def build_network(self):
"""
Build the convnet.
Input is 48x48
3072 nodes in fully connected layer
"""
# Real-time data preprocessing
img_prep = tflearn.ImagePreprocessing()
img_prep.add_featurewise_zero_center(
per_channel=True, mean=[0.53990436, 0.4405486, 0.39328504])
# Real-time data augmentation
img_aug = tflearn.ImageAugmentation()
img_aug.add_random_flip_leftright()
img_aug.add_random_crop([49, 49], padding=4)
# Building Residual Network
self.network = tflearn.input_data(shape=[None, 49, 49, 3],
data_preprocessing=img_prep,
data_augmentation=img_aug)
self.network = tflearn.conv_2d(self.network,
16,
3,
regularizer='L2',
weight_decay=0.0001)
self.network = tflearn.resnext_block(self.network, 5, 16, 32)
self.network = tflearn.resnext_block(self.network,
1,
32,
32,
downsample=True)
self.network = tflearn.resnext_block(self.network, 4, 32, 32)
self.network = tflearn.resnext_block(self.network,
1,
64,
32,
downsample=True)
self.network = tflearn.resnext_block(self.network, 4, 64, 32)
self.network = tflearn.batch_normalization(self.network)
self.network = tflearn.activation(self.network, 'relu')
self.network = tflearn.global_avg_pool(self.network)
# Regression
self.network = tflearn.fully_connected(self.network,
11,
activation='softmax')
opt = tflearn.Momentum(0.1,
lr_decay=0.1,
decay_step=32000,
staircase=True)
self.network = tflearn.regression(self.network,
optimizer=opt,
loss='categorical_crossentropy')
# Training
self.model = tflearn.DNN(self.network,
checkpoint_path='Snapshots/model_resnext',
max_checkpoints=10,
tensorboard_verbose=0,
tensorboard_dir='Logs/',
clip_gradients=0.)
self.load_model()
def network(img_shape, name, LR):
# Building Residual Network
network = tflearn.input_data(shape=img_shape, name=name)
network = tflearn.conv_2d(network,
16,
3,
regularizer='L2',
weight_decay=0.0001)
network = tflearn.resnext_block(network, n, 16, 32)
network = tflearn.resnext_block(network, 1, 32, 32, downsample=True)
network = tflearn.resnext_block(network, n - 1, 32, 32)
network = tflearn.resnext_block(network, 1, 64, 32, downsample=True)
network = tflearn.resnext_block(network, n - 1, 64, 32)
network = tflearn.batch_normalization(network)
network = tflearn.activation(network, 'relu')
network = tflearn.global_avg_pool(network)
# Regression
network = tflearn.fully_connected(network, 2, activation='softmax')
opt = tflearn.Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True)
network = tflearn.regression(network,
optimizer=opt,
name='targets',
loss='categorical_crossentropy')
return network
def resnext_model(X, Y, testX, testY, n_epoch):
n_input = len(X[0])
n_classes = len(Y[0])
X = np.reshape(X, (-1, 1, n_input, 1))
testX = np.reshape(testX, (-1, 1, n_input, 1))
net = tflearn.input_data(shape=[None, 1, n_input, 1])
net = tflearn.conv_2d(net, 16, 3, regularizer='L2', weight_decay=0.0001, activation='selu')
net = tflearn.resnext_block(net, n, 16, 32, activation='selu')
net = tflearn.resnext_block(net, 1, 32, 32, downsample=True, activation='selu')
net = tflearn.resnext_block(net, n-1, 32, 32, activation='selu')
net = tflearn.resnext_block(net, 1, 64, 32, downsample=True, activation='selu')
net = tflearn.resnext_block(net, n-1, 64, 32, activation='selu')
net = tflearn.batch_normalization(net)
net = tflearn.activation(net, 'selu')
# net_p = tflearn.global_avg_pool(net, name='net_p')
net_p = global_max_pool(net, name='net_p')
# Regression
net = tflearn.fully_connected(net_p, 2048, activation='selu')
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, 256, activation='selu')
net = tflearn.dropout(net, 0.7)
net = tflearn.fully_connected(net, n_classes, activation='softmax')
# opt = tflearn.Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True)
net = tflearn.regression(net, optimizer='adam',
loss='categorical_crossentropy')
# Training
model = tflearn.DNN(net, checkpoint_path='../model/model_resnext_grid_adam',
max_checkpoints=10, tensorboard_verbose=0,
clip_gradients=0.)
# model.load('../model/model_resnext_grid_adam-70000')
model.fit(X, Y, n_epoch=n_epoch, validation_set=(testX[:], testY),
snapshot_epoch=False, snapshot_step=50000,
show_metric=True, batch_size=128, shuffle=True,
run_id='resnext')
prob = model.predict(testX)
# pdb.set_trace()
# get the hidden layer value after global average pooling
m2 = tflearn.DNN(net_p, session = model.session)
feature_train = list()
for i in range(4):
feature_train_temp = m2.predict(X[i*4000:(i+1)*4000])
feature_train += list(feature_train_temp)
feature_train_temp = m2.predict(X[16000:])
feature_train = np.array(feature_train+list(feature_train_temp))
feature_test = m2.predict(testX)
# pdb.set_trace()
return prob, feature_train, feature_test
def get_model(model_name):
# First we load the network
print("Setting up neural networks...")
n = 18
# Real-time data preprocessing
print("Doing preprocessing...")
img_prep = tflearn.ImagePreprocessing()
img_prep.add_featurewise_zero_center(
per_channel=True, mean=[0.573364, 0.44924123, 0.39455055])
# Real-time data augmentation
print("Building augmentation...")
img_aug = tflearn.ImageAugmentation()
img_aug.add_random_flip_leftright()
img_aug.add_random_crop([32, 32], padding=4)
#Build the model (for 32 x 32)
print("Shaping input data...")
net = tflearn.input_data(shape=[None, 32, 32, 3],
data_preprocessing=img_prep,
data_augmentation=img_aug)
net = tflearn.conv_2d(net, 16, 3, regularizer='L2', weight_decay=0.0001)
print("Carving Resnext blocks...")
net = tflearn.resnext_block(net, n, 16, 32)
net = tflearn.resnext_block(net, 1, 32, 32, downsample=True)
net = tflearn.resnext_block(net, n - 1, 32, 32)
net = tflearn.resnext_block(net, 1, 64, 32, downsample=True)
net = tflearn.resnext_block(net, n - 1, 64, 32)
print("Erroding Gradient...")
net = tflearn.batch_normalization(net)
net = tflearn.activation(net, 'relu')
net = tflearn.global_avg_pool(net)
net = tflearn.fully_connected(net, 8, activation='softmax')
opt = tflearn.Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True)
net = tflearn.regression(net,
optimizer=opt,
loss='categorical_crossentropy')
print("Structuring model...")
model = tflearn.DNN(net, tensorboard_verbose=0, clip_gradients=0.)
# Load the model from checkpoint
print("Loading the model...")
model.load(model_name)
return model
def inference(x):
n = 5
net = tflearn.conv_2d(x, 16, 3, regularizer='L2', weight_decay=0.0001)
net = tflearn.resnext_block(net, n, 16, 32)
net = tflearn.resnext_block(net, 1, 32, 32, downsample=True)
net = tflearn.resnext_block(net, n - 1, 32, 32)
net = tflearn.resnext_block(net, 1, 64, 32, downsample=True)
net = tflearn.resnext_block(net, n - 1, 64, 32)
net = tflearn.batch_normalization(net)
net = tflearn.activation(net, 'relu')
net = tflearn.global_avg_pool(net)
# Regression
net = tflearn.fully_connected(net, 5, activation=None)
logits = net
net = tflearn.activations.softmax(net)
return logits, net
def ResNext1(network):
# Residual blocks
# 32 layers: n=5, 56 layers: n=9, 110 layers: n=18
n = 5
network = conv_2d(network, 16, 3, regularizer='L2', weight_decay=0.0001)
network = resnext_block(network, n, 16, 32)
network = resnext_block(network, 1, 32, 32, downsample=True)
network = resnext_block(network, n - 1, 32, 32)
network = resnext_block(network, 1, 32, 32, downsample=True)
network = resnext_block(network, n - 1, 32, 32)
network = batch_normalization(network)
network = activation(network, 'relu')
network = global_avg_pool(network)
# Regression
network = fully_connected(network, output_dim, activation='softmax')
return network
def resnext(width, height, frame_count, lr, output=9, model_name = 'sentnet_color.model'):
net = input_data(shape=[None, width, height, 3], name='input')
net = tflearn.conv_2d(net, 16, 3, regularizer='L2', weight_decay=0.0001)
net = tflearn.layers.conv.resnext_block(net, n, 16, 32)
net = tflearn.resnext_block(net, 1, 32, 32, downsample=True)
net = tflearn.resnext_block(net, n-1, 32, 32)
net = tflearn.resnext_block(net, 1, 64, 32, downsample=True)
net = tflearn.resnext_block(net, n-1, 64, 32)
net = tflearn.batch_normalization(net)
net = tflearn.activation(net, 'relu')
net = tflearn.global_avg_pool(net)
# Regression
net = tflearn.fully_connected(net, output, activation='softmax')
opt = tflearn.Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True)
net = tflearn.regression(net, optimizer=opt,
loss='categorical_crossentropy')
model = tflearn.DNN(net,
max_checkpoints=0, tensorboard_verbose=0, tensorboard_dir='log')
return model
def resnext(width, height, frame_count, lr, output=9, model_name = 'sentnet_color.model'):
net = input_data(shape=[None, width, height, 3], name='input')
net = tflearn.conv_2d(net, 16, 3, regularizer='L2', weight_decay=0.0001)
net = tflearn.layers.conv.resnext_block(net, n, 16, 32)
net = tflearn.resnext_block(net, 1, 32, 32, downsample=True)
net = tflearn.resnext_block(net, n-1, 32, 32)
net = tflearn.resnext_block(net, 1, 64, 32, downsample=True)
net = tflearn.resnext_block(net, n-1, 64, 32)
net = tflearn.batch_normalization(net)
net = tflearn.activation(net, 'relu')
net = tflearn.global_avg_pool(net)
# Regression
net = tflearn.fully_connected(net, output, activation='softmax')
opt = tflearn.Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True)
net = tflearn.regression(net, optimizer=opt,
loss='categorical_crossentropy')
model = tflearn.DNN(net,
max_checkpoints=0, tensorboard_verbose=0, tensorboard_dir='log')
return model
def create_resnext_network( input_t, num_classes,
resnext_n, add_color_transfer):
if add_color_transfer:
x = cnn_utils.color_transform_layers(input_t)
else:
x = input_t
x = conv_2d(x, 16, 3, regularizer='L2', weight_decay=0.0001)
x = tflearn.resnext_block(x, resnext_n, 16, 32)
x = tflearn.resnext_block(x, 1, 32, 32, downsample=True)
x = tflearn.resnext_block(x, resnext_n - 1, 32, 32)
x = tflearn.resnext_block(x, 1, 64, 32, downsample=True)
x = tflearn.resnext_block(x, resnext_n - 1, 64, 32)
x = tflearn.batch_normalization(x)
x = tflearn.activation(x, 'relu')
x = tflearn.global_avg_pool(x)
x = fully_connected(x, num_classes, activation='softmax')
return x
def network(img_shape, name, LR):
img_prep = ImagePreprocessing()
img_prep.add_featurewise_zero_center()
img_prep.add_featurewise_stdnorm()
#
# # Real-time data augmentation
img_aug = ImageAugmentation()
img_aug.add_random_blur(sigma_max=3.0)
img_aug.add_random_flip_leftright()
img_aug.add_random_flip_updown()
img_aug.add_random_90degrees_rotation(rotations=[0, 2])
# Building Residual Network
network = tflearn.input_data(shape=img_shape,
name=name,
data_preprocessing=img_prep,
data_augmentation=img_aug)
network = tflearn.conv_2d(network,
16,
3,
regularizer='L2',
weight_decay=0.0001)
network = tflearn.resnext_block(network, n, 16, 32)
network = tflearn.resnext_block(network, 1, 32, 32, downsample=True)
network = tflearn.resnext_block(network, n - 1, 32, 32)
network = tflearn.resnext_block(network, 1, 64, 32, downsample=True)
network = tflearn.resnext_block(network, n - 1, 64, 32)
network = tflearn.batch_normalization(network)
network = tflearn.activation(network, 'relu')
network = tflearn.global_avg_pool(network)
# Regression
network = tflearn.fully_connected(network, 2, activation='softmax')
opt = tflearn.Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True)
network = tflearn.regression(network,
optimizer=opt,
name='targets',
loss='categorical_crossentropy')
return network
# Real-time data preprocessing
img_prep = tflearn.ImagePreprocessing()
img_prep.add_featurewise_zero_center(per_channel=True)
# Real-time data augmentation
img_aug = tflearn.ImageAugmentation()
img_aug.add_random_flip_leftright()
img_aug.add_random_crop([32, 32], padding=4)
# Building Residual Network
net = tflearn.input_data(shape=[None, 32, 32, 3],
data_preprocessing=img_prep,
data_augmentation=img_aug)
net = tflearn.conv_2d(net, 16, 3, regularizer='L2', weight_decay=0.0001)
# Residual blocks
net = tflearn.resnext_block(net, n, 16, 32)
net = tflearn.resnext_block(net, 1, 32, 32, downsample=True)
net = tflearn.resnext_block(net, n - 1, 32, 32)
net = tflearn.resnext_block(net, 1, 64, 32, downsample=True)
net = tflearn.resnext_block(net, n - 1, 64, 32)
net = tflearn.batch_normalization(net)
net = tflearn.activation(net, 'relu')
net = tflearn.global_avg_pool(net)
# Regression
net = tflearn.fully_connected(net, 10, activation='softmax')
mom = tflearn.Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True)
net = tflearn.regression(net, optimizer=mom, loss='categorical_crossentropy')
# Training
model = tflearn.DNN(net,
checkpoint_path='model_resnext_cifar10',
# Real-time data preprocessing
img_prep = tflearn.ImagePreprocessing()
img_prep.add_featurewise_zero_center(per_channel=True)
# Real-time data augmentation
img_aug = tflearn.ImageAugmentation()
img_aug.add_random_flip_leftright()
img_aug.add_random_crop([32, 32], padding=4)
# Building Residual Network
net = tflearn.input_data(shape=[None, 32, 32, 3],
data_preprocessing=img_prep,
data_augmentation=img_aug)
net = tflearn.conv_2d(net, 16, 3, regularizer='L2', weight_decay=0.0001)
net = tflearn.resnext_block(net, n, 16, 32)
net = tflearn.resnext_block(net, 1, 32, 32, downsample=True)
net = tflearn.resnext_block(net, n-1, 32, 32)
net = tflearn.resnext_block(net, 1, 64, 32, downsample=True)
net = tflearn.resnext_block(net, n-1, 64, 32)
net = tflearn.batch_normalization(net)
net = tflearn.activation(net, 'relu')
net = tflearn.global_avg_pool(net)
# Regression
net = tflearn.fully_connected(net, 10, activation='softmax')
opt = tflearn.Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True)
net = tflearn.regression(net, optimizer=opt,
loss='categorical_crossentropy')
# Training
model = tflearn.DNN(net, checkpoint_path='model_resnet_cifar10',
max_checkpoints=10, tensorboard_verbose=0,
# 构建网络
network = tflearn.input_data(shape=[None, 32, 32, 3],
data_augmentation=img_aug,
data_preprocessing=img_prep)
network = tflearn.conv_2d(network,
16,
3,
regularizer='l2',
weight_decay=0.0001)
# 载入残差结构
n = 3
network = tflearn.resnext_block(network,
nb_blocks=n,
out_channels=16,
cardinality=32)
network = tflearn.resnext_block(network,
nb_blocks=1,
out_channels=32,
cardinality=32,
downsample=True) # 残差虚线
network = tflearn.resnext_block(network,
nb_blocks=n,
out_channels=32,
cardinality=32)
network = tflearn.resnext_block(network,
nb_blocks=1,
out_channels=32,
cardinality=32,
