def main(args):
setup_gpu(args.gpu_id)
# Creating embedding function. This corresponds to the function g in the paper.
# You may need to change the network parameters.
model1 = Initialization.Create_Model()
# size of digits 16*16
img_rows, img_cols = 16, 16
input_shape = (img_rows, img_cols, 1)
input_a = Input(shape=input_shape)
input_b = Input(shape=input_shape)
# number of classes for digits classification
nb_classes = 10
# Loss = (1-alpha)Classification_Loss + (alpha)CSA
alpha = .25
# Having two streams. One for source and one for target.
processed_a = model1(input_a)
processed_b = model1(input_b)
# Creating the prediction function. This corresponds to h in the paper.
out1 = Dropout(0.5)(processed_a)
out1 = Dense(nb_classes)(out1)
out1 = Activation('softmax', name='classification')(out1)
distance = Lambda(Initialization.euclidean_distance,
output_shape=Initialization.eucl_dist_output_shape,
name='CSA')([processed_a, processed_b])
model = Model(inputs=[input_a, input_b], outputs=[out1, distance])
model.compile(loss={
'classification': 'categorical_crossentropy',
'CSA': Initialization.contrastive_loss
},
optimizer='adadelta',
loss_weights={
'classification': 1 - alpha,
'CSA': alpha
})
print('Domain Adaptation Task: ' + args.domain_adaptation_task)
# let's create the positive and negative pairs using row data.
# pairs will be saved in ./pairs directory
Initialization.Create_Pairs(args.domain_adaptation_task, args.repetition,
args.sample_per_class)
Acc = Initialization.training_the_model(model, args.domain_adaptation_task,
args.repetition,
args.sample_per_class)
print(
'Best accuracy for {} target sample per class and repetition {} is {}.'
.format(args.sample_per_class, args.repetition, Acc))
from keras.models import Model # let's assume MNIST->USPS task. domain_adaptation_task = 'MNIST_to_USPS' # USPS_to_MNIST is also another option. # let's run the experiments when 1 target sample per calss is available in training. # you can run the experiments for sample_per_class=1, ... , 7. sample_per_class = 1 # Running the experiments for repetition 5. In the paper we reported the average acuracy. # We run the experiments for repetition=0,...,9 and take the average repetition = 2 # Creating embedding function. This corresponds to the function g in the paper. # You may need to change the network parameters. model1 = Initialization.Create_Model() # size of digits 16*16 img_rows, img_cols = 16, 16 input_shape = (img_rows, img_cols, 1) input_a = Input(shape=input_shape) input_b = Input(shape=input_shape) # number of classes for digits classification nb_classes = 10 # Loss = (1-alpha)Classification_Loss + (alpha)CSA alpha = .25 # Having two streams. One for source and one for target. processed_a = model1(input_a)
