def __init__(self):
reg = tf.contrib.layers.l2_regularizer(1e-3)
model = Sequential()
model.add(Lambda(lambda x: x / 255.0 - 0.5, input_shape=(114, 52, 3)))
model.add(
Conv2D(filters=24,
kernel_size=5,
strides=1,
padding='same',
activation='relu',
kernel_regularizer=reg))
model.add(
Conv2D(filters=36,
kernel_size=5,
strides=1,
padding='same',
activation='relu',
kernel_regularizer=reg))
model.add(
Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
kernel_regularizer=reg))
model.add(Flatten())
model.add(Dense(50))
model.add(Dense(10))
model.add(Dense(6, activation='softmax', name=LOGITS))
model.load_weights('./light_classification/weights.h5')
self.model = model
def create_VGG16(num_fc_neurons,
dropout_rate,
num_classes=24,
top_model_weights_path=None,
img_height=224,
img_width=224,
include_loc='all',
activation='softmax'):
# load pre-trained convolutional model
base_model = VGG16(weights='imagenet',
include_top=False,
input_shape=get_input_shape(img_height, img_width))
# build a classifier model to put on top of the convolutional model
top_model = Sequential()
top_model.add(Flatten(input_shape=base_model.output_shape[1:]))
for i in range(6, 8):
top_model.add(Dense(num_fc_neurons, name='fc' + str(i)))
#top_model.add(BatchNormalization(axis=1, name='fc'+str(i)+'_bn'))
top_model.add(Activation('relu', name='fc' + str(i) + '_ac'))
top_model.add(Dropout(dropout_rate))
top_model.add(Dense(num_classes, activation=activation,
name='predictions'))
if top_model_weights_path != None:
top_model.load_weights(top_model_weights_path)
if include_loc == 'base':
model = base_model
elif include_loc == 'top':
model = top_model
elif include_loc == 'all': # add the model on top of the convolutional base
model = Model(inputs=base_model.input,
outputs=top_model(base_model.output))
else:
raise ValueError('Only "base", "top" and "all" can be included.')
return model
def make_vgg16(top=True, weights='imagenet', weight_path=''):
"""
Args: top determines whether to include dense layers at the top. Weights determines
whether to use imagenet weights or pre-trained weights, in which case the filepath
must be specified via weight_path.
Creates a convolutional neural network following the VGG16 structure. There are two options:
the original structure with fully connected layers at the end, or a slimmed down model where the
FC layers are replaced by a global average pooling layer. The latter has far fewer weights.
Returns the model.
"""
model = Sequential()
model.add(ZeroPadding2D((1, 1), input_shape=(224, 224, 3)))
model.add(Convolution2D(64, (3, 3), activation='relu'))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(128, (3, 3), activation='relu'))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(128, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(256, (3, 3), activation='relu'))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(256, (3, 3), activation='relu'))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(256, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(512, (3, 3), activation='relu'))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(512, (3, 3), activation='relu'))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(512, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(512, (3, 3), activation='relu'))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(512, (3, 3), activation='relu'))
model.add(ZeroPadding2D((1, 1)))
model.add(Convolution2D(512, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2), strides=(2, 2)))
if weights == 'imagenet':
model.add(Flatten())
model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(1000, activation='softmax'))
model.load_weights(r'D:/drivers/vgg16_weights.h5')
if top == False:
for i in range(7):
model.layers.pop()
model.outputs = [model.layers[-1].output]
model.layers[-1].outbound_nodes = []
model.add(Convolution2D(1024, (3, 3), activation='relu'))
model.add(AveragePooling2D((14, 14), padding='same'))
model.add(Flatten())
model.add(Dense(10, activation='softmax'))
else:
model.layers.pop()
model.outputs = [model.layers[-1].output]
model.layers[-1].outbound_nodes = []
model.add(Dense(10, activation='softmax'))
elif weights == 'trained':
model.add(Flatten())
model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(10, activation='softmax'))
if top == False:
for i in range(7):
model.layers.pop()
model.outputs = [model.layers[-1].output]
model.layers[-1].outbound_nodes = []
model.add(Convolution2D(1024, (3, 3), activation='relu'))
model.add(AveragePooling2D((14, 14), padding='same'))
model.add(Flatten())
model.add(Dense(10, activation='softmax'))
model.load_weights(weight_path)
sgd = optimizers.SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd,
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
return model
"""
If you need to load weights into a *different* architecture (with some layers in common), for instance for fine-tuning or transfer-learning, you can load weights by *layer name*:
"""
model_yaml.load_weights('to_delete_weights.h5', by_name=True)
# make sure they share the same weights: total 202 parameters
(model_json.get_weights()[0] == model_yaml.get_weights()[0]).sum()
(model_json.get_weights()[1] == model_yaml.get_weights()[1]).sum()
"""
For example
Assume original model looks like this:
"""
model1 = Sequential()
model1.add(Dense(2, input_dim=3, name='dense_1'))
model1.add(Dense(3, name='dense_2'))
model1.save_weights("weights1.h5")
# check out the weights
model1.get_weights()
# new model
model2 = Sequential()
model2.add(Dense(2, input_dim=3, name='dense_1')) # will be loaded
model2.add(Dense(10, name='new_dense')) # will not be loaded
# load weights from first model; will only affect the first layer, dense_1.
model2.load_weights("weights1.h5", by_name=True)
# check out the weights
model2.get_weights()