def AlexNet(weights_path=None, heatmap=False):
if heatmap:
inputs = Input(shape=(3,None,None))
else:
inputs = Input(shape=(3,227,227))
conv_1 = Convolution2D(96, 11, 11,subsample=(4,4),activation='relu',
name='conv_1')(inputs)
conv_2 = MaxPooling2D((3, 3), strides=(2,2))(conv_1)
conv_2 = crosschannelnormalization(name="convpool_1")(conv_2)
conv_2 = ZeroPadding2D((2,2))(conv_2)
conv_2 = merge([
Convolution2D(128,5,5,activation="relu",name='conv_2_'+str(i+1))(
splittensor(ratio_split=2,id_split=i)(conv_2)
) for i in range(2)], mode='concat',concat_axis=1,name="conv_2")
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1,1))(conv_3)
conv_3 = Convolution2D(384,3,3,activation='relu',name='conv_3')(conv_3)
conv_4 = ZeroPadding2D((1,1))(conv_3)
conv_4 = merge([
Convolution2D(192,3,3,activation="relu",name='conv_4_'+str(i+1))(
splittensor(ratio_split=2,id_split=i)(conv_4)
) for i in range(2)], mode='concat',concat_axis=1,name="conv_4")
conv_5 = ZeroPadding2D((1,1))(conv_4)
conv_5 = merge([
Convolution2D(128,3,3,activation="relu",name='conv_5_'+str(i+1))(
splittensor(ratio_split=2,id_split=i)(conv_5)
) for i in range(2)], mode='concat',concat_axis=1,name="conv_5")
dense_1 = MaxPooling2D((3, 3), strides=(2,2),name="convpool_5")(conv_5)
if heatmap:
dense_1 = Convolution2D(4096,6,6,activation="relu",name="dense_1")(dense_1)
dense_2 = Convolution2D(4096,1,1,activation="relu",name="dense_2")(dense_1)
dense_3 = Convolution2D(1000, 1,1,name="dense_3")(dense_2)
prediction = Softmax4D(axis=1,name="softmax")(dense_3)
else:
dense_1 = Flatten(name="flatten")(dense_1)
dense_1 = Dense(4096, activation='relu',name='dense_1')(dense_1)
dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu',name='dense_2')(dense_2)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(1000,name='dense_3')(dense_3)
prediction = Activation("softmax",name="softmax")(dense_3)
model = Model(input=inputs, output=prediction)
if weights_path:
model.load_weights(weights_path)
return model
def AlexNet(weights_path=None, heatmap=False, classes = 1000):
if heatmap:
inputs = Input(shape=(3,None,None))
else:
inputs = Input(shape=(3,227,227))
conv_1 = Convolution2D(96, (11, 11),strides=(4,4),activation='relu',
name='conv_1')(inputs)
conv_2 = MaxPooling2D((3, 3), strides=(2,2))(conv_1)
conv_2 = crosschannelnormalization(name="convpool_1")(conv_2)
conv_2 = ZeroPadding2D((2,2))(conv_2)
conv_2 = merge([
Convolution2D(128,(5,5),activation="relu",name='conv_2_'+str(i+1))(
splittensor(ratio_split=2,id_split=i)(conv_2)
) for i in range(2)], mode='concat',concat_axis=1,name="conv_2")
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1,1))(conv_3)
conv_3 = Convolution2D(384,(3,3),activation='relu',name='conv_3')(conv_3)
conv_4 = ZeroPadding2D((1,1))(conv_3)
conv_4 = merge([
Convolution2D(192,(3,3),activation="relu",name='conv_4_'+str(i+1))(
splittensor(ratio_split=2,id_split=i)(conv_4)
) for i in range(2)], mode='concat',concat_axis=1,name="conv_4")
conv_5 = ZeroPadding2D((1,1))(conv_4)
conv_5 = merge([
Convolution2D(128,(3,3),activation="relu",name='conv_5_'+str(i+1))(
splittensor(ratio_split=2,id_split=i)(conv_5)
) for i in range(2)], mode='concat',concat_axis=1,name="conv_5")
dense_1 = MaxPooling2D((3, 3), strides=(2,2),name="convpool_5")(conv_5)
dense_1 = Flatten(name="flatten")(dense_1)
dense_1 = Dense(4096, activation='relu',name='dense_1')(dense_1)
dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu',name='dense_2')(dense_2)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(classes, name='dense_3a')(dense_3)
prediction = Activation("softmax",name="softmaxa")(dense_3)
model = Model(input=inputs, output=prediction, name='alexnet')
if weights_path:
model.load_weights(weights_path, by_name=True)
# Learning rate is changed to 0.0001
sgd = SGD(lr=1e-4, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd, loss='categorical_crossentropy', metrics=['accuracy'])
return model
def AlexNet(weights_path=None, heatmap=False):
if heatmap:
inputs = Input(shape=(3,None,None))
else:
inputs = Input(shape=(1,227,227))
conv_1 = Convolution2D(96, 11, 11,subsample=(4,4),activation='relu',
name='conv_1')(inputs)
conv_2 = MaxPooling2D((3, 3), strides=(2,2))(conv_1)
conv_2 = crosschannelnormalization(name="convpool_1")(conv_2)
conv_2 = ZeroPadding2D((2,2))(conv_2)
conv_2 = merge([
Convolution2D(128,5,5,activation="relu",name='conv_2_'+str(i+1))(
splittensor(ratio_split=2,id_split=i)(conv_2)
) for i in range(2)], mode='concat',concat_axis=1,name="conv_2")
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1,1))(conv_3)
conv_3 = Convolution2D(384,3,3,activation='relu',name='conv_3')(conv_3)
conv_4 = ZeroPadding2D((1,1))(conv_3)
conv_4 = merge([
Convolution2D(192,3,3,activation="relu",name='conv_4_'+str(i+1))(
splittensor(ratio_split=2,id_split=i)(conv_4)
) for i in range(2)], mode='concat',concat_axis=1,name="conv_4")
conv_5 = ZeroPadding2D((1,1))(conv_4)
conv_5 = merge([
Convolution2D(128,3,3,activation="relu",name='conv_5_'+str(i+1))(
splittensor(ratio_split=2,id_split=i)(conv_5)
) for i in range(2)], mode='concat',concat_axis=1,name="conv_5")
dense_1 = MaxPooling2D((3, 3), strides=(2,2),name="convpool_5")(conv_5)
if heatmap:
dense_1 = Convolution2D(4096,6,6,activation="relu",name="dense_1")(dense_1)
dense_2 = Convolution2D(4096,1,1,activation="relu",name="dense_2")(dense_1)
dense_3 = Convolution2D(1000, 1,1,name="dense_3")(dense_2)
prediction = Softmax4D(axis=1,name="softmax")(dense_3)
else:
dense_1 = Flatten(name="flatten")(dense_1)
dense_1 = Dense(4096, activation='relu',name='dense_1')(dense_1)
dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu',name='dense_2')(dense_2)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(2,name='dense_3')(dense_3)
prediction = Activation("softmax",name="softmax")(dense_3)
model = Model(input=inputs, output=prediction)
def my_model_def(input_dim=(3, 128, 128), nb_classes=3):
# MOdel Definition
inputs = Input(shape=input_dim)
norma1 = BatchNormalization(axis=1)(inputs)
conv1 = Conv2D(96, (7, 7),
padding='same',
activation='relu',
strides=(4, 4),
kernel_initializer='he_normal')(norma1)
pool1 = MaxPool2D(pool_size=(3, 3), strides=(2, 2), padding='same')(conv1)
conv2 = Concatenate(axis=1)([
Conv2D(128, (5, 5),
activation="relu",
strides=(2, 2),
padding='same',
kernel_initializer='he_normal',
name='conv2_' + str(i + 1))(splittensor(ratio_split=2,
id_split=i)(pool1))
for i in range(2)
])
norma2 = BatchNormalization(axis=1)(conv2)
conv2 = Conv2D(128, (5, 5),
activation="relu",
padding='same',
strides=(2, 2),
kernel_initializer='he_normal')(norma2)
pool2 = MaxPool2D(pool_size=(3, 3), strides=(2, 2), padding='same')(conv2)
norma3 = BatchNormalization(axis=1)(pool2)
conv3 = Conv2D(256, (3, 3),
strides=(2, 2),
padding='same',
activation='relu',
kernel_initializer='he_normal')(norma3)
conv3 = Dropout(rate=0.5, seed=10)(conv3)
conv3 = MaxPool2D(pool_size=(3, 3), strides=(2, 2), padding='same')(conv3)
dense1 = Flatten(name="flatten")(conv3)
dense1 = Dense(4096, activation='relu',
kernel_initializer='he_normal')(dense1)
dense1 = Dropout(0.5)(dense1)
dense2 = Dense(4096, activation='relu',
kernel_initializer='he_normal')(dense1)
dense2 = Dropout(rate=0.5)(dense2)
dense3 = Dense(nb_classes,
name='dense3_new',
kernel_initializer='he_normal')(dense2)
prediction = Activation("softmax")(dense3)
model = Model(inputs=inputs, outputs=prediction)
return model
def AlexNet(weights_path=None,
inputs1=None,
inputs2=None,
id='',
trainable=True):
conv_1_op = Convolution2D(96,
11,
11,
subsample=(4, 4),
activation='relu',
name=id + 'conv_1',
trainable=trainable)
conv_1 = conv_1_op(inputs1)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name=id + "convpool_1")(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2_op_a = Convolution2D(128,
5,
5,
activation="relu",
name=id + 'conv_2_1',
trainable=trainable)
conv_2_op_b = Convolution2D(128,
5,
5,
activation="relu",
name=id + 'conv_2_2',
trainable=trainable)
conv_2 = merge([
conv_2_op_a(splittensor(ratio_split=2, id_split=0)(conv_2)),
conv_2_op_b(splittensor(ratio_split=2, id_split=1)(conv_2))
],
mode='concat',
concat_axis=1,
name=id + "conv_2")
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3_op = Convolution2D(384,
3,
3,
activation='relu',
name=id + 'conv_3',
trainable=trainable)
conv_3 = conv_3_op(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4_op_a = Convolution2D(192,
3,
3,
activation="relu",
name=id + 'conv_4_1',
trainable=trainable)
conv_4_op_b = Convolution2D(192,
3,
3,
activation="relu",
name=id + 'conv_4_2',
trainable=trainable)
conv_4 = merge([
conv_4_op_a(splittensor(ratio_split=2, id_split=0)(conv_4)),
conv_4_op_b(splittensor(ratio_split=2, id_split=1)(conv_4))
],
mode='concat',
concat_axis=1,
name=id + "conv_4")
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5_op_a = Convolution2D(128,
3,
3,
activation="relu",
name=id + 'conv_5_1',
trainable=trainable)
conv_5_op_b = Convolution2D(128,
3,
3,
activation="relu",
name=id + 'conv_5_2',
trainable=trainable)
conv_5 = merge([
conv_5_op_a(splittensor(ratio_split=2, id_split=0)(conv_5)),
conv_5_op_b(splittensor(ratio_split=2, id_split=1)(conv_5))
],
mode='concat',
concat_axis=1,
name=id + "conv_5")
dense_1 = MaxPooling2D((3, 3),
strides=(2, 2),
name=id + "convpool_5",
trainable=trainable)(conv_5)
dense_1 = Flatten(name=id + "flatten")(dense_1)
dense_1_op = Dense(4096,
activation='relu',
name=id + 'dense_1',
trainable=trainable)
dense_1_a = dense_1_op(dense_1)
dense_2_op = Dense(4096,
activation='relu',
name=id + 'dense_2_gel',
trainable=True)
dense_2 = dense_2_op(dense_1_a)
dense_3_op = Dense(1000,
name=id + 'dense_3_100',
activation='sigmoid',
trainable=trainable)
dense_3 = dense_3_op(dense_2)
model = Model(input=inputs1, output=dense_3)
if weights_path:
model.load_weights(weights_path, by_name=True)
print "model loaded"
#model.save_weights('../../data/weights/alexnet_weights_finetune.h5')
super_op = Dense(8, name=id + 'supervision', activation='softmax')
dense_4_1 = dense_2
supervision_1 = super_op(dense_4_1)
# FOR INPUTS2
id = '2'
conv_1 = conv_1_op(inputs2)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name=id + "convpool_1")(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
conv_2_op_a(splittensor(ratio_split=2, id_split=0)(conv_2)),
conv_2_op_b(splittensor(ratio_split=2, id_split=1)(conv_2))
],
mode='concat',
concat_axis=1,
name=id + "conv_2")
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = conv_3_op(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
conv_4_op_a(splittensor(ratio_split=2, id_split=0)(conv_4)),
conv_4_op_b(splittensor(ratio_split=2, id_split=1)(conv_4))
],
mode='concat',
concat_axis=1,
name=id + "conv_4")
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
conv_5_op_a(splittensor(ratio_split=2, id_split=0)(conv_5)),
conv_5_op_b(splittensor(ratio_split=2, id_split=1)(conv_5))
],
mode='concat',
concat_axis=1,
name=id + "conv_5")
dense_1 = MaxPooling2D((3, 3), strides=(2, 2),
name=id + "convpool_5")(conv_5)
dense_1 = Flatten(name=id + "flatten")(dense_1)
dense_1_1 = dense_1_op(dense_1)
dense_2 = dense_2_op(dense_1_1)
dense_4_2 = dense_2
super_op = Dense(8, name=id + 'supervision', activation='softmax')
supervision_2 = super_op(dense_4_2)
single_model = Model(input=inputs1, output=dense_4_1)
return dense_4_1, dense_4_2, supervision_1, supervision_2
def get_alex(params, weights=None):
#inputs = Input(shape=(params['nb_channels'], params['im_size'], params['im_size']))
inputs = Input(shape=(3, params['im_size'], params['im_size']))
#model.add(Convolution2D(96, (3, 3), activation='relu', padding='same', name='conv2', strides=(1, 1)))
conv_1 = Convolution2D(96, (11, 11),
strides=(4, 4),
activation='relu',
name='conv_1')(inputs)
conv_2 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name='convpool_1')(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
Convolution2D(
128, 5, 5, activation='relu', name='conv_2_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_2))
for i in range(2)
],
mode='concat',
concat_axis=1,
name='conv_2')
conv_3 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Convolution2D(384, 3, 3, activation='relu', name='conv_3')(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
Convolution2D(
192, 3, 3, activation='relu', name='conv_4_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_4))
for i in range(2)
],
mode='concat',
concat_axis=1,
name='conv_4')
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
Convolution2D(
128, 3, 3, activation='relu', name='conv_5_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_5))
for i in range(2)
],
mode='concat',
concat_axis=1,
name='conv_5')
dense_1 = MaxPooling2D(pool_size=(3, 3), strides=(2, 2),
name='convpool_5')(conv_5)
dense_1 = Flatten(name='flatten')(dense_1)
dense_1 = Dense(4096, activation='relu', name='dense_1')(dense_1)
dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu', name='dense_2')(dense_2)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(1000, name='dense_3')(dense_3)
prediction = Activation('softmax', name='softmax')(dense_3)
model = Model(input=inputs, output=prediction)
if weights:
model.load_weights(weights)
return model
def get_alexnet(input_shape, nb_classes, mean_flag):
# code adapted from https://github.com/heuritech/convnets-keras
inputs = Input(shape=input_shape)
if mean_flag:
mean_subtraction = Lambda(mean_subtract,
name='mean_subtraction')(inputs)
conv_1 = Convolution2D(96,
11,
11,
subsample=(4, 4),
activation='relu',
name='conv_1',
init='he_normal')(mean_subtraction)
else:
conv_1 = Convolution2D(96,
11,
11,
subsample=(4, 4),
activation='relu',
name='conv_1',
init='he_normal')(inputs)
conv_2 = crosschannelnormalization(name="convpool_1")(conv_1)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
Convolution2D(128,
5,
5,
activation="relu",
init='he_normal',
name='conv_2_' + str(i + 1))(splittensor(
ratio_split=2, id_split=i)(conv_2)) for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_2")
conv_3 = crosschannelnormalization()(conv_2)
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Convolution2D(384,
3,
3,
activation='relu',
name='conv_3',
init='he_normal')(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
Convolution2D(192,
3,
3,
activation="relu",
init='he_normal',
name='conv_4_' + str(i + 1))(splittensor(
ratio_split=2, id_split=i)(conv_4)) for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_4")
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
Convolution2D(128,
3,
3,
activation="relu",
init='he_normal',
name='conv_5_' + str(i + 1))(splittensor(
ratio_split=2, id_split=i)(conv_5)) for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_5")
dense_1 = MaxPooling2D((3, 3), strides=(2, 2), name="convpool_5")(conv_5)
dense_1 = Flatten(name="flatten")(dense_1)
dense_1 = Dense(4096, activation='relu', name='dense_1',
init='he_normal')(dense_1)
dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu', name='dense_2',
init='he_normal')(dense_2)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(nb_classes, name='dense_3_new', init='he_normal')(dense_3)
prediction = Activation("softmax", name="softmax")(dense_3)
alexnet = Model(input=inputs, output=prediction)
return alexnet
def alexnet_model(img_rows, img_cols, channels=1, num_classes=None, use_pretraining=True,
pretrained_weights_path=None, fine_tuning_method=END_TO_END_FINE_TUNING, optimizer=None, loss=None,
weights_path=None):
inputs = Input(shape=(channels, img_rows, img_cols))
conv_1 = Convolution2D(96, 11, 11, subsample=(4, 4), activation='relu',
name='conv_1')(inputs)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name='convpool_1')(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
Convolution2D(128, 5, 5, activation='relu', name='conv_2_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_2)
) for i in range(2)], mode='concat', concat_axis=1, name='conv_2')
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Convolution2D(384, 3, 3, activation='relu', name='conv_3')(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
Convolution2D(192, 3, 3, activation='relu', name='conv_4_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_4)
) for i in range(2)], mode='concat', concat_axis=1, name='conv_4')
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
Convolution2D(128, 3, 3, activation='relu', name='conv_5_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_5)
) for i in range(2)], mode='concat', concat_axis=1, name='conv_5')
dense_1 = MaxPooling2D((3, 3), strides=(2, 2), name='convpool_5')(conv_5)
dense_1 = Flatten(name='flatten')(dense_1)
dense_1 = Dense(4096, activation='relu', name='dense_1')(dense_1)
dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu', name='dense_2')(dense_2)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(1000, name='dense_3')(dense_3)
prediction = Activation('softmax', name='softmax')(dense_3)
model = Model(input=inputs, output=prediction)
# Load from ImageNet pretrained weights only if creating a new model
if weights_path is None:
if use_pretraining:
if pretrained_weights_path:
model.load_weights(pretrained_weights_path)
else:
raise Exception('use_pretraining is true but pretrained_weights_path is not specified!')
# Removing the final dense_3 layer and adding the layers with correct classification size
model.layers.pop()
model.layers.pop()
last = model.layers[-1].output
last = Dense(num_classes, name='dense_3')(last)
prediction = Activation('softmax', name='softmax')(last)
model = Model(model.input, prediction)
if fine_tuning_method == FREEZE_INITIAL_LAYERS:
print(get_time_string() + 'Freezing initial 20 layers of the network..')
for layer in model.layers[:20]:
layer.trainable = False
# Load from previously saved weights
if weights_path is not None:
model.load_weights(weights_path)
if optimizer == 'sgd':
optimizer = SGD(lr=1e-3, decay=1e-6, momentum=0.9, nesterov=True)
if loss == 'l1':
loss = get_l1_loss
print(get_time_string() + 'Compiling the model..')
model.compile(optimizer=optimizer, loss=loss)
print('Number of layers: ' + str(len(model.layers)))
print(model.summary())
return model
def alexnet_branches(self, train=True):
p = 0.6
if not train:
p = 0.0
inputs = Input(shape=(3, 227, 227))
self.inputs = inputs
conv_1 = Convolution2D(96,
11,
11,
subsample=(4, 4),
activation='relu',
name='conv_1')(inputs)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name="convpool_1")(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
Convolution2D(
128, 5, 5, activation="relu", name='conv_2_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_2))
for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_2")
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Convolution2D(384, 3, 3, activation='relu',
name='conv_3')(conv_3)
conv_4_0 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
Convolution2D(
192, 3, 3, activation="relu", name='conv_4_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_4_0))
for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_4")
conv_5_0 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
Convolution2D(
128, 3, 3, activation="relu", name='conv_5_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_5_0))
for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_5")
finalmax = MaxPooling2D((3, 3), strides=(2, 2),
name="convpool_5")(conv_5)
dense_1_1 = Flatten(name="flatten")(finalmax)
dense_1 = Dense(4096, activation='relu')(dense_1_1)
dense_2 = Dropout(p)(dense_1)
dense_2 = Dense(4096, activation='relu')(dense_2)
dense_3 = Dropout(p)(dense_2)
dense_4 = Dense(1000)(dense_3)
prediction = Activation("softmax", name="softmax")(dense_4)
model = Model(input=inputs, output=prediction)
if train:
model.load_weights('weights/alexnet_weights.h5')
for i in range(9):
model.layers.pop()
#up1 = UpSampling2D((15,15),dim_ordering='th')(conv_5_0)
branches = []
for i in range(21):
branches.append(self.branchout(conv_5_0, 'uniform'))
prediction = merge(branches, mode='concat', concat_axis=1)
#prediction = Activation('softmax')(prediction)
prediction = Lambda(self.maxnorm,
output_shape=self.maxnormshape)(prediction)
model = Model(input=inputs, output=prediction)
if not train:
model.load_weights('newweights/alexvoc_branchavg.h5')
self.model = model
print(model.summary())
self.model.compile(loss='categorical_crossentropy',
optimizer=sgd,
metrics=['mse'])
return self.model
def alexnetcam(self, train=True):
p = 0.6
if not train:
p = 0.0
inputs = Input(shape=(3, 227, 227))
self.inputs = inputs
conv_1 = Convolution2D(96,
11,
11,
subsample=(4, 4),
activation='relu',
name='conv_1')(inputs)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name="convpool_1")(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
Convolution2D(
128, 5, 5, activation="relu", name='conv_2_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_2))
for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_2")
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Convolution2D(384, 3, 3, activation='relu',
name='conv_3')(conv_3)
conv_4_0 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
Convolution2D(
192, 3, 3, activation="relu", name='conv_4_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_4_0))
for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_4")
conv_5_0 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
Convolution2D(
128, 3, 3, activation="relu", name='conv_5_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_5_0))
for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_5")
finalmax = MaxPooling2D((3, 3), strides=(2, 2),
name="convpool_5")(conv_5)
dense_1_1 = Flatten(name="flatten")(finalmax)
dense_1 = Dense(4096, activation='relu')(dense_1_1)
dense_2 = Dropout(p)(dense_1)
dense_2 = Dense(4096, activation='relu')(dense_2)
dense_3 = Dropout(p)(dense_2)
dense_4 = Dense(1000)(dense_3)
prediction = Activation("softmax", name="softmax")(dense_4)
model = Model(input=inputs, output=prediction)
if train:
model.load_weights('weights/alexnet_weights.h5')
for i in range(9):
model.layers.pop()
for layer in model.layers:
layer.trainable = False
conv_5 = Convolution2D(256,
3,
3,
activation="relu",
border_mode='same',
init='glorot_uniform')(conv_5_0)
#conv_5 = Convolution2D(32,3,3,activation="relu",border_mode='same',init='glorot_uniform')(conv_5_0)
finalmax = AveragePooling2D((15, 15))(conv_5)
dense_3 = Flatten()(finalmax)
dense_4 = Dense(20, init='glorot_uniform')(dense_3)
prediction = Activation("sigmoid", name="softmax")(dense_4)
model = Model(input=inputs, output=prediction)
if not train:
model.load_weights('newweights/alexvoc_cam.h5')
self.model = model
print(model.summary())
self.model.compile(loss='categorical_crossentropy',
optimizer=sgd,
metrics=['mse'])
return self.model
def AlexNet_100(weights_path=None, heatmap=False):
if heatmap:
inputs = Input(shape=(None, None, 3))
else:
inputs = Input(shape=(100, 100, 3))
conv_1 = Conv2D(96, (11, 11),
strides=(4, 4),
activation='relu',
name='conv_1')(inputs)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name='convpool_1')(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
Conv2D(128, (5, 5), activation='relu', name='conv_2_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_2)) for i in range(2)
],
mode='concat',
concat_axis=-1,
name='conv_2')
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Conv2D(384, (3, 3), activation='relu', name='conv_3')(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
Conv2D(192, (3, 3), activation='relu', name='conv_4_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_4)) for i in range(2)
],
mode='concat',
concat_axis=-1,
name='conv_4')
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
Conv2D(128, (3, 3), activation='relu', name='conv_5_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_5)) for i in range(2)
],
mode='concat',
concat_axis=-1,
name='conv_5')
dense_1 = MaxPooling2D((3, 3), strides=(2, 2), name='convpool_5')(conv_5)
if heatmap:
dense_1 = Conv2D(4096, (6, 6), activation='relu',
name='dense_1')(dense_1)
dense_2 = Conv2D(4096, (1, 1), activation='relu',
name='dense_2')(dense_1)
dense_3 = Conv2D(1000, (1, 1), name='dense_3')(dense_2)
prediction = Softmax4D(axis=1, name='softmax')(dense_3)
else:
dense_1 = Flatten(name='flatten')(dense_1)
dense_1 = Dense(4096, activation='relu', name='dense_1')(dense_1)
dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu', name='dense_2')(dense_2)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(1000, name='dense_3')(dense_3)
prediction = Activation('softmax', name='softmax')(dense_3)
model = Model(input=inputs, output=prediction)
if weights_path:
model.load_weights(weights_path)
lastlayer = model.layers[-3].output
if heatmap:
dense_3 = Conv2D(1000, (1, 1), activation='relu',
name='dense_3')(lastlayer)
dense_4 = Conv2D(2, (1, 1), name='dense_4')(dense_3)
prediction = Softmax4D(axis=1, name='softmax')(dense_4)
else:
dense_3 = Dense(1000, activation='relu', name='dense_3')(lastlayer)
dense_4 = Dropout(0.5)(dense_3)
dense_4 = Dense(2, name='dense_4')(dense_4)
prediction = Activation('softmax', name='softmax')(dense_4)
model = Model(inputs=inputs, outputs=prediction)
return model
def AlexNet(weights_path=None,
inputs=None,
id='1',
trainable=True,
finetune=False,
last_layer=None):
if inputs == None:
inputs = Input(shape=(3, 227, 227))
conv_1 = Convolution2D(96,
11,
11,
subsample=(4, 4),
activation='relu',
name=id + 'conv_1',
trainable=trainable)(inputs)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name=id + "convpool_1")(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
Convolution2D(128,
5,
5,
activation="relu",
name=id + 'conv_2_' + str(i + 1),
trainable=trainable)(splittensor(ratio_split=2,
id_split=i)(conv_2))
for i in range(2)
],
mode='concat',
concat_axis=1,
name=id + "conv_2")
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Convolution2D(384,
3,
3,
activation='relu',
name=id + 'conv_3',
trainable=trainable)(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
Convolution2D(192,
3,
3,
activation="relu",
name=id + 'conv_4_' + str(i + 1),
trainable=trainable)(splittensor(ratio_split=2,
id_split=i)(conv_4))
for i in range(2)
],
mode='concat',
concat_axis=1,
name=id + "conv_4")
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
Convolution2D(128,
3,
3,
activation="relu",
name=id + 'conv_5_' + str(i + 1),
trainable=trainable)(splittensor(ratio_split=2,
id_split=i)(conv_5))
for i in range(2)
],
mode='concat',
concat_axis=1,
name=id + "conv_5")
dense_1 = MaxPooling2D((3, 3),
strides=(2, 2),
name=id + "convpool_5",
trainable=trainable)(conv_5)
dense_1 = Flatten(name=id + "flatten")(dense_1)
dense_1 = Dense(4096,
activation='relu',
name=id + 'dense_1',
trainable=trainable)(dense_1)
dense_2 = Dense(4096, name=id + 'dense_4', trainable=True)(dense_1)
dense_2_act = Activation('relu')(dense_2)
dense_2_alex = Dense(4096,
name=id + 'dense_4',
trainable=True,
W_regularizer=l2(0))(dense_1)
dense_2_alex_act = Activation('relu',
name=id + 'dense_4_embedding')(dense_2_alex)
dense_3_alex = Dense(1000,
name=id + 'dense_3_100',
activation='sigmoid',
trainable=trainable)(dense_2_alex_act)
if last_layer == None:
last_layer = Dense(8,
name=id + 'dense_3_100',
activation='softmax',
trainable=True,
W_regularizer=l2(0))
dense_3 = last_layer(dense_2_alex_act)
else:
dense_3 = last_layer(dense_2_alex_act)
if finetune:
model_classification = Model(input=inputs, output=dense_3)
output = dense_2
else:
model_classification = Model(input=inputs, output=dense_3_alex)
output = dense_2_alex
if weights_path:
model_classification.load_weights(weights_path) #, by_name=True)
print "model loaded"
#model.save_weights('../../data/weights/alexnet_weights_finetune.h5')
return dense_3, dense_2_alex_act, last_layer
def AlexNet_base(weights_path=None, input_tensor=None):
if input_tensor is not None:
inputs = input_tensor
else:
inputs = Input(shape=(3, 227, 227))
mean_subtraction = Lambda(mean_subtract, name='mean_subtraction')(inputs)
conv_1 = Convolution2D(96,
11,
11,
subsample=(4, 4),
activation='relu',
name='conv_1')(mean_subtraction)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name="convpool_1")(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
Convolution2D(
128, 5, 5, activation="relu", name='conv_2_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_2))
for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_2")
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Convolution2D(384, 3, 3, activation='relu', name='conv_3')(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
Convolution2D(
192, 3, 3, activation="relu", name='conv_4_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_4))
for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_4")
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
Convolution2D(
128, 3, 3, activation="relu", name='conv_5_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_5))
for i in range(2)
],
mode='concat',
concat_axis=1,
name="conv_5")
dense_1 = MaxPooling2D((3, 3), strides=(2, 2), name="convpool_5")(conv_5)
dense_1 = Flatten(name="flatten")(dense_1)
dense_1 = Dense(4096, activation='relu', name='dense_1')(dense_1)
# dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu', name='dense_2')(dense_1)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(1000, name='dense_3')(dense_3)
prediction = Activation("softmax", name="softmax")(dense_3)
model = Model(inputs=inputs, outputs=prediction)
if weights_path:
model.load_weights(weights_path)
x = model.layers[29].output
new_model = Model(inputs=inputs, outputs=x)
return new_model
def my_model_def(input_dim=(3, 128, 128),
nb_classes=3,
drop_out=0.5,
fmap_size=[128, 500, 4096]):
# MOdel Definition
inputs = Input(shape=input_dim, name='input')
conv1 = Conv2D(96, (7, 7),
padding='same',
activation='relu',
strides=(1, 1),
kernel_initializer='he_normal',
name='conv1')(inputs)
pool1 = MaxPool2D(pool_size=(3, 3),
strides=(2, 2),
padding='same',
name='pool1')(conv1)
concat2 = Concatenate(axis=1, name='concat1')([
Conv2D(48, (5, 5),
activation="relu",
strides=(2, 2),
padding='same',
kernel_initializer='he_normal',
name='conv2_' + str(i + 1))(splittensor(ratio_split=2,
id_split=i)(pool1))
for i in range(2)
])
pool2_1 = MaxPool2D(pool_size=(3, 3),
strides=(2, 2),
padding='same',
name='pool2_1')(concat2)
conv2 = Conv2D(384, (5, 5),
activation="relu",
padding='same',
strides=(2, 2),
kernel_initializer='he_normal',
name='conv2_3')(pool2_1)
pool2_2 = MaxPool2D(pool_size=(3, 3),
strides=(2, 2),
padding='same',
name='pool2_2')(conv2)
concat3 = Concatenate(axis=1, name='concat3')([
Conv2D(fmap_size[0], (5, 5),
activation="relu",
strides=(2, 2),
padding='same',
kernel_initializer='he_normal',
name='conv3_' + str(i + 1))(splittensor(ratio_split=2,
id_split=i)(pool2_2))
for i in range(2)
])
pool3 = MaxPool2D(pool_size=(3, 3),
strides=(2, 2),
padding='same',
name='pool3')(concat3)
# conv3 = Conv2D(250, (3, 3), strides=(2, 2), padding='same', activation='relu',
# kernel_initializer='he_normal', name='conv3_1')(pool2_2)
# pool3 = MaxPool2D(pool_size=(3, 3), strides=(1, 1), padding='same', name='conv3_2')(conv3)
conv4 = Conv2D(fmap_size[1], (3, 3),
strides=(2, 2),
padding='same',
activation='relu',
kernel_initializer='he_normal',
name='conv4_1')(pool3)
pool4 = MaxPool2D(pool_size=(3, 3),
strides=(1, 1),
padding='same',
name='pool4')(conv4)
drop4 = Dropout(rate=drop_out, seed=10, name='drop4')(pool4)
flatten5 = Flatten(name="flatten")(drop4)
dense5 = Dense(fmap_size[2],
activation='relu',
kernel_initializer='he_normal',
name='dense5')(flatten5)
drop5 = Dropout(rate=drop_out - 0.3)(dense5)
dense6 = Dense(fmap_size[2],
activation='relu',
kernel_initializer='he_normal',
name='dense6_1')(drop5)
dense6 = Dense(nb_classes, kernel_initializer='he_normal',
name='dense6_2')(dense6)
prediction = Activation("softmax", name='softmax_act')(dense6)
model = Model(inputs=inputs, outputs=prediction)
return model
def AlexNet():
inputs = Input(shape=(3, 200, 200))
conv_1 = Convolution2D(96,
11,
11,
subsample=(4, 4),
activation='relu',
name='conv_1')(inputs)
conv_2 = LRN2D(name='convpool_1')(conv_1)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
Convolution2D(128,
5,
5,
activation='relu',
init='glorot_uniform',
bias=True,
name='conv_2_' + str(i + 1))(splittensor(
ratio_split=2, id_split=i)(conv_2)) for i in range(2)
],
mode='concat',
concat_axis=1,
name='conv_2')
conv_3 = LRN2D()(conv_2)
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Convolution2D(384,
3,
3,
activation='relu',
init='glorot_uniform',
bias=True,
name='conv_3')(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
Convolution2D(192,
3,
3,
activation='relu',
bias=True,
init='glorot_uniform',
name='conv_4_' + str(i + 1))(splittensor(
ratio_split=2, id_split=i)(conv_4)) for i in range(2)
],
mode='concat',
concat_axis=1,
name='conv_4')
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
Convolution2D(128,
3,
3,
activation='relu',
init='glorot_uniform',
bias=True,
name='conv_5_' + str(i + 1))(splittensor(
ratio_split=2, id_split=i)(conv_5)) for i in range(2)
],
mode='concat',
concat_axis=1,
name='conv_5')
dense_1 = MaxPooling2D((3, 3), strides=(2, 2), name='convpool_5')(conv_5)
dense_1 = Flatten(name='flatten')(dense_1)
dense_1 = Dense(4096, activation='relu', name='dense_1')(dense_1)
dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu', name='dense_2')(dense_2)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(16000, name='dense_3')(dense_3)
reshape = Reshape((400, 40))(dense_3)
prediction = Activation('softmax', name='softmax')(reshape)
model = Model(input=inputs, output=prediction)
return model
def ZFNet(weights_path=None, classes=1000):
inputs = Input(shape=(3, 227, 227))
conv_1 = Convolution2D(96, (7, 7),
strides=(2, 2),
kernel_initializer='random_uniform',
bias_initializer='zeros',
activation='relu',
name='conv_1a')(inputs)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name="convpool_1")(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = concatenate([
Convolution2D(128, (5, 5),
strides=(2, 2),
activation="relu",
name='conv_2' + str(i + 1))(splittensor(
ratio_split=2, id_split=i)(conv_2)) for i in range(2)
],
axis=1,
name="conv_2")
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Convolution2D(384, (3, 3), activation='relu',
name='conv_3')(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = concatenate([
Convolution2D(
192, (3, 3), activation="relu", name='conv_4_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_4))
for i in range(2)
],
axis=1,
name="conv_4")
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = concatenate([
Convolution2D(
128, (3, 3), activation="relu", name='conv_5_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_5))
for i in range(2)
],
axis=1,
name="conv_5")
dense_1 = MaxPooling2D((3, 3), strides=(2, 2), name="convpool_5")(conv_5)
dense_1 = Flatten(name="flatten")(dense_1)
dense_1 = Dense(4096, activation='relu', name='dense_1')(dense_1)
dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu', name='dense_2')(dense_2)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(classes,
kernel_initializer='random_uniform',
bias_initializer='zeros',
name='dense_3a')(dense_3)
prediction = Activation("softmax", name="softmaxa")(dense_3)
model = Model(input=inputs, output=prediction, name='zfnet')
if weights_path:
model.load_weights(weights_path, by_name=True)
sgd = SGD(lr=1e-4, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd,
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
