def BinaryMobilenetV2():
model = Sequential()
model.add(BinaryConv2D(32, kernel_size=kernel_size, strides=(2, 2), H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias,input_shape=(32, 32, 3)))
model.add(BatchNormalization(axis=1))
model.add(Activation(binary_tanh))
model = InvertedResidualBlock(model, 16, (3, 3), t=1, strides=1, n=1)
model = InvertedResidualBlock(model, 24, (3, 3), t=6, strides=2, n=2)
model = InvertedResidualBlock(model, 32, (3, 3), t=6, strides=2, n=3)
model = InvertedResidualBlock(model, 64, (3, 3), t=6, strides=2, n=4)
model = InvertedResidualBlock(model, 96, (3, 3), t=6, strides=1, n=3)
model = InvertedResidualBlock(model, 160, (3, 3), t=6, strides=2, n=3)
model = InvertedResidualBlock(model, 320, (3, 3), t=6, strides=1, n=1)
model.add(BinaryConv2D(1280, H=H,kernel_size=(1,1), kernel_lr_multiplier=kernel_lr_multiplier, use_bias=use_bias,strides=(1, 1)))
model.add(BatchNormalization(axis=1))
model.add(Activation(binary_tanh))
model.add(Flatten())
model.add(BinaryDense(classes, H=H, kernel_lr_multiplier=kernel_lr_multiplier, use_bias=use_bias))
model.add(BatchNormalization(axis=1))
return model
def addBottleneck(model_input, filters, kernel, t, s):
model = model_input
channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
tchannel = K.int_shape(model_input.layers[-1].output)[channel_axis] *t
model.add(BinaryConv2D(tchannel, kernel_size=(1,1), H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias))
model.add(BatchNormalization(axis=1))
model.add(Activation(binary_tanh))
model.add(SeparableBinaryConv2D(filters,kernel_size=kernel_size,strides=(s,s), H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias))
model.add(BatchNormalization(axis=1))
#model.add(Activation(binary_tanh))
return model
def unet(pretrained_weights=None,
batch_size=2,
input_shape=data_shape,
input_size=(30, 32, 32, 1)):
inputs = Input(shape=data_shape, batch_size=batch_size)
bin_conv1 = TimeDistributed(
BinaryConv2D(1,
kernel_size=(32, 32),
input_shape=input_size,
data_format='channels_last',
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
padding='same',
use_bias=use_bias,
name='bin_conv_1'))(inputs)
s = Lambda(streak, output_shape=streak_output_shape)(bin_conv1)
i = Lambda(integrate_ims, output_shape=integrate_ims_output_shape)(s)
f = Flatten()(i)
dense1 = Dense(30720, activation='relu')(f)
resh = Reshape((30, 32, 32, 1))(dense1)
c1 = TimeDistributed(
Conv2D(16, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(resh)
c1 = Dropout(0.1)(c1)
c1 = TimeDistributed(
Conv2D(16, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(c1)
p1 = TimeDistributed(MaxPooling2D((2, 2)))(c1)
c2 = TimeDistributed(
Conv2D(32, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(p1)
c2 = Dropout(0.1)(c2)
c2 = TimeDistributed(
Conv2D(32, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(c2)
p2 = TimeDistributed(MaxPooling2D((2, 2)))(c2)
c3 = TimeDistributed(
Conv2D(64, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(p2)
c3 = Dropout(0.2)(c3)
c3 = TimeDistributed(
Conv2D(64, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(c3)
p3 = TimeDistributed(MaxPooling2D((2, 2)))(c3)
c4 = TimeDistributed(
Conv2D(128, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(p3)
c4 = Dropout(0.2)(c4)
c4 = TimeDistributed(
Conv2D(128, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(c4)
p4 = TimeDistributed(MaxPooling2D(pool_size=(2, 2)))(c4)
c5 = TimeDistributed(
Conv2D(256, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(p4)
c5 = Dropout(0.3)(c5)
c5 = TimeDistributed(
Conv2D(256, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(c5)
u6 = TimeDistributed(
Conv2DTranspose(128, (2, 2), strides=(2, 2), padding='same'))(c5)
u6 = concatenate([u6, c4])
c6 = TimeDistributed(
Conv2D(128, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(u6)
c6 = Dropout(0.2)(c6)
c6 = TimeDistributed(
Conv2D(128, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(c6)
u7 = TimeDistributed(
Conv2DTranspose(64, (2, 2), strides=(2, 2), padding='same'))(c6)
u7 = concatenate([u7, c3])
c7 = TimeDistributed(
Conv2D(64, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(u7)
c7 = Dropout(0.2)(c7)
c7 = TimeDistributed(
Conv2D(64, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(c7)
u8 = TimeDistributed(
Conv2DTranspose(32, (2, 2), strides=(2, 2), padding='same'))(c7)
u8 = concatenate([u8, c2])
c8 = TimeDistributed(
Conv2D(32, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(u8)
c8 = Dropout(0.1)(c8)
c8 = TimeDistributed(
Conv2D(32, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(c8)
u9 = TimeDistributed(
Conv2DTranspose(16, (2, 2), strides=(2, 2), padding='same'))(c8)
c9 = TimeDistributed(
Conv2D(16, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(u9)
c9 = Dropout(0.1)(c9)
c9 = TimeDistributed(
Conv2D(16, (2, 2),
activation='elu',
kernel_initializer='he_normal',
padding='same'))(c9)
outputs = TimeDistributed(Conv2D(1, (1, 1), activation='sigmoid'))(c9)
model = Model(inputs=[inputs], outputs=[outputs])
model.compile(optimizer=Adam(lr=1e-4),
loss='mean_squared_error',
metrics=['accuracy'])
return model
def BinaryVGG16():
model = Sequential()
# 64
model.add(BinaryConv2D(64, kernel_size=kernel_size, input_shape=(32, 32, 3),
data_format='channels_last',
H=H, kernel_lr_multiplier=kernel_lr_multiplier,
padding='same', use_bias=use_bias, name='conv1'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn1'))
model.add(Activation(binary_tanh, name='act1'))
model.add(BinaryConv2D(64, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv2'))
model.add(MaxPooling2D(pool_size=pool_size, name='pool2', data_format='channels_last'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn2'))
model.add(Activation(binary_tanh, name='act2'))
# 128
model.add(BinaryConv2D(128, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv3'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn3'))
model.add(Activation(binary_tanh, name='act3'))
model.add(BinaryConv2D(128, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv4'))
model.add(MaxPooling2D(pool_size=pool_size, name='pool4', data_format='channels_last'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn4'))
model.add(Activation(binary_tanh, name='act4'))
# 256
model.add(BinaryConv2D(256, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv5'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn5'))
model.add(Activation(binary_tanh, name='act5'))
model.add(BinaryConv2D(256, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv6'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn6'))
model.add(Activation(binary_tanh, name='act6'))
model.add(BinaryConv2D(256, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv7'))
model.add(MaxPooling2D(pool_size=pool_size, name='pool7', data_format='channels_last'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn7'))
model.add(Activation(binary_tanh, name='act7'))
# 512
model.add(BinaryConv2D(512, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv8'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn8'))
model.add(Activation(binary_tanh, name='act8'))
model.add(BinaryConv2D(512, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv9'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn9'))
model.add(Activation(binary_tanh, name='act9'))
model.add(BinaryConv2D(512, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv10'))
model.add(MaxPooling2D(pool_size=pool_size, name='pool10', data_format='channels_last'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn10'))
model.add(Activation(binary_tanh, name='act10'))
#512
model.add(BinaryConv2D(512, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv11'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn11'))
model.add(Activation(binary_tanh, name='act11'))
model.add(BinaryConv2D(512, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv12'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn12'))
model.add(Activation(binary_tanh, name='act12'))
model.add(BinaryConv2D(512, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_last',
padding='same', use_bias=use_bias, name='conv13'))
model.add(MaxPooling2D(pool_size=pool_size, name='pool13', data_format='channels_last'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn13'))
model.add(Activation(binary_tanh, name='act13'))
#final
model.add(Flatten())
# dense1
model.add(BinaryDense(4096, H=H, kernel_lr_multiplier=kernel_lr_multiplier, use_bias=use_bias, name='dense1'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn14'))
model.add(Activation(binary_tanh, name='act14'))
# dense2
model.add(BinaryDense(4096, H=H, kernel_lr_multiplier=kernel_lr_multiplier, use_bias=use_bias, name='dense2'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn15'))
model.add(Activation(binary_tanh, name='act15'))
#dense3
model.add(BinaryDense(classes, H=H, kernel_lr_multiplier=kernel_lr_multiplier, use_bias=use_bias, name='dense3'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn16'))
return model
X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255
X_test /= 255
print(X_train.shape[0], 'train samples')
print(X_test.shape[0], 'test samples')
# convert class vectors to binary class matrices
Y_train = np_utils.to_categorical(y_train, classes) * 2 - 1 # -1 or 1 for hinge loss
Y_test = np_utils.to_categorical(y_test, classes) * 2 - 1
model = Sequential()
# conv1
model.add(BinaryConv2D(128, kernel_size=kernel_size, input_shape=(channels, img_rows, img_cols),
data_format='channels_first',
H=H, kernel_lr_multiplier=kernel_lr_multiplier,
padding='same', use_bias=use_bias, name='conv1'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn1'))
model.add(Activation(binary_tanh, name='act1'))
# conv2
model.add(BinaryConv2D(128, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_first',
padding='same', use_bias=use_bias, name='conv2'))
model.add(MaxPooling2D(pool_size=pool_size, name='pool2'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn2'))
model.add(Activation(binary_tanh, name='act2'))
# conv3
model.add(BinaryConv2D(256, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_first',
padding='same', use_bias=use_bias, name='conv3'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn3'))
validate = mask(validate, ims, bk_temp)
validate2 = validate
validate2 = np_streak(validate)
validate3 = np.sum(validate2, axis=1)
model1 = Sequential()
model1.add(Input(shape=(30, 32, 32, 1), batch_size=100))
model1.add(
TimeDistributed(
BinaryConv2D(1,
kernel_size=(32, 32),
input_shape=(30, 32, 32, 1),
data_format='channels_last',
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
padding='same',
use_bias=use_bias,
name='bin_conv_1')))
model1.compile(optimizer=Adam(lr=1),
loss='mean_squared_error',
metrics=['mse'])
model1.summary()
history1 = model1.fit(ims, validate, batch_size=100, epochs=5, verbose=2)
model2 = Sequential()
model2.add(Input(shape=(30, 32, 32, 1), batch_size=100))
model2.add(
#Y_test = np_utils.to_categorical(y_test, classes) * 2 - 1
X_train = np.zeros((80, 256, 256, 1))
for i in range(1, 81):
im = cv2.imread('../data/' + str(i) + '.png', 0)
im = np.reshape(im, (256, 256, 1))
X_train[i - 1, :, :, :] = im
print(np.shape(X_train))
model = Sequential()
cnn = BinaryConv2D(1,
kernel_size=kernel_size,
input_shape=(img_rows, img_cols, channels),
data_format='channels_last',
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
padding='same',
use_bias=use_bias,
name='conv1')
bk = cnn.build((80, 256, 256, 1))
outputs, bk_temp = cnn.call(X_train)
print(np.shape(bk_temp))
plt.figure(0)
plt.imshow(np.reshape(X_train[0], (256, 256)),
interpolation='nearest',
cmap='hot')
plt.figure(1)
plt.imshow(np.reshape(outputs[0], (256, 256)),
interpolation='nearest',
def build_model(kernel_size, channels, img_rows, img_cols, H,
kernel_lr_multiplier, use_bias, epsilon, momentum):
model = Sequential()
# conv1
model.add(
BinaryConv2D(128,
kernel_size=kernel_size,
input_shape=(channels, img_rows, img_cols),
data_format='channels_first',
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
padding='same',
use_bias=use_bias,
name='conv1'))
model.add(
MaxPooling2D(pool_size=pool_size,
name='pool1',
data_format='channels_first'))
model.add(
BatchNormalization(epsilon=epsilon,
momentum=momentum,
axis=1,
name='bn1'))
model.add(Activation(binary_tanh, name='act1'))
# conv2
model.add(
BinaryConv2D(128,
kernel_size=kernel_size,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_first',
padding='same',
use_bias=use_bias,
name='conv2'))
model.add(
MaxPooling2D(pool_size=pool_size,
name='pool2',
data_format='channels_first'))
model.add(
BatchNormalization(epsilon=epsilon,
momentum=momentum,
axis=1,
name='bn2'))
model.add(Activation(binary_tanh, name='act2'))
# conv3
model.add(
BinaryConv2D(128,
kernel_size=kernel_size,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_first',
padding='same',
use_bias=use_bias,
name='conv3'))
model.add(
MaxPooling2D(pool_size=pool_size,
name='pool3',
data_format='channels_first'))
model.add(
BatchNormalization(epsilon=epsilon,
momentum=momentum,
axis=1,
name='bn3'))
model.add(Activation(binary_tanh, name='act3'))
# conv4
model.add(
BinaryConv2D(256,
kernel_size=kernel_size,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_first',
padding='same',
use_bias=use_bias,
name='conv4'))
model.add(
MaxPooling2D(pool_size=pool_size,
name='pool4',
data_format='channels_first'))
model.add(
BatchNormalization(epsilon=epsilon,
momentum=momentum,
axis=1,
name='bn4'))
model.add(Activation(binary_tanh, name='act4'))
model.add(Flatten())
# dense1
model.add(
BinaryDense(256,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
use_bias=use_bias,
name='dense5'))
model.add(
BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn5'))
model.add(Activation(binary_tanh, name='act5'))
# dense2
model.add(
BinaryDense(classes,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
use_bias=use_bias,
name='dense6'))
model.add(
BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn6'))
opt = Adam(lr=lr_start)
model.compile(loss='squared_hinge', optimizer=opt, metrics=['acc'])
model.summary()
return model
X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255
X_test /= 255
print(X_train.shape[0], 'train samples')
print(X_test.shape[0], 'test samples')
# convert class vectors to binary class matrices
Y_train = np_utils.to_categorical(y_train, nb_classes) * 2 - 1 # -1 or 1 for hinge loss
Y_test = np_utils.to_categorical(y_test, nb_classes) * 2 - 1
model = Sequential()
# conv1
model.add(BinaryConv2D(128, 3, 3, input_shape=(nb_channel, img_rows, img_cols),
H=H, W_lr_multiplier=W_lr_multiplier,
border_mode='same', bias=bias, name='conv1'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn1'))
model.add(Activation(binary_tanh, name='act1'))
# conv2
model.add(BinaryConv2D(128, 3, 3, H=H, W_lr_multiplier=W_lr_multiplier,
border_mode='same', bias=bias, name='conv2'))
model.add(MaxPooling2D(pool_size=(2, 2), name='pool2'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn2'))
model.add(Activation(binary_tanh, name='act2'))
# conv3
model.add(BinaryConv2D(256, 3, 3, H=H, W_lr_multiplier=W_lr_multiplier,
border_mode='same', bias=bias, name='conv3'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn3'))
model.add(Activation(binary_tanh, name='act3'))
# conv4
X_train = X_train.astype('float32')
X_test = X_test.astype('float32')
X_train /= 255
X_test /= 255
print(X_train.shape[0], 'train samples')
print(X_test.shape[0], 'test samples')
# convert class vectors to binary class matrices
Y_train = np_utils.to_categorical(y_train, classes) * 2 - 1 # -1 or 1 for hinge loss
Y_test = np_utils.to_categorical(y_test, classes) * 2 - 1
model = Sequential()
# conv1
model.add(BinaryConv2D(32, kernel_size=kernel_size, input_shape=(channels, img_rows, img_cols),
data_format='channels_first',
H=H, kernel_lr_multiplier=kernel_lr_multiplier,
padding='same', use_bias=use_bias, name='conv1'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn1'))
model.add(Activation(binary_tanh, name='act1'))
##############################################################################################################
# conv_dw_2_1
model.add(BinaryConv2D(32, kernel_size=kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
data_format='channels_first',
padding='same', use_bias=use_bias, name='conv_dw_2_1'))
#model.add(MaxPooling2D(pool_size=pool_size, name='pool2', data_format='channels_first'))
model.add(BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='conv_dw_bn_2_1'))
model.add(Activation(binary_tanh, name='conv_dw_act_2_1'))
# conv_1x1_2_1
model.add(BinaryConv2D(32, kernel_size=mini_kernel_size, H=H, kernel_lr_multiplier=kernel_lr_multiplier,
def mse_loss(y_true,y_pred):
return K.mean(K.square(y_pred-y_true))
"""
Fine Tuning Parameters:
Ratio between MSE and SSIM
"""
lrs = [0.00001,0.0001,0.001,0.01]
mse_losses = []
ssim_losses = []
forward_model = Sequential()
forward_model.add(Input(shape=(30,32,32,1),batch_size = 32))
forward_model.add(TimeDistributed(BinaryConv2D(1, kernel_size=(32,32), input_shape=(30,32,32,1),
data_format='channels_last',
H=H, kernel_lr_multiplier=kernel_lr_multiplier,
padding='same', use_bias=use_bias, name='bin_conv_1')))
forward_model.add(Reshape((30,32,32)))
forward_model.add(Lambda(streak,output_shape=streak_output_shape))
forward_model.add(Lambda(integrate_ims, output_shape = integrate_ims_output_shape))
forward_model.add(Flatten())
forward_model.add(Dense(30720, activation = 'relu'))
forward_model.add(Reshape((30,32,32,1)))
forward_model.compile(optimizer = Nadam(0.0001), loss = custom_loss, metrics = ['mean_squared_error',mse_loss])
forward_model.load_weights('../data/model_stuff/forward_weights_4_12.h5')
binary_weights = forward_model.layers[0].get_weights()
inverse_weights = forward_model.layers[5].get_weights()
for i in range(4):
inner_mse_losses=[]
# convert class vectors to binary class matrices
Y_train = np_utils.to_categorical(y_train,
classes) * 2 - 1 # -1 or 1 for hinge loss
Y_test = np_utils.to_categorical(y_test, classes) * 2 - 1
model = Sequential() # This model is a linear stack of layers.
# conv1
model.add(
BinaryConv2D(128,
kernel_size=kernel_size,
input_shape=(channels, img_rows, img_cols),
data_format='channels_first',
kernel_regularizer=regularizers.l2(0.01),
activity_regularizer=regularizer.l1(0.01),
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
padding='same',
use_bias=use_bias,
name='conv1'))
model.add(
BatchNormalization(epsilon=epsilon, momentum=momentum, axis=1, name='bn1'))
model.add(Activation(binary_tanh, name='act1'))
# conv2
model.add(
BinaryConv2D(128,
kernel_size=kernel_size,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
def Binary_Net(kernel_size, img_rows, img_cols, channels, data_format, H,
kernel_lr_multiplier, use_bias, epsilon, momentum, classes,
pool_size):
model = Sequential()
# conv1
model.add(
BinaryConv2D(128,
kernel_size=kernel_size,
input_shape=(img_rows, img_cols, channels),
data_format=data_format,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
padding='same',
use_bias=use_bias,
name='conv1'))
model.add(
BatchNormalization(epsilon=epsilon,
momentum=momentum,
axis=-1,
name='bn1'))
model.add(Activation(binary_tanh, name='act1'))
# conv2
model.add(
BinaryConv2D(128,
kernel_size=kernel_size,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
data_format=data_format,
padding='same',
use_bias=use_bias,
name='conv2'))
model.add(
MaxPooling2D(pool_size=pool_size,
name='pool2',
data_format=data_format))
model.add(
BatchNormalization(epsilon=epsilon,
momentum=momentum,
axis=-1,
name='bn2'))
model.add(Activation(binary_tanh, name='act2'))
# conv3
model.add(
BinaryConv2D(256,
kernel_size=kernel_size,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
data_format=data_format,
padding='same',
use_bias=use_bias,
name='conv3'))
model.add(
BatchNormalization(epsilon=epsilon,
momentum=momentum,
axis=-1,
name='bn3'))
model.add(Activation(binary_tanh, name='act3'))
# conv4
model.add(
BinaryConv2D(256,
kernel_size=kernel_size,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
data_format=data_format,
padding='same',
use_bias=use_bias,
name='conv4'))
model.add(
MaxPooling2D(pool_size=pool_size,
name='pool4',
data_format=data_format))
model.add(
BatchNormalization(epsilon=epsilon,
momentum=momentum,
axis=-1,
name='bn4'))
model.add(Activation(binary_tanh, name='act4'))
model.add(Flatten())
# dense1
model.add(
BinaryDense(1024,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
use_bias=use_bias,
name='dense5'))
model.add(
BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn5'))
model.add(Activation(binary_tanh, name='act5'))
# dense2
model.add(
BinaryDense(classes,
H=H,
kernel_lr_multiplier=kernel_lr_multiplier,
use_bias=use_bias,
name='dense6'))
model.add(
BatchNormalization(epsilon=epsilon, momentum=momentum, name='bn6'))
return model