)
# (None, 450, 64)--->(None,450,64)
conv5 = Conv1D(64, 32, activation='relu', padding='same', kernel_initializer='he_normal')(merge1)
# (None, 450, 64)--->(None,450,64)
conv5 = Conv1D(64, 32, activation='relu', padding='same', kernel_initializer='he_normal')(conv5)
# (None, 450, 64)--->(None,900,32)
up2 = Conv1D(32, 2, activation='relu', padding='same', kernel_initializer = 'he_normal')(UpSampling1D(size=2)(conv5))
# (None, 900, 32)--->(None,900,64)
merge2 = concatenate([up2, conv2], axis=-1)
# (None, 900, 64)--->(None,900,32)
conv6 = Conv1D(32, 32, activation='relu', padding='same', kernel_initializer = 'he_normal')(merge2)
conv6 = Dropout(0.2)(conv6)
# (None, 900, 32)--->(None,900,32)
conv6 = Conv1D(32, 32, activation='relu', padding='same')(conv6)
# (None, 900, 32)--->(None,1800,16)
up3 = Conv1D(16, 2, activation='relu', padding='same', kernel_initializer='he_normal')(UpSampling1D(size=2)(conv6))
#(None,1800,16)--->(None,1800,32)
merge3 = concatenate([up3, conv1], axis=-1)
#(None, 1800, 32)--->(None, 1800, 16)
conv7 = Conv1D(16, 32, activation='relu', padding='same', kernel_initializer='he_normal')(merge3)
#(None, 1800, 16)--->(None, 1800, 16)
conv7 = Conv1D(16, 32, activation='relu', padding='same', kernel_initializer='he_normal')(conv7)
# (None, 1800, 16)--->(None, 1800, classes)
conv8 = Conv1D(nClasses, 1, activation='relu', padding='same', kernel_initializer='he_normal')(conv7)
#(None, 1800, classes)--->(None,3,1800)
conv8 = core.Reshape((nClasses, input_length))(conv8)
#(None, 3, 1800)--->(None, 1800, classes)
conv8 = core.Permute((2, 1))(conv8)
conv9 = core.Activation('softmax')(conv8)
model = Model(inputs=inputs, outputs=conv9)
if not optimizer is None:
model.compile(loss="categorical_crossentropy", optimizer=optimizer, metrics=['accuracy'])
return model
deconv_nir_3 = Conv2DTranspose(num_class*C,(4,4), strides=(2, 2), padding='same', data_format="channels_last", activation='relu',kernel_initializer='glorot_normal')(conv_nir_2)
conv_nir_3 = Conv2D(num_class*C, (3,3), strides=(1,1), padding = 'same', activation='relu', data_format='channels_last')(deconv_nir_3)
deconv_nir_4 = Conv2DTranspose(num_class*C,(4,4), strides=(2, 2), padding='same', data_format="channels_last", activation='relu',kernel_initializer='glorot_normal')(conv_nir_3)
conv_nir_4 = Conv2D(num_class*C, (3,3), strides=(1,1), padding = 'same', activation='relu', data_format='channels_last')(deconv_nir_4)
deconv_nir_5 = Conv2DTranspose(num_class*C,(4,4), strides=(2, 2), padding='same', data_format="channels_last", activation='relu',kernel_initializer='glorot_normal')(conv_nir_4)
# CONACTENATE the ends of RGB & NIR
merge_rgb_nir = keras.layers.concatenate([deconv_rgb_5, deconv_nir_5], axis=-1)
# DECONVOLUTION Layers
deconv_last = Conv2DTranspose(num_class, (1,1), strides=(1, 1), padding='same', data_format="channels_last", activation='relu',kernel_initializer='glorot_normal') (merge_rgb_nir)
#VECTORIZING OUTPUT
out_reshape = core.Reshape((input_dim[0]*input_dim[1],num_class))(deconv_last)
out = core.Activation('softmax')(out_reshape)
# MODAL [INPUTS , OUTPUTS]
model = Model(inputs=[inputs_rgb,inputs_nir], outputs=[out])
print 'compiling'
model.compile(optimizer='sgd',
loss='categorical_crossentropy',
metrics=['accuracy'])
model.summary()
# Save the model according to the conditions
progbar = ProgbarLogger(count_mode='steps')
checkpoint = ModelCheckpoint("nir_rgb_segmentation_2.{epoch:02d}.hdf5", monitor='val_acc', verbose=1, save_best_only=False, save_weights_only=False, mode='auto', period=1)
#early = EarlyStopping(monitor='val_acc', min_delta=0, patience=1, verbose=1, mode='auto')
#haven't specified validation data directory yet
def test_TimeDistributed():
# first, test with Dense layer
model = Sequential()
model.add(wrappers.TimeDistributed(core.Dense(2), input_shape=(3, 4)))
model.add(core.Activation('relu'))
model.compile(optimizer='rmsprop', loss='mse')
model.fit(np.random.random((10, 3, 4)),
np.random.random((10, 3, 2)),
epochs=1,
batch_size=10)
# test config
model.get_config()
# test when specifying a batch_input_shape
test_input = np.random.random((1, 3, 4))
test_output = model.predict(test_input)
weights = model.layers[0].get_weights()
reference = Sequential()
reference.add(
wrappers.TimeDistributed(core.Dense(2), batch_input_shape=(1, 3, 4)))
reference.add(core.Activation('relu'))
reference.compile(optimizer='rmsprop', loss='mse')
reference.layers[0].set_weights(weights)
reference_output = reference.predict(test_input)
assert_allclose(test_output, reference_output, atol=1e-05)
# test with Embedding
model = Sequential()
model.add(
wrappers.TimeDistributed(embeddings.Embedding(5, 6),
batch_input_shape=(10, 3, 4),
dtype='int32'))
model.compile(optimizer='rmsprop', loss='mse')
model.fit(np.random.randint(5, size=(10, 3, 4), dtype='int32'),
np.random.random((10, 3, 4, 6)),
epochs=1,
batch_size=10)
# compare to not using batch_input_shape
test_input = np.random.randint(5, size=(10, 3, 4), dtype='int32')
test_output = model.predict(test_input)
weights = model.layers[0].get_weights()
reference = Sequential()
reference.add(
wrappers.TimeDistributed(embeddings.Embedding(5, 6),
input_shape=(3, 4),
dtype='int32'))
reference.compile(optimizer='rmsprop', loss='mse')
reference.layers[0].set_weights(weights)
reference_output = reference.predict(test_input)
assert_allclose(test_output, reference_output, atol=1e-05)
# test with Conv2D
model = Sequential()
model.add(
wrappers.TimeDistributed(convolutional.Conv2D(5, (2, 2),
padding='same'),
input_shape=(2, 4, 4, 3)))
model.add(core.Activation('relu'))
model.compile(optimizer='rmsprop', loss='mse')
model.train_on_batch(np.random.random((1, 2, 4, 4, 3)),
np.random.random((1, 2, 4, 4, 5)))
model = model_from_json(model.to_json())
model.summary()
# test stacked layers
model = Sequential()
model.add(wrappers.TimeDistributed(core.Dense(2), input_shape=(3, 4)))
model.add(wrappers.TimeDistributed(core.Dense(3)))
model.add(core.Activation('relu'))
model.compile(optimizer='rmsprop', loss='mse')
model.fit(np.random.random((10, 3, 4)),
np.random.random((10, 3, 3)),
epochs=1,
batch_size=10)
# test wrapping Sequential model
model = Sequential()
model.add(core.Dense(3, input_dim=2))
outer_model = Sequential()
outer_model.add(wrappers.TimeDistributed(model, input_shape=(3, 2)))
outer_model.compile(optimizer='rmsprop', loss='mse')
outer_model.fit(np.random.random((10, 3, 2)),
np.random.random((10, 3, 3)),
epochs=1,
batch_size=10)
# test with functional API
x = Input(shape=(3, 2))
y = wrappers.TimeDistributed(model)(x)
outer_model = Model(x, y)
outer_model.compile(optimizer='rmsprop', loss='mse')
outer_model.fit(np.random.random((10, 3, 2)),
np.random.random((10, 3, 3)),
epochs=1,
batch_size=10)
# test with BatchNormalization
model = Sequential()
model.add(
wrappers.TimeDistributed(normalization.BatchNormalization(center=True,
scale=True),
name='bn',
input_shape=(10, 2)))
model.compile(optimizer='rmsprop', loss='mse')
# Assert that mean and variance are 0 and 1.
td = model.layers[0]
assert np.array_equal(td.get_weights()[2], np.array([0, 0]))
assert np.array_equal(td.get_weights()[3], np.array([1, 1]))
# Train
model.train_on_batch(np.random.normal(loc=2, scale=2, size=(1, 10, 2)),
np.broadcast_to(np.array([0, 1]), (1, 10, 2)))
# Assert that mean and variance changed.
assert not np.array_equal(td.get_weights()[2], np.array([0, 0]))
assert not np.array_equal(td.get_weights()[3], np.array([1, 1]))
# Verify input_map has one mapping from inputs to reshaped inputs.
uid = _object_list_uid(model.inputs)
assert len(td._input_map.keys()) == 1
assert uid in td._input_map
assert K.int_shape(td._input_map[uid]) == (None, 2)
def get_patches_unet5(n_ch,patch_height,patch_width):
inputs = Input(shape=(n_ch,patch_height,patch_width))
conv1 = Conv2D(filters=32, kernel_size=(3, 3), activation='relu', padding='same',data_format='channels_first')(inputs)
conv1 = Dropout(0.2)(conv1)
conv1 = Conv2D(filters=32, kernel_size=(3, 3), activation='relu', padding='same',data_format='channels_first')(conv1)
pool1 = MaxPooling2D((2, 2))(conv1)
conv2 = Conv2D(filters=64, kernel_size=(3, 3), activation='relu', padding='same',data_format='channels_first')(pool1)
conv2 = Dropout(0.2)(conv2)
conv2 = Conv2D(filters=64, kernel_size=(3, 3), activation='relu', padding='same',data_format='channels_first')(conv2)
pool2 = MaxPooling2D((2, 2))(conv2)
conv3 = Conv2D(filters=128, kernel_size=(3, 3), activation='relu', padding='same',data_format='channels_first')(pool2)
conv3 = Dropout(0.2)(conv3)
conv3 = Conv2D(filters=128, kernel_size=(3, 3), activation='relu', padding='same',data_format='channels_first')(conv3)
pool3 = MaxPooling2D((2,2))(conv3)
conv4 = Conv2D(filters=256, kernel_size=(3, 3), activation='relu', padding='same',data_format='channels_first')(pool3)
conv4 = Dropout(0.2)(conv4)
conv4 = Conv2D(filters=256, kernel_size=(3, 3), activation='relu', padding='same',data_format='channels_first')(conv4)
pool4 = MaxPooling2D((2,2))(conv4)
conv5 = Conv2D(filters=512, kernel_size=(3, 3), activation='relu', padding='same',data_format='channels_first')(pool4)
conv5 = Dropout(0.2)(conv5)
conv5 = Conv2D(filters=512, kernel_size=(3, 3), activation='relu', padding='same',data_format='channels_first')(conv5)
up1 = UpSampling2D(size=(2, 2))(conv5)
up1 = concatenate([conv4,up1],axis=1)
conv6 = Conv2D(256, (3, 3), activation='relu', padding='same',data_format='channels_first')(up1)
conv6 = Dropout(0.2)(conv6)
conv6 = Conv2D(256, (3, 3), activation='relu', padding='same',data_format='channels_first')(conv6)
up2 = UpSampling2D(size=(2, 2))(conv6)
up2 = concatenate([conv3,up2], axis=1)
conv7 = Conv2D(128, (3, 3), activation='relu', padding='same',data_format='channels_first')(up2)
conv7 = Dropout(0.2)(conv7)
conv7 = Conv2D(128, (3, 3), activation='relu', padding='same',data_format='channels_first')(conv7)
up3 = UpSampling2D(size=(2,2))(conv7)
up3 = concatenate([conv2,up3], axis=1)
conv8 = Conv2D(64, (3, 3), activation='relu', padding='same',data_format='channels_first')(up3)
conv8 = Dropout(0.2)(conv8)
conv8 = Conv2D(64, (3, 3), activation='relu', padding='same',data_format='channels_first')(conv8)
up4 = UpSampling2D(size=(2,2))(conv8)
up4 = concatenate([conv1,up4], axis=1)
conv9 = Conv2D(32, (3, 3), activation='relu', padding='same',data_format='channels_first')(up4)
conv9 = Dropout(0.2)(conv9)
conv9 = Conv2D(32, (3, 3), activation='relu', padding='same',data_format='channels_first')(conv9)
conv10 = Conv2D(2, (1, 1), activation='relu',padding='same',data_format='channels_first')(conv9)
conv10 = core.Reshape((2,patch_height*patch_width))(conv10)
conv10 = core.Permute((2,1))(conv10)
conv10 = core.Activation('softmax')(conv10)
#conv10 = ThresholdedReLU(theta=0.25)(conv10)
model = Model(inputs=inputs, outputs=conv10)
# multi_model = multi_gpu_model(model, gpus=2)
# sgd = SGD(lr=0.01, decay=1e-6, momentum=0.3, nesterov=False)
model.compile(optimizer='sgd', loss='categorical_crossentropy',metrics=['accuracy'])
return model
def get_dilated_unet(n_ch, patch_height, patch_width, dilaterate=3):
inputs = Input(shape=(n_ch, patch_height, patch_width))
conv1 = Conv2D(32, (3, 3), activation='relu', padding='same', dilation_rate=dilaterate,
data_format='channels_first')(inputs)
conv1 = Dropout(0.2)(conv1)
conv1 = Conv2D(32, (3, 3), activation='relu', padding='same', dilation_rate=dilaterate,
data_format='channels_first')(conv1)
pool1 = MaxPooling2D((2, 2))(conv1)
#
conv2 = Conv2D(64, (3, 3), activation='relu', padding='same', dilation_rate=dilaterate,
data_format='channels_first')(pool1)
conv2 = Dropout(0.2)(conv2)
conv2 = Conv2D(64, (3, 3), activation='relu', padding='same', dilation_rate=dilaterate,
data_format='channels_first')(conv2)
pool2 = MaxPooling2D((2, 2))(conv2)
#
conv3 = Conv2D(128, (3, 3), activation='relu', padding='same', dilation_rate=dilaterate,
data_format='channels_first')(pool2)
conv3 = Dropout(0.2)(conv3)
conv3 = Conv2D(128, (3, 3), activation='relu', padding='same', dilation_rate=dilaterate,
data_format='channels_first')(conv3)
up1 = UpSampling2D(size=(2, 2))(conv3)
up1 = concatenate([conv2, up1], axis=1)
conv4 = Conv2D(64, (3, 3), activation='relu', padding='same', data_format='channels_first')(up1)
conv4 = Dropout(0.2)(conv4)
conv4 = Conv2D(64, (3, 3), activation='relu', padding='same', data_format='channels_first')(conv4)
#
up2 = UpSampling2D(size=(2, 2))(conv4)
up2 = concatenate([conv1, up2], axis=1)
conv5 = Conv2D(32, (3, 3), activation='relu', padding='same', data_format='channels_first')(up2)
conv5 = Dropout(0.2)(conv5)
conv5 = Conv2D(32, (3, 3), activation='relu', padding='same', data_format='channels_first')(conv5)
#
conv6 = Conv2D(2, (1, 1), activation='relu', padding='same', data_format='channels_first')(conv5)
conv6 = core.Reshape((2, patch_height * patch_width))(conv6)
conv6 = core.Permute((2, 1))(conv6)
############
conv7 = core.Activation('softmax')(conv6)
model = Model(inputs=inputs, outputs=conv7)
# scheduler = LearningRateScheduler(mlr.lr_scheduler)
sgd = SGD(lr=0.01, decay=2e-5, momentum=0.8, nesterov=False)
model.compile(optimizer=sgd, loss='binary_crossentropy', metrics=['accuracy'])
# adam=optimizers.Adam(lr=0.01, beta_1=0.9, beta_2=0.999, epsilon=1e-07)
# model.compile(optimizer=adam, loss='categorical_crossentropy', metrics=['accuracy'])
# 1、目标函数
# (1）mean_squared_error / mse 均方误差，常用的目标函数，公式为((y_pred-y_true) ** 2).mean()
# (2）mean_absolute_error / mae绝对值均差，公式为( | y_pred - y_true |).mean()
# (3)mean_absolute_percentage_error / mape公式为：(| (y_true - y_pred) / clip((| y_true |), epsilon, infinite) |).mean(axis=-1) * 100，和mae的区别就是，累加的是（预测值与实际值的差）除以（剔除不介于epsilon和infinite之间的实际值)，然后求均值。
# (4)mean_squared_logarithmic_error / msle公式为： (log(clip(y_pred, epsilon, infinite) + 1) - log(clip(y_true, epsilon, infinite) + 1.)) ^ 2.mean(axis=-1)，这个就是加入了log对数，剔除不介于epsilon和infinite之间的预测值与实际值之后，然后取对数，作差，平方，累加求均值。
# (5)squared_hinge公式为：(max(1 - y_truey_pred, 0)) ^ 2.mean(axis=-1)，取1减去预测值与实际值乘积的结果与0比相对大的值的平方的累加均值。
# (6)hinge公式为：(max(1 - y_truey_pred, 0)).mean(axis=-1)，取1减去预测值与实际值乘积的结果与0比相对大的值的的累加均值。
# (7)binary_crossentropy: 常说的逻辑回归, 就是常用的交叉熵函
# (8)categorical_crossentropy: 多分类的逻辑
#
# 2、性能评估函数：
# (1)binary_accuracy: 对二分类问题, 计算在所有预测值上的平均正确率
# (2)categorical_accuracy: 对多分类问题, 计算再所有预测值上的平均正确率
# (3)sparse_categorical_accuracy: 与categorical_accuracy相同, 在对稀疏的目标值预测时有用
# (4)top_k_categorical_accracy: 计算top - k正确率, 当预测值的前k个值中存在目标类别即认为预测正确
# (5)sparse_top_k_categorical_accuracy：与top_k_categorical_accracy作用相同，但适用于稀疏情况
return model
def get_patches_unet4(n_ch,patch_height,patch_width):
inputs = Input(shape=(n_ch,patch_height,patch_width))
conv1 = Conv2D(filters=32, kernel_size=(3, 3), padding='same', data_format='channels_first')(inputs)
conv1 = BatchNormalization(axis=1)(conv1)
conv1 = Activation('relu')(conv1)
conv1 = Dropout(0.2)(conv1)
conv1 = Conv2D(filters=32, kernel_size=(3, 3), padding='same', data_format='channels_first')(conv1)
conv1 = BatchNormalization(axis=1)(conv1)
conv1 = Activation('relu')(conv1)
pool1 = MaxPooling2D((2, 2))(conv1)
conv2 = Conv2D(filters=64, kernel_size=(3, 3), padding='same', data_format='channels_first')(pool1)
conv2 = BatchNormalization(axis=1)(conv2)
conv2 = Activation('relu')(conv2)
conv2 = Dropout(0.2)(conv2)
conv2 = Conv2D(filters=64, kernel_size=(3, 3), padding='same', data_format='channels_first')(conv2)
conv2 = BatchNormalization(axis=1)(conv2)
conv2 = Activation('relu')(conv2)
pool2 = MaxPooling2D((2, 2))(conv2)
conv3 = Conv2D(filters=128, kernel_size=(3, 3), padding='same', data_format='channels_first')(pool2)
conv3 = BatchNormalization(axis=1)(conv3)
conv3 = Activation('relu')(conv3)
conv3 = Dropout(0.2)(conv3)
conv3 = Conv2D(filters=128, kernel_size=(3, 3), padding='same', data_format='channels_first')(conv3)
conv3 = BatchNormalization(axis=1)(conv3)
conv3 = Activation('relu')(conv3)
pool3 = MaxPooling2D((2,2))(conv3)
conv4 = Conv2D(filters=256, kernel_size=(3, 3), padding='same', data_format='channels_first')(pool3)
conv4 = BatchNormalization(axis=1)(conv4)
conv4 = Activation('relu')(conv4)
conv4 = Dropout(0.2)(conv4)
conv4 = Conv2D(filters=256, kernel_size=(3, 3), padding='same', data_format='channels_first')(conv4)
conv4 = BatchNormalization(axis=1)(conv4)
conv4 = Activation('relu')(conv4)
up1 = UpSampling2D(size=(2, 2))(conv4)
up1 = concatenate([conv3,up1], axis=1)
conv5 = Conv2D(128, (3, 3), padding='same', data_format='channels_first')(up1)
conv5 = BatchNormalization(axis=1)(conv5)
conv5 = Activation('relu')(conv5)
conv5 = Dropout(0.2)(conv5)
conv5 = Conv2D(128, (3, 3), padding='same', data_format='channels_first')(conv5)
conv5 = BatchNormalization(axis=1)(conv5)
conv5 = Activation('relu')(conv5)
up2 = UpSampling2D(size=(2,2))(conv5)
up2 = concatenate([conv2,up2], axis=1)
conv6 = Conv2D(64, (3, 3), padding='same', data_format='channels_first')(up2)
conv6 = BatchNormalization(axis=1)(conv6)
conv6 = Activation('relu')(conv6)
conv6 = Dropout(0.2)(conv6)
conv6 = Conv2D(64, (3, 3), padding='same', data_format='channels_first')(conv6)
conv6 = BatchNormalization(axis=1)(conv6)
conv6 = Activation('relu')(conv6)
up3 = UpSampling2D(size=(2,2))(conv6)
up3 = concatenate([conv1,up3], axis=1)
conv7 = Conv2D(32, (3, 3), padding='same', data_format='channels_first')(up3)
conv7 = BatchNormalization(axis=1)(conv7)
conv7 = Activation('relu')(conv7)
conv7 = Dropout(0.2)(conv7)
conv7 = Conv2D(32, (3, 3), padding='same', data_format='channels_first')(conv7)
conv7 = BatchNormalization(axis=1)(conv7)
conv7 = Activation('relu')(conv7)
# # Output shape as entire image
# conv8 = Conv2D(2, (1, 1), padding='same', data_format='channels_first')(conv7)
# conv8 = BatchNormalization(axis=1)(conv8)
# conv8 = Activation('relu')(conv8)
# conv8 = Conv2D(1, (1, 1), padding='same', data_format='channels_first')(conv8)
# conv8 = BatchNormalization(axis=1)(conv8)
# conv8 = Dense(1, data_format='channels_first')(conv8)
# conv8 = core.Activation('softmax')(conv8)
# Output shape as batches, height*width, 2
conv8 = Conv2D(2, (1, 1), activation='relu',padding='same',data_format='channels_first')(conv7)
conv8 = core.Reshape((2,patch_height*patch_width))(conv8)
conv8 = core.Permute((2,1))(conv8)
conv8 = core.Activation('softmax')(conv8)
#
model = Model(inputs=inputs, outputs=conv8)
# multi_model = multi_gpu_model(model, gpus=2)
# sgd = SGD(lr=0.01, decay=1e-6, momentum=0.3, nesterov=False)
model.compile(optimizer='sgd', loss='binary_crossentropy', metrics=['accuracy'])
# model.compile(optimizer='sgd', loss=dice_coef_loss , metrics=['accuracy', dice_coef, jaccard_coef])
return model
data_format='channels_first')(conv6)
conv7 = Conv2D(3, (1, 1),
activation='relu',
padding='same',
data_format='channels_first')(conv6)
conv7 = core.Reshape((3, 128 * 128))(conv7)
#conv6 = core.Permute((3,1))(conv6)
conv7 = core.Flatten()(conv7)
#conv7 = core.Dense(64)(conv7)
#conv7 = core.Activation('relu')(conv7)
#conv7 = Dropout(0.2)(conv7)
conv7 = core.Dense(2)(conv7)
############
conv8 = core.Activation('softmax')(conv7)
model = Model(input=inputs, output=conv8)
# sgd = SGD(lr=0.01, decay=1e-6, momentum=0.3, nesterov=False)
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
########################################
# Viewing model_configuration
model.summary()
model.get_config()
model.layers[0].get_config()
model.layers[0].input_shape
def get_patches_unet3(n_ch,patch_height,patch_width):
inputs = Input(shape=(n_ch,patch_height,patch_width))
conv1 = Conv2D(filters=32, kernel_size=(3, 3), padding='same', data_format='channels_first')(inputs)
conv1 = BatchNormalization(axis=1)(conv1)
conv1 = Activation('relu')(conv1)
conv1 = Dropout(0.2)(conv1)
conv1 = Conv2D(filters=32, kernel_size=(3, 3), padding='same', data_format='channels_first')(conv1)
conv1 = BatchNormalization(axis=1)(conv1)
conv1 = Activation('relu')(conv1)
pool1 = MaxPooling2D((2, 2))(conv1)
#
conv2 = Conv2D(filters=64, kernel_size=(3, 3), padding='same',data_format='channels_first')(pool1)
conv2 = BatchNormalization(axis=1)(conv2)
conv2 = Activation('relu')(conv2)
conv2 = Dropout(0.2)(conv2)
conv2 = Conv2D(filters=64, kernel_size=(3, 3), padding='same',data_format='channels_first')(conv2)
conv2 = BatchNormalization(axis=1)(conv2)
conv2 = Activation('relu')(conv2)
pool2 = MaxPooling2D((2, 2))(conv2)
#
conv3 = Conv2D(filters=128, kernel_size=(3, 3), padding='same',data_format='channels_first')(pool2)
conv3 = BatchNormalization(axis=1)(conv3)
conv3 = Activation('relu')(conv3)
conv3 = Dropout(0.2)(conv3)
conv3 = Conv2D(filters=128, kernel_size=(3, 3), padding='same',data_format='channels_first')(conv3)
conv3 = BatchNormalization(axis=1)(conv3)
conv3 = Activation('relu')(conv3)
up1 = UpSampling2D(size=(2, 2))(conv3)
up1 = concatenate([conv2,up1],axis=1)
conv4 = Conv2D(filters=64, kernel_size=(3, 3), padding='same',data_format='channels_first')(up1)
conv4 = BatchNormalization(axis=1)(conv4)
conv4 = Activation('relu')(conv4)
conv4 = Dropout(0.2)(conv4)
conv4 = Conv2D(filters=64, kernel_size=(3, 3), padding='same',data_format='channels_first')(conv4)
conv4 = BatchNormalization(axis=1)(conv4)
conv4 = Activation('relu')(conv4)
#
up2 = UpSampling2D(size=(2, 2))(conv4)
up2 = concatenate([conv1,up2], axis=1)
conv5 = Conv2D(filters=32, kernel_size=(3, 3),padding='same',data_format='channels_first')(up2)
conv5 = BatchNormalization(axis=1)(conv5)
conv5 = Activation('relu')(conv5)
conv5 = Dropout(0.2)(conv5)
conv5 = Conv2D(filters=32, kernel_size=(3, 3), padding='same',data_format='channels_first')(conv5)
conv5 = BatchNormalization(axis=1)(conv5)
conv5 = Activation('relu')(conv5)
#
conv6 = Conv2D(filters=2, kernel_size=(1, 1), activation='relu',padding='same',data_format='channels_first')(conv5)
conv6 = core.Reshape((2,patch_height*patch_width))(conv6)
conv6 = core.Permute((2,1))(conv6)
############
conv7 = core.Activation('softmax')(conv6)
model = Model(inputs=inputs, outputs=conv7)
# multi_model = multi_gpu_model(model, gpus=2)
# model = make_parallel(model, 2)
# sgd = SGD(lr=0.01, decay=1e-6, momentum=0.3, nesterov=False)
#model.compile(optimizer='sgd', loss='categorical_crossentropy',metrics=['accuracy'])
model.compile(optimizer='sgd', loss='binary_crossentropy',metrics=['accuracy'])
#multi_model.compile(optimizer='sgd', loss='categorical_crossentropy',metrics=['accuracy'])
return model
def build_model(start_neurons,
block="resnet",
DropoutRatio=0.5,
filter_size=32,
nClasses=2):
# 101 -> 50
input_length = 2000
input_layer = Input((input_length, 1))
# conv1 = headneck(input_layer)
# conv1 = Conv1D(start_neurons * 1, filter_size, activation=None, padding="same")(conv1)
conv1 = Conv1D(start_neurons * 1,
filter_size,
activation=None,
padding="same")(input_layer)
conv1 = bottleneck(conv1, start_neurons * 1, block)
conv1 = Activation(ACTIVATION)(conv1)
pool1 = MaxPooling1D((2))(conv1)
pool1 = Dropout(DropoutRatio / 2)(pool1)
# 50 -> 25
conv2 = Conv1D(start_neurons * 2,
filter_size,
activation=None,
padding="same")(pool1)
conv2 = bottleneck(conv2, start_neurons * 2, block)
conv2 = Activation(ACTIVATION)(conv2)
pool2 = MaxPooling1D((2))(conv2)
pool2 = Dropout(DropoutRatio)(pool2)
# 25 -> 12
conv3 = Conv1D(start_neurons * 4,
filter_size,
activation=None,
padding="same")(pool2)
conv3 = bottleneck(conv3, start_neurons * 4, block)
conv3 = Activation(ACTIVATION)(conv3)
pool3 = MaxPooling1D((2))(conv3)
pool3 = Dropout(DropoutRatio)(pool3)
# 12 -> 6
conv4 = Conv1D(start_neurons * 8,
filter_size,
activation=None,
padding="same")(pool3)
# conv4 = residual_block(conv4,start_neurons * 8)
# conv4 = residual_block(conv4,start_neurons * 8)
conv4 = bottleneck(conv4, start_neurons * 8, block)
conv4 = Activation(ACTIVATION)(conv4)
pool4 = MaxPooling1D((2))(conv4)
pool4 = Dropout(DropoutRatio)(pool4)
# Middle
convm = Conv1D(start_neurons * 16,
filter_size,
activation=None,
padding="same")(pool4)
# convm = residual_block(convm,start_neurons * 16)
# convm = residual_block(convm,start_neurons * 16)
convm = bottleneck(convm, start_neurons * 16, block)
convm = Activation(ACTIVATION)(convm)
# 6 -> 12
#deconv4 = Conv2DTranspose(start_neurons * 8, (3, 3), strides=(2, 2), padding="same")(convm)
deconv4 = Conv1D(start_neurons * 8,
filter_size,
activation='relu',
padding='same')(UpSampling1D(
size=2)(convm)) #kernel_initializer='he_normal'
uconv4 = concatenate([deconv4, conv4])
uconv4 = Dropout(DropoutRatio)(uconv4)
uconv4 = Conv1D(start_neurons * 8,
filter_size,
activation=None,
padding="same")(uconv4)
# uconv4 = residual_block(uconv4,start_neurons * 8)
# uconv4 = residual_block(uconv4,start_neurons * 8)
uconv4 = bottleneck(uconv4, start_neurons * 8, block)
uconv4 = Activation(ACTIVATION)(uconv4)
# 12 -> 25
#deconv3 = Conv2DTranspose(start_neurons * 4, (3, 3), strides=(2, 2), padding="same")(uconv4)
#deconv3 = Conv2DTranspose(start_neurons * 4, (3, 3), strides=(2, 2), padding="valid")(uconv4)
deconv3 = Conv1D(
start_neurons * 4,
filter_size,
activation='relu',
padding='same',
)(UpSampling1D(size=2)(uconv4)) #kernel_initializer='he_normal'
uconv3 = concatenate([deconv3, conv3])
uconv3 = Dropout(DropoutRatio)(uconv3)
uconv3 = Conv1D(start_neurons * 4,
filter_size,
activation=None,
padding="same")(uconv3)
# uconv3 = residual_block(uconv3,start_neurons * 4)
# uconv3 = residual_block(uconv3,start_neurons * 4)
uconv3 = bottleneck(uconv3, start_neurons * 4, block)
uconv3 = Activation(ACTIVATION)(uconv3)
# 25 -> 50
#deconv2 = Conv2DTranspose(start_neurons * 2, (3, 3), strides=(2, 2), padding="same")(uconv3)
deconv2 = Conv1D(
start_neurons * 2,
filter_size,
activation='relu',
padding='same',
)(UpSampling1D(size=2)(uconv3)) #kernel_initializer='he_normal'
uconv2 = concatenate([deconv2, conv2])
uconv2 = Dropout(DropoutRatio)(uconv2)
uconv2 = Conv1D(start_neurons * 2,
filter_size,
activation=None,
padding="same")(uconv2)
# uconv2 = residual_block(uconv2,start_neurons * 2)
# uconv2 = residual_block(uconv2,start_neurons * 2)
uconv2 = bottleneck(uconv2, start_neurons * 2, block)
uconv2 = Activation(ACTIVATION)(uconv2)
# 50 -> 101
#deconv1 = Conv2DTranspose(start_neurons * 1, (3, 3), strides=(2, 2), padding="same")(uconv2)
#deconv1 = Conv2DTranspose(start_neurons * 1, (3, 3), strides=(2, 2), padding="valid")(uconv2)
deconv1 = Conv1D(
start_neurons * 1,
filter_size,
activation='relu',
padding='same',
)(UpSampling1D(size=2)(uconv2)) #kernel_initializer='he_normal'
uconv1 = concatenate([deconv1, conv1])
uconv1 = Dropout(DropoutRatio)(uconv1)
uconv1 = Conv1D(start_neurons * 1,
filter_size,
activation=None,
padding="same")(uconv1)
# uconv1 = residual_block(uconv1,start_neurons * 1)
# uconv1 = residual_block(uconv1,start_neurons * 1)
uconv1 = bottleneck(uconv1, start_neurons * 1, block)
uconv1 = Activation(ACTIVATION)(uconv1)
uconv1 = Dropout(DropoutRatio / 2)(uconv1)
#output_layer = Conv1D(1, 1, padding="same", activation="sigmoid")(uconv1)
output_layer = Conv1D(nClasses, 1, activation='relu', padding='same')(
uconv1) #kernel_initializer='he_normal'
#output_layer = core.Reshape((nClasses, input_length))(output_layer)
#output_layer = core.Permute((2, 1))(output_layer)
output_layer = core.Activation('softmax')(output_layer)
#model = Model(inputs=inputs, outputs=conv9)
model = Model(input_layer, output_layer)
return model
ds = dataset.DataSet.load_from_path('usps', '../gmllib/datasets/usps')
# convert to 2D images
x_train = np.reshape(ds.train.x, (ds.train.N, 1, 16, 16))
x_test = np.reshape(ds.test.x, (ds.test.N, 1, 16, 16))
model = kmodel.Sequential()
model.add(
kconv.Convolution2D(nb_filter=4,
nb_row=5,
nb_col=5,
input_shape=(1, 16, 16),
border_mode='valid'))
model.add(klcore.Activation('tanh'))
# instead of average pooling, we use max pooling
model.add(kconv.MaxPooling2D(pool_size=(2, 2)))
# the 12 feature maps in this layer are connected in a specific pattern to the below layer, but it is not possible
# do this in keras easily. in fact, I don't know how keras connects the feature maps in one layer to the next.
model.add(kconv.Convolution2D(nb_filter=12, nb_row=5, nb_col=5))
model.add(klcore.Activation('tanh'))
model.add(kconv.MaxPooling2D(pool_size=(2, 2)))
model.add(klcore.Flatten())
model.add(klcore.Dense(output_dim=10))
model.add(klcore.Activation('softmax'))
model.compile(optimizer='sgd', loss='categorical_crossentropy')
class LayerCorrectnessTest(test_combinations.TestCase):
def setUp(self):
super(LayerCorrectnessTest, self).setUp()
# Set two virtual CPUs to test MirroredStrategy with multiple devices
cpus = tf.config.list_physical_devices('CPU')
tf.config.set_logical_device_configuration(cpus[0], [
tf.config.LogicalDeviceConfiguration(),
tf.config.LogicalDeviceConfiguration(),
])
def _create_model_from_layer(self, layer, input_shapes):
inputs = [layers.Input(batch_input_shape=s) for s in input_shapes]
if len(inputs) == 1:
inputs = inputs[0]
y = layer(inputs)
model = models.Model(inputs, y)
model.compile('sgd', 'mse')
return model
@parameterized.named_parameters(
('LeakyReLU', activation.LeakyReLU, (2, 2)),
('PReLU', activation.PReLU, (2, 2)), ('ELU', activation.ELU, (2, 2)),
('ThresholdedReLU', activation.ThresholdedReLU,
(2, 2)), ('Softmax', activation.Softmax,
(2, 2)), ('ReLU', activation.ReLU, (2, 2)),
('Conv1D', lambda: convolutional.Conv1D(2, 2), (2, 2, 1)),
('Conv2D', lambda: convolutional.Conv2D(2, 2), (2, 2, 2, 1)),
('Conv3D', lambda: convolutional.Conv3D(2, 2), (2, 2, 2, 2, 1)),
('Conv2DTranspose', lambda: convolutional.Conv2DTranspose(2, 2),
(2, 2, 2, 2)),
('SeparableConv2D', lambda: convolutional.SeparableConv2D(2, 2),
(2, 2, 2, 1)),
('DepthwiseConv2D', lambda: convolutional.DepthwiseConv2D(2, 2),
(2, 2, 2, 1)), ('UpSampling2D', reshaping.UpSampling2D, (2, 2, 2, 1)),
('ZeroPadding2D', reshaping.ZeroPadding2D,
(2, 2, 2, 1)), ('Cropping2D', reshaping.Cropping2D, (2, 3, 3, 1)),
('ConvLSTM2D', lambda: conv_lstm2d.ConvLSTM2D(4, kernel_size=(2, 2)),
(4, 4, 4, 4, 4)), ('Dense', lambda: core.Dense(2), (2, 2)),
('Dropout', lambda: regularization.Dropout(0.5), (2, 2)),
('SpatialDropout2D', lambda: regularization.SpatialDropout2D(0.5),
(2, 2, 2, 2)), ('Activation', lambda: core.Activation('sigmoid'),
(2, 2)), ('Reshape', lambda: reshaping.Reshape(
(1, 4, 1)), (2, 2, 2)),
('Permute', lambda: reshaping.Permute(
(2, 1)), (2, 2, 2)), ('Attention', attention.Attention, [
(2, 2, 3), (2, 3, 3), (2, 3, 3)
]), ('AdditiveAttention', attention.AdditiveAttention, [
(2, 2, 3), (2, 3, 3), (2, 3, 3)
]), ('Embedding', lambda: embeddings.Embedding(4, 4),
(2, 4), 2e-3, 2e-3, np.random.randint(4, size=(2, 4))),
('LocallyConnected1D',
lambda: locally_connected.LocallyConnected1D(2, 2),
(2, 2, 1)), ('LocallyConnected2D',
lambda: locally_connected.LocallyConnected2D(2, 2),
(2, 2, 2, 1)), ('Add', merging.Add, [(2, 2), (2, 2)]),
('Subtract', merging.Subtract, [(2, 2), (2, 2)]),
('Multiply', merging.Multiply, [
(2, 2), (2, 2)
]), ('Average', merging.Average, [(2, 2), (2, 2)]),
('Maximum', merging.Maximum, [
(2, 2), (2, 2)
]), ('Minimum', merging.Minimum, [
(2, 2), (2, 2)
]), ('Concatenate', merging.Concatenate, [
(2, 2), (2, 2)
]), ('Dot', lambda: merging.Dot(1), [(2, 2), (2, 2)]),
('GaussianNoise', lambda: regularization.GaussianNoise(0.5), (2, 2)),
('GaussianDropout', lambda: regularization.GaussianDropout(0.5),
(2, 2)), ('AlphaDropout', lambda: regularization.AlphaDropout(0.5),
(2, 2)),
('BatchNormalization', batch_normalization.BatchNormalization,
(2, 2), 1e-2, 1e-2),
('LayerNormalization', layer_normalization.LayerNormalization,
(2, 2)), ('LayerNormalizationUnfused',
lambda: layer_normalization.LayerNormalization(axis=1),
(2, 2, 2)), ('MaxPooling2D', pooling.MaxPooling2D,
(2, 2, 2, 1)),
('AveragePooling2D', pooling.AveragePooling2D,
(2, 2, 2, 1)), ('GlobalMaxPooling2D', pooling.GlobalMaxPooling2D,
(2, 2, 2, 1)),
('GlobalAveragePooling2D', pooling.GlobalAveragePooling2D,
(2, 2, 2, 1)), ('SimpleRNN', lambda: simple_rnn.SimpleRNN(units=4),
(4, 4, 4), 1e-2, 1e-2),
('SimpleRNN_stateful',
lambda: simple_rnn.SimpleRNN(units=4, stateful=True), (4, 4, 4), 1e-2,
1e-2), ('GRU', lambda: gru_v1.GRU(units=4),
(4, 4, 4)), ('LSTM', lambda: lstm_v1.LSTM(units=4),
(4, 4, 4)), ('GRUV2', lambda: gru.GRU(units=4),
(4, 4, 4)),
('GRUV2_stateful', lambda: gru.GRU(units=4, stateful=True),
(4, 4, 4)), ('LSTMV2', lambda: lstm.LSTM(units=4), (4, 4, 4)),
('LSTMV2_stateful', lambda: lstm.LSTM(units=4, stateful=True),
(4, 4, 4)), ('TimeDistributed',
lambda: time_distributed.TimeDistributed(core.Dense(2)),
(2, 2, 2)),
('Bidirectional',
lambda: bidirectional.Bidirectional(simple_rnn.SimpleRNN(units=4)),
(2, 2, 2)),
('AttentionLayerCausal', lambda: attention.Attention(causal=True), [
(2, 2, 3), (2, 3, 3), (2, 3, 3)
]), ('AdditiveAttentionLayerCausal',
lambda: attention.AdditiveAttention(causal=True), [
(2, 3, 4), (2, 3, 4), (2, 3, 4)
]), ('NormalizationAdapt', _create_normalization_layer_with_adapt,
(4, 4)),
('NormalizationNoAdapt', _create_normalization_layer_without_adapt,
(4, 4)), ('Resizing', lambda: image_preprocessing.Resizing(3, 3),
(2, 5, 5, 1)),
('Rescaling', lambda: image_preprocessing.Rescaling(2., 1.),
(6, 6)), ('CenterCrop', lambda: image_preprocessing.CenterCrop(3, 3),
(2, 5, 5, 1)))
def test_layer(self,
f32_layer_fn,
input_shape,
rtol=2e-3,
atol=2e-3,
input_data=None):
"""Tests a layer by comparing the float32 and mixed precision weights.
A float32 layer, a mixed precision layer, and a distributed mixed precision
layer are run. The three layers are identical other than their dtypes and
distribution strategies. The outputs after predict() and weights after fit()
are asserted to be close.
Args:
f32_layer_fn: A function returning a float32 layer. The other two layers
will automatically be created from this
input_shape: The shape of the input to the layer, including the batch
dimension. Or a list of shapes if the layer takes multiple inputs.
rtol: The relative tolerance to be asserted.
atol: The absolute tolerance to be asserted.
input_data: A Numpy array with the data of the input. If None, input data
will be randomly generated
"""
if f32_layer_fn == reshaping.ZeroPadding2D and tf.test.is_built_with_rocm(
):
return
if isinstance(input_shape[0], int):
input_shapes = [input_shape]
else:
input_shapes = input_shape
strategy = create_mirrored_strategy()
f32_layer = f32_layer_fn()
# Create the layers
assert f32_layer.dtype == f32_layer._compute_dtype == 'float32'
config = f32_layer.get_config()
config['dtype'] = policy.Policy('mixed_float16')
mp_layer = f32_layer.__class__.from_config(config)
distributed_mp_layer = f32_layer.__class__.from_config(config)
# Compute per_replica_input_shapes for the distributed model
global_batch_size = input_shapes[0][0]
assert global_batch_size % strategy.num_replicas_in_sync == 0, (
'The number of replicas, %d, does not divide the global batch size of '
'%d' % (strategy.num_replicas_in_sync, global_batch_size))
per_replica_batch_size = (global_batch_size //
strategy.num_replicas_in_sync)
per_replica_input_shapes = [(per_replica_batch_size, ) + s[1:]
for s in input_shapes]
# Create the models
f32_model = self._create_model_from_layer(f32_layer, input_shapes)
mp_model = self._create_model_from_layer(mp_layer, input_shapes)
with strategy.scope():
distributed_mp_model = self._create_model_from_layer(
distributed_mp_layer, per_replica_input_shapes)
# Set all model weights to the same values
f32_weights = f32_model.get_weights()
mp_model.set_weights(f32_weights)
distributed_mp_model.set_weights(f32_weights)
# Generate input data
if input_data is None:
# Cast inputs to float16 to avoid measuring error from having f16 layers
# cast to float16.
input_data = [
np.random.normal(size=s).astype('float16')
for s in input_shapes
]
if len(input_data) == 1:
input_data = input_data[0]
# Assert all models have close outputs.
f32_output = f32_model.predict(input_data)
mp_output = mp_model.predict(input_data)
self.assertAllClose(mp_output, f32_output, rtol=rtol, atol=atol)
self.assertAllClose(distributed_mp_model.predict(input_data),
f32_output,
rtol=rtol,
atol=atol)
# Run fit() on models
output = np.random.normal(
size=f32_model.outputs[0].shape).astype('float16')
for model in f32_model, mp_model, distributed_mp_model:
model.fit(input_data, output, batch_size=global_batch_size)
# Assert all models have close weights
f32_weights = f32_model.get_weights()
self.assertAllClose(mp_model.get_weights(),
f32_weights,
rtol=rtol,
atol=atol)
self.assertAllClose(distributed_mp_model.get_weights(),
f32_weights,
rtol=rtol,
atol=atol)
def get_unet2(n_ch, patch_height, patch_width):
inputs = Input(shape=(n_ch, patch_height, patch_width))
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(inputs)
conv1 = BatchNormalization(axis=1)(conv1)
conv1 = Dropout(0.3)(conv1)
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv1)
conv1 = BatchNormalization(axis=1)(conv1)
pool1 = MaxPooling2D((2, 2), data_format='channels_first')(conv1)
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(pool1)
conv2 = BatchNormalization(axis=1)(conv2)
conv2 = Dropout(0.3)(conv2)
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv2)
conv2 = BatchNormalization(axis=1)(conv2)
pool2 = MaxPooling2D((2, 2), data_format='channels_first')(conv2)
conv3 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(pool2)
conv3 = BatchNormalization(axis=1)(conv3)
conv3 = Dropout(0.3)(conv3)
conv3 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv3)
conv3 = BatchNormalization(axis=1)(conv3)
pool3 = MaxPooling2D((2, 2), data_format='channels_first')(conv3)
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(pool3)
conv4 = BatchNormalization(axis=1)(conv4)
conv4 = Dropout(0.3)(conv4)
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv4)
conv4 = BatchNormalization(axis=1)(conv4)
pool4 = MaxPooling2D((2, 2), data_format='channels_first')(conv4)
conv5 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(pool4)
conv5 = BatchNormalization(axis=1)(conv5)
conv5 = Dropout(0.3)(conv5)
conv5 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv5)
conv5 = BatchNormalization(axis=1)(conv5)
up1 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv5)
up1 = concatenate([conv4, up1], axis=1)
conv6 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up1)
conv6 = BatchNormalization(axis=1)(conv6)
conv6 = Dropout(0.3)(conv6)
conv6 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv6)
conv6 = BatchNormalization(axis=1)(conv6)
up2 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv6)
up2 = concatenate([conv3, up2], axis=1)
conv7 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up2)
conv7 = BatchNormalization(axis=1)(conv7)
conv7 = Dropout(0.3)(conv7)
conv7 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv7)
conv7 = BatchNormalization(axis=1)(conv7)
up3 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv7)
up3 = concatenate([conv2, up3], axis=1)
conv8 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up3)
conv8 = BatchNormalization(axis=1)(conv8)
conv8 = Dropout(0.3)(conv8)
conv8 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv8)
conv8 = BatchNormalization(axis=1)(conv8)
up4 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv8)
up4 = concatenate([conv1, up4], axis=1)
conv9 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up4)
conv9 = BatchNormalization(axis=1)(conv9)
conv9 = Dropout(0.3)(conv9)
conv9 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv9)
conv9 = BatchNormalization(axis=1)(conv9)
conv10 = Conv2D(2, (1, 1),
activation='relu',
padding='same',
data_format='channels_first')(conv9)
conv10 = BatchNormalization(axis=1)(conv10)
conv10 = core.Reshape((2, patch_height * patch_width))(conv10)
conv10 = core.Permute((2, 1))(conv10)
############
conv10 = core.Activation('softmax')(conv10)
model = Model(input=inputs, output=conv10)
adaGrad = Adagrad(lr=1e-7, epsilon=1e-7, decay=1e-6)
model.compile(optimizer='sgd',
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def get_gnet(n_ch, patch_height, patch_width):
inputs = Input((n_ch, patch_height, patch_width))
conv1 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(inputs)
conv1 = Dropout(0.2)(conv1)
conv1 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv1)
up1 = UpSampling2D(size=(2, 2))(conv1)
#
conv2 = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(up1)
conv2 = Dropout(0.2)(conv2)
conv2 = Convolution2D(16, 3, 3, activation='relu',
border_mode='same')(conv2)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv2)
#
conv3 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(pool1)
conv3 = Dropout(0.2)(conv3)
conv3 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv3)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv3)
#
conv4 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(pool2)
conv4 = Dropout(0.2)(conv4)
conv4 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(conv4)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv4)
#
conv5 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(pool3)
conv5 = Dropout(0.2)(conv5)
conv5 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(conv5)
#
up2 = merge([UpSampling2D(size=(2, 2))(conv5), conv4],
mode='concat',
concat_axis=1)
conv6 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(up2)
conv6 = Dropout(0.2)(conv6)
conv6 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(conv6)
#
up3 = merge([UpSampling2D(size=(2, 2))(conv6), conv3],
mode='concat',
concat_axis=1)
conv7 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(up3)
conv7 = Dropout(0.2)(conv7)
conv7 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv7)
#
up4 = merge([UpSampling2D(size=(2, 2))(conv7), conv2],
mode='concat',
concat_axis=1)
conv8 = Convolution2D(16, 3, 3, activation='relu', border_mode='same')(up4)
conv8 = Dropout(0.2)(conv8)
conv8 = Convolution2D(16, 3, 3, activation='relu',
border_mode='same')(conv8)
#
pool4 = MaxPooling2D(pool_size=(2, 2))(conv8)
conv9 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(pool4)
conv9 = Dropout(0.2)(conv9)
conv9 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv9)
#
conv10 = Convolution2D(2, 1, 1, activation='relu',
border_mode='same')(conv9)
conv10 = core.Reshape((2, patch_height * patch_width))(conv10)
conv10 = core.Permute((2, 1))(conv10)
############
conv10 = core.Activation('softmax')(conv10)
model = Model(input=inputs, output=conv10)
# sgd = SGD(lr=0.01, decay=1e-6, momentum=0.3, nesterov=False)
model.compile(optimizer='sgd',
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def get_unet_seg(n_ch, img_rows=480, img_cols=480):
inputs = Input((n_ch, img_rows, img_cols))
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv1_1")(inputs)
conv1 = BatchNormalization(axis=1, name="conv1_2")(conv1)
conv1 = Dropout(0.5, name="conv1_3")(conv1)
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv1_4")(conv1)
conv1 = BatchNormalization(axis=1, name="conv1_5")(conv1)
conv1.trainable = False
pool1 = MaxPooling2D((2, 2), data_format='channels_first',
name="conv1_6")(conv1)
pool1.trainable = False
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv2_1")(pool1)
conv2 = BatchNormalization(axis=1, name="conv2_2")(conv2)
conv2 = Dropout(0.5, name="conv2_3")(conv2)
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv2_4")(conv2)
conv2 = BatchNormalization(axis=1, name="conv2_5")(conv2)
conv2.trainable = False
pool2 = MaxPooling2D((2, 2), data_format='channels_first',
name="conv2_6")(conv2)
pool2.trainable = False
conv3 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv3_1")(pool2)
conv3 = BatchNormalization(axis=1, name="conv3_2")(conv3)
conv3 = Dropout(0.5, name="conv3_3")(conv3)
conv3 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv3_4")(conv3)
conv3 = BatchNormalization(axis=1, name="conv3_5")(conv3)
conv3.trainable = False
pool3 = MaxPooling2D((2, 2), data_format='channels_first',
name="conv3_6")(conv3)
pool3.trainable = False
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv4_1")(pool3)
conv4 = BatchNormalization(axis=1, name="conv4_2")(conv4)
conv4 = Dropout(0.5, name="conv4_3")(conv4)
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv4_4")(conv4)
conv4 = BatchNormalization(axis=1, name="conv4_5")(conv4)
conv4.trainable = False
pool4 = MaxPooling2D((2, 2), data_format='channels_first',
name="conv4_6")(conv4)
pool4.trainable = False
conv5 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv5_1")(pool4)
conv5 = BatchNormalization(axis=1, name="conv5_2")(conv5)
conv5 = Dropout(0.5, name="conv5_3")(conv5)
conv5 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv5_4")(conv5)
conv5 = BatchNormalization(axis=1, name="conv5_5")(conv5)
conv5.trainable = False
up1 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv5)
up1 = concatenate([conv4, up1], axis=1)
conv6 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up1)
conv6 = BatchNormalization(axis=1)(conv6)
conv6 = Dropout(0.3)(conv6)
conv6 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv6)
conv6 = BatchNormalization(axis=1)(conv6)
up2 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv6)
up2 = concatenate([conv3, up2], axis=1)
conv7 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up2)
conv7 = BatchNormalization(axis=1)(conv7)
conv7 = Dropout(0.3)(conv7)
conv7 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv7)
conv7 = BatchNormalization(axis=1)(conv7)
up3 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv7)
up3 = concatenate([conv2, up3], axis=1)
conv8 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up3)
conv8 = BatchNormalization(axis=1)(conv8)
conv8 = Dropout(0.3)(conv8)
conv8 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv8)
conv8 = BatchNormalization(axis=1)(conv8)
up4 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv8)
up4 = concatenate([conv1, up4], axis=1)
conv9 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up4)
conv9 = BatchNormalization(axis=1)(conv9)
conv9 = Dropout(0.3)(conv9)
conv9 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv9)
conv9 = BatchNormalization(axis=1)(conv9)
conv10 = Conv2D(2, (1, 1),
activation='relu',
padding='same',
data_format='channels_first')(conv9)
conv10 = BatchNormalization(axis=1)(conv10)
conv10 = core.Reshape((2, img_rows * img_cols))(conv10)
conv10 = core.Permute((2, 1))(conv10)
############
conv10 = core.Activation('softmax')(conv10)
model = Model(input=inputs, output=conv10)
adaGrad = Adagrad(lr=1e-7, epsilon=1e-7, decay=1e-6)
model.compile(optimizer='sgd',
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def test_TimeDistributed():
# first, test with Dense layer
model = Sequential()
model.add(wrappers.TimeDistributed(core.Dense(2), input_shape=(3, 4)))
model.add(core.Activation('relu'))
model.compile(optimizer='rmsprop', loss='mse')
model.fit(np.random.random((10, 3, 4)),
np.random.random((10, 3, 2)),
nb_epoch=1,
batch_size=10)
# test config
model.get_config()
# compare to TimeDistributedDense
test_input = np.random.random((1, 3, 4))
test_output = model.predict(test_input)
weights = model.layers[0].get_weights()
reference = Sequential()
reference.add(
core.TimeDistributedDense(2, input_shape=(3, 4), weights=weights))
reference.add(core.Activation('relu'))
reference.compile(optimizer='rmsprop', loss='mse')
reference_output = reference.predict(test_input)
assert_allclose(test_output, reference_output, atol=1e-05)
# test when specifying a batch_input_shape
reference = Sequential()
reference.add(
core.TimeDistributedDense(2,
batch_input_shape=(1, 3, 4),
weights=weights))
reference.add(core.Activation('relu'))
reference.compile(optimizer='rmsprop', loss='mse')
reference_output = reference.predict(test_input)
assert_allclose(test_output, reference_output, atol=1e-05)
# test with Convolution2D
model = Sequential()
model.add(
wrappers.TimeDistributed(convolutional.Convolution2D(
5, 2, 2, border_mode='same'),
input_shape=(2, 4, 4, 3)))
model.add(core.Activation('relu'))
model.compile(optimizer='rmsprop', loss='mse')
model.train_on_batch(np.random.random((1, 2, 4, 4, 3)),
np.random.random((1, 2, 4, 4, 5)))
model = model_from_json(model.to_json())
model.summary()
# test stacked layers
model = Sequential()
model.add(wrappers.TimeDistributed(core.Dense(2), input_shape=(3, 4)))
model.add(wrappers.TimeDistributed(core.Dense(3)))
model.add(core.Activation('relu'))
model.compile(optimizer='rmsprop', loss='mse')
model.fit(np.random.random((10, 3, 4)),
np.random.random((10, 3, 3)),
nb_epoch=1,
batch_size=10)
# test wrapping Sequential model
model = Sequential()
model.add(core.Dense(3, input_dim=2))
outer_model = Sequential()
outer_model.add(wrappers.TimeDistributed(model, input_shape=(3, 2)))
outer_model.compile(optimizer='rmsprop', loss='mse')
outer_model.fit(np.random.random((10, 3, 2)),
np.random.random((10, 3, 3)),
nb_epoch=1,
batch_size=10)
# test with functional API
x = Input(shape=(3, 2))
y = wrappers.TimeDistributed(model)(x)
outer_model = Model(x, y)
outer_model.compile(optimizer='rmsprop', loss='mse')
outer_model.fit(np.random.random((10, 3, 2)),
np.random.random((10, 3, 3)),
nb_epoch=1,
batch_size=10)
def get_unet_trainable_seg(n_ch, img_rows=480, img_cols=480):
inputs = Input((n_ch, img_rows, img_cols))
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv1_1")(inputs)
conv1 = BatchNormalization(axis=1, name="conv1_2")(conv1)
conv1 = Dropout(0.5, name="conv1_3")(conv1)
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv1_4")(conv1)
conv1 = BatchNormalization(axis=1, name="conv1_5")(conv1)
conv1.trainable = True
pool1 = MaxPooling2D((2, 2), data_format='channels_first',
name="conv1_6")(conv1)
pool1.trainable = True
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv2_1")(pool1)
conv2 = BatchNormalization(axis=1, name="conv2_2")(conv2)
conv2 = Dropout(0.5, name="conv2_3")(conv2)
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv2_4")(conv2)
conv2 = BatchNormalization(axis=1, name="conv2_5")(conv2)
conv2.trainable = True
pool2 = MaxPooling2D((2, 2), data_format='channels_first',
name="conv2_6")(conv2)
pool2.trainable = True
conv3 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv3_1")(pool2)
conv3 = BatchNormalization(axis=1, name="conv3_2")(conv3)
conv3 = Dropout(0.5, name="conv3_3")(conv3)
conv3 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv3_4")(conv3)
conv3 = BatchNormalization(axis=1, name="conv3_5")(conv3)
conv3.trainable = True
pool3 = MaxPooling2D((2, 2), data_format='channels_first',
name="conv3_6")(conv3)
pool3.trainable = True
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv4_1")(pool3)
conv4 = BatchNormalization(axis=1, name="conv4_2")(conv4)
conv4 = Dropout(0.5, name="conv4_3")(conv4)
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv4_4")(conv4)
conv4 = BatchNormalization(axis=1, name="conv4_5")(conv4)
conv4.trainable = True
pool4 = MaxPooling2D((2, 2), data_format='channels_first',
name="conv4_6")(conv4)
pool4.trainable = True
conv5 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv5_1")(pool4)
conv5 = BatchNormalization(axis=1, name="conv5_2")(conv5)
conv5 = Dropout(0.5, name="conv5_3")(conv5)
conv5 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first',
name="conv5_4")(conv5)
conv5 = BatchNormalization(axis=1, name="conv5_5")(conv5)
conv5.trainable = True
up1 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv5)
up1 = concatenate([conv4, up1], axis=1)
conv6 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up1)
conv6 = BatchNormalization(axis=1)(conv6)
conv6 = Dropout(0.3)(conv6)
conv6 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv6)
conv6 = BatchNormalization(axis=1)(conv6)
up2 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv6)
up2 = concatenate([conv3, up2], axis=1)
conv7 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up2)
conv7 = BatchNormalization(axis=1)(conv7)
conv7 = Dropout(0.3)(conv7)
conv7 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv7)
conv7 = BatchNormalization(axis=1)(conv7)
up3 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv7)
up3 = concatenate([conv2, up3], axis=1)
conv8 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up3)
conv8 = BatchNormalization(axis=1)(conv8)
conv8 = Dropout(0.3)(conv8)
conv8 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv8)
conv8 = BatchNormalization(axis=1)(conv8)
up4 = UpSampling2D(size=(2, 2), data_format='channels_first')(conv8)
up4 = concatenate([conv1, up4], axis=1)
conv9 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up4)
conv9 = BatchNormalization(axis=1)(conv9)
conv9 = Dropout(0.3)(conv9)
conv9 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv9)
conv9 = BatchNormalization(axis=1)(conv9)
conv10 = Conv2D(1, (1, 1),
activation='relu',
padding='same',
data_format='channels_first')(conv9)
conv10 = BatchNormalization(axis=1)(conv10)
############
conv10 = core.Activation('sigmoid')(conv10)
model = Model(input=inputs, output=conv10)
return model
def get_unet(n_ch, patch_height, patch_width):
inputs = Input(shape=(n_ch, patch_height, patch_width))
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(inputs)
conv1 = Dropout(0.2)(conv1)
conv1 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv1)
pool1 = MaxPooling2D((2, 2))(conv1)
#
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(pool1)
conv2 = Dropout(0.2)(conv2)
conv2 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv2)
pool2 = MaxPooling2D((2, 2))(conv2)
#
conv3 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(pool2)
conv3 = Dropout(0.2)(conv3)
conv3 = Conv2D(128, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv3)
up1 = UpSampling2D(size=(2, 2))(conv3)
up1 = concatenate([conv2, up1], axis=1)
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up1)
conv4 = Dropout(0.2)(conv4)
conv4 = Conv2D(64, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv4)
#
up2 = UpSampling2D(size=(2, 2))(conv4)
up2 = concatenate([conv1, up2], axis=1)
conv5 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(up2)
conv5 = Dropout(0.2)(conv5)
conv5 = Conv2D(32, (3, 3),
activation='relu',
padding='same',
data_format='channels_first')(conv5)
#
conv6 = Conv2D(2, (1, 1),
activation='relu',
padding='same',
data_format='channels_first')(conv5)
conv6 = core.Reshape((2, patch_height * patch_width))(conv6)
conv6 = core.Permute((2, 1))(conv6)
############
conv7 = core.Activation('softmax')(conv6)
model = Model(inputs=inputs, outputs=conv7)
# sgd = SGD(lr=0.01, decay=1e-6, momentum=0.3, nesterov=False)
model.compile(optimizer='sgd',
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def att_unet(params, data_format='channels_last'):
inputs = Input((params.img_shape, params.img_shape, 3))
x = inputs
depth = params.depth
features = params.n_filters
skips = []
for i in range(depth):
x = Conv2D(features, (3, 3),
activation='relu',
padding='same',
data_format=data_format)(x)
if params.batchnorm:
x = BatchNormalization()(x)
x = Dropout(params.dropout)(x)
x = Conv2D(features, (3, 3),
activation='relu',
padding='same',
data_format=data_format)(x)
if params.batchnorm:
x = BatchNormalization()(x)
skips.append(x)
x = MaxPooling2D((2, 2), data_format=data_format)(x)
x = Conv2D(features, (3, 3),
activation='relu',
padding='same',
data_format=data_format)(x)
x = Dropout(params.dropout)(x)
x = Conv2D(features, (3, 3),
activation='relu',
padding='same',
data_format=data_format)(x)
for i in reversed(range(depth)):
if i in [0, 1, 2]:
features = features // 2
x = attention_up_and_concate(x,
skips[i],
features,
data_format=data_format)
x = Conv2D(features, (3, 3),
activation='relu',
padding='same',
data_format=data_format)(x)
if params.batchnorm:
x = BatchNormalization()(x)
x = Dropout(params.dropout)(x)
x = Conv2D(features, (3, 3),
activation='relu',
padding='same',
data_format=data_format)(x)
if params.batchnorm:
x = BatchNormalization()(x)
conv6 = Conv2D(params.num_classes, (1, 1),
padding='same',
data_format=data_format)(x)
conv7 = core.Activation('softmax')(conv6)
model = Model(inputs=inputs, outputs=conv7)
return model