h_layer_num = 32 # for both of the hidden layers
l_rate = 0.1
momentum = 0.4
n_epoch = 3000
batch_size = 16
test_rate = 0.214 # 0.214 if BP True
dna_size = 9 * h_layer_num + h_layer_num + h_layer_num * 2 + 2
# [(9, 10), (10,), (10, 2), (2,)]
# bounds = [(-0.1, 0.1)] * (9 * h_layer_num) + [(-0.5, 0.5)] * h_layer_num + [(-0.1, 0.1)] * (2 * h_layer_num) + [(-0.5, 0.5)] * 2
# bounds = [(-1, 1)] * dna_size
bp = ann.BP_ANN(h_layer_num,
l_rate,
momentum,
n_epoch,
batch_size,
test_rate,
opt="momentum",
do=True)
if only_bp:
fit = bp.train_one(None, is_bp=only_bp)
print("Loss and accuracy on the test data: ", fit)
else:
print("---- GENETIC ALGORITHM IS ON ----")
if is_bp:
print("---- BACKPROPAGATION IS ON ----")
sleep(2)
# Fitness function of the GA
def BP_fit_fun(x):
# Parameters
only_bp = False
is_bp = True # If only_bf False and is_bp is True -> DE + BP
h_layer_num = 10
l_rate = 0.01
momentum = 0.4
n_epoch = 10
batch_size = 128
valid_rate = 0.1717 # 0.1717 if BP True
dna_size = 9 * h_layer_num + h_layer_num + h_layer_num + 2
# [(9, 10), (10,), (10, 2), (2,)]
bounds = [(-0.1, 0.1)] * (9 * h_layer_num) + [(-0.5, 0.5)] * h_layer_num + [(-0.1, 0.1)] * (2 * h_layer_num) + [(-0.5, 0.5)] * 2
# bounds = [(-1, 1)] * dna_size
bp = ann.BP_ANN(h_layer_num, l_rate, momentum, n_epoch, batch_size, valid_rate, opt="nesterov", do=only_bp)
if only_bp:
fit = bp.train_one(None, is_bp=only_bp)
print("Global loss at the end of the training: ", fit)
else:
print("---- DIFFERENTIAL EVOLUTION IS ON ----")
if is_bp:
print("---- BACKPROPAGATION IS ON ----")
sleep(2)
# Fitness function of the DE
def BP_fit_fun(x):
x_ravel = np.ravel(x)
fit = bp.train_one(np.array(x_ravel), is_bp=is_bp) # is_bp=False -> only DE and the fitness is the hit ration on the validation set
return fit
