def keypoint_detection():
try:
data = sio.loadmat('data.mat')
except:
load.csv()
data = sio.loadmat('data.mat')
train_x = data['train_x']
train_y = data['train_y']
test_x = data['test_x']
# data normalization
train_x = train_x / 256.0
train_y = (train_y - 48) / 48.0
test_x = test_x / 256.0
sklearn.utils.shuffle(train_x, train_y, random_state=0)
train_x, valid_x = train_x[:-400], train_x[-400:]
train_y, valid_y = train_y[:-400], train_y[-400:]
model = Model(0.01, 0.9, 0.0005, 100, 10000)
model.add_layer(layers.FullConnectedLayer(9216, 256, 1, layers.rectify))
model.add_layer(layers.DropoutLayer(0.5))
model.add_layer(layers.FullConnectedLayer(256, 100, 1, layers.rectify))
model.add_layer(layers.DropoutLayer(0.5))
model.add_layer(layers.FullConnectedLayer(100, 30))
model.set_loss_function(layers.EuclideanLoss)
model.build()
print 'build model complete'
model.train_model(train_x, train_y, valid_x, valid_y)
model.save_test_result(test_x)
def keypoint_detection():
try:
data = sio.loadmat('data.mat')
except:
load.csv()
data = sio.loadmat('data.mat')
train_x = data['train_x']
train_y = data['train_y']
test_x = data['test_x']
# data normalization
train_x = train_x / 256.0
train_y = (train_y - 48) / 48.0
test_x = test_x / 256.0
sklearn.utils.shuffle(train_x, train_y, random_state=0)
train_x, valid_x = train_x[:-400], train_x[-400:]
train_y, valid_y = train_y[:-400], train_y[-400:]
model = Model(0.01, 0.9, 0.0005, 100, 1000)
model.add_layer(layers.ReshapeLayer(1, 96, 96))
model.add_layer(layers.ConvolutionLayer((3, 3), 8, 1, 1, layers.rectify))
model.add_layer(layers.PoolingLayer((2, 2))) # 47 * 47 * 8
model.add_layer(layers.ConvolutionLayer((2, 2), 16, 8, 1, layers.rectify))
model.add_layer(layers.PoolingLayer((2, 2))) # 23 * 23 * 16
model.add_layer(layers.ConvolutionLayer((2, 2), 32, 16, 1, layers.rectify))
model.add_layer(layers.PoolingLayer((2, 2))) # 11 * 11 * 32
model.add_layer(layers.ConvolutionLayer((2, 2), 64, 32, 1, layers.rectify))
model.add_layer(layers.PoolingLayer((2, 2))) # 5 * 5 * 64
model.add_layer(layers.ConvolutionLayer((2, 2), 128, 64, 1, layers.rectify))
model.add_layer(layers.PoolingLayer((2, 2))) # 2 * 2 * 128
model.add_layer(layers.FullConnectedLayer(512, 512, 1, layers.rectify))
model.add_layer(layers.DropoutLayer(0.5))
model.add_layer(layers.FullConnectedLayer(512, 512, 1, layers.rectify))
model.add_layer(layers.DropoutLayer(0.5))
model.add_layer(layers.FullConnectedLayer(512, 30))
model.set_loss_function(layers.EuclideanLoss)
model.build()
print 'build model complete'
model.train_model(train_x, train_y, valid_x, valid_y)
model.save_test_result(test_x)
